ESE535:

Electronic Design Automation

Day 19: March 27, 2013

Sequential Optimization
(FSM Encoding)

1
Penn ESE535 Spring 2013 -- DeHon
 Behavioral
(C, MATLAB, …)

Today Arch. Select

Schedule
RTL
FSM assign
• Encoding Two-level
– Input Multilevel opt.
Covering
– Output Retiming

• State Encoding Gate Netlist

Placement
– “exact” two-level Routing
Layout

Masks
2
Penn ESE535 Spring 2013 -- DeHon
 Input Encoding
• Pick codes for input cases to simplify
logic
• E.g. Instruction Decoding
– ADD, SUB, MUL, OR
• Have freedom in code assigned
• Pick code to minimize logic
– E.g. number of product terms

3
Penn ESE535 Spring 2013 -- DeHon
 Output Encoding
• Opposite problem
• Pick codes for output symbols
• E.g. allocation selection
– Prefer N, Prefer S, Prefer E, Prefer W, No
Preference
• Again, freedom in coding
• Use to maximize sharing
– Common product terms, CSE
4
Penn ESE535 Spring 2013 -- DeHon
 Finite-State Machine
• Logical behavior depends on state
• In response to inputs, may change state

0/1 1/0 -/1

1/0 0/0

5
Penn ESE535 Spring 2013 -- DeHon
 State Encoding
• State encoding is a logical entity
• No a priori reason any particular state
has any particular encoding
• Use freedom to simply logic

6
Penn ESE535 Spring 2013 -- DeHon
 Finite State Machine

0/1 1/0 -/1

1/0 0/0

0 S1 S1 1
1 S1 S2 0
1 S2 S2 0
0 S2 S3 0
1 S3 S3 1
0 S3 S3 1
7
Penn ESE535 Spring 2013 -- DeHon
 Example: Encoding Difference
0 01 01 1
1 01 11 0 0 01 01 1
0 S1 S1 1 S1=01 1 -1 11 0
1 11 11 0
1 S1 S2 0 S2=11 - 11 10 0
0 11 10 0
1 S2 S2 0 S3=10 - 10 10 1
1 10 10 1
0 S2 S3 0
0 10 10 1
1 S3 S3 1 S1+S2 = -1
0 S3 S3 1
0 00 00 1 1 11 01 0
Similar outputs, 1 00 11 0 1 00 11 0
code so S1+S2 1 11 11 0 0 -0 00 1
is simple cube S1=00 0 11 10 0 - 10 00 1
S2=11 1 10 10 1 - 1- 10 0
S3=10 0 10 10 1 8
Penn ESE535 Spring 2013 -- DeHon
 Problem:
• Real: pick state encodings (si’s) so as to
minimize the implementation area
– two-level
– multi-level
• Simplified variants
– minimize product terms
– achieving minimum product terms, minimize state
size
– minimize literals
9
Penn ESE535 Spring 2013 -- DeHon
 Two-Level

• Apla = (2*ins+outs)*prods+ flops*wflop

• inputs = PIs + state_bits
• outputs = state_bits+POs
• products terms (prods)
– depend on state-bit encoding
– this is where we have leverage 10
Penn ESE535 Spring 2013 -- DeHon
 Multilevel
• More sharing  less implementation
area
• Pick encoding to increase sharing
– maximize common sub expressions
– maximize common cubes
• Effects of multi-level minimization hard
to characterize (not predictable)

11
Penn ESE535 Spring 2013 -- DeHon
 Two-Level Optimization
1. Idea: do symbolic minimization of two-level
form
– This represents effects of sharing
2. Generate encoding constraints from this
– Properties code must have to maximize sharing
3. Cover
– Like two-level (mostly…)
4. Select Codes

