You are on page 1of 91

A Recursive approach to Floorplanning

by

Renishkumar V. Ladani

A Thesis Submitted in Partial Fulfilment of the Requirements for the Degree of

Master of Technology

in
Information and Communication Technology
to
Dhirubhai Ambani Institute of Information and Communication Technology

May, 2005

DA-IICT
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Declaration

This is to certify that


(i) the thesis comprises my original work towards the degree of Master of Technology in
Information and Communication Technology at DA-IICT and has not been submitted
elsewhere for a degree,
(ii) due acknowledgement has been made in the text to all other material used.

Renishkumar V. Ladani

i
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Certificate
This is to certify that the thesis work entitled “A Recursive Approach to Floorplanning”
has been carried out by Renishkumar V. Ladani (200311014) for the degree of Master of
Technology in Information and Communication Technology at this Institute under my
supervision.

Prof. Ashok T. Amin


Thesis Supervisor

ii
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Acknowledgements

I am thankful to my guide Prof. Ashok T. Amin for guiding me throughout my thesis work.
His suggestions and constant support during this research are motivating factor for me. I
consider myself very lucky to get an opportunity to work with him. I am thankful to my
co-guide Prof. Amit Bhatt for providing initial insight into field of floorplanning. I am
thankful to my evaluation-committee members, Prof. D. Nag Chaudhary and Prof.
Hemangi Kapoor for providing useful suggestions for my research work. I am thankful to
my colleagues and friends for motivating me for research and providing constant support
in difficult times. I am thankful to J. M. Lin and Y. W. Chang (@cc.ee.ntu.edu.tw) for
making available their floorplanning algorithm implementation and test cases on their home
page. I am thankful to Dr. Hirendu P. Vaishnav of Synapps Corp., USA for his suggestion
of floorplanning as a research topic. I am thankful to DA-IICT for providing me the
resources needed and a favourable environment to carry out my work. I am thankful to my
family for supporting me in all the ways.

iii
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Contents

Page No.
DECLARATION...................................................................................................... .........................I

CERTIFICATE................................................................................................................................ .II

ACKNOWLEDGEMENTS......................................................................................................... .....III

CONTENTS......................................................................................................................... ..........IV

ABSTRACT.................................................................................................................. ................VII

LIST OF PRINCIPAL SYMBOLS AND ACRONYMS............................................................ ......VIII

LIST OF TABLES........................................................................................................................ ..IX

CHAPTER 1....................................................................................................................... ...........1

CHAPTER 2....................................................................................................................... .............7

FLOORPLANNING CONCEPTS AND APPROACHES TO PROBLEM........................................7

CHAPTER 3..................................................................................................................... .............17

A RECURSIVE APPROACH............................................................................................... ..........17

C............................................................................................................................................. ....19

C............................................................................................................................................. ....19

B.............................................................................................................................................. ......19

A...................................................................................................................................... .............19

B.............................................................................................................................................. ......19

A...................................................................................................................................... .............19

C............................................................................................................................................. ....19

A........................................................................................................................................... .........19

B.............................................................................................................................................. ......19

C............................................................................................................................................. ....19

B.............................................................................................................................................. ......19

A........................................................................................................................................... .........19

B.............................................................................................................................................. ......21

A........................................................................................................................................... .........21

A........................................................................................................................................... .........21

iv
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

B.............................................................................................................................................. ......21

D............................................................................................................................................. ....30

C............................................................................................................................................. ....30

D.............................................................................................................................................. ......30

C............................................................................................................................................. ....30

B............................................................................................................................................. ....30

B............................................................................................................................................. ....30

A...................................................................................................................................... ...........30

A...................................................................................................................................... ...........30

D............................................................................................................................................. ....32

C............................................................................................................................................. ....32

B............................................................................................................................................. ....32

A...................................................................................................................................... ...........32

CHAPTER 4..................................................................................................................... .............49

CONCLUSION AND FUTURE WORK ....................................................................................... .49

REFERENCES...................................................................................................................... ........54

APPENDIX..................................................................................................................... ...............56

A.1 CIRCUIT LAYOUT GENERATED BY ALGORITHM-I FOR HARD BLOCKS.......................56

B.1 CIRCUIT LAYOUT GENERATED BY ALGORITHM-I FOR SOFT BLOCKS........................62

v
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

vi
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Abstract

Due to increase in number of components on a chip, floorplanning is important step in


Very Large Scale Integration physical design to ensure quality of design. Various iterative
approaches have been suggested to carryout floorplanning in Electronic Design
Automation tools. Iterative approaches can produce good results but they are slower. In
this thesis, we have taken bottom-up, recursive approach to floorplanning. We have also
suggested efficient exhaustive search procedure for placing two, three or four rectangular
blocks in a floorplan. A rectangular block can either be hard or soft and resultant floorplan
can either be slicing or non-slicing. Further more exhaustive search procedure can also be
extended for five or more rectangular blocks. We have developed two algorithms, which
fall in class of constructive approaches rather than class of iterative approaches. These
algorithms use exhaustive search procedure, works in bottom-up constructive manner and
they are recursive in nature. These algorithms are very fast compared to other search
algorithms and also producing promising results. Complexity of these algorithms is O(n).
Experiments results with MCNC circuits indicate that area utilization of about 85-99% can
be achieved in very less time then iterative algorithms.

vii
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

List of Principal Symbols and Acronyms

VLSI Very Large Scale Integration


EDA Electronic Design Automation
SA Simulated Annealing
GA Genetic Algorithm
SAGA Simulated Annealing and Genetic Algorithm
NPE Normalized Polish Expression
SP Sequence Pair
BSG Bounded Slicing Grid
TCG Transitive Closure Graph
CBL Corner Block List
GPE Generalized Polish Expression

Other minor symbols are defined at first occurrence; where necessary some symbols are
redefined in the text.

viii
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

List of Tables

Page No.

TABLE 2.5 PACKING COMPLEXITY FOR NON-SLICING FLOORPLAN, HERE N IS THE


NUMBER OF BLOCKS IN THE PLACEMENT.....................................................................................16

TABLE 3.4.1 UNIQUE PLACEMENT STRUCTURE AND ITS TWO COMPOSITIONS FOR TWO
BLOCKS PLACEMENT.............................................................................................................................22

TABLE 3.4.1 TWO COMPOSITIONS OF TWO UNIQUE PLACEMENT STRUCTURES FOR


THREE BLOCKS.........................................................................................................................................25

TABLE 3.4.1 TWO COMPOSITIONS OF SIX UNIQUE PLACEMENT STRUCTURES FOR


FOUR BLOCKS...........................................................................................................................................39

TABLE 3.6.1 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-I.....................................43

TABLE 3.6.2 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-II...................................43

TABLE 3.6.3 AREA UTILIZATION AND RUNTIME COMPARISON FOR ALGORITHM-I AND
ALGORITHM-II..........................................................................................................................................43

TABLE 3.6.3A AREA UTILIZATION AND RUNTIME FOR SP AND O-TREE.................................44

TABLE 3.6.3B AREA UTILIZATION AND RUNTIME FOR B*-TREE AND ENHANCED O-TREE
.........................................................................................................................................................................44

TABLE 3.6.3C AREA UTILIZATION AND RUNTIME FOR CBL AND TCG...................................44

TABLE 3.6.3D AREA UTILIZATION AND RUNTIME FOR TCG-S AND FAST-SP.........................45

TABLE 3.6.3E AREA UTILIZATION AND RUNTIME FOR GPE.......................................................45

TABLE 3.6.4 PATTERN OF HARD AND SOFT BLOCKS IN TEST CASE -I..................................46

TABLE 3.6.5 PATTERN OF HARD AND SOFT BLOCKS IN TEST CASE -II...................................46

TABLE 3.6.6 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-I APPLIED ON CASE-I
.........................................................................................................................................................................46

TABLE 3.6.7 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-I APPLIED ON CASE-
II......................................................................................................................................................................47

TABLE 3.6.8 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-I APPLIED ON CASE-
III....................................................................................................................................................................47

TABLE 3.6.9 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-II APPLIED ON CASE-
I.......................................................................................................................................................................47

TABLE 3.6.10 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-II APPLIED ON


CASE-II..........................................................................................................................................................47

TABLE 3.6.11 AREA UTILIZATION AND RUNTIME FOR ALGORITHM-II APPLIED ON


CASE-III........................................................................................................................................................48

TABLE 3.6.12 SUMMARY OF AREA UTILIZATION AND RUNTIME FOR ALGORITHM-I......48

ix
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

TABLE 3.6.13 SUMMARY OF AREA UTILIZATION AND RUNTIME FOR ALGORITHM-II.....48

x
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

List of Figures

Page No.

xi
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Chapter 1

Introduction

1.1 Floorplanning in Context of VLSI Physical Design


VLSI physical design layout can be carried out in bottom up fashion. In this methodology
designer either uses cells from library or designs her/his cells and subsequently compose
the overall layout of the chip by means of placement and routing. But most of time this
leads to poor utilization of the chip area and excessive wiring.

Only a well-conceived design methodology can result in a final design of high quality; one
such methodology is FLOORPLAN-BASED DESIGN METHODOLOGY. It is top-down
design methodology. It advocates that layout aspects should be taken into account in all
design stages. Three design domains in which design stages are classified are behavioral
design domain, structural design domain and physical design domain.

The floorplan-based design methodology can be represented on GAJSKI Y-chart in fig.


1.1.

Fig. 1.1 GAJSKI Y-chart [1]

1
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Advantage of Floorplan-based methodology: Taking layout into account in all design


stages also gives early feed back, thus structural synthesis decision can immediately be
evaluated for their layout consequence and corrected if necessary. The presence of layout
information allows for an estimation of wire lengths. From these lengths one can derive
performance properties of the design such as timing and power consumption. They both
increase when the wire lengths grow.

1.2 Floorplanning Definition


It is easy to deal with layout when structural detail at lowest abstraction is available, one
knows the exact number of transistors in the circuit and the way they are interconnected.
When this type of structural information is not available, one can estimate the area to be
occupied by various sub blocks and together with a precise or estimated interconnection
pattern, try to allocate distinct regions of the integrated circuit to the specific sub blocks.
This process is call floorplanning.

It is important to note that functionally equivalent sub blocks have different shapes and
terminal positions. This is one of the main characteristics of floorplan-based design, one
chooses the shape and terminal positions such that they fit best with the original structure
and assumes that there is a way to design the module satisfying the chosen shape and
terminal position. Above type of blocks are known as flexible or soft blocks. When the
block is flexible one could say that the realization needs an area A. Whichever shape the
block will have its height h and its width w have to obey the constraint hw ≥ A. Other type
of blocks are hard blocks, it means that their shape and terminal positions (pins) are fixed.

It is also important to note that area required for interconnection wiring (Routing) can
either provided by incorporating them in the area estimations for the blocks or in the case
of N-layer metal with over the block routing (wiring), channel less block layouts are the
norm of design.

Example of a structural description of some circuit and possible floorplan and Floorplan
view of PowerPC 604 and Pentium 4 is provided in fig. 1.2.

