Routing

References and Copyright

• Textbooks referred (none required)
 [Mic94] G. De Micheli
“Synthesis and Optimization of Digital Circuits”
McGraw-Hill, 1994.
 [CLR90] T. H. Cormen, C. E. Leiserson, R. L. Rivest
“Introduction to Algorithms”
MIT Press, 1990.
 [Sar96] M. Sarrafzadeh, C. K. Wong
“An Introduction to VLSI Physical Design”
McGraw-Hill, 1996.
 [She99] N. Sherwani
“Algorithms For VLSI Physical Design Automation”
Kluwer Academic Publishers, 3rd edition, 1999.
 References and Copyright (cont.)
• Slides used: (Modified by Kia when necessary)
 [©Sarrafzadeh] © Majid Sarrafzadeh, 2001;
Department of Computer Science, UCLA
 [©Sherwani] © Naveed A. Sherwani, 1992
(companion slides to [She99])
 [©Keutzer] © Kurt Keutzer, Dept. of EECS,
UC-Berekeley
http://www-cad.eecs.berkeley.edu/~niraj/ee244/index.htm
 [©Gupta] © Rajesh Gupta
UC-Irvine
http://www.ics.uci.edu/~rgupta/ics280.html
 [©Kang] © Steve Kang
UIUC
http://www.ece.uiuc.edu/ece482/
 Routing
• Problem
 Given a placement, and a fixed number of metal
layers, find a valid pattern of horizontal and vertical
wires that connect the terminals of the nets
 Levels of abstraction:
o Global routing
o Detailed routing
• Objectives
 Cost components:
o Area (channel width) – min congestion in prev levels helped
o Wire delays – timing minimization in previous levels
o Number of layers (fewer layers  less expensive)
o Additional cost components: number of bends, vias
 Routing Anatomy
3D
Top view
view
Layout
Symbolic

Metal layer 3
Via
Note: Colors used
Metal layer 2 in this slide are not
Metal layer 1 standard
 Global vs. Detailed Routing
• Global routing
 Input: detailed placement, with exact
terminal locations
 Determine “channel” (routing region)
for each net
 Objective: minimize area (congestion),
and timing (approximate)
• Detailed routing
 Input: channels and approximate
routing from the global routing phase
 Determine the exact route and layers
for each net
 Objective: valid routing, minimize area
(congestion), meet timing constraints
 Additional objectives: min via, power
Figs. [©Sherwani]
 Routing Environment
• Routing regions 1 1 4 5 4

 Channel
o Fixed height ? 1,3 4,5
( fixed number of tracks)
o Fixed terminals on top and bottom
o More constrained problem: switchbox. 3 2 2 3 5
Terminals on four sides fixed
 Area routing
o Wires can pass through any region not occupied by cells
(exception: over-the-cell routing)
• Routing layers
 Could be pre-assigned (e.g., M1 horizontal, M2 vert.)
 Different weights might be assigned to layers
 Routing Environment
• Chip architecture
 Full-custom: Feedthroughs
o No constraint on routing regions
 Standard cell: Channel Failed net
o Variable channel height?
o Feed-through cells connect
channels
 FPGA: Tracks
o Fixed channel height
o Limited switchbox connections
o Prefabricated wire segments
have different weights Failed connection

Figs. [©Sherwani]
 Taxonomy of VLSI Routers
Routers

Global Detailed Specialized

Graph Search General Power/Gnd

Restricted
Purpose
Steiner Clock
River Maze
Iterative
Switchbox Line Probe
Maze
Channel Line Expansion

Hierarchical Greedy Left-Edge

[©Keutzer]
 Global Routing
[©Sarrafzadeh]
• Stages
 Routing region definition
M2
 Routing region ordering M1
 Steiner-tree / area routing
• Grid
 Tiles super-imposed on placement M3
 Regular or irregular
 Smaller problem to solve,
higher level of abstraction
 Terminals at center of grid tiles
• Edge capacity
 Number of nets that can pass a
certain grid edge (aka congestion)
 On edge E ,
ij
Capacity(Eij)  Congestion(Eij)
 Grid Graph
• Course or fine-grain
• Vertices: routing regions, edges: route exists?
• Weights on edges
 How costly is to use that edge
 Could vary during the routing (e.g., for congestion)
 Horizontal / vertical might have different weights