12
Penn ESE535 Spring 2013 -- DeHon
 Kinds of Sharing
Input sharing: Output sharing:
encode inputs so share input cubes
cover set to reduce to produce individual
product terms output bits
10 inp1+inp2=01
Out1=11
10 inp1 01 11 inp2+inp3=01
1101 out1 Out2=01
01 inp1 10 Out3=10
10 1- 01 1100 out2
1- inp2 01 Out4=00
Inp1=10 01 10 10 1111 out3
01 inp2 01
Inp2=11 11 –1 01 0000 out4
11 inp3 01 110- 01
Inp3=01 01 11 01 0001 out4
01 inp3 10 11-1 10
01 01 10
13
Penn ESE535 Spring 2013 -- DeHon
 Input Encoding

14
Penn ESE535 Spring 2013 -- DeHon
 Two-Level Input Oriented
• Minimize product rows
– by exploiting common-cube
– next-state expressions

– Does not account for possible sharing of

terms to cover outputs

[DeMicheli+Brayton+SV/TR CAD v4n3p269] 15

Penn ESE535 Spring 2013 -- DeHon
 Outline Two-Level Input
• Represent states as one-hot codes
• Minimize using two-level optimization
– Include: combine compatible next states
• 1 S1 S2 0
• 1 S2 S2 0  1 {S1,S2} S2 0
• Get disjunct on states deriving next state
• Assuming no sharing due to outputs
– gives minimum number of product terms
• Cover to achieve
– Try to do so with minimum number of state bits
16
Penn ESE535 Spring 2013 -- DeHon
 Multiple Valued Input Set
• Treat input states as a multi-valued (not
just 0,1) input variable
• Effectively encode in one-hot form
– One-hot: each state gets a bit, only one on
• Use to merge together input state sets
0 S1 S1 1 0 100 S1 1
1 S1 S2 0 1 100 S2 0
1 S2 S2 0 1 010 S2 0
0 S2 S3 0 0 010 S3 0
1 S3 S3 1 1 001 S3 1
0 S3 S3 1 0 001 S3 1 17
Penn ESE535 Spring 2013 -- DeHon
 One-hot Minimum
• One-hot gives minimum number of
product terms
• i.e. Can always maximally combine
input sets into single product term

18
Penn ESE535 Spring 2013 -- DeHon
 One-hot example

10 100 01
10 inp1 01 One-hot: 01 100 10
01 inp1 10 inp1=100 1- 010 01 10 --0 01
1- inp2 01 inp2=010 01 010 10 11 0-- 01
01 inp2 10 inp3=001 11 001 01 01 --- 10
11 inp3 01 01 001 10
01 inp3 10 -- 11- --
-- 1-1 --
Key: can define a cube to -- -11 --
cover any subset of states -- 000 --
10 --0 01 says 10*(inp1+inp2)  01 19
Penn ESE535 Spring 2013 -- DeHon
 State Combining
• Follows from standard 2-level
optimization with don’t-care
minimization
• Effectively groups together common
predecessor states as shown
• (can define to combine directly)

20
Penn ESE535 Spring 2013 -- DeHon
 Example
0 S s6 00 0 1000000 0000010 00
0 s2 s5 00 0 0100000 0000100 00
0 s3 s5 00 0 0010000 0000100 00
0 s4 s6 00 0 0001000 0000010 00 0 0110001 0000100 00
0 s5 S 10 0 0000100 1000000 10 0 1001000 0000010 00
0 s6 S 01 0 0000010 1000000 01 1 0001001 0000010 10
0 s7 s5 00 0 0000001 0000100 00 0 0000010 1000000 01
1 s6 s2 01 1 0000010 0100000 01 1 0000100 0100000 10
1 s5 s2 10 1 0000100 0100000 10 0 0000100 1000000 10
1 s4 s6 10 1 0001000 0000010 10 1 1000000 0001000 00
1 s7 s6 10 1 0000001 0000010 10 1 0000010 0100000 01
1 S s4 00 1 1000000 0001000 00 1 0100000 0010000 00
1 s2 s3 00 1 0100000 0010000 00 1 0010000 0000001 00
1 s3 s7 00 1 0010000 0000001 00 21
Penn ESE535 Spring 2013 -- DeHon
 Two-Level Input
• One-hot identifies multivalue minimum number of
product terms
• May be fewer product terms if get sharing (don’t
cares) in generating the next state expressions
– (was not part of optimization)
• Encoding places each disjunct on a unique cube
face
– Can distinguish with a single cube
• Can use fewer bits than one-hot
– this part typically heuristic
– Remember one-hot already minimized prod terms

