Guide To FPGA

Lecture Notes in Electrical Engineering
Volume 149
For further volumes:

http://www.springer.com/series/7818
Jean-Pierre Deschamps Gustavo D. Sutter

Enrique Cant
Guide to FPGA
Implementation of
Arithmetic Functions
123
Jean-Pierre Deschamps
University Rovira i Virgili
Tarragona
Spain
Enrique Cant
University Rovira i Virgili
Tarragona
Spain
Gustavo D. Sutter
School of Computer Engineering
Universidad Autonoma de Madrid
Ctra. de Colmenar Km. 15
28049 Madrid
Spain
ISSN 1876-1100
ISBN 978-94-007-2986-5
DOI 10.1007/978-94-007-2987-2
e-ISSN 1876-1119
e-ISBN 978-94-007-2987-2
Springer Dordrecht Heidelberg New York London

Library of Congress Control Number: 2012935019
Springer Science+Business Media Dordrecht 2012
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publishers location, in its current version, and permission for use must always
be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright
Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Preface
Field Programmable Gate Arrays constitute one of the technologies at hand for
developing electronic systems. They form an attractive option for small production
quantities as their non-recurrent engineering costs are much lower than those
corresponding to ASICs. They also offer flexibility and fast time-to-market.
Furthermore, in order to reduce their size and, hence, the unit cost, an interesting
possibility is to reconfigure them at run time so that the same programmable
device can execute different predefined functions.
Complex systems, generally, are made up of processors executing programs,
memories, buses, input-output interfaces, and other peripherals of different types.
Those components are available under the form of Intellectual Property (IP) cores
(synthesizable Hardware Description Language descriptions or even physical
descriptions). Some systems also include specific components implementing
algorithms whose execution on an instruction-set processor is too slow. Typical
examples of such complex algorithms are: long-operand arithmetic operations,
floating-point operations, encoding and processing of different types of signals,
data ciphering, and many others. The way those specific components can be
developed is the central topic of this book. So, it addresses to both, FPGA users
interested in developing new specific componentsgenerally for reducing execution timesand, IP core designers interested in extending their catalog of
specific components.
This book distinguishes itself with the following aspects:
The main topic is circuit synthesis. Given an algorithm executing some complex
function, how can it be translated to a synthesizable circuit description? Which
are the choices that the designer can make in order to reduce the circuit cost,
latency, or power consumption? Thus, this is not a book on algorithms. It is a
book on how to efficiently translate an algorithm to a circuit using, for that
purpose, techniques such as parallelism, pipeline, loop unrolling, and others.
In particular, this is not a new version of a previous book by two of the authors.1
1
Deschamps JP, Bioul G, Sutter G (2006) Synthesis of Arithmetic Circuits. Wiley, New York.
vi
Preface
It is a complement of the cited book; its aim is the generation of efficient

implementations of some of the most useful arithmetic functions.
All throughout this book, numerous examples of FPGA implementation are
described. The circuits are modeled in VHDL. Complete and synthesizable
source files are available at the authors web site www.arithmetic-circuits.org.
This is not a book on Hardware Description Languages, and the reader is
assumed to have a basic knowledge of VHDL.
It is not a book on Logic Circuits either, and the reader is assumed to be
familiarized with the basic concepts of combinational and sequential circuits.
Overview
This book is divided into sixteen chapters. In the first chapter the basic building
blocks of digital systems are briefly reviewed, and their VHDL descriptions are
presented. It constitutes a bridge with previous courses, or books, on Hardware
Description Languages and Logic Circuits.
Chapters 24 constitute a first part whose aim is the description of the basic
principles and methods of algorithm implementation. Chapter 2 describes the
breaking up of a circuit into Data Path and Control Unit, and tackles the scheduling
and resource assignment problems. In Chaps. 3 and 4 some special topics of Data
Path and Control Unit synthesis are presented.
Chapter 5 recalls important electronic concepts that must be taken into account
for getting reliable circuits and Chap. 6 gives information about the main Electronic Design Automation (EDA) tools that are available for developing systems
on FPGAs.
Chapters 713 are dedicated to the main arithmetic operations, namely addition
(Chap. 7), multiplication (Chap. 8), division (Chap. 9), other operations such as
square root, logarithm, exponentiation, trigonometric functions, base conversion
(Chap. 10), decimal arithmetic (Chap. 11), floating-point arithmetic (Chap. 12),
and finite-field arithmetic (Chap. 13). For every operation, several configurations
are considered (combinational, sequential, pipelined, bit serial or parallel), and
several generic models are available, thus, constituting a library of virtual
components.
The development of Systems on Chip (SoC) is the topic of Chaps. 1416. The
main concepts are presented in Chap. 14: embedded processors, memories, buses,
IP components, prototyping boards, and so on. Chapter 15 presents two case
studies, both based on commercial EDA tools and prototyping boards. Chapter 16
is an introduction to dynamic reconfiguration, a technique that allows reducing the
area by modifying the device configuration at run time.
Acknowledgments
The authors thank the people who have helped them in developing this book,
especially Dr. Matthew Copley, for correcting the text. They are grateful to the
following universities for providing them the means for carrying this work through
to a successful conclusion: University Rovira i Virgili (Tarragona, Spain) and
Autonomous University of Madrid (Spain).
vii
Contents
Basic Building Blocks . . . . . . . . . . . . .

1.1 Combinational Components . . . . .
1.1.1
Boolean Equations . . . . .
1.1.2
Tables . . . . . . . . . . . . . .
1.1.3
Controllable Connections.
1.1.4
Arithmetic Circuits . . . . .
1.2 Sequential Components . . . . . . . .
1.2.1
Flip-Flops . . . . . . . . . . .
1.2.2
Registers . . . . . . . . . . . .
1.2.3
Counters . . . . . . . . . . . .
1.2.4
Finite State Machines . . .
1.3 Memory Blocks . . . . . . . . . . . . .
1.4 IO-Port Components . . . . . . . . . .
1.5 VHDL Models . . . . . . . . . . . . . .
1.6 Exercises . . . . . . . . . . . . . . . . . .
Reference . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
1
3
5
8
10
10
11
11
13
17
19
21
21
22
Architecture of Digital Circuits . . . .

2.1 Introductory Example . . . . . . .
2.2 Data Path and Control Unit . . .
2.3 Operation Scheduling . . . . . . .
2.3.1
Introductory Example .
2.3.2
Precedence Graph. . . .
2.3.3
Optimization Problems
2.4 Resource Assignment . . . . . . .
2.5 Final Example . . . . . . . . . . . .
2.6 Exercises . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
23
23
31
34
34
37
40
44
47
53
54
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ix
Contents
Special Topics of Data Path Synthesis . . . . . . . . . . . . . .

3.1 Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
Introductory Example . . . . . . . . . . . . . . . .
3.1.2
Segmentation . . . . . . . . . . . . . . . . . . . . . .
3.1.3
Combinational to Pipelined Transformation .
3.1.4
Interconnection of Pipelined Components . .
3.2 Loop Unrolling and Digit-Serial Processing . . . . . . .
3.3 Data Path Connectivity . . . . . . . . . . . . . . . . . . . . .
3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
55
55
55
58
62
66
72
77
78
81
Control Unit Synthesis . . . . . . . .

4.1 Command Encoding . . . . . .
4.2 Hierarchical Control Unit . .
4.3 Variable-Latency Operations
4.4 Exercises . . . . . . . . . . . . . .
Reference . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
83
83
86
91
94
94
Electronic Aspects of Digital Design . . . . . . . . . . . . . . . .

5.1 Basic Electronic Aspects of Digital Design . . . . . . . .
5.1.1
Basic Concepts. . . . . . . . . . . . . . . . . . . . . .
5.1.2
Propagation DelayTransition Time . . . . . .
5.1.3
Glitches in Digital Circuits . . . . . . . . . . . . .
5.2 Synchronous Design Issues . . . . . . . . . . . . . . . . . . .
5.2.1
Edge Sensitive and Level Sensitive Registers.
5.2.2
Temporal Parameters of Flip-Flops . . . . . . . .
5.2.3
Metastability . . . . . . . . . . . . . . . . . . . . . . .
5.3 Clock Distribution Network . . . . . . . . . . . . . . . . . . .
5.3.1
Clock Skew . . . . . . . . . . . . . . . . . . . . . . . .
5.3.2
Clock Jitter . . . . . . . . . . . . . . . . . . . . . . . .
5.3.3
Clock Gating . . . . . . . . . . . . . . . . . . . . . . .
5.3.4
Clock Managers . . . . . . . . . . . . . . . . . . . . .
5.3.5
Interfacing Different Clock Domains . . . . . .
5.4 Power Consumption . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1
Sources of Power Consumption . . . . . . . . . .
5.4.2
Power Reduction Techniques . . . . . . . . . . . .
5.4.3
Power Measurement and Estimation . . . . . . .
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
95
95
95
99
101
103
104
104
105
108
110
112
113
113
115
119
120
122
123
123
125
EDA Tools . . . . . . . . . . . . . . . . . . . . . .
6.1 Design Flow in FPGA EDA Tools .
6.1.1
Design Entry . . . . . . . . . .
6.1.2
Synthesis . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
127
127
128
130
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
xi
6.1.3
Implementation (Mapping, Placement

