Wire Length Prediction-based Technology Mapping and Fanout Optimization

1
Wire Length Prediction-based Technology Mapping
and Fanout Optimization

• Qinghua Liu
• Malgorzata Marek-Sadowska
• VLSI Design Automation Lab
• UC-Santa Barbara

2
Outline

• Motivation and previous work
• Pre-layout wire length prediction
• Technology mapping with wire-length prediction
• Fanout optimization with wire-length prediction
• Experimental results
• Conclusions and future work

3
Motivation

• Traditional logic synthesis does not consider
  accurate layout information
• Placement quality depends on
• netlist structure
• placement algorithm

4
Previous work

• Logic and physical co-synthesis
• Layout-driven logic synthesis
• Local netlist transformations
• Metric-driven structural logic synthesis
• Adhesion
• Distance

5
Pre-layout wire-length prediction

• Previous work
• Statistical wire-length prediction
• Lou Sheffer et al. Why Interconnect Prediction
  Doesnt work? SLIP00
• Individual wire-length prediction
• Qinghua Liu et al. Wire Length Prediction in
  Constraint Driven Placement SLIP03
• Semi-individual wire-length prediction
• Predict that nets have a tendency to be long or
  short
• Qinghua Liu et al. Pre-layout Wire Length and
  Congestion Estimation DAC04

6
Summary of the semi-individual wire length
prediction technique

• Predict lengths of connections
• Mutual contraction
• Predict lengths of multi-pin nets by
• Net range

7
Mutual contraction
B.Hu and M.Marek-Sadowska, Wire length
prediction based clustering and its application
in placement DAC03
v
y
u
x
8
Relative weight of a connection
v
Wr(u, v) 0.71
u
EQ1
y
EQ2
Wr(x, y) 0.5
x
9
Mutual contraction of a connection
Cp(x, y) Wr(x, y) Wr(y, x)
EQ3
Wr(u, v) 0.71 Wr(v, u) 0.33 Cp(u, v) 0.234
Wr(x, y) 0.71 Wr(y, x) 0.6 Cp(x, y) 0.426
y
v
j
i
x
u
10
Predictions on connections
(a)
(b)
Mutual contraction vs. Connection length
11
Net range
0 1 2 3 4 5 6 7
8 9 10 11
Circuit depth
Example of net range
12
Predictions on multi-pin nets
Net range vs. average length for multi-pin nets
13
Technology mapping with wire-length prediction
(WP-Map)

• Node Decomposition
• Technology Mapping

14
Node decomposition
a
b
c
G
a
b
c
a
b
c
T.Kutzschebauch and L.Stok, Congestion aware
layout driven logic synthesis, ICCAD01
15
Greedy node decomposition algorithm
CurrentPinNumn
N
CurrentPinNum CurrentPinNum-1
Done
Y
(n1,n2)two input nets with largest mutual
contraction
Update mutual contraction
Decompose(G,n1,n2)
Remove n1 and n2, insert new net
Decompose n-input gate G with wire length
prediction
16
Correlation between mutual contraction and
interconnection complexity
Average mutual contraction vs. Rents exponent
17
Technology mapping
EQ4
18
Fanout optimization with wire-length prediction
(WP-Fanout)

• Net selection
• Select all large-degree nets
• Select small-degree nets with large net range
• Net decomposition

LT-tree
Balanced tree
Circuit depth
19
Experiment setting

• LGSyn93 benchmark suite
• Optimized by script.rugged
• Mapped with 0.13um industrial standard cell
  library
• Placement is done by mPL4
• Global routing is done by Labyrinth

20
Experimental results

• Compare with the traditional area-driven
  technology mapping algorithm implemented in SIS
• Results of the WP-Map algorithm
• Results of combined WP-Map and WP-Fanout algorithm

21
Compare WP-Map with SIS
Compare mapped netlists
22
Compare WP-Map with SIS (cont.)
Average cut number distribution of C6288
23
Compare WP-Map with SIS (cont.)
Results after placement and global routing
24
Compare WP-Map WP-Fanout with SIS
Results after placement and global routing
25
Conclusions

• Wire length can be predicted in structural level
• Mutual contraction
• Net range
• Wire length prediction technique can be applied
  into technology mapping and fanout optimization
• 8.7 improvement on average congestion
• 17.2 improvement on peak congestion

26
Future work

• Logic extraction with wire-length and congestion
  prediction
• Timing-driven technology mapping with wire-length
  prediction