1. A DSP-Based Platform for Wireless Video Compression Patrick Murphy, Vinay Bharadwaj, Erik Welsh & J. Patrick Frantz Rice University November 18, 2002

2. 1 Motivations & Requirements Wireless video communication Demand expected to explode Much more challenging than voice Standards not yet established Need modular hardware & software Testbed should approximate real system Similar hardware Portable code

3. 2 Testbed Description Source coding on Texas Instruments DSP Similar to processors in most wireless devices Mature software development environment Bluetooth wireless interface Datarate comparable to future mobile systems Decoding performed on separate PC

4. 3 Testbed Hardware Custom hardware Inspired by TI Imaging Developers Kit Support for lower resolution imaging Integrates wireless interface & host processor Daughtercard to TI DSP Starter Kits Adheres to DSK interface specification Backward and forward compatible with DSKs

5. 4 Testbed Hardware Custom daughtercard on C6416 DSP board

6. 5 Hardware Details: Processor Axis Etrax 100LX 32-bit 100 MIPS RISC processor Runs full Linux operating system Uses 2.4.x kernels and glibc Open source Bluetooth stack Originally developed by Axis Now available under GPL Multichip Module Integrated flash, SDRAM and Ethernet interface

7. 6 Hardware Details: Camera Omnivision CMOS imager 16-bit digital interface Supports YUV 4:2:2 and raw RGB On-board image preprocessor Color space conversion White balance Configuration by I 2 C Easily controlled by Etrax processor

8. 7 Hardware Details: Bluetooth Support for variety of Bluetooth hardware Up to 1 Mbps serial connection Standard UART interface SiliconWave and Ericsson modules tested Flexible application-level interface Simple (virtual) serial port TCP/IP via PPP Custom protocols possible

9. 8 Hardware Details: DSPs Daughtercard adheres to TI spec Compatible with C62x, C67x & C64x boards Possible future use with C5x systems Minimal resource requirements Uses just one external interrupt All transfers via external memory interface Allows additional daughtercards to be used

10. 9 Software Tools DSP software Development in TI Code Composer Compatible with CC 1.2x and 2.x Only requires DSK tools Uses dsplib and imglib No third-party libraries or extensions required Etrax software Standard Linux development in C/C++ GNU tools (gcc, gdb, etc.)

11. 10 Image Capture Dataflow Each pixel read is read from the camera Daughtercard buffers one line EDMA interrupted for each line Buffer emptied by EDMA burst read EDMA buffers full frame DSP interrupted for full frame Frame used only if DSP is ready

12. 11 Transmission Dataflow DSP writes coded data to memory EDMA constantly writes to EMIF Etrax on daughtercard reads data Data packaged and transmitted

13. 12 Testbed Software Generic framework for dataflow EDMA setup optimized for input & output Two coding schemes implemented MPEG-4 Advanced video compression Designed for low-datarate applications JPEG2000 Advanced still image compression Extremely good compression vs. quality

14. 13 MPEG-4 Encoder Structure based on open source x86 code Heavily optimized for C6x processors Extensive use of TI assembly routines Implements MPEG-4 Simple Profile Only one previous frame required Operates on 8 x 8 macroblocks Minimizes internal data memory requirements Core uses 57 KB of program memory

15. 14 MPEG-4 Encoder Performance Depends heavily on cache configuration L2 Cache Size0KB16KB32KB48KB64KB SRAM Size64KB48KB32KB16KB0KB Cache Associativity-1-way2-way3-way4-way Frames/second11.914.517.618.519.2 Encoder Performance on C6711

16. 15 JPEG2000 Image Compression New version of popular JPEG standard Continuous tone still image compression Wavelet based algorithm More computational intensive than JPEG’s DCT Much better compression ratios Useful features Lossless and lossy options Random access - no inter-frame dependence Good control over quality vs. compression ratio

17. 16 JPEG2000 Encoder Based on JasPer codec Open source reference implementation Project of Image Power & U of British Columbia Very modular and abstracted design Easy addition of features and formats Great for PC execution Bad for embedded DSP execution DSP optimizations Reduce function calls in core code Eliminate some extraneous format flexibilities

18. 17 JPEG2000 Encoder Results Operation on 32 x 32 pixel tiles Constrained by DSP’s internal data memory Total code size around 300KB More internal program memory boosts performance CIFQCIF C67110.170.46 C64166.216.5 Encoder Performance (frames/sec)

19. 18 Future Work: Hardware Interface to new C55x DSP systems Dependent on availability from TI Further integrate components Design single board system Migrate wireless processing to DSP TI’s OMAP platform is good candidate Support full-duplex operation Add audio support Integrate video display

20. 19 Future Work: Software Implement MPEG-4 decoder Relatively easy compared to encoder Further optimize JPEG2000 encoder Reduce number of data structures Minimize function calls Migrate core to optimized assembly routines Increase tile size to reduce DMA transfers Investigate better error protection Supplement or replace Bluetooth error control

21. 20 Conclusions Modular & standards-based testbed Realistic platform Resembles capabilities of future mobile devices It works! Hardware manufactured and tested Software framework validated Two compression standards implemented Ready for future research

22. 21 Questions ?