Recursive Bisection Placement*: feng shui 5.0

1
Recursive Bisection Placement feng shui 5.0
We abandoned recursive bisection about 6 months
ago

• Ameya R. Agnihotri Satoshi Ono Patrick H.
  MaddenSUNY Binghamton CSD, FAIS, University of
  Kitakyushu
• (with code by Mehmet Can Yildiz and Ateen
  Khatkhate, and with help from Purdues Cheng-Kok
  Koh and Chen Li)

This work was supported by an IBM Faculty
Partnership Award, SRC Project 947.1, an
equipment grant from Intel, NYSTAR MDC, and by
funds from the Japanese Ministry of ECSST via
Kitakyushu and Fukuoka knowledge-based cluster
project.
2
Where is Kitakyushu?
3
feng shui versionshttp//vlsicad.cs.binghamton.ed
u

• 1.0 basic bisection engine
• ISPD99,GLSVLSI01 iterative deletion, cycling,
  terminal propagation work.
• 1.5 cut sequence optimization
• DAC01 paper with Rents rule.
• 2.0 Fractional cut
• ICCAD03 results close to annealing possible to
  place very very dense.
• 2.4, 2.6 Mixed size placement paper
• ISPD04 results improved by 26 on average over
  the closest competing tool.
• 3.0 Structural (beacon) placer
• ASPDAC05 new method designed to have fun with
  the PEKO benchmarks currently within 22 of
  optimal.
• 5.0 Integration of tools
• GUI front end, GLSVSLI00 non-Manhattan global
  router, DAC03 global router, ISCAS05 delay
  analysis, PowerARC delay library, support for
  LEF/DEF, gate sizing tool Versions of the
  stand-alone placer are available in both hMetis
  and MLPart flavors.
• 5 or so improvements on mixed size placement
  (with minor tuning).
• The bisection based placement engine is now in
  maintainence mode well fix minor problems, but
  its not our main focus.

4
Bisection Based Placement
Logic elements
Semiconductor chip
5
Fractional Cut Bisection
6
Mixed Size Fractional Cut
7
IBM01 before legalization
8
Placement Legalization

• Standard cells
• Dynamic programming approaches ICCAD03, and an
  unpublished method.
• Moving macro blocks
• ISPD04 Extension of the tetris method by Hill.
• Fixed macro blocks
• Fractured row legalization Khatkhate MS thesis.

9
ISPD04 Experimental Results
Capo 1 ISPD02 Capo 2 ICCAD03 Capo 3 ICCAD03 mPG-ms ASPDAC03 Feng Shui 2.4 ISPD04
Ibm01 3.96 3.36 3.05 3.01 2.41
Ibm02 8.37 8.23 6.83 7.42 5.34
Ibm03 12.16 11.53 10.38 11.20 7.51
Ibm04 13.48 11.93 10.11 10.50 7.96
Ibm05 11.51 11.20 11.10 10.90 10.10
Ibm06 10.25 9.63 9.94 9.21 6.82
Ibm07 15.75 15.80 15.25 13.70 11.71

Average 1.29 1.29 1.29 1.26 1
As much as 51 better on some benchmarks.
Closest is around 8, for the design that doesnt
have macro blocks.
10
Now about the contest

• The designs have very large amounts of internal
  white space (apples vs. oranges)
• Space MUST be distributed internally to have any
  chance of a good result (due to fixed locations
  of the macro blocks)
• We have no support at all for internal white
  space distribution
• Or, to be more precise The correct location for
  white space is usually outside of the core.

11
Example
IBM
12
Example
13
How Much White Space Do You Really Need?

• If you had 100X more space, would you really want
  to stretch that far?
• Why do the industry guys do this?
• Put space in early for sizing and routability
• Or even better, use something like.
• ICCAD04 cut line shifting for routability.
• ASPDAC05, same idea for gate sizing and buffer
  insertion.
• Weve integrated our global router and sizing
  tools with the placer to avoid needing to insert
  excess space.

14
Placement Contest, Round 2

• Were going to lose round 1 thats for sure.
• Apples vs. oranges problem some of the apple
  placement methods are going to look very bad
• Whos up for another round?
• Moving macro blocks?
• Rotation of blocks?
• Mirroring cells?
• Or round 1 again, but give me some time to finish
  the move away from recursive bisection!
• Routing benchmark announcement
• Patrick Groeneveld, Jurjen Westra, and myself,
  have been working on global routing benchmarking.
  Well talk about this at EDP. WANTED more
  research groups in routing!

15
Better Visualization (feng shui 5.0 GUI version)

• Start with a reference placement
• For the PEKOs, this is the known optimal
• Map pixel colors from an image onto cells in the
  reference placement
• Rearrange the cells to match a placement under
  study
• Keep the cell colors from the reference

16
Fun with the PEKOs
If the placement is optimal, the mapped image
should be identical to the reference image.
17
feng shui 5.0Suboptimality Rogues Gallery
mPL
Dragon
Capo
feng shui
Beacon
26 HPWL improvement on mixed size, fractional
cut, non-Manhattan, and what do people ask
about? The funny pictures from the GUI
Prof. Jason Cong
Prof. Majid Sarrafzadeh
Prof. Igor Markov
Prof. Patrick H. Madden
Satoshi Ono