and Routing) . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.4
Programming File Generation and Programming
6.2 Implementation Constraints . . . . . . . . . . . . . . . . . . . . .
6.2.1
Timing Constrains . . . . . . . . . . . . . . . . . . . . .
6.2.2
Placement and Other Constrains . . . . . . . . . . . .
6.3 System Verification . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.1
Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3.2
Formal Verification . . . . . . . . . . . . . . . . . . . .
6.3.3
In-Circuit Co-Simulation . . . . . . . . . . . . . . . . .
6.3.4
In-Circuit Testing and Debugging. . . . . . . . . . .
6.3.5
Design for Test . . . . . . . . . . . . . . . . . . . . . . .
6.4 Timing Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5 Power Consumption Estimation . . . . . . . . . . . . . . . . . .
6.5.1
Reducing the Power Consumption . . . . . . . . . .
6.6 Example of EDA Tool Usage . . . . . . . . . . . . . . . . . . .
6.6.1
Simple Example Using Xilinx ISE . . . . . . . . . .
6.6.2
Simple Example Using Altera Quartus II. . . . . .
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
132
133
133
134
135
135
135
137
137
137
138
139
140
141
141
142
148
149
151
Adders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1 Addition of Natural Numbers. . . . . . . . . . . .
7.2 Binary Adder . . . . . . . . . . . . . . . . . . . . . . .
7.3 Radix-2k Adder . . . . . . . . . . . . . . . . . . . . .
7.4 Carry Select Adders . . . . . . . . . . . . . . . . . .
7.5 Logarithmic Adders . . . . . . . . . . . . . . . . . .
7.6 Long-Operand Adder . . . . . . . . . . . . . . . . .
7.7 Multioperand Adders . . . . . . . . . . . . . . . . .
7.7.1
Sequential Multioperand Adders. . . .
7.7.2
Combinational Multioperand Adders.
7.7.3
Parallel Counters . . . . . . . . . . . . . .
7.8 Subtractors and AdderSubtractors . . . . . . . .
7.9 FPGA Implementations . . . . . . . . . . . . . . . .
7.9.1
Binary Adder . . . . . . . . . . . . . . . . .
7.9.2
Radix 2k Adders. . . . . . . . . . . . . . .
7.9.3
Carry Select Adder . . . . . . . . . . . . .
7.9.4
Logarithmic Adders . . . . . . . . . . . .
7.9.5
Long Operand Adder . . . . . . . . . . .
7.9.6
Sequential Multioperand Adders. . . .
7.9.7
Combinational Multioperand Adders.
7.9.8
Comparison . . . . . . . . . . . . . . . . . .
7.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
153
153
155
155
159
160
162
164
164
168
171
175
177
178
178
179
179
179
180
180
181
181
182
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
xii
Contents
Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1 Basic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2 Combinational Multipliers . . . . . . . . . . . . . . . . . . . . .
8.2.1
Ripple-Carry Parallel Multiplier . . . . . . . . . . .
8.2.2
Carry-Save Parallel Multiplier . . . . . . . . . . . .
8.2.3
Multipliers Based on Multioperand Adders . . .
8.2.4
Radix-2k and Mixed-Radix Parallel Multipliers
8.3 Sequential Multipliers . . . . . . . . . . . . . . . . . . . . . . . .
8.3.1
Shift and Add Multiplier . . . . . . . . . . . . . . . .
8.3.2
Shift and Add Multiplier with CSA . . . . . . . .
8.4 Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.4.1
Mod 2Bn+m Multiplication . . . . . . . . . . . . . . .
8.4.2
Modified Shift and Add Algorithm . . . . . . . . .
8.4.3
Post Correction Multiplication . . . . . . . . . . . .
8.4.4
Booth Multiplier. . . . . . . . . . . . . . . . . . . . . .
8.5 Constant Multipliers . . . . . . . . . . . . . . . . . . . . . . . . .
8.6 FPGA Implementations . . . . . . . . . . . . . . . . . . . . . . .
8.6.1
Combinational Multipliers . . . . . . . . . . . . . . .
8.6.2
Radix-2k Parallel Multipliers . . . . . . . . . . . . .
8.6.3
Sequential Multipliers . . . . . . . . . . . . . . . . . .
8.6.4
Combinational Multipliers for Integers . . . . . .
8.6.5
Sequential Multipliers for Integers . . . . . . . . .
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
183
183
185
185
187
189
191
195
195
198
199
200
202
206
208
211
215
215
216
216
217
219
219
220
Dividers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1 Basic Digit-Recurrence Algorithm . . . . . . . .
9.2 Radix-2 Division . . . . . . . . . . . . . . . . . . . .
9.2.1
Non-Restoring Divider . . . . . . . . . .
9.2.2
Restoring Divider . . . . . . . . . . . . . .
9.2.3
Binary SRT Divider . . . . . . . . . . . .
9.2.4
Binary SRT Divider with Carry-Save
9.2.5
Radix-2k SRT Dividers . . . . . . . . . .
9.3 Radix-B Dividers . . . . . . . . . . . . . . . . . . . .
9.4 Convergence Algorithms . . . . . . . . . . . . . . .
9.5 FPGA Implementations . . . . . . . . . . . . . . . .
9.5.1
Digit-Recurrence Algorithms . . . . . .
9.5.2
Convergence Algorithms . . . . . . . . .
9.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
221
221
224
224
227
229
232
234
239
245
246
247
248
248
249
10 Other Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.1 Binary to Radix-B Conversion (B even) . . . . . . . . . . . . . . . .
10.2 Radix-B to Binary Conversion (B even) . . . . . . . . . . . . . . . .
251
251
252
.....
.....
.....
.....
.....
.....
Adder
.....
.....
.....
.....
.....
.....
.....
.....
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
10.3 Square Rooters . . . . . . . . . . . . . . . . . . . . . . . . . .

10.3.1 Restoring Algorithm . . . . . . . . . . . . . . . .
10.3.2 Non-Restoring Algorithm . . . . . . . . . . . .
10.3.3 Fractional Numbers. . . . . . . . . . . . . . . . .
10.3.4 Convergence Methods (NewtonRaphson) .
10.4 Logarithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 Exponential . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6 Trigonometric Functions . . . . . . . . . . . . . . . . . . .
10.7 FPGA Implementations . . . . . . . . . . . . . . . . . . . .
10.7.1 Converters . . . . . . . . . . . . . . . . . . . . . . .
10.7.2 Square Rooters . . . . . . . . . . . . . . . . . . . .
10.7.3 Logarithm and Exponential . . . . . . . . . . .
10.7.4 Trigonometric Functions . . . . . . . . . . . . .
10.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xiii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
254
254
258
260
260
262
264
268
272
272
273
274
274
274
275
11 Decimal Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.1 Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.1.1 Decimal Ripple-Carry Adders . . . . . . . . . . .
11.1.2 Base-B Carry-Chain Adders. . . . . . . . . . . . .
11.1.3 Base-10 Carry-Chain Adders . . . . . . . . . . . .
11.1.4 FPGA Implementation of the Base-10
Carry-Chain Adders . . . . . . . . . . . . . . . . . .
11.2 Base-10 Complement and Addition: Subtration . . . . .
11.2.1 Tens Complement Numeration System. . . . .
11.2.2 Tens Complement Sign Change . . . . . . . . .
11.2.3 10s Complement BCD Carry-Chain
Adder-Subtractor . . . . . . . . . . . . . . . . . . . .
11.2.4 FPGA Implementations of Adder Subtractors.
11.3 Decimal Multiplication . . . . . . . . . . . . . . . . . . . . . .
11.3.1 One-Digit by One-Digit BCD Multiplication .
11.3.2 N by One BCD Digit Multiplier. . . . . . . . . .
11.3.3 N by M Digits Multiplier . . . . . . . . . . . . . .
11.4 Decimal Division . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4.1 Non-Restoring Division Algorithm . . . . . . . .
11.4.2 An SRT-Like Division Algorithm. . . . . . . . .
11.4.3 Other Methods for Decimal Division . . . . . .
11.5 FPGA Implementation Results . . . . . . . . . . . . . . . . .
11.5.1 Adder-Subtractor Implementations . . . . . . . .
11.5.2 Multiplier Implementations . . . . . . . . . . . . .
11.5.3 Decimal Division Implementations . . . . . . . .
11.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
277
277
277
278
280
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
281
282
282
283
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
283
284
286
286
288
289
290
291
293
297
298
298
299
301
303
303
xiv
Contents
12 Floating Point Arithmetic . . . . . . . . . . . . . . . . . .

12.1 IEEE 754-2008 Standard . . . . . . . . . . . . . . .
12.1.1 Formats . . . . . . . . . . . . . . . . . . . . .
12.1.2 Arithmetic and Interchange Formats .
12.2 Arithmetic Operations . . . . . . . . . . . . . . . . .
12.2.1 Addition of Positive Numbers . . . . .
12.2.2 Difference of Positive Numbers . . . .
12.2.3 Addition and Subtraction . . . . . . . . .
12.2.4 Multiplication . . . . . . . . . . . . . . . .
12.2.5 Division . . . . . . . . . . . . . . . . . . . .
12.2.6 Square Root . . . . . . . . . . . . . . . . . .
12.3 Rounding Schemes . . . . . . . . . . . . . . . . . . .
12.3.1 Rounding Schemes in IEEE 754. . . .
12.4 Guard Digits . . . . . . . . . . . . . . . . . . . . . . .
12.5 Arithmetic Circuits . . . . . . . . . . . . . . . . . . .
12.5.1 AdderSubtractor . . . . . . . . . . . . . .
12.5.2 Multiplier . . . . . . . . . . . . . . . . . . .
12.5.3 Divider . . . . . . . . . . . . . . . . . . . . .
12.5.4 Square Root . . . . . . . . . . . . . . . . . .
12.5.5 Implementation Results . . . . . . . . . .
12.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . .
Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
305
305
306
307
309
309
311
313
314
315
316
318
319
320
321
322
327
331
332
334
335
336
13 Finite-Field Arithmetic . . . . . . . . . . . . . . . . . . . . . . .
13.1 Operations Modulo m . . . . . . . . . . . . . . . . . . . .
13.1.1 Addition and Subtraction Mod m . . . . . .
13.1.2 Multiplication Mod m . . . . . . . . . . . . . .
13.2 Division Modulo p . . . . . . . . . . . . . . . . . . . . . .
13.3 Operations Over Z2 x=fx . . . . . . . . . . . . . . . .
13.3.1 Addition and Subtraction of Polynomials.
13.3.2 Multiplication Modulo f(x). . . . . . . . . . .
13.4 Division Over GF(2m) . . . . . . . . . . . . . . . . . . . .
13.5 FPGA Implementations . . . . . . . . . . . . . . . . . . .
13.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
337
337
338
339
346
347
347
347
351
353
354
355
14 Systems on Chip. . . . . . . . . . . . . . . . . . .
14.1 System on Chip . . . . . . . . . . . . . . .
14.2 Intellectual Property Cores. . . . . . . .
14.3 Embedded Systems . . . . . . . . . . . . .
14.3.1 Embedded Microprocessors .
14.3.2 Peripherals. . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
357
357
357
358
359
362
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contents
14.3.3
14.3.4
14.3.5
References .
xv
Coprocessors
Memory . . .
Busses. . . . .
...........
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
365
366
368
369
15 Embedded Systems Development: Case Studies . . . . . . . . . . .

15.1 Introduction to Xilinx EDK . . . . . . . . . . . . . . . . . . . . . .
15.1.1 Case Study 1-A: A Basic Embedded System . . . .
15.1.2 Hardware. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.1.3 Programming and Debugging. . . . . . . . . . . . . . .
15.2 Case Study 1-B: Creating a Custom Peripheral . . . . . . . .
15.2.1 Design of a Custom Peripheral. . . . . . . . . . . . . .
15.2.2 Systems Hardware/Software Modification . . . . .
15.2.3 Functional Simulation . . . . . . . . . . . . . . . . . . . .
15.3 Case Study 2: Implementation of a Custom Coprocessor. .
15.3.1 A Brief Introduction to the AES-128 Cipher . . . .
15.3.2 Software Implementation of the AES-128 Cipher .
15.3.3 Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.3.4 Coprocessor Design . . . . . . . . . . . . . . . . . . . . .
15.3.5 Modification of the Embedded System . . . . . . . .
15.3.6 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15.3.7 Experimental Results . . . . . . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
371
371
374
375
394
397
397
406
409
414
416
417
422
424
429
431
434
434
16 Partial Reconfiguration on Xilinx FPGAs . . . . . . . . . . .

