VLSI CAD: Lecture 6.

1
Logic to Layout Logic Synthesis:
Multilevel Logic and
the Boolean Network Model
Rob A. Rutenbar
University of Illinois
Why Multi-level Logic?
•  2-level forms are too restrictive: specific area vs. delay tradeoff
•  Area = gates + literals (wires), i.e., things that take space on a chip
•  Delay = maximum levels of logic gates required to compute function
•  2-level is minimum gate delay possible, but usually worst on area

DELAY Multi-level designs =

fewer gates, but > 2 levels

slower, >2 levels

Typical 2-level design =
many gates, but only 2 levels
of logic, so fastest possible
fastest, 2 levels
AREA
small, big,
Slide 2 few gates+literals many gates+literals © 2013, R.A. Rutenbar
Why Multilevel?
•  Rarely see 2-level designs for really big things…
•  We use 2-level logic mostly for pieces of bigger things
•  Even smallish things routinely done as multi-level

1 2 3 4 999
?
~1000 gate
“block” of logic
1000
This is just NOT going to be the
preferred logic network structure...

Slide 3 © 2013, R.A. Rutenbar

Real Multilevel Example
•  …this is a small design, done by commercial synthesis tool

Levels of
logic in
1 2 3 4 5 6 7 8 9 10 11 network

Slide 4 © 2013, R.A. Rutenbar

Boolean Logic Network Model
•  Need more sophisticated model: Boolean Logic Network
•  Idea: it’s a graph of connected blocks, like any logic diagram, but now individual
component blocks can be 2-level Boolean functions in SOP form

…now as a Boolean logic network,

Ordinary gate logic x, y are now Boolean functions
a x a
b x=ab x Bug fix:
y b x+c, not b+c
c
y=x+c y
c

primary internal primary

inputs vertices outputs
Slide 5 © 2013, R.A. Rutenbar
Multilevel Logic: What to Optimize?
•  This is simplistic but surprisingly useful: Total literal count
•  Count every appearance of every variable on right hand side of “=“ in every node
•  (Yes, delays matter too, but for this class, only focus on logic complexity)

#Literals =
a
b Q = aX
Y = b+c Q
c
X = Yd + Z
Z = b•c

Slide 6 © 2013, R.A. Rutenbar

Optimizing Multilevel Logic: Big Ideas
•  Again: this is a data structure. What operators do we need?

•  3 basic kinds of operators

•  Simplify network nodes: no change in # of nodes, just simplify insides, SOP form
•  à You already know this! This is 2-level synthesis, this is ESPRESSO!
•  Remove network nodes: take “too small” nodes, substitute them back into fanouts
•  à This is not too hard. This is mostly manipulating the graph, simple SOP edits.
•  Add new network nodes: this is factoring. Take big nodes, split into smaller nodes.
•  à This is a big deal. This is new. This is what we need to teach you…

Slide 7 © 2013, R.A. Rutenbar

Optimizing Boolean Logic Network
•  Simplifying a node •  Removing a node
•  Just run ESPRESSO on 2-level form •  Typical case is you have a “small”
inside the node, to reduce # literals factor which doesn’t seem to be worth
making it a separate node
•  As structural changes happen across
network, “insides” of nodes may •  “Push” it back into its fanouts, make
present opportunity to simplify those nodes bigger, and hope you
can use 2-level simplification on them

X = a +ab +bc X = qZ + stuff X = qab + stuff

Z = ab
X = a + bc Y = cdZ + stuff Y = cdab + stuff

Slide 8 © 2013, R.A. Rutenbar

Optimizing Boolean Logic Network
•  Adding new node(s): this is Factoring, this is new, and hard
•  Look at existing nodes, identify common divisors, extract them, connect as fanins
•  Tradeoff: more delay (another level of logic), but fewer literals (less gate area)

