1 / 32

Handling Complexities in Modern Large-Scale Mixed-Size Placement

Handling Complexities in Modern Large-Scale Mixed-Size Placement. Jackey Z. Yan Natarajan Viswanathan Chris Chu. Dept. of Electrical & Computer Engineering Iowa State University, Ames, IA 50010. Mixed-Size Placement. Millions of standard cells

lynne
Download Presentation

Handling Complexities in Modern Large-Scale Mixed-Size Placement

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. Handling Complexities in Modern Large-Scale Mixed-Size Placement Jackey Z. Yan Natarajan Viswanathan Chris Chu Dept. of Electrical & Computer Engineering Iowa State University, Ames, IA 50010

  2. Mixed-Size Placement • Millions of standard cells • Hundreds of macros, e.g., IP cores or memory blocks • All objects are movable with hundreds of fixed I/O objects

  3. Previous Work: Two-Stage Approach • First fix macros, then place standard cells Stage 1 Stage 2 Initial Placement Macro Placer / Legalizer Standard-cell Placer • Capo+Parquet+Capo [S. N. Adya et al., ISPD 2002] • XDP [J. Cong et al., ASP-DAC 2006] • MP-trees based macro placer [T.-C. Chen et al., DAC 2007] • TCG based macro placer [H.-C. Chen et al., ICCAD 2008] • Macro positions are fixed based on an initial placement, which • is generated using some inaccurate WL model • contains many overlaps among the objects • Standard cell positions in initial placement are discarded

  4. Previous Work: Simultaneous Approach • Simultaneous placement of macros and standard cells • Dragon[T. Taghavi et al., ISPD 2006] • Capo Floorplacement [J. A. Roy et al., ISPD 2006] • FastPlace3[N. Viswanathan et al., ASPDAC 2007] • APlace[A. B. Kahng et al., ISPD 2006] • mPL6[T. Chan et al., ISPD 2006] • Kraftwerk[P. Spindler et al., ICCAD 2006] • NTUplace3[T.-C. Chen et al., ICCAD 2006] State of the art: Analytical Placers • Wirelength is approximated (e.g., log-sum-exp, quadratic) • Many overlaps among the objects  Legalization is very hard • Scalability is an issue • Cannot optimize the macro orientations

  5. FLOP: Floorplan-Guided Placer • Effectively integrate the floorplanning and incremental placement algorithms • Efficiently handle large-scale mixed-size designs with all movable objects and fixed I/O objects • Derived Modern Mixed-Size (MMS) placement benchmarks from ISPD05/06 placement benchmarks • Experimental results • State-of-the-art mixed-size placers (simultaneous approach) • Leading macro placers (two-stage approach) Best Wirelength

  6. Cut down problem size • Optimize positions of small objects New Algorithm Flow Block Formation • Direct HPWL minimization • Precise object distribution Floorplanning for overlap-free layout • Optimize macro orientation Wirelength-driven Shifting • Optimize HPWL by shifting • blocks at floorplan level Final positions of big macros Initial positions of small objects Incremental Placement

  7. Advantages of New Algorithm Flow Block Formation • Compared with analytical placer: • Exact HPWL minimization (Steps 1-3) • More precise object distribution (Step 2) • Better scalability (Step 1) • Capability of handling constraints (Step 2) • Avoid legalization of big macros (Step 2) Floorplanning Wirelength-driven Shifting • Compared with two-stage approach: • Do not rely on low-quality initial placement • Respect positions of small objects provided by previous step Incremental Placement

  8. Fixed- outline region Background: Fixed-outline Floorplanning DeFer [Yan et al. DAC 2008] 500 1. Partitioning 240 260 2. Combining 120 140 128 112 3. Back-Tracing 4. Swapping 5 8 7 9 10 9

  9. Original Circuit Block Formation • Integrated into partitioning step of DeFer • Recursively cut the original circuit • By hMetis2.0 • Stopping Criteria: Area% of a subcircuit ≤ 0.15% OR # of objects ≤ 10

