Synthesis of Digital Systems
Part 4: FSM Encoding
Instructor: Preeti Ranjan Panda
Department of Computer Science and Engineering
Indian Institute of Technology Delhi
HLS Output – FSM + Datapath
a b c d
Control
Signals mux
S1 S2
+ +
Status
Signals mux
S3 e
f,g
FSM Datapath
(C) P. R. Panda, I.I.T Delhi, 2017 2
FSM Encoding / State Encoding /
State Assignment
• Assign codes to symbolic states in an FSM
– no 2 states should have the same code
10/00 10/00
S1 S2 00 01
0-, -1/01 1-/00 0-, -1/01 11/11 1-/00
11/11
10/01 10/01 0-/01
0-/01
S3 State 10
Encoding 0-/01
0-/01
(C) P. R. Panda, I.I.T Delhi, 2017 3
FSM and State Table
• One row in State Table for every FSM Transition
Inputs Present Next Outputs
10/00
State State
S1 S2 1 0 S1 S2 0 0
0-, -1/01 1-/00 0 - S1 S1 0 1
11/11
- 1 S1 S1 0 1
10/01 1 - S2 S2 0 0
0-/01
S3 0 - S2 S3 0 1
1 1 S3 S2 1 1
0-/01 0 - S3 S3 0 1
1 0 S3 S1 0 1
FSM State Table
(C) P. R. Panda, I.I.T Delhi, 2017 4
Why Encode an FSM?
• So we can implement it
• #bits in encoding DFFs in State Register Present State Lines
• Truth Table Next State and Output Logic
I/P PS NS O/P I/P PS NS O/P
State
1 0 S1 S2 0 0 1 0 00 01 0 0 Register Next
0 - S1 S1 0 1 0 - 00 00 0 1 State
- 1 S1 S1 0 1 - 1 00 00 0 1 Lines
1 - S2 S2 0 0 1 - 01 01 0 0
0 - S2 S3 0 1 0 - 01 10 0 1 Next State
1 1 S3 S2 1 1 1 1 10 01 1 1 and
0 - S3 S3 0 1 0 - 10 10 0 1 Output Logic Outputs
1 0 S3 S1 0 1 1 0 10 00 0 1
State Table Truth Table Inputs
(C) P. R. Panda, I.I.T Delhi, 2017 5
Encoding an FSM
inp PS NS outp
• Convert State Table to
00 st1 st2 101
Truth Table 01 st1 st3 --1
• Truth table can then be 1- st1 st1 111
-1 st2 st3 0--
optimised using standard -0 st2 st1 000
logic minimisation -- st3 st1 11-
techniques Encoding
st1: 00 inp PS NS outp
• Problem: Find encoding
st2: 01 00 00 01 101
that minimises area of st3: 10 01 00 10 --1
final circuit 1- 00 00 111
-1 01 10 0--
-0 01 00 000
-- 10 00 11-
(C) P. R. Panda, I.I.T Delhi, 2017 6
FSM Implementation Target: PLA
AND-plane OR-plane
State
FF FF
Register
Inputs Outputs
(C) P. R. Panda, I.I.T Delhi, 2017 7
FSM Implementation Target:
Multi-level Logic
Inputs Outputs
FF
FF State
Register
(C) P. R. Panda, I.I.T Delhi, 2017 8
FSM Encoding: What to Optimise?
• PLA implementation
– #rows x #cols in PLA
• Multi-level logic implementation
– Number of gates
• number of literals (ac’ + bc + a’b 6 literals)
(C) P. R. Panda, I.I.T Delhi, 2017 9
State Assignment Options
• Use minimum number of bits (log2 n)
– s0 (00), s1 (01), s2 (10), s3 (11)
– minimum number of flip flops
• One-hot encoding
– s0 (0001), s1 (0010), s2 (0100), s3 (1000)
– more flip flops, but may simplify logic
• no need of decode, possibly faster
• Any number of bits between minimum (log2 n)
and n
(C) P. R. Panda, I.I.T Delhi, 2017 10
Brute-force Search
• Generate all possible encodings
• Synthesise each case
• Choose the best
How many encodings for n-state FSM?