X = ab + c + r X=Q+r

Y = abd + cd Q = ab+c Y = Qd

divisor
Z = abrs+crs Z = Qrs
16 literals 10 literals
Slide 9 © 2013, R.A. Rutenbar
VLSI CAD: Lecture 6.2
Logic to Layout Logic Synthesis:
Multilevel Logic:
Algebraic Model for
Factoring
Rob A. Rutenbar
University of Illinois
Aside: Multilevel Synthesis Scripts
•  Multilevel synthesis–like 2-level synthesis–is heuristic
•  …but it’s also more complex. Write scripts of basic operators
•  Do several passes of different optimizations over the Boolean logic network
•  Do some “cleanup” steps to get rid of “too small” nodes (remove them)
•  Look for “easy” factors, just sitting around as existing nodes, and try to use them
•  Look for “hard” factors, do the work to extract them, try them, keep the good ones
•  Do 2-level optimization of insides of each logic node in network (ESPRESSO)
•  Lots of “art” in the engineering of these scripts

•  For us, the one big thing you don’t know: How to factor…

Slide 11 © 2013, R.A. Rutenbar

Another New Model: Algebraic Model
•  Factoring: How do we really do it?
•  Develop another model for Boolean functions, cleverly designed to let us do this
•  Tradeoff: lose some “expressivity” -- some aspects of Boolean behavior and some
Boolean optimizations we just cannot do, but we gain practical factoring.

•  New model: Algebraic model

•  Surprise: Term “algebraic” comes from pretending that Boolean expressions
behave like polynomials of real numbers, not like Boolean algebra
•  Big new Boolean operator: Algebraic Division (or, also “Weak” Division)

Slide 12 © 2013, R.A. Rutenbar

Algebraic Model
•  Idea: keep just those rules (axioms) that work for polynomials of
reals AND Boolean algebra, dump the rest
Real numbers Boolean algebra
a•b = b•a a•b = b•a
a+b = b+a a+b = b+a
a•(b•c) = (a•b)•c a•(b•c) = (a•b)•c
a+(b+c) = (a+b)+c SAME a+(b+c) = (a+b)+c
a•(b+c) = a•b + a•c a•(b+c) = a•b+a•c
a•1 = a a•0 = 0 a•1 = a a•0 = 0
a+0 = a a+0 = a

x
a+a’ = 1 a•a’ = 0
NOT a•a = a a+a = a
ALLOWED a+1 = 1
a+(b•c) = (a+b)•(a+c)

Slide 13 © 2013, R.A. Rutenbar

Algebraic Model
•  If we only get to use algebra rules from real numbers….
•  Consequence: A variable and its complement must be treated as totally unrelated
•  Aside: this is one of the losses of “expressive power” for Booleans we just tolerate

F = ab + a’x + b’y
Let R = a’
Let S = b’
ab + Rx + Sy
•  Idea
•  Boolean functions manipulated as SOP expressions–like polynomials
•  Each product term in such an expression is just a set of variables, e.g., abRy
•  SOP expression itself is just a list of these products (cubes), e.g. ab + Rx = ab,Rx
Slide 14 © 2013, R.A. Rutenbar
Algebraic Division: Our Model for Factoring
•  Given function D we want to factor as:

F = D•Q + R
divisor remainder (if =0, then we say the divisor
quotient is ‘technically’ called a Factor)

Example with REAL Example with BOOLEAN SOP

15 = 7•2 + 1 F = ac + ad + bc + bd + e
7 = D = divisor = (a+b)•(c+d) + e
2 = Q = quotient
1 = R = remainder (a+b) = D = divisor
(c+d) = Q = quotient
e = R = remainder
Slide 15 © 2013, R.A. Rutenbar
Algebraic Division
•  Example F = ac + ad + bc + bd + e want F = D • Q + R
•  Reminder: Quotient is only called a “Factor” if remainder is 0

Divisors (D) Quotient (Q) Remainder (R) Factor?

ac+ad+bc+bd+e 1 0 Yes (trivial)
Bug fix:
a+b c+d Q e No
c+d a+b e No
a c+d bc+bd+e No
b c+d ac+ad+e No
c a+b ad+bd+e No
d a+b ac+bc+e No
e 1 ac+ad+bc+bd No

Slide 16 © 2013, R.A. Rutenbar

VLSI CAD: Lecture 6.3
Logic to Layout Logic Synthesis:
Multilevel Logic:
Algebraic Division
Rob A. Rutenbar
University of Illinois
Algebraic Division: Very Nice Algorithm
•  Inputs
•  A Boolean expression F and a divisor (to divide by) D, represented as lists of cubes
(and each cube a set of literals)