16.1 Partial Reconfiguration on Xilinx FPGAs . . . . . . . .
16.2 Design Flow for Partial Reconfiguration . . . . . . . . .
16.3 Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16.3.1 Hardware Design . . . . . . . . . . . . . . . . . . .
16.3.2 Partial Reconfiguration Flow on PlanAhead.
16.3.3 Software Development . . . . . . . . . . . . . . .
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
435
435
437
437
439
445
452
459
About the Authors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
461
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
463
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Chapter 1
Basic Building Blocks
Digital circuits are no longer defined by logical schemes but by Hardware

Description Language programs [1]. The translation of this kind of definition to an
actual implementation is realized by Electronic Automation Design tools (Chap. 5).
All along this book the chosen language is VHDL. In this chapter the most useful
constructions are presented. For all of the proposed examples, the complete source
code is available at the Authors web page.
1.1 Combinational Components

1.1.1 Boolean Equations
Among the predefined operations of any Hardware Description Language are the
basic Boolean operations. Boolean functions can easily be defined. Obviously, all
the classical logic gates can be defined. Some of them are considered in the
following example.
Example 1.1
The following VHDL instructions define the logic gates NOT, AND2, OR2,
NAND2, NOR2, XOR2, XNOR2, NAND3, NOR3, XOR3, XNOR3 (Fig. 1.1).
J.-P. Deschamps et al., Guide to FPGA Implementation of Arithmetic Functions,

Lecture Notes in Electrical Engineering 149, DOI: 10.1007/978-94-007-2987-2_1,
1 Basic Building Blocks
NOT
AND2
OR2
NAND3
NOR3
NAND2
NOR2
XOR3
XNOR3
XOR2
XNOR2
Fig. 1.1 Logic gates
The same basic Boolean operations can be used to define sets of logic gates
working in parallel. As an example, assume that signals a = (an-1, an-2,, a0) and
b = (bn-1, bn-2,, b0) are n-bit vectors. The following assignation defines a set of
n AND2 gates that compute c = (an-1bn-1, an-2bn-2,, a0b0):
More complex Boolean functions can also be defined. It is worthwhile to indicate

that within most Field Programmable Gate Arrays (FPGA) the basic combinational
components are not 2-input logic gates but Look Up Tables (LUT) allowing the
implementation of any Boolean function of a few numbers (4, 5, 6) of variables
(Chap. 5). Hence, it makes sense to consider the possibility of defining small
combinational components by the set of Boolean functions they implement.
Example 1.2
The following VHDL instruction defines an ANDORI gate (a 4-input 1-output
component implementing the complement of ab v cd, Fig. 1.2)
and the two following instructions define a 1-bit full adder (a 3-input 2-output
component implementing a+b+cin mod 2 and ab v acin v bcin, Fig. 1.3).
Fig. 1.2 ANDORI gate
Fig. 1.3 One-bit full adder
cout
FA
cin
1.1.2 Tables
Combinational components can also be defined by their truth tables, without the
necessity to translate their definition to Boolean equations. As a first example,
consider a 1-digit to 7-segment decoder, that is a 4-input 7-output combinational
circuit.
Example 1.3
The behavior of a 4-to-7 decoder (Fig. 1.4) can be described by a conditional
assignment instruction
The last choice of a WITH SELECT construction must be WHEN OTHERS in

order to avoid the inference of an additional latch, and it is the same for other
multiple choice instructions such as CASE.
As mentioned above, small Look Up Tables are basic components of most
FPGA families. The following example is a generic 4-input 1-output LUT.

segment 6
segment 4
segment5
4
decoder
digit
segment 3
7
segments
segment2
segment 1
segment 0
Fig. 1.4 Four-digit to seven-segment decoder
Fig. 1.5 Four-input Look Up

Table
16-bit ROM
a 3,a 2,a 1,a 0
4
stored data:
truth vector of f
b = f(a 3,a 2,a1,a0)
Example 1.4
The following entity defines a 4-input Boolean function whose truth vector is
stored within a generic parameter (Fig. 1.5). Library declarations are omitted.
Then the following component instantiation defines a 4-input XOR function.
Fig. 1.6 Multiplexers
sel
2
sel
a
b
n
n
a
n
1
MUX2-1
b
c
d
n
n
n
n
00
01
10
11
MUX4-1
Comment 1.1
The VHDL model of the preceding Example 1.4 can be used for simulation
purposes, independently of the chosen FPGA vendor. Nevertheless, in order to
implement an actual circuit, the corresponding vendors primitive component
should be used instead (Chap. 5).
1.1.3 Controllable Connections

Multiplexers are the basic components for implementing controllable connections.
Conditional assignments are used for defining them. Some of them are considered
in the following example.
Example 1.5
The following conditional assignments define a 2-to-1 and a 4-to-1 multiplexer
(Fig. 1.6). The signal types must be compatible with the conditional assignments:
a, b, c, d and e are assumed to be n-bit vectors for some constant value n.
Demultiplexers, address decoders and tri-state buffers are other components frequently used for implementing controllable connections such as buses.
Example 1.6
The following equations define a 1-bit 1-to-2 demultiplexer (Fig. 1.7a)
and the following conditional assignments a 1-to-4 demultiplexer (Fig. 1.7b)
6
Fig. 1.7 Demultiplexers
(a)
(b)
sel
0
00
a
1
sel
2
DEMUX1-2
01
10
11
n
n
n
n
b
c
c
d
DEMUX1-4
In the second case the signal types must be compatible with the conditional
assignments: a, b, c, d and e are assumed to be n-bit vectors for some constant
value n.
An address decoder is a particular case of 1-bit demultiplexer whose input is 1.
Example 1.7
The following equations define a 3-input address decoder (Fig. 1.8).
Three-state buffers implement controllable switches.
Fig. 1.8 Address decoder
row(0)
row(1)
row(2)
row(3)
row(4)
000
001
010
011
100
101
address 3
row(5)
row(6)
110
111
Fig. 1.9 Three-state buffer
row(7)
a n
enable
Fig. 1.10 Example of a data

bus
y2y 1
2
11
1
c
8
enable_b
10
1
0
y0
d
8
enable_c
01
00
b
8
enable_a
enable_d
8
enable_e
8
data_bus
Example 1.8
The following conditional assignment defines a tri-state buffer (Fig. 1.9): when the
enable signal is equal to 1, output b is equal to input a, and when the enable signal
is equal to 0, output b is disconnected (high impedance state). The signal types
must be compatible with the conditional assignments: a and b are assumed to be nbit vectors for some constant value n.
An example of the use of demultiplexers and tri-state buffers, namely a data bus, is
shown in Fig. 1.10.
Example 1.9
The circuit of Fig. 1.10, made up of demultiplexers and three-state buffers, can be
described by Boolean equations (the multiplexers) and conditional assignments
(the three-sate buffers).
Fig. 1.11 Data bus, second

version
e
n
d
n
a
n
000 001 01- 10- 11y 2y1y 0

n
data bus
Nevertheless, the same functionality can be implemented by an 8-bit 5-to-1

multiplexer (Fig. 1.11).
Generally, this second implementation is considered safer than the first one. In
fact, tri-state buffers should not be used within the circuit core. They should only
be used within I/O-port components (Sect. 1.4).
1.1.4 Arithmetic Circuits

Among the predefined operations of any Hardware Description Language there are
also the basic arithmetic operations. The translation of this kind of description to

Fig. 1.12 n-bit adders
(a)
(b)
mod 2n adder
a
0
c in
c out
mod 2n+1 adder
c in
Fig. 1.13 n-bit by m-bit

multiplier
a
b
n
m
n-bit by m-bit
multiplier
n+m
product
actual implementations, using special purpose FPGA resources (carry logic,

multiplier blocks), is realized by Electronic Automation Design tools (Chap. 5).
Example 1.10
The following arithmetic equation defines an adder mod 2n, being a, b and s n-bit
vectors and cIN an input carry (Fig. 1.12a).
By adding a most significant bit 0 to one (or both) of the n-bit operands a and b, an
n-bit adder with output carry cOUT can be defined (Fig. 1.12b). The internal signal
sum is an (n+1)-bit vector.
The following arithmetic equation defines an n-bit by m-bit multiplier where a is

an n-bit vector, b an m-bit vector and product an (n+m)-bit vector (Fig. 1.13).
Comment 1.2
In most VHDL models available at the Authors web page, the type unsigned has
been used, so that bit-vectors can be treated as natural numbers. In some cases, it
could be better to use the signed type, for example when bit-vectors are interpreted
as 2s complement integers, and when magnitude comparisons or sign-bit extensions are performed.
10
Fig. 1.14 D flip-flops
(a)
(b)
D
(clk)
Q
(clkb)
PRESET
D
Q
Qb
CLEARb
1.2 Sequential Components

1.2.1 Flip-Flops
The basic sequential component is the D-flip-flop. In fact, several types of D-flipflop can be considered: positive edge triggered, negative edge triggered, with
asynchronous input(s) and with complemented output.
Example 1.11
The following component is a D-flip-flop triggered by the positive edge of clk
(Fig. 1.14a),
while the following, which is controlled by the negative edge of clkb, has two
asynchronous inputs clearb (active at low level) and preset (active at high
level), having clearb priority, and has two complementary outputs q and qb
(Fig. 1.14b).
Comment 1.3
The use of level-controlled, instead of edge-controlled, components is not recommendable. Nevertheless, if it were necessary, a D-latch could be modeled as
follows:
where en is the enable signal and d the data input.
11
Fig. 1.15 Parallel register
d
n
CE
register
(clk)
n
q
1.2.2 Registers
Registers are sets of D-flip-flops controlled by the same synchronization and
control signals, and connected according to some regular scheme (parallel, left or
right shift, bidirectional shift).
Example 1.12
The following component is a parallel register with ce (clock enable) input,
triggered by the positive edge of clk (Fig. 1.15).
As a second example (Fig. 1.16), the next component is a right shift register with
parallel input (controlled by load) and serial input (controlled by shift).
1.2.3 Counters
A combination of registers and arithmetic operations permits the definition of
counters.
12
Fig. 1.16 Right shift register
parallel_in
n
serial_in
right shift register
load
shift
(clk)
n
q
Fig. 1.17 Up/down counter
parallel_in
n
load
count
up/down counter upb_down

(clk)
n
q
Example 1.13
This defines an up/down counter (Fig. 1.17) with control signals load (input
parallel_in), count (update the state of the counter) and upb_down (0: count up, 1:
count down).
The following component is a down counter (Fig. 1.18) with control signals load
(input parallel_in) and count (update the state of the counter). An additional binary
output equal_zero is raised when the state of the counter is zero (all 0s vector).
13
Fig. 1.18 Down counter
parallel _in
n
down counter
equal_zero
load
count
(clk)
n
q
combinational
circuit 1
(t1)
next state
internal state
combinational
circuit 2
(t2)
output state
input state
(tSUinput )
(clk)
clk
internal state
input state
output state
next state
tSUinput
t1
t2
Fig. 1.19 Moore machine
1.2.4 Finite State Machines

Hardware Description Languages allow us to define finite state machines at an
input/output behavioral level. The translation to an actual implementation
including registers and combinational circuitsa classical problem of traditional
switching theoryis realized by Electronic Automation Design tools (Chap. 5).
In a Moore machine, the output state only depends on the current internal state
(Fig. 1.19) while in a Mealy machine the output state depends on both the input
state and the current internal state (Fig. 1.20). Let tSUinput be the maximum set up
time of the input state with respect to the positive clock edge, t1 the maximum
delay of the combinational block that computes the next internal state, and t2 the
maximum delay of the combinational block that computes the output state. Then,
in the case of a Moore machine, the following conditions must hold
tSUinput t1 \TCLK and t2 \TCLK;
1:1
and in the case of a Mealy machine

tSUinput t1 \TCLK and tSUinput t2 \TCLK :
1:2
14
combinational
circuit 1
(t1)
next state
internal state
input state
(tSUinput )
combinational
circuit 2
(t2)
output state
(clk)
clk
internal state
input state
output state
next state
tSUinput
t2
t1
Fig. 1.20 Mealy machine
The set up and hold times of the register (Chap. 6) have not been taken into
account.
Example 1.14
A Moore machine is shown in Fig. 1.21. It is the control unit of a programmable
timer (Exercise 2.6.2). It has seven internal states, three binary input signals start,
zero and reference, and two output signals operation (2 bits) and done. It can be
described by the following processes.
15
start = 1
start = 0
reset
start = 0
start = 1
zero = 1
reference = 1
zero = 0
reference = 0
reference = 0
4
reference = 1
5
state
operation
done
0
1
2
3
4
5
6
00
00
11
00
00
00
01
1
1
0
0
0
0
0
Fig. 1.21 An example of a Moore machine
Example 1.15
Consider the Mealy machine of Table 1.1. It has four internal states, two
binary inputs x1 and x0, and one binary output z. Assume that x0 and x1 are
16
Table 1.1 A Mealy
machine: next state/z
A
B
C
D
X1 x0 : 00
01
10
11
A/0
B/1
B/1
A/0
B/0
B/0
C/1
C/1
A/1
A/1
D/0
D/0
D/1
C/0
C/0
D/1
periodic, but out of phase, signals. Then the machine detects if x0 changes
before x1 or if x1 changes before x0. In the first case the sequence of internal
states is A B C D A B and z = 0. In the second case the sequence is D C B
A D C and z = 1.
It can be described by the following processes.
17
1.3 Memory Blocks