2
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. 1.2 Structural description of some circuit and its floorplan, and Floorplan view of
PowerPC 604 and Pentium 4 [2]

1.3 Floorplan Problem Description


Given a set of blocks B = {b1, b2,…, bn}. Each block bi is rectangular and has fixed width
and height. The outputs of algorithm are coordinates of blocks (the absolute coordinates
of the lower left corner of the block). The objectives of floorplan optimization problem are
to minimize the area of B and reduce wire lengths of interconnects subject to the
constraints that no pair of blocks overlaps. There may be other objectives such as
maximize routability (minimize congestion), delay of critical path, noise, heat dissipation,
etc. But either they are not of much interest or in some way they are related to reduction
of wire lengths of interconnects [4].

In addition to above problem description, other then rectangular block study of L-shaped
and U-shaped blocks has been carried out [14]. Also, Flexible blocks have been not
addressed in above problem description. There also exist some representations and

3
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

algorithms, which addressed floorplan problem with flexible blocks, e.g., Normalized
Polish Expression (NPE) [5], SP [10], Fast-SP [13], O-tree [1] and B*-tree [4]. For such
type of representations and algorithms following problem formulation would provide more
insight.

Let B = {b1,b2, …, bn} be a set of n rectangular blocks. Each block bi ∈ B is associated


with a three tuple (hi, wi, ai), where hi, wi, and ai denote the width, height, and aspect ratio
of Bi, respectively. The area Ai of Bi is given by hi * wi, and the aspect ratio ai of Bi is given
by hi/wi, Let ri,min and ri,max be the minimum and maximum aspect ratios, i.e., hi/wi ∈ [ri,min,
ri,max]. Here both soft(flexible) and hard blocks are being considered. A hard module is not
flexible in its shape, but free to rotate. A soft module is free to rotate and change its shape
within range [ri,min, ri,max]. Output of foorplanning is a placement (floorplan) P = {(xi, yi) |
bi ∈ B} is an assignment of rectangular blocks with the coordinates of their bottom-left
corners being assigned to (xi, yi)’s so that no two blocks overlap (and Hi/wi ∈ [ri,min; ri,max],
∀ i ). As previously describe in problem description, the objective of floorplanning is to
minimize a specified cost metric such as a combination of the area Atot and wire length Wtot
induced by the assignment of bi’s, where Atot is measured by the final enclosing rectangle
of P and Wtot the summation of half the bounding box of pins for each net.
Cost = α*Atot + β*Wtot
Where,
Atot = Total area of the packing.
Wtot = Total wire length of packing.
α and β = User specified constant.
Here wire length estimation is to be done because exact wire length of each net is not
known until routing is done and also pin positions are not known yet. Two possible ways
of wire length estimation are center-to-center estimation and half-perimeter estimation.

Fig. 1.3. Center-to-center estimation and half-perimeter estimation.


1.3.1 Floorplan Sizing: A optimization problem in Floorplanning

4
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

The availability of flexible blocks implies the possibility of having different shapes for the
same hardware units. It’s therefore possible to choose a suitable shape for each flexible
block such that the resulting floorplan is optimal in some sense (e.g. minimal area).

1.3.2 Some Constraints in Floorplanning


In floorplanning, it is important to allow users to specify placement constraints. Three
common types of placement constraints are preplaced constraint, boundary constraint, and
range constraint. For preplaced constraint, we require a block to be placed exactly at a
certain position in the final packing. For boundary constraint, we require a block to be
placed along one particular side of the final floorplan: on the left, on the right, at the
bottom, or at the top. This is useful when users want to place some specific block along
the boundary for input–output connections. For range constraint, we require a module to
be placed within a given rectangular region in the final packing. This is indeed a more
general formulation of the placement constraint problem and any preplaced constraint can
be written as a range constraint by specifying the rectangular region such that it has the
same size as the module itself. Some representations and algorithms for floorplan are
extended for above given constraints.

1.4 Motivation
Due to the growth in design complexity, circuit sizes are getting larger. To cope with the
increasing design complexity, hierarchical design and IP modules are widely used. The
trend makes module floorplanning much more critical to the quality of a VLSI design. And
with current EDA tools with practice we can create good initial placement by
floorplanning hints and a pictorial display. This is one area where the human ability to
recognized patterns and spatial relations is currently superior to a computer program’s
ability. Thus practically floorplanning is not fully automated till now date.

1.5 Organization of the Thesis


The rest of the report is organized as follows. The second chapter starts with different
approaches to floorplanning problem with different representation of floorplan as well as

5
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

available algorithms. The second chapter end with previous work that has been done in this
particular direction and its comparison. The third chapter provides description of our
work. It includes introduction and description of our suggested recursive bottom-up
algorithms. It also includes how these algorithms use exhaustive search procedure for
placing two; three or four rectangular blocks in a floorplan. The chapter 3 ends with
experiments results and resultant floorplan view of MCNC benchmark suite. The forth
chapter contains conclusion to our thesis work and scope of future work.

6
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Chapter 2
Floorplanning Concepts and Approaches to Problem

2.1 Background
The floorplan problem is known to be NP-complete [11]. Various heuristic approaches
have been taken to solve this problem. These approaches can be categorized in Simulated
Annealing (SA), Genetic Algorithm (GA) and Hybrid approach (SAGA: simulated
annealing and genetic algorithm). This type of algorithm searches through the feasible
solution space for floorplan. Evaluate each solution at each stage to know its cost or
fitness compare it with earlier available results. Keep it or discard it according to
strategies. Carry out different moves to obtain different feasible solutions from a available
feasible solution.

In Genetic algorithms [15] moves are crossover, mutation and inversion. Similar types of
moves exist for simulated annealing. Hence these algorithms depend on representation of
feasible solution space. Representation for floorplan can be categorized in slicing floorplan
representation and non-slicing floorplan representation.

2.1.1 Slicing Structure


A rectangle dissection is a subdivision of a given rectangle by horizontal and vertical line
segments into a finite number of non-overlapping rectangles. The non-overlapping
rectangles are called basic rectangles. By slicing a rectangle, we mean to divide the
rectangle into two rectangles by a vertical or horizontal line. A slicing structure is a
rectangle dissection that can be obtained by recursively slicing rectangles into smaller
rectangles (see Fig. 2.1.1a).
The hierarchical structure of a slicing structure can be described by an oriented rooted
binary tree, called a slicing tree (see Fig. 2.1.1b). A Slicing tree is essentially a top down
description of a slicing structure. It specifies bow a given rectangle is cut into smaller
rectangles by horizontal and vertical slicing lines. Each internal node of the tree is labelled
either * or +, corresponding to either a vertical or a horizontal cut, respectively. Each leaf
corresponds to a basic rectangle and is labelled by a number between 1 and n when the
slicing structure has n basic rectangles. Wong and Liu proposed an algorithm for slicing
floorplan designs using a normalized polish expression [5] to represent a slicing structure.

7
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. 2.1.1a A slicing floorplan Fig. 2.1.1b A slicing tree

2.1.2 Non-Slicing Structure


Not all floorplans are slicing. If the basic rectangles corresponding to leaf nodes in slicing
structures can’t be obtained by recursive cutting rectangles into smaller rectangles then the
floorplan has non-slicing structure (See Fig. Fig. 2.1.2) and represented in different ways.
The representation are sequence pair [6], bounded slicing grid (BSG) [7], O-tree [1],
Transitive Closure Graph (TCG) [2], Corner Block List (CBL) [3] and B* Trees [4].

Fig. 2.1.2 A non-slicing structure

2.1.3 Normalized Polish Expression


A binary sequence b1,b2, …, bm, is a balloting sequence iff for any k, 1 <= k <=m, the
number of 0 ‘s in b1, …, bk, is less then the number of the 1 ‘s in b1, …, bk. Let σ be a
function σ: {l,2 ,..., n,*,+} -> {0,1} defined by σ (i) = 1, 1<= i <=n, and σ(*) = σ (+) =
0.
A sequence α1α2 … α2n-1 of elements from {1, 2, ..., n, *, +} is a Polish expression of
length 2n -1 iff
(1) Every i appears exactly once in the sequence, 1 <= i <= 2n -1,
(2) σ(α1) σ(α2) … σ(α2n-1) is a balloting sequence.

8
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

A Polish expression α1α2 … α2n-1 is said to be normalized iff there is no consecutive *‘s
or +‘s in the sequence. (e.g. 1 2 + 4 3 * + is a normalized Polish expression.)

In general, there might be two or more Polish expressions (slicing trees) that correspond
to a given slicing structure (see Fig.3.2e). The number of Polish expressions corresponding
to a slicing structure can vary from slicing structure to slicing structure. This makes Polish
expressions an undesirable choice for representation of solutions in a simulated annealing
setting for the following reasons: 1. There is an unnecessary increase in the number of
states. 2. The set of slicing structures is unevenly distributed over the set of Polish
expressions, which might lead to unintentional and undesirable biases toward some slicing
structures. It is observation that given any slicing structure, it can be described by a unique
skewed slicing tree by performing the cuts always from right to left and from top to
bottom. Hence, the set of normalized Polish expressions as the solution space in our
simulated annealing algorithm. The Polish expression in fact is the Polish postfix notation
for this “arithmetic expression”.

Fig. 2.1.3 Two different slicing trees for the same slicing structure.

2.1.4 Neighbourhood Structures


We define three types of moves that can be used to modify a given normalized Polish
expression.
M1. Swap two adjacent operands.
M2. Complement some chain of nonzero length.
M3. Swap two adjacent operand and operator.
Two normalized Polish expressions are said to be neighbours if one can be obtained from
the other via one of these three moves. We also want to make sure that the move selected
will also produce a normalized Polish expression.

2.1.5 The Cost Function

9
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Cost = α*Atot + β*Wtot


Where,
Atot = Total area of the packing.
Wtot = Total wire length of packing.
α and β = User specified constant.

2.1.6 Comparisons between slicing and non-slicing approach


Slicing representation has some advantages such as smaller encoding cost and solution
space bringing faster runtime for packing. Furthermore it is flexible to deal with hard,
preplaced, soft and rectilinear blocks. However in real designs optimal solution might not
be in the solution space of slicing structure. While with non-slicing representation optimal
solution might be achieved but it needs more evaluating runtime for packing then slicing
approach.

2.2 Algorithmic Approaches


2.2.1 Simulated annealing
Simulated annealing is a well-known high performance optimization technique for
combinatorial problems. The simulated annealing algorithm is presented below:

01 Temperature = Initial Temperature;


02 Current placement = Random initial placement;

10
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

03 Current score = Score (Current placement);


04 While equilibrium at temperature not reached Do
05 Selected component = Select (at random);
06 Trail placement = Move (selected component);
07 Trail score = Score (trail placement);
08 If trail score < current score then
09 Current score = trial score;
10 Current placement = trail placement;
11 else
12 if uniform random(0,1) < e-(trail score – current score)/temperature then
13 Current score = trial score;
14 Current placement = trail placement;
15 temperature = temperature * Alpha; // alpha ~ 0.95

The temperature in initialised to a relatively high value and its slowly decrease until a
freezing point is reached. At each temperature, components are selected for possible
movement until equilibrium is reached. If movement of the selected components results in
an improved placement, the movement is performed. Otherwise the movement is
performed with a probability that decrease exponentially with temperature. Components
are typically selected randomly for pair wise exchange.