•  Output
•  Quotient Q = F/D = cubes in quotient, or 0 (empty) if none
•  Remainder R = cubes in remainder, or 0 (empty) if D was a factor
•  i.e., figures out Q, R so that F = D•Q+ R = D•(F/D) + R

•  Strategy
•  Cubewise walk thru cubes in divisor D, trying to divide each of them into F
•  ...being careful to track which cubes do divide into F

Slide 18 © 2013, R.A. Rutenbar

 Algebraic Division Algorithm
AlgebraicDivision( F, D) { // divide D into F Example:
Cube xyzw contains
for ( each cube d in divisor D ) { product term “yz”
let C = { cubes in F that contain this product term “d” };
if ( C is empty ) {
return ( quotient = 0, remainder = F);
}
let C = cross out literals of cube “d” in each cube of C;
if ( d is the first cube we have looked at in divisor D )
let Q = C;
else Q = Q ∩ C;
} Example:
R = F - ( Q • D ); Suppose C = xyz + yzw +pqyz
return ( quotient = Q, remainder = R) and d = “yz”. Then crossing
} out all the “yz” parts yields
Example: x + w + pq
Suppose Q = xyz + yzw +pqyz
and C = “ xyz”. Then Q ∩ C
is just the cubes that are in both Q and C
Slide 19 © 2013, R.A. Rutenbar
in this case: xyz
 Algebraic Division: Example
F/D: F = axc + axd + axe + bc + bd + de D = ax + b
D cube: ax D cube: b Easiest way manually is to make this table:
F cube
1 C=… 3 C=… one row per cube in F, one column per cube in D,
axc axcàaxcàc -- bottom row to evolve Quotient Q and, when done,
remember to get remainder R
axd axdàaxdàd --
AlgebraicDivision( F, D) { // divide D into F
axe axeàaxeàe -- for ( each cube d in divisor D ) {
bc -- bcàbcàc let C = { cubes in F that contain this product term “d” };
if ( C is empty ) {
Bug fix: return ( quotient = 0, remainder = F);
bd axe -- bdàbdàd }
let C = cross out literals of cube “d” in each cube of C;
if ( d is the first cube we have looked at in divisor D )
de -- -- let Q = C;
else Q = Q ∩ C;

2 Q = c+d+e 4 Q = (c+d) ∩ }
R = F - ( Q • D );
(c+d+e) }
return ( quotient = Q, remainder = R)

= c+d

Remainder R = F–Q•D: R = (axc + axd + axe + bc + bd + de) – (c+d)•(ax+b) 5

(axc + axd + axe + bc + bd + de) – (axc + axd + bc + bd)
“—” means remove cubes in
Q•D that appear same in F = (axe + de) = remainder
Slide 20 © 2013, R.A. Rutenbar
Algebraic Division: Warning
•  Remember: No “Boolean” simplification, only “algebraic”
•  So what? Well, suppose you have this

F = ab’c’ + ab + ac + bc G = ab + c’ want F / G

•  You must transform it to something like this...

Let X=b’, Y=c’

à F = aXY + ab + ac + bc G = ab + Y then can do F / G

•  Because must treat the true and complement forms of variable as different

Slide 21 © 2013, R.A. Rutenbar

One More Constraint: Redundant Cubes
•  To do F/D, function F must have no redundant cubes
•  Technical term is: “minimal with respect to single-cube containment”
•  In words: no one cube is completely covered by other cubes in SOP cover

F = a + ab + bc is redundant quotient ab
D = a is the divisor c 00 01 11 10
Now: compute F / D, i.e., F / a 0 1 1
use our algebraic division algorithm
1 1 1 1
à  Problem: F / D = 1 + b, remainder = bc
à  “1+b” is not operation in Algebraic model!

Slide 22 © 2013, R.A. Rutenbar

Multilevel Synthesis Models: Where are We?
•  For Boolean F, D, can compute F = Q•D + R via algebraic model
•  This is great—but it’s still not enough. Don’t know how to find these divisors.
•  Real problem: n functions F1, F2, … Fn, find a set of good common divisors di

F1 = ab + c + x F1 = d1 + x

factor d1 = ab+c
F2 = abx + cx + q F2 = (d1)x+q

F3 = ab + q F3 = ab+q

•  What are we looking for?