With regards to memory blocks, a previous comment similar to Comment 1.1 must
be outlined: VHDL models can be generated for simulation purposes; nevertheless,
in order to implement an actual circuit, the corresponding vendors primitive
component should be used instead (Chap. 5).
Example 1.16
The following entity defines a Read Only Memory storing 2n m-bit words. The
stored data is defined by a generic parameter (Fig. 1.22). Library declarations are
omitted.
18
Fig. 1.22 Read only

memory
ROM
address
stored data :
generic parameter
m
word
Fig. 1.23 Random access

memory
data_in
m
address
write
clk
RAM
data_out
Then the following component instantiation defines a ROM storing 16 4-bit words,
namely
Example 1.17
The following entity defines a synchronous Random Access Memory storing 2n
m-bit words (Fig. 1.23). A write input enables the writing operation. Functionally,
it is equivalent to a Register File made up of 2n m-bit registers whose clock enable
inputs are connected to write, plus an address decoder. Library declarations are
omitted.
1.3 Memory Blocks
19
The following component instantiation defines a synchronous RAM storing 16

4-bit words.
1.4 IO-Port Components

Once again, a previous comment similar to Comment 1.1 must be outlined: VHDL
models can be generated for simulating input and output amplifiers; nevertheless,
in order to implement an actual circuit, the corresponding vendors I/O component
should be used instead (Chap. 5).
20
(a)
(d)
b
a
(b)
(e)
c
b
a
(c)
a
b
en
en
Fig. 1.24 I/O ports
In order to generate VHDL models of I/O amplifiers, it is convenient to

understand the meaning of the STD_LOGIC type elements, that is
X: forcing unknown, 0 : forcing 0, 1 : forcing 1,
W: weak unknown, L : weak 0, H : weak 1,
Z: high impedance
(uninitialized and dont care states have no sense in the case of I/O
amplifiers).
An amplifier generates low-impedance (forcing) signals when enabled and high
impedance signals when disabled. Thus the possible outputs generated by an
amplifier are X, 0, 1 and Z. In the following example several types of input,
output and bidirectional amplifiers are defined.
Example 1.18
The following conditional assignment defines an input buffer (Fig. 1.24a):
An input buffer with a pull-up resistor can be defined as follows (Fig. 1.24b).
1.4 IO-Port Components
21
The definition of a tri-state output buffer (Fig. 1.24c) is the same as in example 1.8,
that is
An open-drain output can be defined as follows (Fig. 1.24d).
As an example of hierarchical description, a bidirectional I/O buffer can be defined

by instantiating an input buffer and a tri-state output buffer (Fig. 1.24e).
1.5 VHDL Models

The following complete VHDL models are available at the Authors web page
www.arithmetic-circuits.org:
logic_gates.vhd (Sects. 1.1.1, 1.1.2 and 1.1.3),
arithmetic_blocks.vhd (Sects. 1.1.4, 1.2.1 and 1.2.2),
sequential_components.vhd (Sects. 1.2.1, 1.2.2 and 1.2.3),
fnite_state_machines.vhd (Sect. 1.2.4),
memories.vhd (Sect. 1.3),
input_output.vhd (Sect. 1.4).
1.6 Exercises
1. Generate the VHDL model of a circuit that computes y = ax where a is a bit,
and x and y are n-bit vectors, so that y = (axn-1, axn-2,, ax0).
2. Generate several models of a 1-bit full subtractor (Boolean equations, table,
LUT instantiation).
3. Generate a generic model of an n-bit 8-to-1 multiplexer.
22
4. Generate a generic model of an n-input address decoder.

5. Design an n-bit magnitude comparator: given two n-bit naturals a and b, it
generates a 1-bit output gt equal to 1 if a C b and equal to 0 if a \ b.
6. Design a 60-state up counter with reset and count control inputs.
7. Design a finite state machine with two binary inputs x and y and a binary output
z defined as follows: if the input sequence is (x, t) = 00 01 11 10 00 01 11 then
z = 0, and if the input sequence is (x, y) = 00 10 11 01 00 10 11 then z = 1.
Reference
1. Hamblen JO, Hall TS, Furman MD (2008) Rapid prototyping of digital systems. Springer,
New York
Chapter 2
Architecture of Digital Circuits
This chapter describes the classical architecture of many digital circuits and presents,
by means of several examples, the conventional techniques that digital circuit
designers can use to translate an initial algorithmic description to an actual circuit. The
main topics are the decomposition of a circuit into Data Path and Control Unit and the
solution of two related problems, namely scheduling and resource assignment.
In fact, modern Electronic Design Automation tools have the capacity to directly
generate circuits from algorithmic descriptions, with performanceslatency, cost,
consumptioncomparable with those obtained using more traditional methods.
Those development tools are one of the main topics of Chap. 5. So, it is possible
that, in the future, the concepts and methods presented in this chapter will no longer
be of interest to circuit designers, allowing them to concentrate on algorithmic
innovative aspects rather than on scheduling and resource assignment optimization.
2.1 Introductory Example

As a first example, a naive method for computing the square root of a natural
x is considered. The following algorithm sequentially computes all the pairs
[r, s = (r ? 1)2] with r = 0, 1, 2, etc.:
Initially r = 0 and thus s = 1. Then, at each step, the pair [r ? 1, (r ? 2)2] is

computed in function of r and s = (r ? 1)2:
r 22 r 1 12 r 12 2 r 1 1 s 2 r 1 1:

23
24
2 Architecture of Digital Circuits
The same method can be used for computing the square root of x. For that, the loop
execution is controlled by the condition s B x.
Algorithm 2.1: Square root
The loop is executed as long as s B x, that is (r ? 1)2 B x. Thus, at the end of the
loop execution,
r 2 x\ r 12 :
Obviously, this is not a good algorithm as its computation time is proportional to
the square root itself, so that for great values of x (x % 2n) the number of steps is
of the order of 2n/2. Efficient algorithms are described in Chap. 10.
In order to implement Algorithm 2.1, the list of operations executed at each
clock cycle must be defined. In this case, each iteration step includes three operations: evaluation of the condition s B x, s ? 2(r ? 1) ? 1 and r ? 1. They can
be executed in parallel. On the other hand, the successive values of r and s must be
stored at each step. For that, two registers are used. Their initial values (0 and 1
respectively) are controlled by a common load signal, and their updating at the end
of each step by a common ce (clock enable) signal. The circuit is shown in Fig. 2.1.
To complete the circuit, a control unit in charge of generating the load and ce
signals must be added. It is a finite state machine with one input greater (detection
of the loop execution end) and two outputs, load and ce. A start input and a done
output are added in order to allow the communication with other circuits. The
finite state machine is shown in Fig. 2.2.
The circuit of Fig. 2.1 is made up of five blocks whose VHDL models are the
following:
computation of next_r:
computation of next_s:
(multiplying by 2 is the same as shifting one position to the right)
25
adder
r+1
x2
adder
next_s
next_ r
load
ce
register
initial value : 0
s+2(r+1)+1
load
ce
register
initial value : 1
s
x
comparator
greater
Fig. 2.1 Square root computation: data path
Fig. 2.2 Square root

computation: control unit
start=1/
nop
reset
start=0/
nop
start=0/
nop
greater=0/
update
start=1/
begin
1
greater=1/
nop
nop
begin
update
ce
0
0
1
load
0
1
0
done
1
0
0
26
register r:
register s:
end of loop detection:
The control unit is a Mealy finite state machine that can be modeled as follows:
next state computation:
output state computation:
27
The circuit of Fig. 2.1 includes three n-bit adders: a half adder for computing
next_r, a full adder for computing next_s and another full adder (actually a subtractor) for detecting the condition s [ x. Another option is to use one adder and to
decompose each iteration step into three clock cycles. For that, Algorithm 2.1 is
slightly modified.
Algorithm 2.2: Square root, version 2
A circuit able to execute the three operations, that is r ? 1, s ? 2r ? 1 and

evaluation of the condition s [ x must be defined. The condition s [ x is equivalent
to s C x ? 1 or s ? 2n - 1 - x C 2n. The binary representation of 2n - 1 - x is
obtained by complementing the bits of the binary representation of x. So, the
condition s [ x is equivalent to s ? not(x) C 2n. Thus, the three operations amount
to additions: r ? 1, s ? 2r ? 1 and s ? not(x). In the latter case, the output carry
defines the value of greater. The corresponding circuit is shown in Fig. 2.3. It is an
example of programmable computation resource: under the control of a 2-bit
command operation, it can execute the three previously defined operations.
The corresponding VHDL description is the following:
28
Fig. 2.3 Square root

computation: programmable
computation resource
x2
1,2
operand1
adder
operation (1..0)
operand 2
not(operation(1))
result(n)
result(n-1..0)
The complete circuit is shown in Fig. 2.4.

A control unit must be added. It is a finite state machine with one input greater
and five outputs load, ce_r, ce_s, ce_greater. As before, a start input and a done
output are added in order to allow the communication with other circuits. The
finite state machine is shown in Fig. 2.5.
The building blocks of the circuit of Fig. 2.4 (apart from the programmable
resource) are the following:
register r:
register s:
29
x
ce_greater
FF
operation
programmable
resource
greater
load
ce_r
register
initial value: 0
load
ce_s
register
initial value: 1
Fig. 2.4 Square root computation, second version: data path
start=0/
nop
start=1/
nop
reset
start=0/
nop
start=1/
begin
greater=1/
nop
commands
nop
begin
update_r
update_s
update_greater
greater=0/
update_r
-/
update_g
3
-/
update_s
ce_r, ce_s, ce_ greater, load, operation , done

0
0
1
0
0
0
0
0
1
0
0
0
0
0
1
0
1
0
0
0
Fig. 2.5 Square root computation, second version: control unit
0
0
0
1
2
1
0
0
0
0
30
flip-flop greater
The control unit is a Mealy finite state machine whose VHDL model is the
following:
next state computation:
output state computation:
31
Complete VHDL models (square_root.vhd) of both circuits (Figs. 2.1, 2.4) are
available at the Authors web page.
2.2 Data Path and Control Unit