2.2.2 Genetic Algorithm


The original GA and its many variants collectively known as genetic algorithms are
computational procedure that mimics the natural process of evolution. Darwin observed
that as variations are introduced into a population with each new generation the less fit
individuals are tend to die off in the competition for food and this survival of the fittest
principle leads to improvement in the species.

GAs has also applied to optimisation problems, and the applications like floorplanning in
EDA tools falls into this category. The objective of the GA is then to find an optimal
solution to a problem. Since Gas are heuristic procedure, they are not guaranteed to find
the optimum but experience has shown that they are able to find very good solutions for
wide range of problems.

GAs work by evolving a population of individual in the population where the fitness
computation depend s on the application. For each generation individuals are selected from
the population for reproduction, the individuals are crossed to generate new individuals
and the new individuals are muted with some low mutation probability. The new individual
may completely replace the old individuals in the population with distinct generation

11
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

evolved; alternatively the new individuals may be combined with old individuals in the
population.
Since selection is biased towards more highly fit individuals, the average fitness of the
population tends to improve from one generation to the next. The fitness of the best
individuals is also expected to improve overtime, and the best individual may be chosen as
a solution after several generations.

Simple GA: Also referred to as total replacement algorithm. Flowchart of this simple
genetic algorithm is available in fig. 2.2.2. [15]

Stopping Criteria: The GA may be limited to a fixed number of generations or it may be


terminated when all individuals in the population coverage to the same string or no
improvements in fitness values are found after given number of generation.

Since selection is biased towards more highly fit individuals the fitness of the overall
population is expected to increase in successive generations. However, the best individual
may appear in any generation.

Generate initial population

Evaluate each individual

Select Np individuals with repetition, such that


the probability of selection of each individual is
proportional to its fitness
With a high probability, Pc, perform crossover on
the pair s to generate two offspring. If crossover
Perform inversion 12 the offspring with
is not performed,
Replace thenonthe
all individuals parents
of are copied
the previous
probability
Mutate
Pair the
the p1 ifwith
individuals
offspring
unchanged
generation the
to algorithm
with
randomly
a small
the
the to calls parents
for it Pm
probability,
form
offspring.
Np offspring
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

No

Stoppi
ng
criteria

Yes
Fig. 2.2.2 Flowchart of the simple genetic algorithm

2.2.3 SAGA
Rather than simply using a GA for floorplanning, its better to use a new stochastic
optimization algorithm called SAGA, Which is combination of genetic algorithm and
simulated annealing algorithm applied to floorplanning. The aim of this idea is to improve
the typical convergence rate of the pure GA by combining it with simulated annealing.
The typical GA convergence curve is shown in fig Fig. 2.2.3.

Cost

Runtime

13
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. 2.2.3. Typical convergence of a GA

Initially the solution cost improves very rapidly, however obtaining further improvement
soon becomes difficult and the majority of runtime is spent in the later phase of the process
in which small improvements are obtained very slowly, while in case of simulated annealing
algorithm. The typical convergence curve of SA is very different from that of the GA.
Initially SA converges much slower but in the late phase of the process, SA may be able to
obtain improvement faster than the GA. The unified algorithm called SAGA (an acronym
for simulated annealing and genetic algorithm) is designed in such a way that the initial fast
convergence of the GA is combined with the faster convergence of SA in the late phase.
The SAGA algorithm is application independent and highly adaptive. When applied to the
floorplanning SAGA perform better than a pure GA.

2.2.4 Comparisons between SA and GA


Both simulated annealing and the genetic algorithm are computation intensive. One
difference is that simulated annealing operates on only one solution at a time while genetic
algorithm maintains a large population of solutions which are optimized simultaneously.
Thus the genetic algorithm takes advantages of the experience gained in the past
exploration of the solution space. Both simulated annealing and the genetic algorithm have
mechanisms for avoiding entrapment at local optima. In simulated annealing this is
accomplished by occasionally discarding a superior solution and accepting and inferior
one. The genetic algorithm also relies on inferior individuals as a means of avoiding false
optima, but, since it has whole population of individuals, the genetic algorithm can keep
and process inferior individuals without losing the best one.
Simulated annealing is an inherently serial algorithm while genetic algorithm can be
parallelized on such loosely coupled distributed computer network with 100% processor
utilization.

2.5 State-of-art in floorplan representations


VLSI floorplans are often grouped into two categories, the slicing structure [5] and the
non-slicing structure [1, 2, 3, 4, 6, 7]. A binary tree whose leaves denote blocks can
represent a slicing structure, and internal nodes specify horizontal or vertical cut lines.
Wong and Liu proposed an algorithm for slicing floorplan designs [5]. They presented a
normalized Polish expression to represent a slicing structure, enabling the speed-up of its
search procedure. However, this representation cannot handle non-slicing floorplans. It

14
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

takes only O(n) time to derive a floorplan from a representation. Recently, proposed
several representations such as sequence pair [6], bounded slicing grid (BSG) [7], O-tree
[1], Transitive Closure Graph (TCG) [2], Corner Block List (CBL) [3] and B* Trees [4]
can handle non-slicing floorplans. Table 2.2.6 shows packing complexity for non-slicing
floorplan.

GPE Recently, a new representation for VLSI floorplan problem has been published [11].
They proposed a new and easy representation for VLSI floorplan and building block
problem. The representation effectively inherits the useful property of normalized polish
expression [5] and is able to present non-slicing floorplan. The test using MCNC
benchmarks and the experiments give promising results. The time complexity to transform
a GPE to a corresponding placement is also O(n). Results of GPE suggest that it achieves
better area utilization compared to previous non-slicing representation Fast-SP and
Enhance O-tree.

Flooplan sizing (shaping) as defined previously can be done optimally and efficiently for
slicing floorplans. It can also be done optimally for some non-slicing floorplans, but its
very time consuming. “Shape Curve Computation” is used for Shaping in slicing floorplans
[14] and the sizing algorithm runs in polynomial time for slicing floorpalns. Langrangian
Relaxation method used for shaping in non-slicing floorplan. But it is not efficient and
applicable to only non-slicing floorplans, which are using Constraints graphs for packing
such as SP [10], Fast-SP [13], O-tree [1] and B*-tree [4].

Representation Runtime for packing

SP O(n2)

Fast-SP O(n lg n lg n)

BSG O(n2)

O-tree O(n)

B*-tree O(n)

CBL O(n)

15
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

TCG O(n2)

GPE O(n)

Table 2.5 Packing complexity for non-slicing floorplan, here n is the number of blocks in
the placement

16
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Chapter 3
A Recursive Approach

3.1 Introduction
Algorithms for floorplanning are classified in two classes of approaches, iterative ap-
proaches and constructive approaches. Iterative approaches produce floorplan with better
areas utilization but they are slower then constructive algorithms. An iterative approach
starts with one initial solution, evaluate it and then generate more such solutions from
available solution. At each stage, an iterative approach evaluates new solution and com-
pares it with earlier available results and keeps only promising solutions. In these ap-
proaches, an algorithm run up to either reaching timeout or based on some criteria such as
no more improvement in results. While in case of a constructive approach a feasible solu-
tion is generated gradually from available inputs using some techniques and principles. We
propose and investigate two constructive algorithms based on the notion that grouping
blocks having nearly same area in a floorplan produce better results than placing blocks
having wide difference in area.

In both algorithms, exhaustive search procedure is carried out at each step to place four or
less blocks at a time to get a floorplan having best area utilization. This exhaustive search
procedure is repeated in bottom up to construct a floorplan. .

Objectives of floorplanning problem is either area optimisation; wire length optimisation or


both. Although wire length optimisation is also critical to VLSI physical design but we will
focus on only area optimisation.

3.2 Problem Definition


Suppose, we are given a set of n blocks or rectangular objects b1, b2, …, bn. A block can be
of a fixed type or a flexible type. A fixed block has fixed height and width. A flexible block
has constant area but can have height and width ratio, called aspect ratio, from a given set
of possible values.

These blocks are to be placed in a rectangular area in non-overlapping manner. A block


may be rotated by + 90os. The problem is to arrange n blocks inside a rectangle of
minimum possible area.

17
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

With n blocks b1, b2, …, bn, we are given a list of n quadruplets of numbers (A1, r1, s1, d1),
(A2, r2, s2, d2), …, (An, rn, sn, dn). This quadruplet of number (Ai, ri, si, di), with ri ≤ si,
specifies the area and the shape constrains for module i. In fact, if we let wi be the width of
module i and hi be the height of module i, we must have wi * hi = Ai and ri ≤ hi/wi ≤ si.
Thus ri and si are our lower and upper limit of aspect ratio. Block i is a rigid (hard) block if
ri = si, otherwise its is a soft (flexible) block. If a block is hard then di has no meaning to it
and it’s just don’t care value. But if a block is soft, di specifies all possible shapes for a
flexible block, having aspect ratios as ri, ri + di, ri + 2*di, …, si.

A solution of the floorplan design problem consist of an enveloping rectangle R which


contains blocks b1, b2, …, bn in non overlapping manner and floorplan F = {(x1i, y1i, x2i,
y2i) | 1 ≤ i ≤ n}, indicating that placement of block bi with its bottom-left corner being at
(x1i, y1i) and top-right corner being at (x2i, y2i).

3.3 Terminology And Concepts


Minimum Area: Minimum Area (MA) is a summation of area of n blocks b1, b2, …, bn.

Floorplan Area: Floorplan Area (FA) is area of minimum possible of rectangle which
accommodates n blocks b1, b2, …, bn in non-overlapping manner. Clearly, FA ≥ MA.

Dead Area: A minimum possible rectangle which can accommodate n blocks in non-
overlapping manner has some area not occupied by any blocks. It is known as Dead Area
(DA) and measured in percentage of FA, namely DA = (FA-MA)/FA*100. Area utilization
factor is defined to be 100- DA.

L-compact: A floorplan L-compact if and only if there is no block that can shift left from
its original position with other components fixed.

B-compact: A floorplan is B-compact if and only if there is no block that can shift bottom
from its original position with other components fixed.

LB-compact: A floorplan is LB-Compact if and only if it’s both L-compact and B-


compact.

18
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

These types of floorplans are illustrated in fig 3.3.1.

C B C
A A
B

Fig. 3.3.1 (a) A Floorplan Fig. 3.3.1 (b) B-Compact Floorplan

C B C
A A
B

Fig. 3.3.1 (c) L-Compact Floorplan Fig. 3.3.1 (d) LB-Compact Floorplan

3.4 Exhaustive Search Procedure


Suppose, we are given a set of 2, 3 or 4 blocks say {A, B}, {A, B, C} and {A, B, C, D},
Where A, B, C and D are hard blocks. Here, (Ah, Aw), (Bh, Bw), (Ch, Cw) are (Dh , Dw) are
height and width of A, B, C, and D respectively. Let F is a floorplan generated after
placing either {A, B}, {A, B, C} or {A, B, C, D}. Here, (Fh, Fw) is height and width of
floorplan and FA = Fh * Fw , represent floorplan area. In this section, we present an
efficient way of searching a floorplan which has best area utilization or say minimum dead
area from all possible placement of 2, 3 or 4 blocks. For generation of all possible
placements of 2, 3 or 4 blocks first we generate the set PO of all possible ordering of
blocks. Let say for Set of 2 blocks {A, B}, PO = {AB, BA} represent all possible ordering
of two blocks. For three blocks {A, B, C}, PO = {ABC, ACB, BAC, BCA, CAB, CBA}
represent all possible ordering of three blocks. And similarly set PO for 4 blocks is also
generated. Since A block can be rotated by + 90os, we have two orientation vertical and
horizontal orientation for a block. So count of all possible pattern of placement (PPP) for
two blocks is equal to 2! * 22 = 8, Let say PPP = {AB, BA, AB’, B’A, A’B, BA’, A’B’,
B’A’} represents possible pattern of placement. Here A’ and B’ represent rotation by +
90os for Block A and B respectively and similarly count of all possible pattern of placement
(PPP) for three blocks equal to 3! * 23 = 48 and its equal to 4! * 24 = 384 for four
blocks.