22
Penn ESE535 Spring 2013 -- DeHon
 0 S s6 00
Encoding Example
0 s2 s5 00
S 010
0 s3 s5 00
0 0110001 0000100 00 s2 110
0 s4 s6 00
0 1001000 0000010 00 s3 101
0 s5 S 10
1 0001001 0000010 10 s4 000
0 s6 S 01
0 0000010 1000000 01 s5 001
0 s7 s5 00
1 0000100 0100000 10 s6 011
1 S s4 01
0 0000100 1000000 10 s7 100
1 s2 s3 10
1 s3 s7 10 1 1000000 0001000 00
1 0000010 0100000 01 s2+s3+s7=1--
1 s4 s6 10
1 0100000 0010000 00 No 111 code
1 s5 s2 00
1 s6 s2 00 1 0010000 0000001 00
1 s7 s6 00 23
Penn ESE535 Spring 2013 -- DeHon
 Encoding Example
S 010
0 0110001 0000100 00 0 1-- 001 00
s2 110
0 1001000 0000010 00 0 0-0 011 00
s3 101
1 0001001 0000010 10 1 -00 011 10
s4 000
0 0000010 1000000 01 0 011 010 01
s5 001
1 0000100 0100000 10 1 001 110 10
s6 011
0 0000100 1000000 10 0 001 010 10
s7 100
1 1000000 0001000 00 1 010 000 00
1 0000010 0100000 01 1 011 110 01
s4+s7=-00 1 110 101 00
1 0100000 0010000 00
1 0010000 0000001 00 S+s4=0-0 1 101 100 00

s2+s3+s7=1--
24
Penn ESE535 Spring 2013 -- DeHon (no 111 code)
 Input and Output

Skip?

25
Penn ESE535 Spring 2013 -- DeHon
 General Problem
• Track both input and output encoding
constraints

26
Penn ESE535 Spring 2013 -- DeHon
 General Two-Level Strategy
1. Generate “Generalized” Prime
Implicants
2. Extract/identify encoding constraints
3. Cover with minimum number of GPIs
that makes encodeable
4. Encode symbolic values

[Devadas+Newton/TR CAD v10n1p13] 27

Penn ESE535 Spring 2013 -- DeHon
 Output Symbolic Sets
• Maintain output state, PIs as a set
• Represent inputs one-hot as before

0 S1 S1 1 0 100 S1 1 0 100 (S1) (o1)

1 S1 S2 0 1 100 S2 0 1 100 (S2) ()
1 S2 S2 0 1 010 S2 0 1 010 (S2) ()
0 S2 S3 0 0 010 S3 0 0 010 (S3) ()
1 S3 S3 1 1 001 S3 1 1 001 (S3) (o1)
0 S3 S3 1 0 001 S3 1 0 001 (S3) (o1)

28
Penn ESE535 Spring 2013 -- DeHon
 Generate GPIs
• Same basic idea as PI generation
– Quine-McKlusky
• …but different

29
Penn ESE535 Spring 2013 -- DeHon
 Merging
• Cubes merge if
– distance one in input
• 000 100
• 001 100  00- 100
– inputs same, differ in multi-valued input
(state)
• 000 100
• 000 010  000 110