  10. Shape Curve Generation for Blocks H H H A’ A A A A W W W Un-rotatable Hard Block Rotatable Hard Block Soft Block

  11. Exact Net Model in Min-Cut Placement Traditional min-cut placement minimizes the wirelength by minimizing the cut value (# of interconnections) Circuit M N Placement Region Cut Nets Longer Nets Exact Net Model Min Cut Min Wirelength ≈ = [Chen et al. ICCAD 2005]

  12. A B V A B A B H B A A B V-cut H-cut B A B A Usage of Net Model in FLOP and H/V H/V H/V βLevels(β= 3 by default) H/V H/V H/V H/V Better Cut ? or

  13. Cut Value Comparison in Net Model Net Bounding Box H W

  14. After Before Wirelength-driven Shifting • Floorplan provides good initial positions for blocks • WL can be further improved by shifting blocks • Formulated optimally as a Linear Programming (LP) problem • Solved by QSopt • Hard block positions are fixed after this step • Soft block positions determine initial small objects positions for following incremental placer

  15. Incremental Analytical Placement Similar to FastPlace3 [Viswanathan et al., ASPDAC 2007] All Objects Positions Physical & Netlist based clustering Coarse global placement Refinement of fine-grain clusters Refinement of flat netlist Legalization and detailed placement Final Placement

  16. Modern Mixed-Size (MMS) Benchmarks • 16 mixed-size placement benchmarks • Derived from ISPD 05/06 Placement Benchmarks • Recover the complexities of modern mixed-size designs • One main change has been made on the original circuits • Macros are freed from the original positions • Publicly available at http://www.public.iastate.edu/~zijunyan

  17. Statistics of MMS Benchmarks

  18. Comparing MMS & Original Benchmarks (* results are cited from RQL paper [DAC 2007])

  19. Experimental Setup • All experiments run on AMD Opteron 2.6 GHz CPU and 8 GB memory • Four groups of experiments: • Comparing FLOP with state-of-the-art mixed-size placers (simultaneous approach) on MMS Benchmarks • HPWL on ISPD05 circuits • Scaled HPWL on ISPD06 circuits • Comparing FLOP with leading macro placers (two-stage approach) on modified ISPD06 circuits • Comparing FLOP with and without macro rotation • Comparing FLOP with and without incremental placement

  20. HPWL Results on MMS Benchmarks Binaries are best-available versions on MMS Benchmarks

  21. Norm HPWL 1.44 1.26 1.08 1.02 1 Normalized Results (Runtime & HPWL)

  22. Results of Non-Rotatable FLOP

  23. Results of Non-Incremental FLOP

  24. HPWL Comparison with Macro Placers The benchmarks are modified ISPD06 circuits (All area I/O pads are made movable. Only newblue1 has fixed I/O pads) [Chen et al. 2008 ICCAD].

  25. HPWL Improvement by Moving Macros Best HPWL: 210.96 (RQL) HPWL:175.99 (FLOP) 19.9% longer

  26. HPWL Improvement by Moving Macros Best Scaled HPWL: 432.58 (NTUplace2) Scaled HPWL: 357.83 (FLOP) 20.89% longer

  27. Conclusion • FLOP – a robust, efficient and high-quality mixed-size placement algorithm • FLOP achieves the best wirelength among all state-of-the-art mixed-size placers and leading macro placers

  28. Runtime Breakdown Partitioning Fixed-Outline Floorplanning Shape Curve Combining Swapping LP-based Shifting Global Placement Incremental Placement Detailed Placement

  29. Future Work • Propose a better stand-alone clustering/block-formation technique • Develop a faster shifting technique to substitute LP-based shifting • Extend current technique to handle fixed macros • Try other floorplanning and incremental placement engines

  30. Thank you!

  31. L L L ( I ) ( II ) ( III ) b l r Exact Net Model Assumption 1: Movable pins are placed at region centers Assumption 2: HPWL is used to measure wirelength

  32. H H H r l m Three Hyperedges in Net Model

More Related