The general structure of a digital circuit is shown in Fig. 2.6. It consists of a data
path and a control unit. The data path (leftmost part of Fig. 2.6) includes computation resources executing the algorithm operations, registers storing the algorithm variables, and programmable connections (for example multiplexers, not
represented in Fig. 2.6) between resource outputs and register inputs, and between
register outputs and resource inputs. The control unit (rightmost part of Fig. 2.6) is
a finite state machine. It controls the sequence of data path operations by means of
a set of control signals (commands) such as clock enables of registers, programming of computation resources and multiplexers, and so on. It receives from the
data path some feedback information (conditions) corresponding to the algorithm
control statements (loop, if, case).
In fact, the data path could also be considered as being a finite state machine. Its
internal states are all the possible register contents, the next-state computation is
performed by the computation resources, and the output states are all the possible
values of conditions. Nevertheless, the number of internal states is enormous and
there is generally no sense in using a finite state machine model for the data path.
However, it is interesting to observe that the data path of Fig. 2.6 is a Moore
32

data_out
data_in
start done
computation
resources
(t3)
command
generation
(t2)
commands
internal state
next_data
registers
next-state
computation
(t1)
(clk)
data
(t4)
next state
(clk)
conditions
clk
internal_state
data
conditions
commands
next_state
next_data
t4
t2
t3
t1
Fig. 2.6 Structure of a digital circuit: data path and control unit
machine (the output state only depends on the internal state) while the control unit
could be a Moore or a Mealy machine. An important point is that, when two finite
state machines are interconnected, one of them must be a Moore machine in order
to avoid combinational loops.
According to the chronograms of Fig. 2.6, there are two critical paths: from the
data registers to the internal state register, and from the data registers to the data
registers. The corresponding delays are
Tdatastate t4 t1
2:1
2.2 Data Path and Control Unit
33
and
Tdatadata t4 t2 t3 ;
2:2
where t1 is the computation time of the next internal state, t2 the computation time
of the commands, t3 the maximum delay of the computation resources and t4 the
computation time of the conditions (the set up and hold times of the registers have
not been taken into account).
The clock period must satisfy
Tclk [ maxft4 t1 ; t4 t2 t3 g:
2:3
If the control unit were a Moore machine, there would be no direct path from the
data registers to the data registers, so that (2.2) and (2.3) should be replaced by
Tstatedata t2 t3
2:4
Tclk [ maxft4 t1 ; t2 t3 g:
2:5
and
In fact, it is always possible to use a Moore machine for the control unit. Generally
it has more internal states than an equivalent Mealy machine and the algorithm
execution needs more clock cycles. If the values of t1 to t4 do not substantially
vary, the conclusion could be that the Moore approach needs more, but shorter,
clock cycles. Many designers also consider that Moore machines are safer than
Mealy machines.
In order to increase the maximum frequency, an interesting option is to insert a
command register at the output of the command generation block. Then relation
(2.2) is substituted by
Tdatacommands t4 t2
and Tcommandsdata t3 ;
2:6
so that
Tclk [ maxft4 t1 ; t4 t2 ; t3 g:
2:7
With this type of registered Mealy machine, the commands are available one cycle
later than with a non-registered machine, so that additional cycles must be
sometimes inserted in order that the data path and its control unit remain
synchronized.
To summarize, the implementation of an algorithm is based upon a decomposition of the circuit into a data path and a control unit. The data path is in charge
of the algorithm operations and can be roughly defined in the following way:
associate registers to the algorithm variables, implement resources able to execute
the algorithm operations, and insert programmable connections (multiplexers)
between the register outputs (the operands) and the resource inputs, and between
the resource outputs (the results) and the register inputs. The control unit is a finite
state machine whose internal states roughly correspond to the algorithm steps, the
34
input states are conditions (flags) generated by the data path, and the output states
are commands transmitted to the data path.
In fact, the definition of a data path poses a series of optimization problems,
some of them being dealt with in the next sections, for example: scheduling of the
operations, assignment of computation resources to operations, and assignment of
registers to variables. It is also important to notice that minor algorithm modifications sometimes yield major circuit optimizations.
2.3 Operation Scheduling

Operation scheduling consists in defining which particular operations are in the
process of execution during every clock cycle. For that purpose, an important
concept is that of precedence relation. It defines which of the operations must be
completed before starting a new one: if some result r of an operation A is an initial
operand of some operation B, the computation of r must be completed before the
execution of B starts. So, the execution of A must be scheduled before the execution of B.
2.3.1 Introductory Example

A carry-save adder or 3-to-2 counter (Sect. 7.7) is a circuit with 3 inputs and 2
outputs. The inputs xi and the outputs yj are naturals. Its behavior is defined by the
following relation:
x1 x2 x3 y1 y2 :
2:8
It is made up of 1-bit full adders working in parallel. An example where x1, x2 and
x3 are 4-bit numbers, and y1 and y2 are 5-bit numbers, is shown in Fig. 2.7.
The delay of a carry-save adder is equal to the delay TFA of a 1-bit full adder,
independently of the number of bits of the operands. Let CSA be the function
associated to (2.8), that is
y1 ; y2 CSAx1 ; x2 ; x3 :
2:9
Using carry-save adders as computation resources, a 7-to-3 counter can be

implemented. It allows expressing the sum of seven naturals under the form of the
sum of three naturals, that is
x1 x2 x3 x4 x5 x6 x7 y1 y2 y3 :
In order to compute y1, y2 and y3, the following operations are executed (op1 to op4
are labels):
35
x13 x23
y14 y24
x33 x12 x22
x32 x11 x21
x31 x10 x20
FA
FA
FA
FA
y13 y23
y12 y22
y11 y21
y10 y20
x30
Fig. 2.7 Carry-save adder
op1 : a1 ; a2 CSAx1 ; x2 ; x3 ;
op2 : b1 ; b2 CSAx4 ; x5 ; x6 ;
op3 : c1 ; c2 CSAa2 ; b2 ; x7 ;
op4 : d1 ; d2 CSAa1 ; b1 ; c1 :
2:10
According to (2.10) and the definition of CSA

a1 a2 x1 x2 x3 ;
b1 b2 x4 x5 x6 ;
c1 c2 a2 b2 x7 ;
d1 d2 a1 b1 c 1 ;
so that
c1 c2 d1 d2 a2 b2 x7 a1 b1 c1
x1 x2 x3 x4 x5 x6 x7 c1 :
Thus
c2 d1 d2 a2 b2 x7 a1 b1 c1 x1 x2 x3 x4 x5 x6 x7
and y1, y2 and y3 can be defined as follows:
y1 d1 ; y2 d2 ; y3 c2 :
The corresponding precedence relation is defined by the graph of Fig. 2.8,
according to which op1 and op2 must be executed before op3, and op3 before op4.
Thus, the minimum computation time is equal to 3TFA.
For implementing (2.10) the following options could be considered:
1. A combinational circuit, made up of four carry-save adders, whose structure is
the same as that of the graph of Fig. 2.8. Its computation time is equal to 3TFA
and its cost to 4CCSA, being CCSA the cost of a carry-save adder. This is
probably a bad solution because the cost is high (4 carry-save adders) and the
delay is long (3 full-adders) so that the minimum clock cycle of a synchronous
circuit including this 7-to-3 counter should be greater than 3TFA.
2. A data path including two carry-save adders and several registers (Sect. 2.5).
The computation is executed in three cycles:
36
Fig. 2.8 Precedence relation

of a 7-to-3 counter
x1,x2,x3
x4,x5,x6
op1
op 2
a1
a2 b 1
x7
b2
op3
c1
op4
c2
y3
d1,d2
y1,y2
The computation time is equal to 3Tclk, where Tclk [ TFA, and the cost equal to
2CCSA, plus the cost of the additional registers, controllable connections and
control unit.
3. A data path including one carry-save adder and several registers. The computation is executed in four cycles:
The computation time is equal to 4Tclk, where Tclk [ TFA, and the cost equal to
CCSA, plus the cost of the additional registers, controllable connections and
control unit.
In conclusion, there are several implementations, with different costs and delays,
corresponding to the set of operations in (2.10). In order to get an optimized
circuit, according to some predefined criteria, the space for possible implementations must be explored. For that, optimization methods must be used.
37
2.3.2 Precedence Graph

Consider a computation scheme, that is to say, an algorithm without branches and
loops. Formally it can be defined by a set of operations
opJ : xi ; xk ; . . . f xl ; xm ; . . .;
2:11
where xi, xk, xl, xm, are variables of the algorithm and f one of the algorithm
operation types (computation primitives). Then, the precedence graph (or data flow
graph) is defined as follows:
associate a vertex to each operation opJ,
draw an arc between vertices opJ and opM if one of the results generated by opJ
is used by opM.
An example was given in Sect. 2.3.1 (operations (2.10) and Fig. 2.8).
Assume that the computation times of all operations are known. Let tJM be the
computation time, expressed in number of clock cycles, of the result(s) generated
by opJ and used by opM. Then, a schedule of the algorithm is an application Sch
from the set of vertices to the set of naturals that defines the number Sch(opJ) of
the cycle at the beginning of which the computation of opJ starts. A necessary
condition is that
SchopM SchopJ tJM
2:12
if there is an arc from opJ to opM.

As an example, if the clock period is greater than the delay of a full adder, then,
in the computation scheme (2.10), all the delays are equal to 1 and two admissible
schedules are
Schop1 1; Schop2 1; Schop3 2; Schop4 3;
2:13
Schop1 1; Schop2 2; Schop3 3; Schop4 4:
2:14
They correspond to the options 2 and 3 of Sect. 2.3.1.

The definition of an admissible schedule is an easy task. As an example, the
following algorithm defines an ASAP (as soon as possible) schedule:
initial step: Sch(opJ) = 1 for all initial (without antecessor) vertices opJ;
step number n ? 1: choose an unscheduled vertex opM whose total amount of
antecessors, say opP, opQ, have already been scheduled, and define
Sch(opM) = maximum{Sch(opP) ? tPM, Sch(opQ) ? tQM,}.
Applied to (2.10) the ASAP algorithm gives (2.13). The corresponding data flow
graph is shown in Fig. 2.9a.
An ALAP (as late as possible) schedule can also be defined. For that, assume
that the latest admissible starting cycle for all the final vertices (without successor)
has been previously specified:
38
(a) x1,x2,x3
(b)
x4,x5,x6
op1
op2 1
a1
a2 b 1
x1,x2,x3
2
x7
b2
op1
a1
op2 2
a2 b1
op3 2
c1
op4
y3
d1,d2
y1,y2
c2
(c)
x4,x5,x6
x1,x2,x3
1
x7
b2
x4,x5,x6
op1
a1
op2 2
a 2 b1
op3 3
c1
op4
d1,d2
c2
y3
x7
b2
op3 3
c1
op4
c2
y3
d1,d2
y1,y2
Fig. 2.9 7-to-3 counter: a ASAP schedule. b ALAP schedule. c Admissible schedule
initial step: Sch(opM) = latest admissible starting cycle of opM for all final vertices opM;
step number n ? 1: choose an unscheduled vertex opJ whose all successors, say
opP, opQ, have already been scheduled, and define Sch(opJ) = minimum{Sch(opP) - tJP, Sch(opQ) - tJQ,}.
Applied to (2.10), with Sch(op4) = 4, the ALAP algorithm generates the data
flow graph of Fig. 2.9b.
Let ASAP_Sch and ALAP_Sch be ASAP and ALAP schedules, respectively.
Obviously, if opM is a final operation, the previously specified value ALAP_Sch(opM) must be greater than or equal to ASAP_Sch(opM). More generally,
assuming that the latest admissible starting cycle for all the final operations has been
previously specified, for any admissible schedule Sch the following relation holds:
ASAP SchopJ SchopJ ALAP SchopJ ; 8opJ :
2:15
Along with (2.12), relation (2.15) defines the admissible schedules.

