1 / 30

Partial Reconfigurable Designs

Partial Reconfigurable Designs. Multi-layer Floorplanning Max Walton. Outline. Introduction Challenges Floorplanning Model Proposed Floorplanner Data Representation Cost Functions Moves Matching Results. Introduction. Partial Reconfiguration Difference-based (used in this paper)

allan
Download Presentation

Partial Reconfigurable Designs

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. Partial Reconfigurable Designs Multi-layer Floorplanning Max Walton

  2. Outline • Introduction • Challenges • Floorplanning • Model • Proposed Floorplanner • Data Representation • Cost Functions • Moves • Matching • Results Max Walton

  3. Introduction • Partial Reconfiguration • Difference-based (used in this paper) • Module-based • Issues • Reconfiguration overhead • Time issues • Placement Max Walton

  4. Floorplanning Terminology • Static Module • AKA: fixed module • A module that will not be reconfigured • Non-static module • AKA: reconfigured modules • Modules that will be reconfigured between designs Max Walton

  5. Introduction: Challenges • Reuse Matching • Which components overlap during configuration? • Reuse Placement • Where do the components need to be placed? • Reuse Interface • How are such components connected to reconfigurable regions? Max Walton

  6. Floorplanning • Placing components on a chip • Differs from placement by only placing large sized components on chip • Does not look at logic • Complements Placement • Outputs coordinates defining positions of block on device Max Walton

  7. Floorplanning • Three types of floorplanning • Independent • Dependent • Combined Max Walton

  8. Floorplanning • Simulated Annealing Based • Fixed Outline Floorplanning • Constrain design in rectangular shapes of fixed aspect ratio • Parquet • Area • Wirelength • Aspect Ratio Max Walton

  9. Model • Frames • Span n columns • Module spans contiguous set of frames (hor/vert) • Time to reconfigure linear function of number of frames to be reconfigured • Minimize number of frames by placing fixed and reconfigurable parts in separate frames Max Walton

  10. Assumptions • Complete Sequence of Designs is known • No Data Dependency Between Designs / Input and Output Buffered in Static region • Soft blocks • Changing of aspect ratio is allowed • Block can be placed anywhere on device • Heterogeneous floorplanning out of scope • All designs are timing critical Max Walton

  11. Reusable Components • Keep interconnects outside of static regions • Use of whitespace for interfaces • Maximize A1,2 and Areused Max Walton

  12. Proposed Floorplanner (FFPR) • Built from Parquet floorplanner • Routing congestion • Total Frames • Handles multiple designs simultaneously Max Walton

  13. Definition • Given design D1 with a set of modules M1 = {m1, …, mn1} and corresponding connectivity, • Given design D2 with a set of modules M2 = {m1, …, mn2} and corresponding connectivity, • Given a set of common modules between the two designs M12, • Floorplan each design such that the total area and wirelength in each design is minimized as well as total reconfiguration area is minimized. • Extensible to k>2 reconfigurable designs Max Walton

  14. FFPR Max Walton

  15. FFPR: Data Representation • Placement by Sequence Pairs • Exact placement found with horizontal and vertical graphs • Algorithm runs in O(n2) Max Walton

  16. FFPR: Data Representation • Two-layer Sequence pair • Non-static have no left-right or up-down relationship between each other • Horizontal and vertical graphs are connected through static nodes only Max Walton

  17. FFPR: Data Representation Max Walton

  18. FFPR: Cost Functions • Scaling factors • Sum to 1 • Represent respective weights of area, aspect ratio, frames, wirelength, congestion • Area • Minimum bounding box encompassing all designs • Negative if new area is less than current Max Walton

  19. FFPR: Cost Functions • Aspect Ratio • Computed as a penalty function • Computes change in cost of the aspect ratio • Wirelength • Adds wirelength of each interconnect in design • Half-perimeter bounding box is used for each interconnect Max Walton

  20. FFPR: Cost Functions • Congestion Cost • Probability Congestion model • 2D array of bins (CLBs in Virtex 4) • A pin lies in only one bin • A bin may contain multiple pins • Sum of probabilites of all the paths that pass through bin • Bin is congested if its congestion exceeds a threshold • Calculated as sum of excess congestion of each bin Max Walton

  21. FFPR: Cost Functions • Reconfiguration Frames Cost • Computed by looking at the fixed and reconfigurable regions compared with next design • Consecutive design frames are added together to get total number Max Walton

  22. FFPR: Moves • Moves on Blocks • Changing orientation of a block • Changing aspect ratio • Changing whitespace along the border • Moves on Data Representation • Compaction: swapping random modules • Compaction: moving block left/right or up/down • Matching Max Walton

  23. Moves: Changing the Whitespace • Add four offsets to the blocks • n, e, w, s • Range {0 .. 5} Max Walton

  24. FFPR: Matching • 2 designs is equivalent to bipartite matching • Matching for multiple designs • Leads to many cases Max Walton

  25. FFPR: Matching Max Walton

  26. Results • Matching: • 50% savings on frames (50% partial matching) Max Walton

  27. Results • Direction of matching can impact design • D1  D2 vs. D2  D1 • Dependent Mode as much as 3X wirelength of combined/independent • 50% higher on average • Combined 9% more wirelength than independent • Multilayer vs. traditional floorplanner • 12% better clock period on average • Reduces place and route time Max Walton

  28. References • “Multi-layer floorplanning for reconfigurable designs”, L. Singhal and E. Bozorgzadeh, IET Comput. Digit. Tech., 2007, 1, pp. 276-294 Max Walton

  29. Project • Create a Scatter Search implementation in Celoxica Handel-C • Search out better performance from Handel-C version of SS • Use Handel-C constructs to gain better performance • Attempt multiple approaches of implementation (time permits) Max Walton

  30. Project Status • Used C code for 0-1 Knapsack Problem from “Scatter Search: Methodology and Implementations in C” by Laguna and Marti • Currently converting to avoid pointer use and use less complex data structures more inherent in hardware Max Walton

More Related