•  Case 1: divisors d that are just 1 cube (1 product term), e.g., d = ab
•  Case 2: “bigger” multiple-cube divisors, e.g. d = ab + cd + e
Slide 23 © 2013, R.A. Rutenbar
VLSI CAD: Lecture 6.4
Logic to Layout Logic Synthesis:
Multilevel Logic:
Role of Kernels and
Co-Kernels in Factoring
Rob A. Rutenbar
University of Illinois
Where To Look For Good Divisors?
•  Surprisingly, the Algebraic Model has a beautiful answer
•  One more reason we like it: Has some surprising and elegant “deep structure”

•  Where to look for divisors of function F? In the kernels of F

•  Denoted K(F). Pronounced like “Colonel.” Spelled with TWO e’s (remember!)
•  K(F) is another set of two-level SOP forms which are the special, foundational
structure of any function F, being interpreted in our algebraic model.

•  How to find a kernel k∈K(F)? Algebraically divide F by one of

its co-kernels, c.
•  Wow – lots of new concepts and terms here. Let us go develop these ideas…

Slide 25 © 2013, R.A. Rutenbar

Kernels and Co-Kernels of Function F
•  Kernel of a Boolean expression F is:
•  A cube-free quotient k obtained by (algebraically) dividing F by a single cube c
•  This single cube c also has a name: it is a co-kernel of function F

quotient Q kernel k if cube-free

divisor D expression F c = 1 cube expression F

remainder R remainder R
F = D•Q + R F = c•k + R

Slide 26 © 2013, R.A. Rutenbar

 Kernels Are Cube-Free…
•  Cube-free means...?
•  You cannot factor out a single cube (product term) divisor that leaves no remainder
•  Technically -- has no one cube (product) that is a factor of expression
•  Do F / cube, look at result, if you can ‘cross out’ some cube in each term, not a kernel

Expression F F=d•Q+R Cube-free?

a a(1)+0 No
a+b -- Yes
ab + ac a(b+c)+0 No

abc + abd ab(c+d)+0 No

ab + acd + bd -- Yes

Slide 27 © 2013, R.A. Rutenbar

Some Kernel Examples
•  Kernels of F denoted K(F). Suppose F = abc + abd + bcd

Divisor cube d F= d•Q + R Is it a Kernel of f?

1 (1)(abc+abd+bcd)+0 No, has cube = b as factor
a (a)(bc+bd)+bcd No, also has cube = b as factor
b (b)(ac+ad+cd)+0 Yes, a kernel; co-kernel = (b)
ab (ab)(c+d)+bcd Yes, a kernel; co-kernel = (ab)
… etc

•  So, any Boolean F can have many different kernels k∈K(F)

Slide 28 © 2013, R.A. Rutenbar

Kernels: Why Are They Important?
•  …, if they are important, how do we actually compute them?
•  Big result: Brayton & McMullen Theorem
•  From: R. Brayton and C. McMullen, “The decomposition and factorization of Boolean
expressions. In IEEE International Symposium on Circuits and Systems, pages 49–54, 1982.

Expressions F, G have a common multiple-cube divisor d

if and only if

there are kernels k1∈K(F), k2 ∈K(G)

such that d = k1 ∩ k2 (ie, SOP form with common cubes in it)
and d is an expression with at least 2 cubes in it

Slide 29 © 2013, R.A. Rutenbar

Multiple-Cube Divisors and Kernels
•  In words:
•  The only place to look for multiple-cube divisors is in the intersection of kernels!
•  Further: this intersection of kernels is the divisor, there are no others

Kernels of F Intersect?
k3
≥2 cubes?
k1 F = new
F = stuff
k2
k3 k5
d = ab+c+…
k4 k5 Yes! d=… G = new
G = stuff Multi-cube
k6 k7
divisor extracted!
Kernels of G

Slide 30 © 2013, R.A. Rutenbar

Brayton-McMullen: Informal Illustration
•  Remember: kernel1, kernel2 are cube-free…
F = cube1 • kernel1 + remainder1 F = cube1 • [X + Y + stuff1] + remainder1
G = cube2 • kernel2 + remainder2 G = cube2 • [X + Y + stuff2] + remainder2

F = cube1 • [X + Y] + [cube1•stuff1 + remainder1]

G = cube2 • [X + Y] + [cube2•stuff2 + remainder2]