An example of admissible schedule is defined by (2.14), to which corresponds
the data flow graph of Fig. 2.9c.
A second, more realistic, example is now presented. It corresponds to part of an
Elliptic Curve Cryptography algorithm.
Example 2.1
Given a point P = (xP, yP) of an elliptic curve and a natural k, the scalar product
kP = P ? P+ ? P can be defined [1, 2]. In the case of the curve
y2 ? xy = x3 ? ax ? 1 over the binary field, the following formal algorithm [3]
computes kP. The initial data are the scalar k = km - 1 km - 2k0 and the xcoordinate xP of P. All the algorithm variables are elements of the Galois field
GF(2m), that is, polynomials of degree m over the binary field GF(2) (Chap. 13).
39
Algorithm 2.3: Scalar product, projective coordinates
In fact, the preceding algorithm computes the value of four variables xA, zA, xB and
zB in function of k and xP. A final, not included, step would be to compute the
coordinates of kP in function of the coordinates of P (xP and yP) and of the final
values of xA, zA, xB and zB.
Consider one step of the main iteration of Algorithm 2.3, and assume that
km - i = 0. The following computation scheme computes the new values of xA, zA,
xB and zB in function of their initial values and of xP. The available
computation primitives are the addition, multiplication and squaring in GF(2m)
(Chap. 13).
40
Fig. 2.10 Example 2.1:

precedence graph
xA,zB
xB,zA
xA,zA
xA,zA
b
+
c
squ
xP
squ
squ
k
squ
e
+
The updated values of xA, zA, xB and zB are xA = l, zA = i, xB = g, zB = d. The

corresponding data flow graph is shown in Fig. 2.10. The operation type corresponding to every vertex is indicated (instead of the operation label). If km - i = 1
the computation scheme is the same but for the interchange of indexes A and B.
Addition and squaring in GF(2m) are relatively simple one-cycle operations,
while multiplication is a much more complex operation whose maximum computation time is tm 1. In what follows it is assumed that tm = 300 cycles.
An ASAP schedule is shown in Fig. 2.11. The computation of g starts at the
beginning of cycle 603 so that all the final results are available at the beginning of
cycle 604. The corresponding circuit must include three multipliers as the
computations of a, b and h start at the same time.
The computation scheme includes 5 multiplications. Thus, in order to execute
the algorithm with only one multiplier, the minimum computation time is 1,500.
More precisely, one of the multiplications e, f or h cannot start before cycle 1,201,
so that the next operation (g or i) cannot start before cycle 1,501. An ALAP
schedule, assuming that the computations of g and i start at the beginning of cycle
1,501, is shown in Fig. 2.12.
2.3.3 Optimization Problems

Assuming that the latest admissible starting cycle for all the final operations has
been previously specified then any schedule, such that (2.12) and (2.15) hold true,
can be chosen. This poses optimization problems. For example:
1. Assuming that the maximum computation time has been previously specified,
look for a schedule that minimizes the number of computation resources of
each type.
41

ASAP schedule
xA,zB
1
xB,zA
1
301
squ
squ
k
3
303
squ
xP
h
301
301
xA,zA
1
c
302
xA,zA
squ
e
603

ALAP schedule with
Sch(g) = 1501
xA,zB
899
xB,zA
899
1199
1200
b
+
1201
xA,zA
1499
h
1501 squ
i
j
1500
squ
squ
xP
1201
1201
xA,zA
1501
squ
e
1501
2. Assuming that the number of available computation resources of each type has
been previously specified, minimize the computation time.
An important concept is the computation width w(f) with respect to the computation primitive (operation type) f. First define the activity intervals of f. Assume
that f is the primitive corresponding to the operation opJ, that is
opJ : xi ; xk ; . . . f xl ; xm ; . . .:
Then
SchopJ ; SchopJ maximumftJM g
42
is an activity interval of f. This means that a resource of type f must be available

from the beginning of cycle Sch(opJ) to the end of cycle Sch(opJ) ? tJM for all
M such that there is an arc from opJ to opM. An incompatibility relation over the set of
activity intervals of f can be defined: two intervals are incompatible if they overlap. If
two intervals overlap, it is obvious that the corresponding operations cannot be
executed by the same computation resource. Thus, a particular resource of type f must
be associated to each activity interval of f in such a way that if two intervals overlap,
then two distinct resources of the same type must be used. The minimum number of
computation resources of type f is the computation width w(f).
The following graphical method can be used for computing w(f).
Associate a vertex to every activity interval.
Draw an edge between two vertices if the corresponding intervals overlap.
Color the vertices in such a way that two vertices connected by an edge have
different colors (a classical problem of graph theory).
Then, w(f) is the number of different colors, and every color defines a particular
resource assigned to all edges (activity intervals) with this color.
Example 2.2
Consider the scheduled precedence graph of Fig. 2.11. The activity intervals of the
multiplication are
a : 1; 300; b : 1; 300; h : 1; 300; f : 301; 600; e : 303; 602:
The corresponding incompatibility graph is shown in Fig. 2.13a. It can be colored
with three colors (c1, c2 and c3 in Fig. 2.13a). Thus, the computation width with
respect to the multiplication is equal to 3.
If the scheduled precedence graph of Fig. 2.12 is considered, then the activity
intervals of the multiplication are
a : 899; 1198; b : 899; 1198; h : 1201; 1500; f : 1201; 1500; e : 1201; 1500:
The corresponding incompatibility graph is shown in Fig. 2.13b. It can be colored
with three colors. Thus, the computation width with respect to the multiplication is
still equal to 3.
Nevertheless, other schedules can be defined. According to (2.15) and
Figs. 2.11 and 2.12, the time intervals during which the five multiplications can
start are the following:
a : 1; 899; b : 1; 899; h : 1; 1201; f : 301; 1201; e : 303; 1201:
As an example, consider the admissible schedule of Fig. 2.14. The activity
intervals of the multiplication operation are
a : 1; 300 ; b : 301; 600; h : 601; 900; f : 901; 1200; e : 1201; 1500:
43
(a)
(b)
c1
c1
a
f
c2
c1
c1
c2
c3
a
f
b
c2
c1
c3
Fig. 2.13 ColoringComputation width: graph coloring

admissible schedule using
only one multiplier
xA,zB
1
301
a
601
901
squ
xP
1201
c
602
xB,zA
xA,zA
601
xA,zA
1
h
901
squ
j
2
squ
i
3
squ
e
1501
They do not overlap hence the incompatibility graph does not include any edge
and can be colored with one color. The computation width with respect to the
multiplication is equal to 1.
Thus, the two optimization problems mentioned above can be expressed in
terms of computation widths:
1. Assuming that the maximum computation time has been previously specified,
look for a schedule that minimizes some cost function
C c1 wf 1 c2 wf 2 cm wf m
2:16
where f1, f2,, fm are the computation primitives and c1, c2,, cm their corresponding costs.
2. Assuming that the maximum computation width w(f) with respect to every
computation primitive f has been previously specified, look for a schedule that
minimizes the computation time.
Both are classical problems of scheduling theory. They can be expressed in
terms of integer linear programming problems whose variables are xIt for all
44
schedule corresponding to the
first optimization problem

xA,zB
1
xB,zA
299
a
599
b
+
c
600
squ
xP
601
601
xA,zA
601
xA,zA
1
901 squ
i
j
2
squ
k
3
squ
e
901
operation indices I and all possible cycle numbers t: xIt = 1 if Sch(eI) = t, 0

otherwise. Nevertheless, except for small computation schemesgenerally tractable by handthe so obtained linear programs are intractable. Modern Electronic
Design Automation tools execute several types of heuristic algorithms applied to
different optimization problems (not only to schedule optimization). Some of the
more common heuristic strategies are list scheduling, simulated annealing and
genetic algorithms.
Example 2.3
The list scheduling algorithm, applied to the graph of Fig. 2.10, with tm = 300 and
assuming that the latest admissible starting cycle for all the final operations is
cycle number 901 (first optimization problem), would generate the schedule of
Fig. 2.15. The list scheduling algorithm, applied to the same graph of Fig. 2.10,
with tm = 300 and assuming that the computation width is equal to 1 for all
operations (second optimization problem), would generate the schedule of
Fig. 2.14.
2.4 Resource Assignment

Once the operation schedule has been defined, several decisions must be taken.
The number w(f) of resources of type f is known, but it remains to decide which
particular computation resource executes each operation. Furthermore the definition of multifunctional programmable resources could also be considered.
As regards the storing resources, a simple solution is to assign a particular
register to every variable. Nevertheless, in some cases the same register can be
used for storing different variables.
2.4 Resource Assignment
45
A key concept for assigning registers to variables is the lifetime [tI, tJ] of every
variable: tI is the number of the cycle during which its value is generated, and tJ is
the number of the last cycle during which its value is used.
Example 2.4
Consider the computation scheme of Example 2.1 and the schedule of Fig. 2.14.
The computation width is equal to 1 for all primitives (multiplication, addition and
squaring). The computation is executed as follows:
In order to compute the variable lifetimes, it is assumed that the multiplier reads
the values of the operands during some initial cycle, say number I, and generates
the result during cycle number I ? tm - 1 (or sooner), so that this result can be
stored at the end of cycle number I ? tm - 1 and is available for any operation
beginning at cycle number I ? tm (or later). As regards the variables xA, zA, xB and
zB, in charge of passing values from one iteration step to the next (Algorithm 2.3),
their initial values must be available from the first cycle up to the last cycle during
which those values are used. At the end of the computation scheme execution they
must be updated with their new values. The lifetime intervals are given in
Table 2.1.
The definition of a minimum number of registers can be expressed as a graph
coloring problem. For that, associate a vertex to every variable and draw an edge
between two variables if their lifetime intervals are incompatible, which means
that they have more than one common cycle. As an example, the lifetime intervals
of j and k are compatible, while the lifetime intervals of b and d are not.
The following groups of variables have compatible lifetime intervals:
46
Table 2.1 Lifetime intervals
a
j
k
l
b
h
c
d
f
i
e
g
xA
zA
xB
zB
[300, 901]
[1, 2]
[2, 3]
[3, final]
[600, 901]
[900, 901]
[601, 602]
[602, final]
[1200, 1501]
[901, final]
[1500, 1501]
[1501, final]
[initial, 601]
[initial, 601]
[initial, 301]
[initial, 1]
zB initial ! 1; j1 ! 2; k2 ! 3; l3 ! final;
xB initial ! 301; b600 ! 901; f 1200 ! 1501; g1501 ! final;
zA initial ! 601; c601 ! 602; d602 ! final;
xA initial ! 601; h900 ! 901; e1500 ! 1501;
a300 ! 901; i901 ! final:
Thus, the computing scheme can be executed with five registers, namely xA, zA, xB,
zB and R:
2.5 Final Example
47
2.5 Final Example