19
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

In the following section we identify unique structure that can hold 2, 3 or 4 blocks in LB
compact floorplan. We have removed other redundant structures that always produce a
floorplan with same floorplan area, FA when placed with possible pattern of placement
(PPP) for two, three or four blocks. When blocks according to possible pattern of
placement are placed in unique structures we received set of all possible placement say PP
= {F1, F2, …, FK}. Here we have one unique structure for two blocks, two unique
structures for three blocks and six unique structures for four blocks. Next three sections
describe how we have identified the unique structures. Now for set PP we have, | PP | =
k = 2! * 22 * 1 = 8, for two blocks, | PP | = k = 3! * 23 * 2 = 96, for three blocks and |
PP | = k = 4! * 24 * 6 = 2304.

From set PP we search for floorplan Fi which has smallest area, where, 0 ≤ i ≤ k. if two
or more floorplans have equal and minimum area then a floorplan with aspect ratio near to
1.0 is selected. Thus | PP | represent number placement to be considered before selecting
one.

In case of soft blocks, blocks A, B, C, D can take any one of the shape form its given set
of aspect ratios, which increases number of possible placement. Let say AAR is set of aspect
ratios for block A, BAR is set of aspect ratios for block B, CAR is set of aspect ratios for
block C, DAR is set of aspect ratios for block D then size of possible placement set PP, get
scaled proportional to value of | AAR |, | BAR |, | CAR |, | DAR |. Thus size of possible
placement set PP, | PP | = k = 2! * 22 * 1 * | AAR | * | BAR |, for two blocks, | PP | = k
= 3! * 23 * 2 * | AAR | * | BAR | * | CAR |, for three blocks and | PP | = k = 4! * 24 * 6 *
| AAR | * | BAR | * | CAR | * | DAR |. Once a soft block get placed in floorplan of 2, 3 or 4
blocks, its aspect ratio get fixed and it’s no longer a soft blocks now. And floorplan F that
we received after placing 2,3 or 4 blocks together has also fixed aspect ratio because we
are selecting floorplan F from set PP according to it smallest area value and if two or more
floorplans have equal and minimum area then a floorplan with aspect ratio nearer to 1.0 is
selected.

3.4.1 Two Block Placements


In this section we have identifies LB-compact unique structure for placing 2 blocks
together in a floorplan. While placing two blocks together we can only have slicing
structures. Non-slicing structure can not possible for placing two blocks together. For

20
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

producing all possible slicing structures for blocks, we have used a binary tree with a root
node and two children. A root node is operator and its children are two blocks. This
suggests placement of two children in a way that placement of right child is with respect to
left child and according to operator in LB-compact manner. In slicing structure we have
two-operator horizontal placement operator say H and vertical placement operator say V.
Let O is set of operator for slicing structure then set O is define as O = {H, V}. Here
horizontal placement means two blokes are placed in side-by-side or adjacent in LB-
compact manner. And vertical placements mean two blocks are placed one above other in
LB-compact manner.

In fig. 3.4.1 (a) shows a binary tree of two blocks, while in fig 3.4.1 (b) and fig. 3.4.1 (c)
show horizontal and vertical placement respectively derived from binary by placing value
of operator as O1 = {H, V}.

O1

A B

Fig. 3.4.1 (a)

A A
B

Fh = max (Ah, Bh) Fh = Ah + Bh


Fw = Aw + Bw Fw = max (Aw, Bw)
Fig. 3.4.1 (b) Fig. 3.4.1 (c)

Under condition of exhaustive search with all possible ordering of blocks A and B with for
each block + 90os rotation allowed both of structures from fig 3.4.1 (b) and fig. 3.4.1 (c)
produce same minimum floorplan area FA = Fh * Fw. But difference is that one is horizontal

21
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

composition and other is vertical composition which achieved by floorplan rotation by +


90os and rearranging each block in floorplan once again with 90os. Thus we identified only
one unique placement structure and its two compositions for two blocks placement. Table
3.4.1 presents a unique placement structure and its two compositions for two blocks
placement.

Vertical Composition Horizontal composition

B
A A B

Fh = Ah + Bh
Fh = max (Ah, Bh)
Fw = max (Aw, Bw)
Fw = Aw + Bw

Table 3.4.1 unique placement structure and its two compositions for two blocks placement

3.4.2 Three Block Placements


In this section we have identifies LB-compact unique structure for placing 3 blocks
together in a floorplan. While placing three blocks together we can only have slicing
structures. Non-slicing structure can’t be possible for placing three blocks together. For
producing all possible slicing structures for three blocks, we have used two binary trees
with leaf nodes represent block and all other internal nodes are operator. These binary
trees have two operators to arrange three blocks. Fig. 3.4.2 (a) and fig. 3.4.2 (b) show a
binary tree of three blocks, we have two operators O1 = {H, V} and O2 = {H, V}

Fig 3.4.2 (c), (d), (e) and (f) show placement derived from a binary tree (in Fig. 3.4.2 (a))
by placing value of operator as O1 O2 = {HH, HV, VH, VV} and similarly Fig 3.4.2 (g), (h),
(i) and (j) show placement derived from a binary tree available in Fig. 3.4.2 (b).

22
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

O2 O1

C A O2
O1

B C
A B

Fig. 3.4.2 (a) Fig. 3.4.2 (b)

A B C A B

Fh = max (Ah, Bh, Ch) Fh = max (Ah, Bh) + Ch


Fw = Aw + Bw + Cw Fw = max (Aw + Bw, Cw)
Fig. 3.4.2 (d)
Fig. 3.4.2 (c)

C
B
C B
A
A
Fh = max (Ah + Bh, Ch)
Fw = max (Aw + Bw) + Cw Fh = Ah + Bh + Ch
Fw = max (Aw, Bw, Cw
Fig. 3.4.2 (e)
Fig. 3.4.2 (f)

23
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

C
A B C A
B

Fh = max (Ah, Bh, Ch) Fh = max (Ah, Bh + Ch)


Fw = Aw + Bw + Cw Fw = Aw + max (Bw, Cw)
Fig. 3.4.2 (g) Fig. 3.4.2 (h)

C
B C
B
A A

Fh = Ah + max (Bh, Ch) Fh = Ah + Bh + Ch


Fw = max (Aw, Bw + Cw) Fw = max (Aw, Bw, Cw)
Fig. 3.4.2 (i) Fig. 3.4.2 (j)

Under condition of exhaustive search with all possible ordering of blocks A, B and C with
each block + 90os rotation allowed there are few redundant structures from fig 3.4.2 (c) to
fig. 3.4.2 (i) always produce same floorplan area FA = Fh * Fw. Thus we identified two
unique placement structures and its two compositions for three blocks placement. Table
3.4.2 presents two compositions of two unique placement structures for three blocks
placement.

Vertical Composition Horizontal composition

24
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

C
B A B C
A
Fh = Ah + Bh + Ch Fh = max (Ah, Bh, Ch)
Fw = max (Aw, Bw, Cw) Fw = Aw + Bw + Cw

B C C
A
B
A
Fh = max (Ah, Bh + Ch)
Fh = Ah + max (Bh, Ch) Fw = Aw + max (Bw, Cw)
Fw = max (Aw, Bw + Cw)
Table 3.4.1 two compositions of two unique placement structures for three blocks

3.4.3 Four Block placements


In this section we have identifies LB-compact unique structure for placing 4 blocks
together in a floorplan. While placing for blocks together we have slicing structures as well
as Non-slicing structure. For producing all possible slicing structures for four blocks, we
have used five different binary trees with leaf nodes represent block and all other internal
nodes are operator. These binary trees have three operators to arrange four blocks. Fig.
3.4.3 (a), Fig. 3.4.3 (b), Fig. 3.4.3 (c), Fig. 3.4.3 (d) and fig. 3.4.3 (e) show a binary tree
of four blocks, we have three operators O1 = {H, V}, O2 = {H, V} and O3 = {H, V}.

25
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

There exist no such procedure for producing all possible for non-slicing structures. But we
have one LB-compact unique non-slicing structure possible for placing four blocks.

Fig 3.4.3 (a1) to Fig 3.4.3 (a8) show placement derived from a binary tree in Fig. 3.4.3 (a)
and similarly Fig 3.4.3 (b1) to Fig 3.4.3 (b8) show placement derived from a binary tree in
Fig. 3.4.3 (b) and then so on up to a binary tree in Fig. 3.4.3 (e). These placements are
derived after placing value of operator as O1 O2 O3= {HHH, HHV, HVH, HHVV, VHH,
VHV, VVH, VVV}.

O2

O1 O3

A D
B C

Fig. 3.4.3 (a)

26
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
A B C D A B
C
Fh = max (Ah, Bh, Ch, Dh) Fh = max (Ah, Bh, Ch + Dh)
Fw = Aw + Bw + Cw + Dw Fw = Aw + Bw + max (Cw, Dw)
Fig. 3.4.3 (a1) Fig. 3.4.3 (a2)

D
C D
C

A B A B

Fh = max (Ah, Bh) + max (Ch, Dh) Fh = max (Ah, Bh) + Ch + Dh


Fw = max (Aw + Bw, Cw + Dw ) Fw = max (Aw + Bw, Cw, Dw )
Fig. 3.4.3 (a3) Fig. 3.4.3 (a4)

27
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
C D
C
B B
A A

Fh = Ah + Bh + max (Ch + Dh) Fh = Ah + Bh + Ch + Dh


Fw = max (Aw, Bw, Cw + Dw) Fw = max (Aw, Bw, Cw, Dw)
Fig. 3.4.3 (a7) Fig. 3.4.3 (a8)

B B D
C D
A A C
Fh = max (Ah + Bh, Ch, Dh) Fh = max (Ah + Bh, Ch + Dh)
Fw = max (Aw + Bw) + Cw + Dw Fw = max (Aw + Bw) + max (Cw, Dw)
Fig. 3.4.3 (a5) Fig. 3.4.3 (a6)

28
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

O3

O2 D

O1 C

A B

Fig. 3.4.3 (b)

D
A B C D
A B C

Fh = max (Ah, Bh, Ch, Dh)


Fh = max (Ah, Bh, Ch) + Dh
Fw = Aw + Bw + Cw + Dw
Fw = max (Aw + Bw + Cw, Dw)
Fig. 3.4.3 (b1) Fig. 3.4.3 (b2)

