Interconnect Resource-Aware Placement for Hierarchical FPGAs

1
Interconnect Resource-Aware Placementfor
Hierarchical FPGAs

• Amit Singh, Ganapathy Parthasarathy, Malgorzata
  Marek-SadowskaDepartment of Electrical and
  Computer Engineering University
• of California, Santa Barbara, Santa Barbara, CA
  93106, USAICCAD 2001
• Speaker??? 2004/04/09

2
Outline

• Rents Rule
• Rents Parameter
• Local Rents Parameter
• Application
• Problem Definition
• Logic Clustering using Rents rule
• Placement using Rents rule
• Experimental Results

3
Rents Rule
4
Rents Parameter

• Rents Parameter
• Is a measure of the interconnection complexity of
  the design
• Allow us to choose the best physical architecture
  to implement the design
• Rents Rule

5
Local Rents Parameter

• Local Rents Parameter
• Rents Rule doesnt reflect local fluctuations of
  the interconnection complexity in the design
• For a placed circuit that minimizes the total
  wire-length, the Local Rents parameter of a
  design can be determined from the slope of the
  log-log plot of number of interconnections from a
  block versus length of interconnection (Net
  Length Distribution)
• For a given block vi,
• Count the number of interconnections,Nj,of length
  lj,from block vi
• Find the best-fitting line on the log(Nj) versus
  log(lj)plot
• Local Rents parameter gt slope of this log-log
  plot

6
Example

• For a good placement in a D-dimensional
  Manhattan grid , the interconnection length
  distribution is of
• the form

7
Application
8
Problem Definition

• Given a hierarchical 2-D mesh FPGA chip with
  Rents parameter, Pa, and a design mapped to
  k-LUTs, having Rents parameter Pd,
• Fit the design in the chip so that the area of
  the mapped, placed and routed design is
  minimized, subject to the constraint of a given
  bounding box aspect-ratio.

9
Hierarchical FPGA Architecture
10
Find the Rents parameter of hierarchical FPGA
architecture(Pa)

• Find Pa
• Use gnl to create uniform complexity benchmarks
  circuits having values of Pd ranging from 0.2 to
  0.9
• Do a bottom-up clustering of blocks having a
  pre-defined Rents parameter(Pd)
• Placement Routing
• Measure the routing resource utilization
• Find the value(Pd) that gives the highest network
  utilization

11
Example for Pa
12
Logic Clustering using Rents rule

• A trial and error clustering and placement
  implementation method may lead to unnecessary
  congestion and too much unused routing resources
• to alleviate this problem by depopulating logic
  clusters
• Clustering technique guarantees
• a fair amount of uniformity in the clustered
  designs complexity as measured by Rents
  parameter
• gives our placement tool, a fairly uniform
  depopulated hierarchical circuit where the worst
  case Pd of any logic cluster is equal to Pa

13
Two Phases of Bottom-up Clustering

• Find the upper bound on the cluster pin size
  based on Pa
• Pd of any logic cluster in the design Pa
• Clustering using modified T-VPack
• Clustering based on the new set of cluster pin
  constraints

14
Placement using Rents rule

• Simulated Annealing
• Placement Cost Function
• Using local Rents parameter
• Recognizes the variations in the spatial
  complexities of the clusters and generates
  placement solutions which result in fewer routing
  tracks and shorter wirelengths after routing

15
Placement Cost Function
16
Experimental Results
13
21.3
9.4
17
Experimental Results
17.7
11