30
Penn ESE535 Spring 2013 -- DeHon
 Merging
• When merge
– binary valued output contain outputs
asserted in both (and)
• 000 100 (foo) (o1,o2)
• 001 100 (bar) (o1,o3)  00- 100 ? (o1)
– next state tag is union of states in merged
cubes
• 000 100 (foo) (o1,o2)
• 001 100 (bar) (o1,o3)  00- 100 (foo,bar) (o1)

31
Penn ESE535 Spring 2013 -- DeHon
 Merged Outputs
• Merged outputs
– Set of things asserted by this input
– States would like to turn on together
• 000 100 (foo) (o1,o2)
• 001 100 (bar) (o1,o3)  00- 100 (foo,bar) (o1)

32
Penn ESE535 Spring 2013 -- DeHon
 Cancellation
• K+1 cube cancels k-cube only if
– multivalued input is identical
– AND next state and output identical
• 000 100 (foo) (o1)
• 001 100 (foo) (o1)
– Also cancel if multivalued input contains all inputs
• 000 111 (foo) (o1)
• Discard cube with next state containing all
symbolic states and null output
– 111 100 (foo,bar,baz…) ()  does nothing

33
Penn ESE535 Spring 2013 -- DeHon
 Example
(copy to board…work;
Note inclass exercise, back of preclass)
0 100 (S1) (o1)
1 100 (S2) ()
1 010 (S2) ()
0 010 (S3) ()
1 001 (S3) (o1)
0 001 (S3) (o1)

34
Penn ESE535 Spring 2013 -- DeHon
 Cancellation
• K+1 cube cancels k-cube only if
– multivalued input is identical
– AND next state and output identical
• 000 100 (foo) (o1) 00- 100 (foo) (o1)
• 001 100 (foo) (o1)
– Also cancel if multivalued input contains all inputs
• 000 111 (foo) (o1)
• Discard cube with next state containing all
symbolic states and null output
– 111 100 (foo,bar,baz….) ()  does nothing

35
Penn ESE535 Spring 2013 -- DeHon
 Example

0 100 (S1) (o1) - 100 (S1,S2) () 0 111 (S1,S3) ()

1 100 (S2) () 0 110 (S1,S3) () x - 011 (S2,S3) ()
1 010 (S2) () 0 101 (S1,S3) (o1) 1 111 (S2,S3) ()
0 010 (S3) () 1 110 (S2) - 110 (S1,S2,S3) () x
1 001 (S3) (o1) x 1 101 (S2,S3) () x
0 001 (S3) (o1) x - 010 (S2,S3) ()
1 011 (S2,S3) () x
0 011 (S3) ()
- 001 (S3) (o1)

36
Penn ESE535 Spring 2013 -- DeHon
 Covering
• Cover with branch-and-bound similar to two-
level
– row dominance only if
• tags of two GPIs are identical
• OR tag of first is subset of second
• Once cover, check encodeability
– [talk about next]
• If fail, branch-and-bound again on additional
GPIs to add to satisfy encodeability

37
Penn ESE535 Spring 2013 -- DeHon
 Encoding Constraints
0 S1 S1 1
• Minterm to symbolic state v 1 S1 S2 0
– should assert v 1 S2 S2 0
0 S2 S3 0
1 S3 S3 1
0 S3 S3 1

• For all minterms m

 all GPIs [(all symbolic tags) e(tag state)] = e(v)

38
Penn ESE535 Spring 2013 -- DeHon
 Example
all GPIs [(all symbolic tags) e(tag state)] = e(v)
Consider 1101 (out1)
covered by
1101 out1 110- (out1,out2) 110- (out1,out2)
1100 out2 11-1 (out1,out3) 11-1 (out1,out3)
1111 out3 000- (out4)
x 0000 out4 110-  e(out1)  e(out2)
x 0001 out4 11-1  e(out1)  e(out3)
OR-plane gives me OR of these two

Want output to be e(out1)