t1 t2 t3 t1 t2 t3
t4 t4

1 1 1

t1 t2 t3 2 2 1 1
t4 1 1 1
[©Sherwani]
 Global Routing – Graph Search
• Good for two-terminal nets
• Build grid graph (Coarse? Fine?)
• Use graph search algorithms, e.g., Dijkstra
• Iterative: route nets one by one
• How to handle:
 Congestion?
 Critical nets?
• Order of the nets to route?
 Net criticality
 Half-perimeter of the bounding box
 Number of terminals
 Global Routing – Maze Routing
• Similar to breadth-first search 5
 Very simple algorithm 5 4 5
 Works on grid graph 4 3 t5 4 5
 Time complexity: grid size (NxN) 3 2 3 4 5
• Algorithm 2 1 s 1 2 3 4 5
 Propagate a “wave” from source 3 2 1 2 3 4 5
until hit the sink
(implemented using a queue)
 Trace back to find the path
• Guaranteed to find the optimal solution
 Usually multiple optimal solutions exist
• More than two terminals?
 For the third terminal, use the path between the first
two as the source of the wave
 Maze Routing
• Key to popularity:
 Simplicity
 Guaranteed to find the optimal solution
 Can realize more complex cost functions too
(e.g., number of bends in a path)
• Weakness:
 Multiple terminals not handled efficiently
 Dependent on grid, a two dimensional data structure
• Different variations exist
 Soukup’s alg:
o First use DFS, when get to an obstacle, use BFS to get around
o No guarantee to find the shortest path
 Multiple Terminal Nets: Steiner Tree
• Steiner tree (aka Rectilinear Steiner Tree – RST):
 A tree connecting multiple terminals
o Original points: “demand points” – set D
o Added points: “Steiner points” – set S
 Edges horizontal or vertical only
• Steiner Minimum Tree (SMT)
 Similar to minimum spanning tree (MST)
 But finding SMT is NP-complete
 Many good heuristics introduced to find SMT
• Algorithm
 Find MST
 Pass horizontal and vertical lines from each terminal to
get the Hannan grid (optimal solution is on this grid)
 Convert each edge of the MST to an L-shaped route
on Hannan grid (add a Steiner point at the corner of L)
 Steiner Tree
• Hannan grid reduces solution space (smaller grid)
 For min length RST, Steiner points always on Hannan grid
• Convert MST to rectilinear paths
 Length bounded by 1.5 times optimal SMT length
• Use alternate “L” routes to find the minimum tree

Steiner
point

MSP (length=11) Steiner tree (len=13)

[©Sherwani]
 Steiner Tree Routing
• Can apply different costs to different regions
(or horizontal/vertical preference)
• Order of the nets
 Sequential
o Use # of terminals, criticality, etc. to determine order
 Parallel
o Divide the chip into large regions, perform the routing in
parallel
• Key to popularity
 Fast (not theoretically, but practically)
 Bounded solution quality
• Shortcomings
 Difficult to predict or avoid congestion
 Global Routing Approaches
• A combination of different approaches
might be used in chip-level routing
 Route simple nets (2-3 pins in local area) directly
(e.g., L-shaped or Z-shaped)
 Use a “close to optimal” Steiner Tree algorithms to
route nets of intermediate length
 Route remaining “big” nets using a maze router
• Ordering
 Some ordering is chosen, if can route all, then done,
otherwise:
 Rip-up and Re-route

[©Keutzer]
 Taxonomy of VLSI Routers
Routers

Global Detailed Specialized

Graph Search General Power/Gnd

Restricted
Purpose
Steiner Clock
River Maze
Iterative
Switchbox Line Probe
Maze
Channel Line Expansion

Hierarchical Greedy Left-Edge

[©Keutzer]
 Detailed Routing: Channel vs. Switchbox
a) Channels have no conflicts
b) Conflicting channels
c) Conflict resolved using L-shaped channels
• Order matters
d) Switchbox used to resolve the conflict
• Order matters
• Harder problem (compared to channel routing)

Switch-
1 3
box
3 1 1
4 2
2 2 2
1
(c) (d)
(a) (b) [©Sherwani]
 Channel Routing Problem
• Input
 Fixed terminal locations on the top and bottom
 Possibly floating terminals on the left and right
 Possibly fixed channel capacity constraint
Grid-based
(capacity = max # of horizontal wires between
the top and bottom boundaries of the channel)
 Either gridless (aka area-based) or grid-based
• In the algorithms we consider
 Constraints:
o Grid structure
o Two routing layers (one for H, another for V) Gridless
Still relevant?
 Minimize
o # tracks (channel height)
o Total wire length
o # vias [©Sherwani]
[©Keutzer]
Channel Routing Algorithms and Terminology
• General case is NP-Complete
• Algorithms
 Simple case: left-edge algorithm (P)
 General case: NP  heuristics
o Problem defined using horizontal constraints graph
and vertical constraints graph
o Either graph-based algorithms or
greedy algorithms are used
nets
• Terminology 2 2 branch 1 1
via

tracks
trunk pins

[©Keutzer]
2 1
 Channel Routing Example
• Problem instance:
1 4 5 1 6 7 4 9 10 10

2 3 5 3 5 2 6 8 9 8 7 9
• Solution:
1 4 5 1 6 7 4 9 10 10

2 3 5 3 5 2 6 8 9 8 7 9 [©Keutzer]
 Vertical Constraint Graph
• Represents the relative vertical positions of
different net trunks (tracks) 1 4 5