Let 2m-1 <n ≤ 2m
Assume m bits in state register (use minimum #bits)
# encodings = 2m x (2m-1) x (2m-2) x...x (2m-n+1)
(C) P. R. Panda, I.I.T Delhi, 2017 11
FSM Encoding for Multi-level Logic
• Given a State Table
• Generate Truth Table
• Minimise expected gate count in implementation
– without actually implementing
– predict the effect of logic optimisation
(C) P. R. Panda, I.I.T Delhi, 2017 12
Typical Multi-level Logic
Optimisations
Factoring ab + bc + d b (a + c) + d
s=a+b
Common ace + bce + de sce + de
Sub-expression ade + bde + af sde + af
t = bef
u = ae
Common Cube ace + bcef + de
tc + uc + de
(cube = product term) ade + bdef + af
td + ud + af
Objective: Maximise the number of available
common cubes (extracted later by logic optimisation)
(C) P. R. Panda, I.I.T Delhi, 2017 13
Common Next State
I PS NS O Let state bits = “abcde”
xy z Encoding:
S0 – 10011
-0 S0 S0 0 S1 – 11100
11 S0 S0 0 S2 – 10101
01 S0 S1 - S0 01/-
0- S1 S1 1
11 S1 S0 0 S1
10 S1 S2 1 xy abcde ABCDE z
1- S2 S2 1 S2 00/1 ...
00 S2 S1 1 01 10011 11100 -
01 S2 S3 1 00 10101 11100 1
0- S3 S3 1 ...
11 S3 S2 1
Truth Table
State Table
(C) P. R. Panda, I.I.T Delhi, 2017 14
Logic Equations after State
Assignment
I PS NS O
abcde
xy z
... State Reg
01 S0 S1 - ABCDE
00 S2 S1 1
... Comb. z
Logic
xy abcde ABCDE z xy
...
01 10011 11100 - A = x’yab’c’de + x’y’ab’cd’e + ... D = ...
00 10101 11100 1 B = x’yab’c’de + x’y’ab’cd’e + ... E = ...
... C = x’yab’c’de + x’y’ab’cd’e + ... z = x’y’ab’cd’e + ...
(C) P. R. Panda, I.I.T Delhi, 2017 15
Identifying Common Cubes
A = x’yab’c’de + x’y’ab’cd’e + ... s = ab’e
B = x’yab’c’de + x’y’ab’cd’e + ... A = x’yc’ds + x’y’cd’s + ...
C = x’yab’c’de + x’y’ab’cd’e + ... B = x’yc’ds + x’y’cd’s + ...
D = ... C = x’yc’ds + x’y’cd’s + ...
E = ... D = ...
z = x’y’ab’cd’e + ... E = ...
z = x’y’ab’cd’e + ...
Encoding (N bits):
S0 01/- S0 – 10011
S1 – 11100
S1 S2 – 10101
Hamming Distance H (S0, S2) = 2
S2 00/1 Lines for A, B, C have common cube of size N-H = 5-2 = 3
Similar observation for O/Ps
(C) P. R. Panda, I.I.T Delhi, 2017 16
Common Present State
I PS NS O Let state bits = “abcde”
xy z Encoding:
S0 – 10011
-0 S0 S0 0 S1 – 11100
11 S0 S0 0 S0 S2 – 10101
01 S0 S1 -
11/0
0- S1 S1 1 S1
11 S1 S0 0
10 S1 S2 1 10/1 S2 xy abcde ABCDE z
1- S2 S2 1 ...
00 S2 S1 1 11 11100 10011 0
01 S2 S3 1 10 11100 10101 1
0- S3 S3 1 ...
11 S3 S2 1
Truth Table
State Table
(C) P. R. Panda, I.I.T Delhi, 2017 17
Common Cubes after State
Assignment
I PS NS O
A = xyabcd’e’ + xy’abcd’e’ + ... S0
xy z
B = ...
C = xy’abcd’e’ + ... 11/0
... S1
D = xyabcd’e’ + ...
11 S1 S0 0
E = xyabcd’e’ + xy’abcd’e’ + ...
10 S1 S2 1 10/1 S2
... z = xy’abcd’e’ + ...
Encoding (N bits):
S0 – 10011
s = abcd’e’ S1 – 11100
A = xys + xy’s + ... S2 – 10101
xy abcde ABCDE z
B = ... Hamming Distance (S0, S2) = H
...
C = xy’s + ... Up to N-H = 3 Next State Lines (A, E
11 11100 10011 0
D = xys + ... in this example) have common cube
10 11100 10101 1
E = xys + xy’s + ... of size N = 5
...