29
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
C
C
A B D A B

Fh = max (max (Ah, Bh) + Ch, Dh) Fh = max (Ah, Bh) + Ch + Dh


Fw = max (Aw + Bw, Cw) + Dw Fw = max (Aw + Bw, Cw, Dw)
Fig. 3.4.3 (b3) Fig. 3.4.3 (b4)

D
B B
C D C
A A

Fh = max (Ah + Bh, Ch, Dh) Fh = max (Ah + Bh, Ch) + Dh


Fw = max (Aw + Bw) + Cw + Dw Fw = max (Aw + Bw, Cw, Dw)
Fig. 3.4.3 (b5) Fig. 3.4.3 (b6)

D
C C
D
B B
A A

Fh = max (Ah + Bh + Ch, Dh) Fh = Ah + Bh + Ch + Dh


Fw = max (Aw + Bw, Cw) + Dw Fw = max (Aw, Bw, Cw, Dw)
Fig. 3.4.3 (b7) Fig. 3.4.3 (b8)

30
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

O1

A O2

B O3

D
C

Fig. 3.4.3 (c)

D
A B C D A B
C

Fh = max (Ah, Bh, Ch, Dh) Fh = max (Ah, Bh, Ch + Dh)


Fw = Aw + Bw + Cw + Dw Fw = Aw + Bw + max (Cw, Dw)
Fig. 3.4.3 (c1) Fig. 3.4.3 (c2)

C D D
C
A A
B B
Fh = max (Ah, Bh + max (Ch, Dh)) Fh = max (Ah, Bh + Ch + Dh)
Fw = Aw + max (Bw, Cw + Dw) Fw = Aw + max (Bw, Cw, Dw)
Fig. 3.4.3 (c3) Fig. 3.4.3 (c4)

31
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
B C D B
C
A A
Fh = Ah + max (Bh, Ch, Dh) Fh = Ah + max (Bh, Ch + Dh)
Fw = max
Fig. (Aw(c5)
3.4.3 , Bw + Cw + Dw) Fw = max (Aw, Bw + max (Cw, Dw))
Fig. 3.4.3 (c5) Fig. 3.4.3 (c6)

D
C D
C
B B
A A
Fh = Ah + Bh + max (Ch, Dh) Fh = Ah + Bh + Ch + Dh
Fw =Fig.
max3.4.3
(Aw,(c7)
Bw, Cw + Dw) Fw Fig.
= max (Aw(c8)
3.4.3 , Bw, Cw, Dw)
Fig. 3.4.3 (c7) Fig. 3.4.3 (c8)

32
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

O3

O1

A
O2

B C

Fig. 3.4.3 (d)

A B C D A B C

Fh = max (Ah, Bh, Ch, Dh) Fh = max (Ah, Bh, Ch) + Dh


Fw = Aw + Bw + Cw + Dw Fw = max (Aw + Bw + Cw, Dw)
Fig. 3.4.3 (d1)
Fig. 3.4.3 (d2)

33
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
C C
A D A
B B
Fh = max (Ah, Bh + Ch, Dh) Fh = max (Ah, Bh + Ch) + Dh
Fw =Aw + max (Bw, Cw) + Dw Fw = max (Aw + max (Bw, Cw), Dw)
Fig. 3.4.3 (d3) Fig. 3.4.3 (d4)

B C B C
D
A A

Fh = max (Ah + max (Bh, Ch), Dh) Fh = Ah + max (Bh, Ch) + Dh


Fw = max (Aw + Bw + Cw) + Dw Fw = max (Aw, Bw + Cw, Dw)
Fig. 3.4.3 (d5) Fig. 3.4.3 (d6)

34
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
C C
B B
A D
A
Fh = max (Ah + Bh + Ch, Dh) Fh = Ah + Bh + Ch + Dh
Fw = max (Aw, Bw, Cw) + Dw Fw = max (Aw, Bw, Cw, Dw)

Fig. 3.4.3 (d7) Fig. 3.4.3 (d8)

O1

A O3

D
O2

B
C

Fig. 3.4.3 (e)

35
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
A
A B C D B C

Fh = max (Ah, Bh, Ch, Dh) Fh = max (Ah, max (Bh, Ch) + Dh)
Fw = Aw + Bw + Cw + Dw Fw = Aw + max (Bw + Cw, Dw)
Fig. 3.4.3 (e1) Fig. 3.4.3 (e2)

D
C C
A D A
B B
Fh = max (Ah, Bh + Ch, Dh) Fh = max (Ah, Bh + Ch + Dh)
Fw =Aw + max (Bw, Cw) + Dw Fw = Aw + max (Bw, Cw, Dw)
Fig. 3.4.3 (e3) Fig. 3.4.3 (e4)

D
B C D B C

A A
Fh = Ah + max (Bh, Ch, Dh) Fh = Ah + max (Bh, Ch) + Dh
Fw = max (Aw, Bw + Cw + Dw) Fw = max (Aw, Bw + Cw, Dw)
Fig. 3.4.3 (e5) Fig. 3.4.3 (e6)

36
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
C C
D
B B
A A

Fh = Ah + max (Bh + Ch, Dh) Fh = Ah + Bh + Ch + Dh


Fw = max (Aw, max (Bw, Cw) + Dw ) Fw = max (Aw, Bw, Cw, Dw)

Fig. 3.4.3 (e7) Fig. 3.4.3 (e8)

Under condition of exhaustive search with all possible ordering of blocks A, B, C and D
with for each block + 90os rotation allowed there are few redundant structures from fig
3.4.3 (a, a1-a8) to fig. 3.4.3 (e, e1-a8) always produce same floorplan area FA = Fh * Fw.
Thus we identified five unique LB-compact placement structures and its two
compositions. In addition to this we have one more unique non-slicing LB-compact
placement structure its two compositions. Table 3.4.3 presents two compositions of six
unique placement structures four blocks placement.
Vertical Composition Horizontal composition

D
C
B
A B C D
A

Fh = Ah + Bh + Ch + Dh Fh = max (Ah, Bh, Ch, Dh)


Fw = max (Aw, Bw, Cw, Dw) Fw = Aw + Bw + Cw + Dw

C D
D
A B
B C
A
Fh = max (Ah, Bh, Ch + Dh)
Fh = Ah + Bh + max (Ch, Dh) Fw = Aw + Bw + max (Cw, Dw)
Fw = max (Aw, Bw, Cw + Dw)

37
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

D
B C D
C
A
A B
Fh = Ah + max (Bh, Ch, Dh) Fh = max (Ah, Bh + Ch + Dh)
Fw = max (Aw, Bw + Cw + Dw) Fw = Aw + max (Bw, Cw, Dw)

C D
B D
A B
A C
Fh = max (Ah, Bh) + max (Ch, Dh)
Fh = max (Ah + Bh, Ch + Dh)
Fw = max (Aw + Bw, Cw + Dw )
Fw = max (Aw + Bw) + max (Cw, Dw)

B D C D
A
C
B
A B
Fh = max (Ah, Bh + max (Ch, Dh))
Fh = Ah + max (Bh, Ch + Dh)
Fw = Aw + max (Bw, Cw + Dw)
Fw = max (Aw, Bw + max (Cw, Dw))

Non-slicing

C
B
D
D
C
A B A

Fh =max (Bh + Dh, max (Ah, Bh) + Ch ) Fh =max (Ah + Bh, max (Ah, Ch) + Dh )
Fw = max (max (Aw, Cw) + Bw, Cw + Dw ) Fw = max (max (Aw, Bw) + Cw, Bw + Dw )

38
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Table 3.4.1 two compositions of six unique placement structures for four blocks

3.5 Algorithm
We propose two algorithms for floorplanning. Both having complexity O(n) but first
algorithm requires recursive call in order of log4 n while second algorithm required
recursive call in order of n. We have taken bottom-up, recursive approach in these
algorithms. These algorithms use efficient exhaustive search procedure as explain in last
section for placing two, three or four rectangular blocks in a floorplan. Both algorithms
fall in class of constructive algorithm rather than class of iterative algorithm and work in
bottom-up constructive manner and they are recursive by nature. These algorithms
designed with concept that In case of placing few blocks together in non overlapping
manner, we can achieve better area utilization if blocks are having their area value in
neighbourhood if area values are arrange in order.

3.5.1 Algorithm-I
This algorithm starts with given blocks b1, b2, …, bn, before initiating recursive call, first
blocks are arranged in ascending order according to their area. Let say ordered list of
blocks as ab1, ab2, …, abn. Then list of composite blocks is generated from ordered list of
blocks as ab1, ab2, …, abn.. Here a composite block is a block that which generate after
placing 2, 3 or 4 blocks together using exhaustive search procedure. A composite block
also generated from placing 2, 3 or 4 composite blocks together. Let say list of composite
blocks as cb1, cb2, …, cbk. Here k = n / 4 if n mod 4 = 0 otherwise k = n / 4+1. In list of
composite blocks cb1 generated from first four blocks of order list ab1, ab2, …, abn, cb1
generated from next four blocks and so on up to cbk, generated from last four blocks from
our order list ab1, ab2, …, abn, if n mod 4 = 0 otherwise cbk generated from {abn-2, abn-1,
abn} if n mod 4 = 3, { abn-1, abn} if n mod 4 = 2 or {abn} if n mod 4 = 1. Here for
generating composite blocks list, blocks are selected in-group of four from order list of
blocks starting from smallest area and then up to end of list. So last the composite block

39
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

may have 4,3, 2 or 1 blocks or block according to number of blocks in list. Then the
selected group of four blocks are place using exhaustive search procedure, which
generates a composite block. A composite block has same property as a hard block define
previously. Thus this approach once again applies to new list of composite blocks. Before
initiating same recursive procedure, composite blocks are ordered according to area in
aviable list. This recursive procedure is stooped when only one composite block remains in
the list. And this composite block is our floorplan rectangle, which envelops n blocks in
non-overlapping manner.

This bottom up constructive approach provides us floorplan rectangle but exact co-
ordinates of each blocks has been not assigned. So with each returning from recursive call
in top-down way each composite block assign co-ordinated to it’s constitute blocks or
composite blocks according to rotation, ordering of blocks and LB-compact unique
structure used to generate that composite block.

At the end of algorithm we have rectangle R which contains blocks b1, b2, …, bn in non
overlapping manner and floorplan F = {(x1i, y1i, x2i, y2i) | 1 ≤ I ≤ n}, means each block
has bottom-left corners being assigned to (x1i, y1i) and top-right corners being assigned to
(x2i, y2i).

Algorithm: algorithm-I (listOfBlocks)


Input: listOfBlocks – blocks with height, width and aspect ratio range in case
of soft blocks.
Output: listOfBlocks – with each block having fix co-ordinates and aspect
ratio.
FloorplanH – Floorplan Height.
FloorplanW – Floorplan Width.