Each iteration step of Algorithm 2.3 consists of executing a computation scheme,
either the preceding one when km - i = 0, or a similar one when km - i = 1. Thus,
Algorithm 2.3 is equivalent to the following algorithm 2.4 in which sentences
separated by commas are executed in parallel.
Algorithm 2.4: Scalar product, projective coordinates (scheduled version)
The data processed by Algorithm 2.4 are m-bit vectors (polynomials of degree
m over the binary field GF(2)) and the computation resources are field multiplication, addition and squaring. Field addition amounts to bit-by-bit modulo 2
additions (XOR functions). On the other hand, VHDL models of computation
resources executing field squaring and multiplication are available at the Authors
web page, namely classic_squarer.vhd and interleaved_mult.vhd (Chap. 13). The
classic_squarer component is a combinational circuit. The interleaved_mult
component reads and internally stores the input operands during the first cycle
after detecting a positive edge on start_mult and raises an output flag mult_done
when the multiplication result is available.
48
The operations executed by the multiplier are

x A zB ; x B zA ; x A zA ; R x B ; x P zA ; x B zB ; R x A ; x P zB :
An incompatibility relation can be defined over the set of involved variables: two
variables are incompatible if they are operands of a same operation. As an
example, xA and zB are incompatible, as xAzB is one of the operations. The corresponding graph can be colored with two colors corresponding to the sets
fxA ; xB ; xP g and fzA ; zB ; Rg:
The first set of variables can be assigned to the leftmost multiplier input and the
other to the rightmost input.
The operations executed by the adder are
xA zA ; R xB ; xA xB ; xB zB ; R xA ; xB xA :
The incompatibility graph can be colored with three colors corresponding to the
sets
fxA ; zB g; fxB ; zA g and fRg:
The first one is assigned to the leftmost adder input, the second to the rightmost
input, and R to both inputs.
Finally, the operations realized by the squaring primitive are
z2B ; z2A ; x2A ; x2B :
The part of the data path corresponding to the computation resources and the
multiplexers that select their input data is shown in Fig. 2.16. The corresponding
VHDL model can easily be generated. As an example, the multiplier, with its input
multiplexers, can be described as follows.
Consider now the storing resources. Assuming that xP and k remain available
during the whole algorithm execution, there remain five variables that must be
internally stored: xA, xB, zA, zB and R. The origin of the data stored in every register
must be defined. For example, the operations that update xA are
2.5 Final Example

x A xB x P
49
zA zB R
sel_p1
xA zB R
sel_p2
xB zA R
sel_a1
sel_a2
start_mult
XOR gates
interleaved_mult
mult_done
product
z A z B xA x B
adder_out
sel_sq
classic_squarer
square
Fig. 2.16 Example 2.4: computation resources
So, the updated value can be 1 (initial value), product, adder_out or zB. A similar
analysis must be done for the other registers. Finally, the part of the data path
corresponding to the registers and the multiplexers that select their input data is
shown in Fig. 2.17. The corresponding VHDL model is easy to generate. As an
example, the xA register, with its input multiplexers, can be described as follows.
50

adder_ out
adder_ out
zB
product
zA
product
sel_xA
load
en_xA
register
initially: 1
register
initially:xP
square
sel_xB
load
en_xB
register
initially: 0
xB
xA
R zB
adder_ out
sel_zA
load
en_zA
zA
square
zA R
adder_ out
register
initially: 1
zB
product square
sel_zB
load
en_zB
register
sel_R
en_R
Fig. 2.17 Example 2.5: data registers
A complete model of the data path scalar_product_data_path.vhd is available at

the Authors web page.
The complete circuit is defined by the following entity.
It is made up of
the data path;
a shift register allowing sequential reading of the values of km
a counter for controlling the loop execution;
- i;
2.5 Final Example
51
a finite state machine in charge of generating all the control signals, that is
start_mult, load, shift, en_xA, en_xB, en_zA, en_zB, en_R, sel_p1, sel_p2,
sel_a1, sel_a2, sel_sq, sel_xA, sel_xB, sel_zA, sel_zB and sel_R. In particular,
the control of the multiplier operations is performed as follows: the control unit
generates a positive edge on the start_mult signal, along with the values of
sel_p1 and sel_p2 that select the input operands; then, it enters a wait loop until
the mult_done flag is raised (instead of waiting for a constant time, namely 300
cycles, as was done for scheduling purpose); during the wait loop the start_mult
is lowered while the sel_p1 and sel_p2 values are maintained; finally, itgenerates
the signals for updating the register that stores the result. As an example, assume that
the execution of the fourth instruction of the main loop, that is xB:= xBzA, starts at
state 6 and uses identifiers start4, wait4 and end4 for representing the corresponding
commands. The corresponding part of the next-state function is
and the corresponding part of the output function is
a command decoder (Chap. 4). Command identifiers have been used in the
definition of the finite state machine output function, so that a command decoder
must be used to generate the actual control signal values in function of the
identifiers. For example, the command start4 initializes the execution of
xB:= xBzA and is decoded as follows:
In the case of operations such as the first of the main loop, that is R:= xAzB,
zB:= xA ? zA, the 1-cycle operation zB:= xA ? zA is executed in parallel with the
final cycle of R:= xAzB and not in parallel with the initial cycle. This makes the
algorithm execution a few cycles (3) longer, but this is not significant as tm is
generally much greater than 3. Thus, the control signal values corresponding to
the identifier end1 are:
52
The control unit also detects the start signal and generates the done flag. A complete model scalar_product.vhd is available at the Authors web page.
Comment 2.1
The interleaved_mult component is also made up of a data path and a control unit,
while the classic_squarer component is a combinational circuit. An alternative
solution is the definition of a data path able to execute all the operations, including
those corresponding to the interleaved_mult and classic_squarer components. The
so-obtained circuit could be more efficient than the proposed one as some computation resources could be shared between the three algorithms (field multiplication, squaring and scalar product). Nevertheless, the hierarchical approach
consisting of using pre-existing components is probably safer and allows a
reduction in the development times.
Instead of explicitly disassembling the circuit into a data path and a control unit,
another option is to describe the operations that must be executed at each cycle,
and to let the synthesis tool define all the details of the final circuit. A complete
model scalar_product_DF2.vhd is available at the Authors web page.
Comment 2.2
Algorithm 2.4 does not compute the scalar product kP. A final step is missing:
The design of a circuit that executes this final step is left as an exercise.
2.6 Exercises
53
2.6 Exercises
1. Generate several VHDL models of a 7-to-3 counter. For that purpose use the
three options proposed in Sect. 2.3.1.
2. Generate the VHDL model of a circuit executing the final step of the scalar
product algorithm (Comment 2.2). For that purpose, the following entity,
available at the Authors web page, is used:
It computes z = gh-1 over GF(2m). Several architectures can be considered.

3. Design a circuit to compute the greatest common divisor of two natural
numbers, based on the following simplified Euclidean algorithm.
4. Design a circuit for computing the greatest common divisor of two natural
numbers, based on the following Euclidean algorithm.
5. The distance d between two points (x1, y1) and (x2, y2) of the (x, y)-plane is
equal to d = ((x1 - x2)2 ? (y1 - y2)2)0.5. Design a circuit that computes
d with only one subtractor and one multiplier.
6. Design a circuit that, within a three-dimensional space, computes the distance
between two points (x1, y1, z1) and (x2, y2, z2).
7. Given a point (x, y, z) of the three-dimensional space, design a circuit that
computes the following transformation.
2 3 2
3 2 3
xt
a11 a21 a31
x
4 yt 5 4 a21 a22 a32 5 4 y 5
z
zt
a31 a32 a11
8. Design a circuit for computing z = ex using the formula
54
ex 1
x x2 x3

1! 2! 3!
9. Design a circuit for computing xn, where n is a natural, using the following
relations: x0 = 1; if n is even then xn = (xn/2)2, and if n is odd then
xn = x(x(n-1)/2)2.
10. Algorithm 2.4 (scalar product) can be implemented using more than one
interleaved_multiplier. How many multipliers can operate in parallel? Define
the corresponding schedule.
References
1. Hankerson D, Menezes A, Vanstone S (2004) Guide to elliptic curve cryptography. Springer,
New York
2. Deschamps JP, Imaa JL, Sutter G (2009) Hardware implementation of finite-field arithmetic.
McGraw-Hill, New York
3. Lpez J, Dahab R (1999) Improved algorithm for elliptic curve arithmetic in GF(2n). Lect
Notes Comput Sci 1556:201212
Chapter 3
Special Topics of Data Path Synthesis
Several important implementation techniques are presented in this chapter. The

first one is pipelining, a very commonly used method in systems that process great
volumes of data. Self-timing is the topic of the second section. To some extent it
can be considered as an extension of the pipelining concept and is especially
attractive in the case of very big circuits. The third section is an introduction to a
circuit level, or even algorithm level, transformation known as loop unrolling.
It permits the exploration of different costperformance tradeoffs, from combinational iterative circuits to completely sequential circuits. Finally, the last section
tackles the problem of reducing the number of connection resources.
3.1 Pipeline
A very useful implementation technique, especially for signal processing circuits,
is pipelining [1, 2]. It consists of inserting additional registers so that the maximum
clock frequency and input data throughput are increased. Furthermore, in the case
of FPGA implementations, the insertion of pipeline registers has a positive effect
on the power consumption.
3.1.1 Introductory Example

Consider the introductory example of Sect. 2.3.1. The set of Eq. (2.10) can be
implemented by a combinational circuit (option 1. of Sect. 2.3.1) made up of four
carry-save adders, with a computation time equal to 3TFA. That means that the
minimum clock period of a synchronous circuit including this 7-to-3 counter
should be greater than 3TFA, and that the introduction interval between successive
data inputs should also be greater than 3TFA. The corresponding circuit is shown

55
56
3 Special Topics of Data Path Synthesis
(a)
(b)
x1 x2 x3
x4 x 5 x6
CSA
CSA
a1
a2 b1
b2
x7
x 1 x2 x 3
x 4 x 5 x6
CSA
CSA
a1
a2 b1
x7
b2
clk
CSA
c1
CSA
c2
c1
c2
clk
CSA
d1
y1
CSA
d2
y2
d1
y3
y1
d2
y2
y3
Fig. 3.1 a Combinational circuit. b Pipelined circuit
in Fig. 3.1a. As previously commented, this is probably a bad circuit because its
cost is high and its maximum clock frequency is low.
Consider now the circuit of Fig. 3.1b in which registers have been inserted in
such a way that operations scheduled in successive cycles, according to the ASAP
schedule of Fig. 2.9a, are separated by a register. The circuit still includes four
carry-save adders, but the minimum clock period of a synchronous circuit
including this counter must be greater than TFA, plus the set-up and hold times of
the registers, instead of 3TFA. Furthermore, the minimum data introduction
interval is now equal to Tclk: as soon as a1, a2, b1 and b2 have been computed, their
values are stored within the corresponding output register, and a new computation,
with other input data, can start; at the same time, new computations of c1 and c2,
and of d1 and d2 can also start. Thus, at time t, three operations are executed in
parallel:
To summarize, assuming that the set-up and hold times are negligible,
Tclk [ TFA ; latency 3 Tclk ; r Tclk ;
3.1 Pipeline
57
Fig. 3.2 Two-stage twocycle implementation
x1x2x3
x4x5x6
x7
stage 1
CSA
a1a2
x7
b1b2
conexiones
a2b2x7
a1b1c1
stage 2
CSA
c1
y1
y2
y3
where latency is the total computation time and r is the minimum data introduction
interval.
Another implementation, based on the admissible schedule of Fig. 2.9c is show
in Fig. 3.2. In this case the circuit is made up of two stages separated by a pipeline
register. Within every stage the operations are executed in two cycles. During the
first cycle the following operations are executed
and during the second cycle, the following ones are executed
The circuit of Fig. 3.2 includes two carry-save adders instead of four, and its
timing constraints are the following:
58
Tclk [ TFA Tmultiplexor ; latency 4 Tclk ; r 2 Tclk :