z = xy’s + ... (irrespective of Encoding for S1)
Similar observation for I/Ps
(C) P. R. Panda, I.I.T Delhi, 2017 18
Strategy
• For an FSM, we can identify many such
relationships
– distance of codes vs. size/number of common cubes
• Estimate reduction in literals (or “gain”) if two
states are assigned close codes
• Assign codes that maximise the number/size of
common cubes
• Assume that logic optimiser will find and exploit
the common cubes
(C) P. R. Panda, I.I.T Delhi, 2017 19
Formulation
• Build Undirected Graph
– Nodes: states
– Edge Weight: gain if states are assigned
close codes
• Assign an encoding that maximises total
gain
(C) P. R. Panda, I.I.T Delhi, 2017 20
Building the Graph:
a Fanout-oriented Algorithm
• PS’s with the same NS and O/P get
high weights
– leads to close codes eventually
– maximise size of common cubes
ab PS NS t
---------- high s0: 0000
01 s0 s1 0 s0 s2 s2: 0001
00 s2 s1 1
(C) P. R. Panda, I.I.T Delhi, 2017 21
Fanout: Constructing Output Sets
• If 2 PS’s assert I PS NS O
xy z
same o/p,
common cube -0 S0 S0 0
can be extracted 11 S0 S0 0
01 S0 S1 -
OUTPUT_SETZ
• Construct No 0- S1 S1 1
= {S1 (2), S2 (3), S3 (2)}
different o/p sets 11 S1 S0 0
10 S1 S2 1
• Assign weight = 1- S2 S2 1
#occurences of 00 S2 S1 1
01 S2 S3 1
the PS for same 0- S3 S3 1
o/p 11 S3 S2 1
(C) P. R. Panda, I.I.T Delhi, 2017 22
Fanout: Constructing NS Sets
• If 2 PS’s produce
same NS, common I PS NS O
cube can be xy z
extracted
– intersection of the -0 S0 S0 0
2 PS encodings NS_SETS0 = {S0 (2), S1 (1)}
11 S0 S0 0
• How many common 01 S0 S1 - NS_SETS1 = {S0 (1), S1 (1), S2 (1)}
cubes? 0- S1 S1 1
– # 1’s in Next State 11 S1 S0 0 NS_SETS2 = {S1 (1), S2 (1), S3 (1)}
encoding 10 S1 S2 1
– ≅ Nb / 2 1- S2 S2 1 NS_SETS3 = {S2 (1), S3 (1)}
• Construct Ns 00 S2 S1 1
different NS sets 01 S2 S3 1
0- S3 S3 1
• Assign weight =
11 S3 S2 1
#occurences of the
PS for same NS
(C) P. R. Panda, I.I.T Delhi, 2017 23
Fanout: Constructing the Graph
OUTPUT_SETZ = {S1 (2), S2 (3), S3 (2)}
• Node = state
NS_SETS0 = {S0 (2), S1 (1)}
• Edge Weight (a-b) NS_SETS1 = {S0 (1), S1 (1), S2 (1)}
– multiply node weights in NS_SETS2 = {S1 (1), S2 (1), S3 (1)}
each OUTPUT_SET
NS_SETS3 = {S2 (1), S3 (1)}
• if a, b are in same set
• each pair of transitions For Nb = 2:
has a common cube Edge Weight (S2 - S3) =
3x2 + (1x1 + 1x1) x 2/2 = 8
– multiply node weights in
each NS_SET 3
• if a, b are in same set S0 S1
0
• scale by Nb / 2 1 8 5
S2 S3
8
(C) P. R. Panda, I.I.T Delhi, 2017 24
Graph Embedding
• Start from Graph we 3
S0 S1
constructed 0
8
1 5
– fanin
– fanout S2 S3
2
• Assign codes to
minimise Cost Cost = Σedges a→b wt(a→b) × HD (a→b)
HD = Hamming Distance
Encoding 1 Encoding 2
S0: 00 Cost = S0: 00 Cost =
S1: 01 3x1 + 5x1 + 2x1 + S1: 11 3x2 + 5x1 + 2x2 +
S2: 10 1x1 + 0x2 + 8x2 = 27 S2: 01 1x1 + 0x1 + 8x1 = 24
S3: 11 S3: 10
(C) P. R. Panda, I.I.T Delhi, 2017 25
Solution to Graph Embedding
Problem
• Well studied problem
• Practical observation
– for large FSMs, clusters of nodes with high edge
weights exist
• Heuristic – in what order to choose nodes for
assignment?
– identify heavy clusters and assign codes to their
states
– remove these states from consideration
– proceed with remaining graph
(C) P. R. Panda, I.I.T Delhi, 2017 26
Embedding Algorithm
1. Select node n with maximum sum Nb = 3
of Nb incident edge weight 4
a b
2. Assign some code to n and
3 4 0
minimally distant codes to the Nb 2
6
adjacent nodes e c
1 2
allows all Nb nodes to be at distance 1
3 2
from n d
3. Remove node n (and all incident
Select b (max sum: 4+6+2)
edges) from graph
b – 000
4. Repeat (Step 1) with smaller a – 001 (distance from b = 1)
graph d – 010 (distance from b = 1)
e – 100 (distance from b = 1)
(C) P. R. Panda, I.I.T Delhi, 2017 27
Embedding Algorithm - Example
4 001 a
a b
3 3 4
2 4 0 2
6 1
e c 100 e c
1 2
3 2 3 2
d d
010
Select b (max sum: 4+6+2) Select a (max sum: 3+2+4)
b – 000 c – 011 (distance from a = 1)
a – 001
d – 010
e – 100
(C) P. R. Panda, I.I.T Delhi, 2017 28