01 ArrangeBlocksInAscOrderOfArea (listOfBlocks);
02 If NumberOfBlocks (listOfBlocks) = 1 then
03 SetCordinateOfSubBlocks (listOfBlocks);
04 FloorplanH = firstBlock (listOfBlocks).Height;
05 FloorplanW = firstBlock (listOfBlocks).Width;
06 Return;
07 End If
08 newListOfCompositeBlocks = CreateCompositeBlocks (listOfBlocks);
09 Call algorithm-I (newListOfCompositeBlocks);
10 SetCordinateOfSubBlocks (newListOfCompositeBlocks);

3.5.2 Algorithm-II
This algorithm starts with given blocks b1, b2, …, bn, before initiating recursive call, first
blocks are arranged in ascending order according to their area. Let say ordered list of

40
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

blocks as ab1, ab2, …, abn. Then from list of blocks first four blocks are selected and a new
composite block is generated from it. Let say cb1234 as it is generated from b1, b2, b3, and b4
after exhaustive search procedure. The composite block added to list of order blocks after
replacing it’s constituted in order list. The composite block is inserted in order list
according to its area so that order is maintained in the list. Since a composite block has
property same as a hard block. Thus this approach once again applies to new list available
after adding new composite block. Thus with each recursive call 4 blocks are replaced with
1 composite block, hence size of list reduce by 3 at each recursive call. This recursive
procedure is stooped when only one composite block remains in the list. And this
composite block is our floorplan rectangle, which envelops n blocks in non-overlapping
manner.

This bottom up constructive approach provides us floorplan rectangle but exact co-
ordinates of each blocks has been not assigned. So with each returning from recursive call
in top-down way each composite block assign co-ordinated to it’s constitute blocks or
composite blocks according to rotation, ordering of blocks and LB-compact unique
structure used to generate that composite block.

At the end of algorithm we have rectangle R which contains blocks b1, b2, …, bn in non
overlapping manner and floorplan F = {(x1i, y1i, x2i, y2i) | 1 ≤ I ≤ n}, means each block
has bottom-left corners being assigned to (x1i, y1i) and top-right corners being assigned to
(x2i, y2i).

Algorithm: algorithm-II (listOfBlocks)


Input: listOfBlocks – blocks with height, width and aspect ratio range in case
of soft blocks.
Output: listOfBlocks – with each block having fix co-ordinates and aspect
ratio.
FloorplanH – Floorplan Height.
FloorplanW – Floorplan Width.

01 ArrangeBlocksInAscOrderOfArea (listOfBlocks);
02 If NumberOfBlocks (listOfBlocks) = 1 then
03 SetCordinateOfSubBlocks (listOfBlocks);
04 FloorplanH = firstBlock (listOfBlocks).Height;
05 FloorplanW = firstBlock (listOfBlocks).Width;
06 Return;
07 End If
08 newCompositeBlock = CreateOneCompositeBlock
(getFirstFourOrLessBlocks (listOfBlocks));
09 InsertNewBlockInList (newCompositeBlock, listOfBlocks);
10 Call algorithm-II (listOfBlocks);

41
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

11 SetCordinateOfSubBlocks (listOfBlocks);

3.5.3 Comparisons between Algorithm-I and Algorithm-II


In algorithm-I, numbers of recursive call are in order of log4 n while in algorithm-II,
numbers of recursive calls are equal to (n-2)/3 hence in order of n. If both algorithms are
evaluated with respect to numbers of exhaustive search procedures required, both are
same in this respect. Numbers of exhaustive search procedures required is equal to (n-2)/3
thus it is in order of n (O (n)).

In case of algorithm-II, composite block is inserted in order list of blocks according to its
area, while in Algorithm-II sorting is used to arrange the list of composite blocks.

Algorithm-I can easily implemented on distributed environment for better performance


because in this algorithm. We can divide our problem size n in two problems of each size
n/2.

3.6 Experimental Results


We have implemented the algorithm-I and algorithm-II in the C++ programming language
on a PC with Intel PIV 1.8 GHz CPU and 256 MB memory. We have compared algorithm-
I and algorithm-II with SP [6], O-tree [1], B*-tree [4], Enhanced O-tree [9], CBL [3],
TCG [2], TCG-S, FAST-SP [13] and GPE [11] based on the five MSNC benchmark
circuits. All of these algorithms are iterative algorithm. So they are taking much more time
then our algorithm and also producing better results in area utilization, while our algorithm
producing satisfactory results in area utilization and taking very less time. Here we have
compared algorithms only for hard blocks placement.
The area and runtime comparisons among SP [6] (on SUN Sparc Ultra-I), O-tree [1] (on a
200 MHz SUN Sparc Ultra-I workstation with 521 MB memory), B*-tree [4] (on a 200
MHz SUN Sparc Ultra-I workstation with 256 MB memory), Enhanced O-tree [9] (on a
SUN Sparc Ultra-60), CBL [3] (on a SUN Sparc Ultra-20), TCG [2] (on a 433 MHz SUN
Sparc Ultra-60 workstation with 1GB memory), TCG-S [16] (on a 433 MHz SUN Sparc
Ultra-60 workstation with 1GB memory), FAST-SP [13] (on ultra1) and GPE [11] (on a
PC with Intel PIII 800 MHz CPU and 128 MB memory) is provided from Table 3.6.3a to
Table 3.6.3e.

42
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Area utilization and runtimes for algorithm-I and algorithm-II are shown in Table 3.6.1 and
Table 3.6.2 respectively. Their comparisons are available in Table 3.6.3. Appendix A.1 and
Appendix A.2 contain circuit layout generated by algorithm-I and algorithm-II
respectively.

M Modul Width(mm Height(mm Floorplan Total Dead Time


CNC e ) ) Area(mm2 Modules Area(% (s)
Circuit Count ) Areas(mm2 )
)
ami49 49 10.4720 3.8220 40.0240 35.4454 11.4395 <1
ami33 33 1.8410 0.7000 1.2887 1.1564 10.2624 <1
Hp 11 2.2680 4.1160 9.3351 8.8306 5.4044 <1
Xerox 10 2.6670 7.7140 20.5732 19.3503 5.9443 <1
Apte 9 25.6140 1.8320 46.9248 46.5616 0.7740 <1
Table 3.6.1 Area utilization and runtime for Algorithm-I

MCNC Modul Width(mm Height(mm Floorplan Total Dead Time


Circuit e ) ) Area(mm2 Modules Area(% (s)
Count ) Areas(mm2 )
)
ami49 49 5.1520 7.8400 40.3917 35.4454 12.2457 <1
ami33 33 1.9600 0.7560 1.4818 1.1564 21.9544 <1
Hp 11 3.2200 3.3040 10.6389 8.8306 16.9971 <1
Xerox 10 2.6670 7.7140 20.5732 19.3503 5.9443 <1
Apte 9 25.6140 1.8320 46.9248 46.5616 0.7740 <1
Table 3.6.2 Area utilization and runtime for Algorithm-II

MCNC Module Minimum Algo-I Algo-II


Circuit Count Area(mm2)
Floorplan Dead Floorplan Dead Time
Area Area Time Area Area (s)
(mm2) (%) (s) (mm2) (%)
ami49 49 35.4454 40.0240 11.4395 <1 40.3917 12.2457 <1
ami33 33 1.1564 1.2887 10.2624 <1 1.4818 21.9544 <1
hp 11 8.8306 9.3351 5.4044 <1 10.6389 16.9971 <1
xerox 10 19.3503 20.5732 5.9443 <1 20.5732 5.9443 <1
apte 9 46.5616 46.9248 0.7740 <1 46.9248 0.7740 <1
Table 3.6.3 Area utilization and runtime comparison for Algorithm-I and Algorithm-II

43
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

MCNC Module Minimum SP O-tree


Circuit Count Area(mm2
Floorplan Dead Floorplan Dead Time
)
Area Area Time Area Area (s)
(mm2) (%) (s) (mm2) (%)
ami49 49 35.4454 38.842 8.7446 1580 37.6 5.7303 7428
ami33 33 1.1564 1.22 5.2131 676 1.25 7.488 1430
hp 11 8.8306 9.93 11.071 5 9.21 4.1194 57
xerox 10 19.3503 20.69 6.4751 15 20.1 3.7298 118
apte 9 46.5616 48.12 3.2385 13 47.1 1.1430 38
Table 3.6.3a Area utilization and runtime for SP and O-tree

MCNC Module Minimum B*-tree Enhanced O-tree


Circuit Count Area(mm2
Floorplan Dead Floorplan Dead Time
)
Area Area Time Area Area (s)
(mm2) (%) (s) (mm2) (%)
ami49 49 35.4454 36.80 3.6809 4752 37.73 6.0551 406
ami33 33 1.1564 1.27 8.9448 3417 1.24 6.7419 118
hp 11 8.8306 8.947 1.3009 55 9.16 3.5960 19
xerox 10 19.3503 19.83 2.4190 25 20.16 4.0163 38
apte 9 46.5616 46.92 0.7638 7 46.92 0.7638 11
Table 3.6.3b Area utilization and runtime for B*-tree and Enhanced O-tree

MCNC Module Minimum CBL TCG


Circuit Count Area(mm2
Floorplan Dead Floorplan Dead T
)
Area Area Time Area Area ime
(mm2) (%) (s) (mm2) (%) (s)
ami49 49 35.4454 38.58 8.1249 65 36.77 3.6023 434
ami33 33 1.1564 1.20 3.63333 36 1.20 3.6333 306
hp 11 8.8306 NA NA NA 8.947 1.3009 20
xerox 10 19.3503 20.96 7.6798 30 19.83 2.4190 18
apte 9 46.5616 NA NA NA 46.92 0.7638 1
Table 3.6.3c Area utilization and runtime for CBL and TCG

44
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

MCNC Module Minimum TCG-S FAST-SP


Circuit Count Area(mm2
Floorplan Dead Floorplan Dead Time
)
Area Area Time Area Area (s)
(mm2) (%) (s) (mm2) (%)
ami49 49 35.4454 36.40 2.6225 369 36.50 2.8893 31
ami33 33 1.1564 1.185 2.4135 84 1.205 4.0331 20
hp 11 8.8306 8.947 1.3009 7 8.947 1.3009 6
xerox 10 19.3503 19.796 2.2514 5 19.80 2.2712 14
apte 9 46.5616 46.92 0.7638 1 46.92 0.7638 1
Table 3.6.3d Area utilization and runtime for TCG-S and FAST-SP

MCNC Module Minimum GPE


Circuit Count Area(mm2
Floorplan Dead
)
Area Area(% Time
(mm2) ) (s)
ami49 49 35.4454 36.45 2.7561 247
ami33 33 1.1564 1.18 2 81
hp 11 8.8306 9.12 3.1732 2
xerox 10 19.3503 20.14 3.9210 2
apte 9 46.5616 46.90 0.7215 1

Table 3.6.3e Area utilization and runtime for GPE

We have also generated three test cases for checking our algorithm for placement of soft
blocks. In case-I, half numbers of blocks are soft and they are selected randomly. Range of
their aspect ratio is from 1.0 to 2.0 (with + 90os rotation allowed) with 0.1 as increment.
Table 3.6.4 shows pattern of hard and soft Blocks in test CASE –I. In Case-II, half
numbers of blocks are soft and they are complement of blocks in Case-I. It means those
blocks, which are soft in case-I, are hard in case-II and visa versa. Range of their aspect
ratio is from 1.0 to 2.0 (with + 90os rotation allowed) with 0.1 as increment. Table 3.6.5
shows pattern of hard and soft Blocks in test CASE –II. In Case-III, all blocks are soft.
Range of their aspect ratio is from 1.0 to 2.0 (with + 90os rotation allowed) with 0.1 as
increment.

