1 / 27

Hongtao Du

Hongtao Du. AICIP Research ECE Department University of Tennessee Feb 23, 2005. Background. Blind Source Separation (BSS) Motivation: “cocktail party problem” BSS Model: (Mixing) (Unmixing) BSS Algorithms ICA LCNN Pixel level processing.

shae
Télécharger la présentation

Hongtao Du

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Hongtao Du AICIP Research ECE Department University of Tennessee Feb 23, 2005

  2. Background • Blind Source Separation (BSS) Motivation: “cocktail party problem” • BSS Model: (Mixing) (Unmixing) • BSS Algorithms • ICA • LCNN • Pixel level processing the observed signal (pixel) the source signal (pure pixel or noise) weight matrix or unmixing matrix

  3. Synthesis Structures • Serial Processing • Processing pixel-by-pixel in a serial sequence • Parallel Processing • Using SIMD structure • Multiple pixels in, multiple pixels out • Depending on hardware constraints • Segment Processing • Pipeline structure • Parallel processing

  4. Contrast Stretching s, r : grey level of input pixel and output pixel

  5. Component Contrast

  6. Component Contrast - RTL

  7. Component Contrast - Schematic

  8. Top-level - Schematic

  9. Pre-layout Simulation

  10. Pre-layout Simulation – Small Signal

  11. Pre-layout Simulation – Reset

  12. Pre-layout Simulation – Write Enable

  13. Contrast Stretching (32-bit) – FPGA layout

  14. Contrast Stretching (8-bit) – FPGA layout

  15. Constraint Requested Frequency Estimated Delay Actual TS_Clk = PERIOD TIMEGRP "Clk" 100 nS HIGH 50 nS 100.000 ns 14.464 ns 69.14MHz 16.63 ns 32-bit TS_Clk = PERIOD TIMEGRP "Clk" 100 nS HIGH 50 nS 100.000 ns 7.450 ns 134.23MHz 11.33 ns Comparison 32-bit v.s. 8-bit • Device utilization summary: • 32-bit • Number of External IOBs 132 out of 158 83% • Number of Occupied SLICEs 605 out of 12288 4% • 8-bit • Number of External IOBs 36 out of 158 22% • Number of Occupied SLICEs 53 out of 12288 1% • Clock Report 8-bit

  16. Parallel Contrast- Schematic

  17. Pre-layout Simulation

  18. Parallel Contrast Stretching – FPGA layout

  19. Constraint Requested Frequency Estimated Delay Actual TS_Clk = PERIOD TIMEGRP "Clk" 100 nS HIGH 50 nS 100.000 ns / / 16.63 ns Constraint • Device utilization summary: • 32-bit • Number of External IOBs 580 out of 158 367% • Number of Occupied SLICEs 4838 out of 12288 39% • Too many required IOBs, exceeding the target FPGA capacity • 8-bit • Number of External IOBs 148 out of 158 93% • Number of Occupied SLICEs 422 out of 12288 3% • Clock Report 32-bit TS_Clk = PERIOD TIMEGRP "Clk" 100 nS HIGH 50 nS 100.000 ns 6.748 ns 148.19MHz 8-bit 11.63 ns

  20. Pipeline Contrast- Schematic

  21. Top-level - Schematic

  22. Pre-layout Simulation

  23. Pre-layout Simulation - threshold

  24. Pipeline Contrast Stretching – FPGA layout

  25. Constraint Requested Frequency Estimated Delay Actual TS_Clk = PERIOD TIMEGRP "Clk" 100 nS HIGH 50 nS 100.000 ns 20.944ns 47.75MHz 11.63 ns Synthesis Performance Synthesis Performance (8-bit) Device: Xilinx V1000EHQ-6 • Device utilization summary: • Number of External IOBs 156 out of 158 98% • Number of Occupied SLICEs 586 out of 12288 4% • Total equivalent gate count for design 13,474 • Clock Report

  26. Structure Requested Estimated Delay Actual Frequency 100.000 ns 7.450 ns 134.23MHz 11.33 ns Serial Parallel 100.000 ns 6.748 ns 148.19MHz 11.63 ns 100.000 ns 20.944ns 47.75MHz Pipeline 11.63 ns Serial v.s. Parallel • Serial processing should have the minimum delay, but actually not. • Parallel processing is the fastest structure • Pipeline is the most efficient structure, but very slow.

  27. Serial Parallel Pipeline

More Related