 Node: represents a net

 Edge (x, y): if at the same column,
x has a terminal on the upper edge and 2 3 5 3
y has a terminal on the lower edge 2 1 4
 (a,b) means that net “a” has to be
above “b” – why? 3 5

• Lower bound:
b a
 # tracks  longest path in VCG
? a
• VCG may have a cycle!
 What shall we do?? b
a b
 Horizontal / Vertical Constraint Graphs
• Channel routing problem 1 4 5 1 6 7 4 9 1010
can be completely
characterized by VCG and
HCG
2 3 5 3 5 2 6 8 9 8 7 9

1 4 10 1
3 2

5 9 7 5 6
4
6 7
3 8
8
2 9 10
Vertical constraint graph (VCG) Horizontal constraint graph (HCG)
[©Keutzer]
 Channel Density
• A net extends from its leftmost 1 4 5 1 6 7
terminal to its rightmost one
• Local density at column C
 ld (C) = # nets split by column C
2 3 5 3 5 2 6 8
• Channel density
 d = max ld(c) over all C 1
• Relationship to HCG? 3 2
 Local density  clique in HCG 5 6
 d  size of maximum clique in HCG 4
7
• Lower bound: 8
 # tracks  d
[©Keutzer]
 Left-edge Channel Routing Algorithm
• Used when no VCG edges
• Finds the optimal solution (# tracks = d)
• Nets are sorted according to their left endpoints

• Algorithm:
Create an initial track t
For all nets ni in the order of their left endpoints
if feasible to place the net on an existing track tj,
assign net ni to track tj.
else create a new track tnew and assign ni to it.

• Time complexity: O(n log n)

 Doglegs
• Doglegs can reduce the longest path in VCG
a a b c a a a b c a c2
c2
b
c1 b d
c
a b c d d c1
a b c d d d

• Doglegs break cycles in VCG

b a b a b1
a b1
?
a
b2
b
a b a b b2

[©Keutzer]
 Greedy Channel Router
• Many greedy algorithms for channel routing exist
• Example: Rivest and Fiduccia DAC’82
 Simple, linear algorithm
 Guarantees routing of all nets
 Uses doglegs (both restricted and unrestricted)
 BUT may extend to right hand side of the channel
• Other techniques
 Hierarchical: divide the channel into two smaller
channels, route each small channel, merge
 VCG reduction: nets that can be placed on the same
track merged into one VCG node to reduce VCG size
 Greedy Router: Rivest and Fiduccia

• Proceed column by column (left to right)

• Make connections to all pins in that column
• Free up tracks by collapsing as many tracks as possible to
collapse nets
• Shrink range of rows occupied by a net by using doglegs
• If a pin cannot enter a channel, add a track
• O(pins) time
[©Keutzer]
 Rivest and Fiduccia: Example
• Example output:

• Observations:
 Always succeeds
(even if cyclic conflict
is present)
 Allows unrestricted
doglegs
 Allows a net to occupy
more than 1 track at a
given column
 May use a few
columns off the edge [©Keutzer]
 Switchbox Routing
• Much harder problem than channel routing
 Two-dimensional problem
(channel routing was in a way one dimensional)
 Need to solve in a hierarchical flow
(split a channel
into two, route one
first, and route the
second as a
switchbox)
• A number of
complex heuristic
algorithms exist

[©Sarrafzadeh]
 FPGA Architecture - Layout
• Island FPGAs
 Array of functional units
 Horizontal and vertical routing
channels connecting the
functional units
 Versatile switch boxes
 Example: Xilinx, Altera
• Row-based FPGAs
 Like standard cell design
 Rows of logic blocks
 Routing channels (fixed width)
between rows of logic
 Example: Actel FPGAs
 FPGA Programmable Switch Elements
• Used in connecting:
 The I/O of functional units
to the wires

 A horizontal wire to a
vertical wire

 Two wire segments to

form a longer wire
segment
 FPGA Routing Channels Architecture
• Note: fixed channel widths (tracks)
• Should “predict” all possible connectivity
requirements when designing the FPGA chip
• Channel -> track -> segment
segment

track channel

• Segment length?
 Long: carry the signal longer,
less “concatenation” switches, but might waste track
 Short: local connections, slow for longer connections
 FPGA Switch Boxes
• Ideally, provide switches
for all possible connections

• Trade-off:
 Too many switches:
o Large area
o Complex to program
 Too few switches:
o Cannot route signals

One possible
solution
Xilinx 4000
 FPGA Routing
• Routing resources pre-fabricated
 100% routability using existing channels
 If fail to route all nets, redo placement
• FPGA architectural issues
 Careful balance between number of logic blocks and
routing resources (100% logic area utilization?)
 Designing flexible switchboxes and channels
(conflicts with high clock speeds)
• FPGA routing algorithms
 Graph search algorithms
o Convert the wire segments to graph nodes, and switch
elements to edges
 Bin packing heuristics (nets as objects, tracks as bins)
 Combination of maze routing and graph search
algorithms