1101 e(out1) e(out2)  e(out1) e(out3)=e(out1)
39
Penn ESE535 Spring 2013 -- DeHon
 Example
all GPIs [(all symbolic tags) e(tag state)] = e(v)
Sample Solution:
out1=11
1101 out1 out2=01 110- 01
110- (out1,out2)
1100 out2 out3=10 11-1 10
11-1 (out1,out3)
1111 out3 out4=00
000- (out4)
x 0000 out4 Think about PLA
x 0001 out4
1101 e(out1) e(out2)  e(out1) e(out3)=e(out1)
1100 e(out1) e(out2)=e(out2)
1111 e(out1) e(out3)=e(out3)
0000 e(out4)=e(out4)
0001 e(out4)=e(out4)
40
Penn ESE535 Spring 2013 -- DeHon
 To Satisfy
• Dominance and disjunctive relationships from
encoding constraints
> Means strictly
• e.g. more bits on
– e(out1)  e(out2)  e(out1)  e(out3)=e(out1)
– one of:
• e(out2)>e(out1) [i.e. e(out1) e(out2)=e(out1)]
• e(out3)>e(out1) [i.e. e(out1) e(out3)=e(out1)]
• e(out2)|e(out3) e(out1)

41
Penn ESE535 Spring 2013 -- DeHon
 Encodeability Graph
1100 One of:
e(out2)>e(out1)
out1 out2 e(out3)>e(out1)
out4 e(out1)e(out2)|e(out3)
1111

out3 1101 e(out1) e(out2)  e(out1) e(out3)=e(out1)

1100 e(out1) e(out2)=e(out2)
1111 e(out1) e(out3)=e(out3)
0000 e(out4)=e(out4)
0001 e(out4)=e(out4) 42
Penn ESE535 Spring 2013 -- DeHon
 Encoding Constraints
• No directed cycles (proper dominance)
• Siblings in disjunctive have no directed
paths between
– (one cannot dominate other)
• No two disjunctive equality can have
exactly the same siblings for different
parents
• Parent of disjunctive should not
dominate all sibling arcs
43
Penn ESE535 Spring 2013 -- DeHon
 Encodeability Graph
1100 One of:
cycle e(out2)>e(out1)
out1 X out2 cycle e(out3)>e(out1)
e(out1)e(out2)|e(out3)
1111 X
out4 No cycles  encodeable
out3 1101 e(out1) e(out2)  e(out1) e(out3)=e(out1)
1100 e(out1) e(out2)=e(out2)
1111 e(out1) e(out3)=e(out3)
0000 e(out4)=e(out4)
0001 e(out4)=e(out4) 44
Penn ESE535 Spring 2013 -- DeHon
 Determining Encoding
• Can turn into boolean satisfiability
problem for a target code length
• All selected encoding constraints
become boolean expressions
• Also uniqueness constraints

45
Penn ESE535 Spring 2013 -- DeHon
 What we’ve done
• Define another problem
– Constrained coding
• This identifies the necessary coding
constraints
– Solve optimally with SAT solver
– Or attack heuristically

46
Penn ESE535 Spring 2013 -- DeHon
 Summary
• Encoding can have a big effect on area
• Freedom in encoding allows us to maximize
opportunities for sharing
• Can do minimization around unencoded to
understand structure in problem outside of
encoding
• Can adapt two-level covering to include and
generate constraints
• Multilevel limited by our understanding of
structure we can find in expressions
– heuristics try to maximize expected structure

47
Penn ESE535 Spring 2013 -- DeHon
 Today’s Big Ideas
• Exploit freedom
• Bounding solutions
• Dominators
• Formulation and Reduction
• Technique:
– branch and bound
– SAT
– Understanding structure of problem
• Creating structure in the problem 48
Penn ESE535 Spring 2013 -- DeHon
 Admin
• Assignment 6, 7 out
– For Assignment 6 you essentially write the
assignment for 7
• Monday Reading on Web
• Makeup office hour today (W) at 5pm

49
Penn ESE535 Spring 2013 -- DeHon