F = cube1•d + …
kernel1 ∩ kernel2 = [X + Y]
d = X+Y
= a multicube divisor of F & G
G = cube2•d + …

Slide 31 © 2013, R.A. Rutenbar

Kernels: Real Example
•  Consider this F, G

F = ae + be + cde + ab G= ad + ae + bd + be + bc
K(F) Kernel Co-kernel K(G) Kernel Co-kernel
a+b+cd e a+b d or e
b+e a d+e a or b
a+e b d+e+c b
ae+be+cde+ab 1 ad+ae+bd+be+bc 1

Intersecting these 2 kernels: (a+b+cd) ∩ (a+b) =

So, this is workable multicube divisor we can consider for both F,G

Slide 32 © 2013, R.A. Rutenbar

VLSI CAD: Lecture 6.5
Logic to Layout Logic Synthesis:
Multilevel Logic:
Finding the Kernels
Rob A. Rutenbar
University of Illinois
Kernels: Very Useful, But How To Find?
•  Another recursive algorithm (are we surprised...?)
•  There are 2 more useful properties of kernels we need to see first…

•  Start with a function F and a kernel k1 in K(F)

F = cube1 • k1 + remainder1

•  Then: a new, interesting question: what about K( k1) ??

•  k1 is a perfectly nice Boolean expression, so it has got its own kernels
•  Do these k1’s kernels have anything interesting to say about K(F)...?

Slide 34 © 2013, R.A. Rutenbar

How K(k1) Relates to K(F)…
•  We know this: F = cube1•k1 + remainder1
•  Suppose k2 is a kernel in K( k1 ), then we know
•  k1 = cube2•k2 + remainder2
•  Substitute this expression for k1 in original expression for F
•  F = cube1•[cube2•k2 + remainder2] + remainder1
•  Neat trick: cube1•cube2 is itself just another single cube, rewrite to emphasize:
•  F = (cube1•cube2)•[ k2 ] + [ cube1•remainder2 + remainder1]

•  Lovely result: k2 also a kernel of original F (with co-kernel cube1•cube2)

Slide 35 © 2013, R.A. Rutenbar

 There is a Hierarchy of Kernels Inside F
•  Terminology: k∈K(F) is:
F = stuff
•  A level-0 kernel if it contains no kernels
inside it except itself
•  In words: Only cube you can pull Level-2 kernel
out, get a cube-free quotient is ‘1’
•  A level-n kernel if it contains at least Level-1 kernel Level-1 kernel
one level-(n-1) kernel, and no other
level-n kernels except itself Level-0 kernel
Level-0 kernel Level-0 kernel
•  In words: a level-1 kernel only has
level-0 kernels inside it.
A level-2 kernel only has level-1 and
level-0 kernels in it, etc…

Slide 36 © 2013, R.A. Rutenbar

Kernel Hierarchy
•  2nd useful result [Brayton et al]
•  Co-kernels of a Boolean expression in SOP form correspond to
intersections of 2 or more of the cubes in this constituent SOP form

•  Note: Intersections here means specifically that we regard a

cube as a set of literals, and look at common subsets of literals
•  This is not like “AND” for products. This is simple common sub-expressions.
•  Example ace + bce + de + g

ace ∩ bce = ce à ce is a potential co-kernel

ace ∩ bce ∩ de = e à e is a potential co-kernel

Slide 37 © 2013, R.A. Rutenbar

Kernel Hierarchy
•  How do we use these 2 results?
•  Find the kernels recursively. Whenever we find one:
•  Call FindKernels( ) on it, to find (if any) lower level kernels are inside

•  Use algebraic division to divide function by potential co-kernels, to drive recursion

•  …but be smart: co-kernels are intersections of the cubes
•  ...if there’s at least 2 cubes, then look at the intersection of those cubes, and use
that intersected result as our potential co-kernel cube