MCNC Hard, Soft Pattern of Hard and Soft Blocks.


Circuit Blocks Arranged in ascending order of area, 1 and 0 represents hard

45
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Count and soft blocks respectively.


ami49 (25,24) {1000001011010110000110010101101110111010111001000}
ami33 (17,16) {001101100101010000101111100101001}
hp (6,5) {000001111011}
xerox (5,5) {0001111100}
apte (5,4) {1000001111}
Table 3.6.4 Pattern of Hard and Soft Blocks in test CASE -I

MCNC Hard, Soft Pattern of Hard and Soft Blocks.


Circuit Blocks Arranged in ascending order of area, 1 and 0 represents hard
Count and soft blocks respectively.
ami49 (24,25) {0111110100101001111001101010010001000101000110111}
ami33 (16,17) {110010011010101111010000011010110}
hp (5,6) {111110000100}
xerox (5,5) {1110000011}
apte (4,5) {0111110000}
Table 3.6.5 Pattern of Hard and Soft Blocks in test CASE -II

Area utilization and runtime of algorithm-I for case-I, case-II and case-III are shown in
Table 3.6.6, Table 3.6.7 and Table 3.6.8 respectively. And for algorithm-II its available in
Table 3.6.9, Table 3.6.10 and Table 3.6.11. The comparisons between algorithm-I and
algorithm II with respect to case-I, case-II and case-III are available in Table 3.6.12 and
Table 3.6.13. Appendix B.1 and Appendix B.2 contain circuit layout generated by
algorithm-I and algorithm-II respectively for case-I, case-II and case-III.

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead T


e Area(mm) Modules Area (%) ime
Count Areas(mm2) (s)
ami49 49 9.7980 4.1190 40.3580 35.4454 12.1724 9
ami33 33 1.5030 0.9960 1.4970 1.1564 22.7483 31
hp 11 5.8330 1.6090 9.3853 8.8306 5.9104 2
xerox 10 9.0030 2.4380 21.9493 19.3503 11.8410 31
apte 9 5.5230 9.7810 54.0205 46.5616 13.8074 0
Table 3.6.6 Area utilization and runtime for algorithm-I applied on case-I

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead Time


e Area(mm) Modules Area(% (s)
Count Areas(mm2) )
ami49 49 10.6370 3.7150 39.5165 35.4454 10.3021 9

46
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

ami33 33 1.8160 0.7320 1.3293 1.1564 13.0039 6


hp 11 10.4830 1.0080 10.5669 8.8306 16.4314 30
xerox 10 3.2070 7.6440 24.5143 19.3503 21.0653 3
apte 9 5.3780 10.4530 56.2162 46.5616 17.1741 2
Table 3.6.7 Area utilization and runtime for algorithm-I applied on case-II

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead Time


e Area(mm) Modules Area(% (s)
Count Areas(mm2) )
ami49 49 9.9360 3.9850 39.5950 35.4454 10.4800 367
ami33 33 1.8590 0.7300 1.3571 1.1564 14.7834 245
hp 11 3.4400 2.6520 9.1229 8.8306 3.2040 61
xerox 10 8.8570 2.4380 21.5934 19.3503 10.3878 62
apte 9 12.2180 3.9850 48.6887 46.5616 4.3688 60
Table 3.6.8 Area utilization and runtime for algorithm-I applied on case-III

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead Time


e Area(mm) Modules Area(% (s)
Count Areas(mm2) )
ami49 49 9.4570 4.1160 38.9250 35.4454 8.9392 8
ami33 33 1.8120 0.7310 1.3246 1.1564 12.6926 3
hp 11 4.5130 2.0440 9.2246 8.8306 4.2711 3
xerox 10 8.3230 2.5340 21.0905 19.3503 8.2510 30
apte 9 5.5230 9.7810 54.0205 46.5616 13.8074 1
Table 3.6.9 Area utilization and runtime for algorithm-II applied on case-I

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead Time


e Area(mm) Modules Area(% (s)
Count Areas(mm2) )
ami49 49 5.1320 7.6930 39.4805 35.4454 10.2204 7
ami33 33 1.5030 0.9780 1.4699 1.1564 21.3265 5
hp 11 2.3580 4.5300 10.6817 8.8306 17.3301 31
xerox 10 2.8830 7.6440 22.0377 19.3503 12.1944 2
apte 9 5.3780 10.4530 56.2162 46.5616 17.1741 2
Table 3.6.10 Area utilization and runtime for algorithm-II applied on case-II

Circuit Modul Width(mm) Height(mm) Floorplan Total Dead Time


e Area(mm) Modules Area(% (s)
Count Areas(mm2) )
ami49 49 8.9700 4.1840 37.5305 35.4454 5.5556 312
ami33 33 1.1930 1.3010 1.5521 1.1564 25.4910 216
hp 11 4.0700 2.2980 9.3529 8.8306 5.5841 63
xerox 10 8.1090 2.4470 19.8427 19.3503 2.4817 61

47
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

apte 9 12.2180 3.9850 48.6887 46.5616 4.3688 61


Table 3.6.11 Area utilization and runtime for algorithm-II applied on case-III

MCNC Module CASE I CASE II CASE III


Circuit Count
Dead Dead Time (s) Dead Time
Area(%) Time (s) Area(%) Area(%) (s)
ami49 49 12.1724 9 10.3021 9 10.4800 367
ami33 33 22.7483 31 13.0039 6 14.7834 245
hp 11 5.9104 2 16.4314 30 3.2040 61
xerox 10 11.8410 31 21.0653 3 10.3878 62
apte 9 13.8074 0 17.1741 2 4.3688 60
Table 3.6.12 Summary of area utilization and runtime for algorithm-I

MCNC Module CASE I CASE II CASE III


Circuit Count
Dead Dead Time (s) Dead Time
Area(%) Time (s) Area(%) Area(%) (s)
ami49 49 8.9392 8 10.2204 7 5.5556 312
ami33 33 12.6926 3 21.3265 5 25.4910 216
hp 11 4.2711 3 17.3301 31 5.5841 63
xerox 10 8.2510 30 12.1944 2 2.4817 61
apte 9 13.8074 1 17.1741 2 4.3688 61
Table 3.6.13 Summary of area utilization and runtime for algorithm-II

48
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Chapter 4

Conclusion and Future Work

4.1 Conclusion

In this thesis, we presented a bottom-up recursive approach to floorplanning using an


efficient exhaustive search procedure for placing two, three, or four rectangular blocks in a
floorplan. Exhaustive search procedure is also applicable to soft block. We considered
slicing as well as non-slicing structures. We developed two algorithms, which fall in class
of constructive approach rather than class of iterative approach. Algorithm-1 and and
algorithm-II are very fast compare to other iterative approach and also producing
promising results. Complexity of these algorithms is O(n). Experiments results with
MCNC circuits indicate that area utilization of about 85-99% can be achieved in very less
time then iterative algorithms. Drawback of these algorithms is that dead area gets
accumulated with each recursive call because, once a composite block is created from 2, 3
or 4 blocks we are considering this composite block as hard block and further using it for
creating higher order composite blocks. Thus due to above reason in few cases algorithms
are not performing with respect to area utilization. These algorithms are also applicable to
soft blocks. But in case of wide range of possible aspect ratios for soft blocks and with
four soft blocks out of four blocks under Exhaustive search procedure increases search
iteration tremendously. Thus under above conditions, algorithms do not perform well with
respect to runtime.

4.1 Future Work


Drawback of these algorithms can be eliminated with help of concepts provided in
following section.

4.1.1 Initial Arrangement of Soft blocks


Evenly distributing soft blocks among available hard blocks according to area, reduces
chance of getting all four consecutive soft blocks under exhaustive search procedure for
placements. Because a soft block can have a range of shapes as per the range of its aspect
ratio, with availability of some soft blocks in each exhaustive search procedure for
placement increases chances of a floorplan with better area utilization.

49
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

4.1.2 Extending Floorplan Sizing


An exhaustive search procedure for placement of 2, 3 or 4 blocks, produces a composite
block. We are considering this composite block as a hard block for further placements in
both recursive algorithms. But if this composite block is considered as soft block if while
creating this floorplan we have more than one floorplan option avialable which are having
equal area. Currently, if two or more floorplans have the same minimum area, then a
floorplan with aspect ratio nearer to 1.0 is selected. Further more this concept can be
extended with acceptable range of values of dead area for placements of 2, 3, 4 blocks. In
this case we are considering range for dead area in place of only minimum dead areas as
our selection criteria.

4.1.3 Exhaustive Search Procedure Extension


In this thesis, we have identified unique placement structure for at most 4 blocks
placement at a time. But this work is further extends for 5 blocks at a time once unique
placement structures for 5 blocks are derived. With five blocks a time provide more option
of placements and hence less dead area per each composition.

4.1.4 Iterative Algorithm


The complexity of presented algorithms are in order of O(n). So they can update to
iterative algorithm using simulated annealing, genetic algorithm or any other approaches.
An iterative algorithm with timeout will eliminate chance of not performing well in area
utilization.

4.1.5 A New Algorithm: Greedy Approach to Floorplanning


This algorithm generates list of composite blocks from given list of blocks. A composite
block is a block composed of two or more then two blocks. In this algorithm only slicing
structure has been used. This one is recursive algorithm. In this algorithm, at each pass,
according to specified selection criteria two blocks are selected from available list of
blocks. A block in list can either be simple block or a composite block. Then a composite
block is constructed from selected two blocks according composition rule specified. The
composite block further added to same list of blocks and its constituent blocks are
removed from list of blocks. Above procedure is repeated until it reaches to a termination
condition.

50
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

There are two variety of flows available from this propose algorithm. First flow output is
list of composite blocks while second flow ends with only one composite block and that
one is our final floorplan. Varity of flow is due to some input criteria, which is defined
further in this chapter.

This algorithm can be described using following there step:


1. Input Preparation
2. Core Recursive Algorithm
3. Output preparation

1. Input preparation:
We are given a set of n blocks or rectangular objects b1, b2, …, bn. and their respective
height and width are (h1, w1), (h2, w2), …, (hn , wn). Let A1, A2, …, An is respective areas of
blocks. We prepare a list { {E1, { h1, A1, C1}},{E1, { w1, A1, C1}, {E2, { h2, A2, C2}}, {E2, { w2, A2,
C2}}, …, {En, { hn, An, Cn}}, {En, { wn, An, Cn}}}. Where {E1, { h1, A1, C1}} is one node of list and
these is two such similar nodes in a list for each block.

Here in this list E1, E2, …, En are expression which illustrate floorplan composition of that
node in form of slicing tree. Since blocks are not composite blocks initially its respective
values are b1, b2, …, bn..

In case of Composite blocks, E1 may have values like b1’ b2’+ , b1, b2*, etc. Here b1’
b2’+ means blocks b1 and b2 are rotated and placed horizontally side by side and b1, b2*
mean b1 and b2 are placed vertically one above other but without rotation.

Here in this list C1, C2, …, Cn are composition index which specifies at which iteration this
composite block is created. These values are initialized to zero. At each pass we add only
one composite block in the list so every block has either composition index set to zero
value or distinct value on later stage.

We define three function size() , area() and composition index(). The function size() is
applicable to node and its value is equal to size of edge of block available in that node.
E.g. size({E1, { h1, A1, C1}}) = h1 and size({E1, { w1, A1, C1}) = w1. The function area() is

51
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

applicable to node and its value is equal to area of block available in that node. E.g.
area({E1, { h1, A1, C1}}) = A1 and area({E1, { w1, A1, C1}) = A1.

The function composition_ index() is applicable to node and its value is equal to
composition index of that node. E.g. composition_index ({E1, { h1, A1, C1}}) = C1 and
composition_ index ({E1, { w1, A1, C1}) = A1.

We prepare an order list having an invariance properties and it is then feed to our
algorithm. The invariance properties defined below is maintain by list through out
algorithm during any operations on list such as insertion of composite block . . The
invariance properties followed by list is that it always maintain order size (node1) ≤ size
(node2) ≤ … ≤ size (noden), If size (node1) = size (node2) then it should follow area
(node1) ≤ area (node2) and If area (node1) = area (node2) then it should follow
composition_ index (node1) ≤ composition_ index (node2).

In addition to order list we have more Input to algorithm and it is


Acceptable_Size_Range. It mean we can only create creating composite block from two
blocks if edge by which we are going for composition, should have length difference less
then Acceptable_Size_Range Acceptable_Size_Range decides the flow of algorithm if its
set to maximum of integer then output of our algorithm is only one composite block at
end. And if it sets to zero then there may be more then one composite blocks at end of
algorithm but dead area accumulation in each composite block would be zero. Then
further these types of composites blocks are feed to other algorithm for final
flooorplanning. And advantage in this algorithm is that we can achieve as many as possible
composite blocks with out any accumulation of dead area.

2. Core Recursive Algorithm:


Input:
OL: Order List as defined above
Acceptable_Size_Range: as defined above.
Composition_Count: Iteration count of this recursive algorithm, it
specifies composition_index, initially it’s equal to zero.

Output:
OL: may be with one composite block or more then one composite
block depends on Acceptable_Size_Rang
.
01 minAdjDiff = searchMinimumAdjacentNodeSizeDiffIgnoreSelfNode (OL);

52
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

02 If minAdjDiff <= Acceptable_Size_Range then


03 firstNode = getFirstNodeFromStart(OL, minAdjDiff);
04 secondNode = getSecondNodeFromStart(OL, minAdjDiff);
05 firstNodeSubling = getSubling (OL, firstNode);
06 secondNodeSubling = getSubling (OL, firstNodeSubling);
07 newCompositeBlock = CreateCB(firstNode, firstNodeSubling,
secondNode, secondNodeSubling);
08 SetCompositionIndex(newCompositeBlock, Composition_Count);
09 removeNodesfromOL(OL, firstNode, firstNodeSubling, secondNode,
secondNodeSubling);
10 insertNewCBinOL(OL, newCompositeBlock);
11 Composition_Count = Composition_Count + 1;
12 Call recursively same algorithm (OL, Acceptable_Size_Range,
Composition_Count);

3. Output Preparation:
Since two nodes are there in OL for each block or a composite block. In case of only one
composite block Floorplan Height and Floorplan Width is calculated from node.

FloorplanH = size (node1)


FloorplanW = size (node2)

And E1 and E2 which are equal because node1 and node2 are from same composite block.
Since E1 represent slicing structure and we also have height and width of each block as
(h1, w1), (h2, w2), …, (hn , wn) respectively. The co-ordinates of each block are calculated on
the base of above thing. If there is more then one composite block at the end of algorithm
then their height and width are extracted from OL and a new list of composite blocks is
created from it.

53
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

References

[1] P.-N. Guo, C.-K. Cheng, and T. Yoshimura, “Floorplanning Using a Tree
Representation,” IEEE TCAD February 2001, pp.281-289.
[2] Jai-Ming Lin and Yao-Wen Chang “TCG: A Transitive Closure Graph-Based
Representation for Non-Slicing Floorplans,” Proc. DAC, pp. 764–769, June2001.
[3] X. Hong, G. Huang, Y. Cai, S. Dong, C.-K. Cheng, and J. Gu, “Corner Block List: An
effective and efficient topological representation of non-slicing floorplan,” Proc. ICCAD,
pp. 8–12, Nov. 2000.
[4] Y.-C. Chang, Y.-W. Chang, G.-M. Wu, and S.-W. Wu, “B*-trees: A new representation
for nonslicing floorplans,” Proc. DAC, pp. 458–463, June 2000.
[5] D. F. Wong, and C.-L. Liu, “A new algorithm for floorplan design,” Proc. DAC, pp.
101–107, June 1986.
[6]H. Murata, K. Fujiyoshi, S. Nakatake, and Y. Kajitani, “Rectangle -packing based
module placement,” Proc. ICCAD, pp. 472–479, Nov. 1995.
[7] S. Nakatake, K. Fujiyoshi, H. Murata, and Y. Kajitani, “Module placement on BSG-
structure and IC layout applications,” Proc. ICCAD, pp. 484–491, Nov. 1996.
[8] H. Onodera, Y. Taniquchi, and K. Tamaru, “Branch-and-bound placement for building
block layout,” Proc. DAC, pp. 433–439, 1991.
[9] Y.-Pang, C.-K. Cheng, and T. Yoshimura, “An enhanced perturbing algorithm for
floorplan design using the O-tree representation,” Proc. ISPD, pp. 168-173, April 2000.
[10] J. Xu, P.N. Guo, C.K. Cheng, “Sequence Pair Approach for Rectilinear Module
Placement,” IEEE TCAD April 1999, pp.484-493
[11] Chang-Tzu Lin, De-Sheng Chen and Yi-Wen Wang, “GPE: A New Representation for
VLSI Floorplan Problem,” Proc. ICCD, pp. 42-44, 2002.
[12] S Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, “Optimization by simulated annealing,”
Science, pp.671–680, 1983.
[13] X. Tang and D. F. Wong, ”FAST-SP: A Fast Algorithm for Block Placement based on
Sequence Pair,” Proc. ASP-DAC, pp. 521-526, 2001.
[14] S. H. Gerez, “Algorithms for VLSI Design Automation”, John Wiley & Sons, 2000.
[15] Pinaki Mazumdar and Elizabeth M. Rudnick, “Genetic Algorithms for VLSI Design,
Layout & Test Automation”, Addison Wesley Longman, 2000.
[16] J.-M. Lin and Y.-W. Chang, .TCG-S: Orthogonal Coupling of P*-admissible
Representations for General Floorplans,. DAC 2002, pp. 842.847.

54
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

55
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Appendix

A.1 Circuit Layout Generated by Algorithm-I for Hard Blocks

Fig. A.1a to A.1e show layout of MCNC benchmark circuits ami49, ami33, hp, xerox and
apte respectively. These layouts are output of algorithm-I.

Fig. A.1a ami49 layout (hard blocks, algorithm-I)

56
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. A.1b ami33 layout (hard blocks, algorithm-I)

Fig. A.1c hp layout (hard blocks, algorithm-I)

57
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. A.1d xerox layout (hard blocks, algorithm-I)

Fig. A.1e apte layout (hard blocks, algorithm-I)

58
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

A.2 Circuit Layout Generated by Algorithm-II for Hard Blocks


Fig. A.2a to A2e show layout of MCNC benchmark circuits ami49, ami33, hp, xerox and
apte respectively. These layouts are output of algorithm-II.

Fig. A.2a ami49 layout (hard blocks, algorithm-II)

59
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. A.2b ami33 layout (hard blocks, algorithm-II)

Fig. A.2c hp layout (hard blocks, algorithm-II)

60
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. A.2d xerox layout (hard blocks, algorithm-II)

Fig. A.2e apte layout (hard blocks, algorithm-II)

61
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

B.1 Circuit Layout Generated by Algorithm-I for Soft Blocks

Fig. B.1a to B1e show layout of MCNC benchmark circuits ami49, ami33, hp, xerox and
apte respectively for case-I. Similarly fig. B.1f to B1j show layout for case-II and fig. B.1k
to B1o show layout for case-III. These layouts are output of algorithm-I.

Fig. B.1a ami49 layout (soft blocks, hard blocks, case-I, algorithm-I)

62
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1b ami33 layout (soft blocks, hard blocks, case-I, algorithm-I)

Fig. B.1c hp layout (soft blocks, hard blocks, case-I, algorithm-I)

63
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1d xerox layout (soft blocks, hard blocks, case-I, algorithm-I)

Fig. B.1e apte layout (soft blocks, hard blocks, case-I, algorithm-I)

64
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1f ami49 layout (soft blocks, hard blocks, case-II, algorithm-I)

Fig. B.1g ami33 layout (soft blocks, hard blocks, case-II, algorithm-I)

65
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1h hp layout (soft blocks, hard blocks, case-II, algorithm-I)

Fig. B.1i xerox layout (soft blocks, hard blocks, case-II, algorithm-I)

66
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1j apte layout (soft blocks, hard blocks, case-II, algorithm-I)

Fig. B.1k ami49 layout (soft blocks, hard blocks, case-III, algorithm-I)

67
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1l ami33 layout (soft blocks, hard blocks, case-III, algorithm-I)

Fig. B.1m hp layout (soft blocks, hard blocks, case-III, algorithm-I)

68
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.1n xerox layout (soft blocks, hard blocks, case-III, algorithm-I)

Fig. B.1o apte layout (soft blocks, hard blocks, case-III, algorithm-I)

69
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

B.2 Circuit Layout Generated by Algorithm-II for Soft Blocks


Fig. B.2a to B2e show layout of MCNC benchmark circuits ami49, ami33, hp, xerox and
apte respectively for case-I. Similarly fig. B.2f to B2j show layout for case-II and fig. B.2k
to B2o show layout for case-III. These layouts are output of algorithm-II.

Fig. B.2a ami49 layout (soft blocks, hard blocks, case-I, algorithm-II)

70
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2b ami33 layout (soft blocks, hard blocks, case-I, algorithm-II)

Fig. B.2c hp layout (soft blocks, hard blocks, case-I, algorithm-II)

71
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2d xerox layout (soft blocks, hard blocks, case-I, algorithm-II)

Fig. B.2e apte layout (soft blocks, hard blocks, case-I, algorithm-II)

72
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2f ami49 layout (soft blocks, hard blocks, case-II, algorithm-II)

Fig. B.2g ami33 layout (soft blocks, hard blocks, case-II, algorithm-II)

73
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2h hp layout (soft blocks, hard blocks, case-II, algorithm-II)

Fig. B.2i xerox layout (soft blocks, hard blocks, case-II, algorithm-II)

74
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2j apte layout (soft blocks, hard blocks, case-II, algorithm-II)

Fig. B.2k ami49 layout (soft blocks, hard blocks, case-III, algorithm-II)

75
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2l ami33 layout (soft blocks, hard blocks, case-III, algorithm-II)

76
Sponsored By: Bonrix Software Systems, Ahmedabad, India.
www.bonrix.net, www.bonrix.co.in

Fig. B.2m hp layout (soft blocks, hard blocks, case-III, algorithm-II)

Fig. B.2n xerox layout (soft blocks, hard blocks, case-III, algorithm-II)

Fig. B.2o apte layout (soft blocks, hard blocks, case-III, algorithm-II)

77

You might also like