•  One technical point: need to start with a cube-free function F to make things work
right. If not cube-free, just divide by the biggest common cube to simplify F…
Slide 38 © 2013, R.A. Rutenbar
Kernel Algorithm
•  Algorithm is then...
FindKernels( cube-free SOP expression F) {
K = empty;
for ( each variable x in F ) {
if ( there are at least 2 cubes in F that have variable x ) {
let S = { cubes in F that have variable x in them };
let co = cube that results from intersection of all cubes in S,
this will be the product of just those literals
that appear in each of these cubes in S;
K = K ∪ FindKernels( F / co) ;
} F/co = a kernel, returned at end.
} This is algebraic division, but
K=K∪F; simpler since it always
return( K ) Cube-free F is always its just divides by exactly
} own kernel, with 1 cube, a simple product term
trivial co-kernel = 1
Slide 39 © 2013, R.A. Rutenbar
 Kerneling Example
•  To start, divide F by each of the variables, and use to recurse
•  We are looking for co-kernels that start with this ONE variable in them
•  But—be smart, it cannot be a co-kernel unless it’s in at least 2 cubes
F = ace + bce + de + g
a b c d e g
no work no work cubes w/ c = no work cubes w/ e = no work
only 1 only 1 only 1 only 1
ace,bce ace,bce,de
cube w/ a cube w/ b cube w/ d cube w/ g
co =∩= co =∩=
ce e
Bug fix:
F/co = F/co = ac+bc+d
a+b ac+bc+d
Aside: “w/” = “with” recurse on: a+b recurse on: ac+bc+d
Slide 40 © 2013, R.A. Rutenbar
Kernel Hierarchy, Example Continued
•  With this algorithm, overall recursion tree looks like this
F = ace + bce + de + g
a b c d e g
no work no work cubes= ace,bce no work cubes= ace,bce,de no work
co = ce co = e
F/co = a+b F/co = ac+bc+d

Kernel = ac+bc+d
no work for any variable
Kernel = a+b Co-kernel = e
Co-kernel = ce
a b c d e g
no work no work cubes= ac,bc no work no work no work
Kernel = a+b co =∩= c
Co-kernel = ce F/co = a+b
(get co-kernel by ANDing
co cubes up recursion tree no work for any variable
Slide 41 © 2013, R.A. Rutenbar
Kernel Hierarchy
•  With this algorithm...
•  Can find all the kernels (and co-kernels too)
•  Get co-kernels by ANDing the divisor co cubes up recursion tree

•  One tiny problem

•  Will revisit same kernel multiple times

•  Solution: remember which variables already tried in co-kernels

•  Problem: kernel you get for co-kernel abc is same as for cba, but current
algorithm doesn’t know this and will find same kernel for both cubes
•  A little extra book keeping solves this

Slide 42 © 2013, R.A. Rutenbar

Using Kernels and Co-Kernels
•  These are exactly the right component pieces for...
•  Extraction of a single-cube divisor from multiple expressions
•  Extraction of a multiple-cube divisor from multiple expressions

F F = d•Q1 + R1
d
G G = d•Q2 +R 2

•  When you want a single-cube divisor: Go look for co-kernels

•  When you want a multiple-cube divisor: Go look for kernels

Slide 43 © 2013, R.A. Rutenbar

Multilevel Synthesis Models: Summary
•  Boolean network model
•  Like a gate network, but each node in network is an SOP form
•  Supports many operations to add, reduce, simplify nodes in network
•  Algebraic model & algebraic division
•  Simplifies Boolean functions to behave like polynomials of real numbers
•  Divides one Boolean function by another: F = (divisor D )•(quotient Q) + remainder R
•  Kernels / Co-kernels of a function F
•  Kernel = cube-free quotient obtained by dividing by a single cube (co-kernel)
•  Intersections of F, G kernels à all multiple-cube common divisors (Brayton-McMullen)
•  Next: what are best common divisors to get, given these ideas?

Slide 44 © 2013, R.A. Rutenbar

Notes:
•  The algebraic model (and division) are not the only options
•  There are also “Boolean division” models and algorithms that don’t lose expressivity
•  ..but they are more complex.
•  Rich universe of models & methods here.

•  Good references to read about all these ideas

•  R.K. Brayton, R. Rudell, A. Sangiovanni-Vincentelli, A.R. Wang, “MIS: A Multiple-
Level Logic Optimization System,” IEEE Transactions on CAD of ICs, vol. CAD-6, no.
6, November 1987, pp. 1062-1081.
•  Giovanni De Micheli, Synthesis and Optimization of Digital Circuits, McGraw-Hill,
1994.

Slide 45 © 2013, R.A. Rutenbar