To summarize, the main parameters of a pipelined circuit are the following.
Latency (also called delay, response time): total delay between the introduction
of a new set of input data and the generation of the corresponding output results.
It is equal to nTclk where n is the number of pipeline stages and Tclk the clock
period.
Pipeline rate (also called pipeline period): data introduction interval.
Throughput (also called speed, bandwidth, production): number of input data
processed per time unit. For great numbers of processed data, it is the inverse of
the pipeline rate r.
Assuming that the combinational delay of stage number i is equal to ti, and that
the register set-up and hold times are negligible, the minimum clock period is the
maximum of all tis. In the first example (Fig. 3.1a) t1 = t2 = t3 = TFA, while in
the second example (Fig. 3.2) t1 = t2 = TFA ? Tmultiplexor.
A very common situation is that of the first example (Fig. 3.1). An initial
combinational circuit has a delay equal to C, so that it is able to process 1/C input
data per time unit. It is divided up into n pipeline stages, all of them with the same
delay C/n (balanced stages). Then
Tclk C=n; r 1=Tclk n=C; latency n Tclk C; T m
n Tclk m 1 Tclk ;
where T(m) is the time necessary to process m input data. Thus, the average number
of input data processed per time unit is equal to m/T(m) = m/(n ? m - 1)
Tclk % mn/(n ? m - 1)C. For great values of m, the number of input data
processed per time unit is equal to n/C. Thus, with respect to the initial combinational
circuit, the throughput has been multiplied by n.
The actual speedup factor is smaller if the connection and register delays are
taken into account. Assume that those additional delays are equal to d time units.
Then, the minimum clock period is equal to Tclk = C/n ? d, r = 1/Tclk = n/
(C ? nd), latency = nTclk = C ? nd, T(m) = (n ? m-1)Tclk = (n ? m-1)(C/
n ? d) % m(C/n ? d), m/T(m) % 1/(C/n ? d) = n/(C ? nd). Hence, the
throughput increase factor is equal to nC/(C ? nd) = n/(1 ? a) where a = nd/C.
3.1.2 Segmentation
Given a computation scheme and its precedence graph G, a segmentation of G is
an ordered partition {S1, S2,, Sk} of G. The segmentation is admissible if it
respects the precedence relationship. This means that if there is an arc from opJ [
Si to opM then either opM belongs to the same segment Si or it belongs to a different
segment Sj with j [ i. Two examples are shown in Fig. 3.3 in which the segments
are separated by dotted lines.
3.1 Pipeline
59
(a)
(b)
op1
op1
S1
op2
op2
S1
op3
op5
S2
op3
op5
S2
op4
op4
S3
S3
op6
op6
Fig. 3.3 a Admissible segmentation. b Non-admissible segmentation
The segmentation of Fig. 3.3a, that is S1 = {op1, op2}, S2 = {op3, op4},

S3 = {op5, op6}, is admissible, while that of Fig. 3.3b, that is S1 = {op1, op3},
S2 = {op2, op5}, S3 = {op4, op6}, is not (there is an arc op2 ? op3 from S2
to S1).
Once an admissible partition has been defined, every segment can be synthesized separately, using the same methods as before (scheduling, resource assignment). In order to assemble the complete circuit, additional registers are inserted:
if an arc of the precedence graph crosses the line that separates segments i and
i ? 1, then a register must be inserted; it will store the output data generated by
segment i that in turn are input data to segment i ? 1. As an example, the structure
of the circuit corresponding to Fig. 3.3a is shown in Fig. 3.4.
Assume that Ci and Ti are the cost and computation time of segment i. The cost
and latency of the complete circuit are
C C1 C2 Ck Cregisters and T T1 T2 Tk Tregisters
where Cregisters represents the total cost of the pipeline registers and Tregisters the
additional delay they introduce. The time interval d between successive data inputs is
d maxfT1 ; T2 ; . . .; Tk g TSU TP maxfT1 ; T2 ; . . .; Tk g
where TSU and TP are the set-up and propagation times of the used flip-flops.
60
Fig. 3.4 Pipelined circuit
stage 1:
op1 and op2
stage 2:
op3 and op4
stage 3:
op5 and op6
Fig. 3.5 An admissible

segmentation
xP
xA,zB
xB,zA
xA,zA
xA,zA
S1
a
b
+
S2
squ
f
h
squ
squ
d
S3
e
S4
k
squ
l
S5
A second, more realistic, example is now presented. It corresponds to part of an

Elliptic Curve Cryptography algorithm (Example 2.1).
Example 3.1
An admissible segmentation of the precedence graph of Fig. 2.10 is shown in
Fig. 3.5.
The operations corresponding to each segment are the following.
3.1 Pipeline
61
Segment 1:
Segment 2:
Segment 3:
Segment 4:
Segment 5:
So, every segment includes a product over a finite field plus some additional
1-cycle operations (finite field additions and squares) in segments 2, 4 and 5. The
corresponding pipelined circuit, in which it is assumed that the output results are
g, d, l and i, is shown in Fig. 3.6.
A finite field product is a complex operation whose maximum computation time
tm, expressed in number of clock cycles, is much [1. Thus, the latency T and the
time interval d between successive data inputs of the complete circuit are
T 5tm and d tm :
The cost of the circuit is very high. It includes five multipliers, three adders, four
squarers and four pipeline registers. Furthermore, if used within the scalar product
circuit of Sect. 2.5, the fact that the time interval between successive data inputs has
been reduced (d % tm) does not reduce the execution time of Algorithm 2.4 as the
input variables xA, zA, xB and zB are updated at the end of every main loop execution.
As regards the control of the pipeline, several options can be considered. A simple
one is to previously calculate the maximum multiplier computation time tm and to
choose d [ tm ? 2 (computation time of segment 2). The control unit updates the
pipeline registers and sends a start pulse to all multipliers every d cycles. In the
following VHDL process, time_out is used to enable the pipeline register clock every
delta cycles and sync is a synchronization procedure (Sect. 2.5):
62
In order to describe the complete circuit, five multipliers, three adders and four
squarers are instantiated, and every pipeline register, for example the segment 1
output register, can be described as follows:
A complete VHDL model pipeline_DF2.vhd is available at the Authors web page.

Another interesting option could be a self-timed circuit (Example 3.4).
3.1.3 Combinational to Pipelined Transformation

A very common situation is the following: a combinational circuit made up of
relatively small blocks, all of them with nearly equal delays, has been designed,
and its computation time is equal to T seconds. If this combinational circuit is used
as a computation resource of a synchronous circuit, then the clock cycle must be
greater than T, and in some cases it could be an over extended time (a too low
frequency). In order to increase the clock frequency, as well as to reduce the
minimum time interval between successive data inputs, the solution is pipelining.
As the combinational version already exists, it is no longer necessary to use the
general method of Sect. 3.1.2. The combinational circuit can be directly segmented into stages.
Consider a generic example. The iterative circuit of Fig. 3.7 is made up twelve
identical blocks, each of them with a maximum delay of tcell seconds. The maximum propagation time of every connection is equal to tconnection seconds. Thus,
the computation time of this circuit is equal to T = 6tcell ? 7tconnection (input and
output connections included). Assume that this circuit is part of a synchronous
3.1 Pipeline
Fig. 3.6 Pipelined circuit
(V s: s0 = s(t - 1),
s00 = s(t - 2), s000 =
s(t - 3), siv = s(t - 4))
63
xP
xA zB
xB zA
*
a
xP
xB
zA
xA
b zA
xA
*
b
+
c
sq
d
d
xP
a
*
e
e
zA
xA
*
f
+
g
xAiv
zAiv
giv
liv
div
sq
+
kiv
sq
jiv
*
hiv
sq
i iv
circuit and that all inputs come from register outputs and all outputs go to register
inputs. Then the minimum clock cycle TCLK is defined by the following relation:
TCLK [ 6tcell 7tconnection tSU tP ;
3:1
where tSU and tP are the minimum set-up and propagation times of the registers
(Chap. 6).
If the period defined by condition (3.1) is too long, the circuit must be segmented. A 2-stage segmentation is shown in Fig. 3.8. Registers must be inserted in
64
Fig. 3.7 Combinational

circuit
Fig. 3.8 2-stage

segmentation
all positions where a connection crosses the dotted line. Thus, seven registers must
be added. Assuming that the propagation time of every part of a segmented
connection is still equal to tconnection, the following condition must hold:
TCLK [ 3tcell 4tconnection tSU tP :
3:2
A 5-stage segmentation is shown in Fig. 3.9. In this case, 32 registers must be

added and the following condition must hold:
TCLK [ tcell 2tconnection tSU tP :
3:3
Consider a practical example.

Example 3.2
Implement a 128-bit adder made up of four 32-bit adders. A combinational
implementation is described in Fig. 3.10. The computation time T of the circuit is
equal to 4Tadder, where Tadder is the computation time of a 32-bit adder.
3.1 Pipeline
65
Fig. 3.9 5-stage segmentation

x12796 y12796
cout
x9564
y9564
x6332
y6332
x310
y310
32-bit
adder
32-bit
adder
32-bit
adder
32-bit
adder
z12796
z9564
z6332
z310
cin
Fig. 3.10 128-bit adder
x12796 y12796
cout
x9564
y9564
x6332
y6332
x310
y310
32-bit
adder
32-bit
adder
32-bit
adder
32-bit
adder
z12796
z9564
z6332
z310
cin
Fig. 3.11 4-stage segmentation
A 4-stage segmentation is shown in Fig. 3.11. Every stage includes one 32-bit
adder so that the minimum clock cycle, as well as the minimum time interval
between successive data inputs, is equal to Tadder. The corresponding circuit is
shown in Fig. 3.12. In total, (732 ? 1) ? (632 ? 1) ? (532 ? 1) = 579 additional flip-flops are necessary in order to separate the pipeline stages.
66

x12796 y12796
x9564
y9564
x6332
y6332
x310
y310
32-bit
adder
cin
32-bit
adder
32-bit
adder
32-bit
adder
cout
z12796
z9564
z6332
z310
Fig. 3.12 Pipelined 128-bit adder
Comments 3.1
The extra cost of the pipeline registers could appear to be prohibitive. Nevertheless, the basic cell of a field programmable gate array includes a flip-flop,
so that the insertion of pipeline registers does not necessarily increase the total
cost, computed in terms of used basic cells. The pipeline registers could consist
of flip-flops not used in the non-pipelined version.
Most FPGA families also permit implementing with LUTs those registers that
do not need reset signals. This can be another cost-effective option.
The insertion of pipeline registers also has a positive effect on the power consumption: the presence of synchronization barriers all along the circuit drastically reduces the number of generated spikes.
3.1.4 Interconnection of Pipelined Components

Assume that several pipelined circuits are used as computational resources for
generating a new pipelined circuit. For example, consider a circuit that computes
g = ab and f = (ab ? c)d, and uses a 2-stage pipelined multiplier and a 3-stage
3.1 Pipeline
67
Fig. 3.13 Interconnection of

pipelined components:
scheduling
S1
S2
S3
S4
S5
S6
*
S7
g
pipelined adder, both of them working at the same frequency 1/Tclk and with the
same pipeline rate r = 1. The operations can be scheduled as shown in Fig. 3.13.
Some inputs and outputs must be delayed: input c must be delayed 2 cycles, input
d must be delayed 5 cycles, and output g must be delayed 5 cycles. The corresponding additional registers, which maintain the correct synchronization of the
data, are sometimes called skewing (c and d) and deskewing (g) registers.
An alternative solution, especially in the case of large circuits, is self-timing.
As a generic example, consider the pipelined circuit of Fig. 3.14a. To each
stage, for example number i, are associated a maximum delay tMAX(i) and an
average delay tAV(i). The minimum time interval between successive data inputs is
d maxftMAX 1; tMAX 2; . . .; tMAX ng;
3:4
and the minimum circuit latency T is

T n maxftMAX 1; tMAX 2; . . .; tMAX ng:
3:5
A self-timed version of the same circuit is shown in Fig. 3.14b. The control is
based on a Request/Acknowledge handshaking protocol:
a req_in signal to stage 1 is raised by an external circuit; if stage 1 is free, the
input data is registered (ce = 1), and an ack_out signal is issued;
the start signal of stage 1 is raised; after some amount of time, the done signal of
stage 1 elevates indicating the completion of the computation;
a req_out signal to stage 2 is issued by stage 1; if stage 2 is free, the output of
stage 1 is registered and an ack_out signal to stage 1 is issued; and so on.
If the probability distribution of the internal data were uniform, inequalities (3.4)
and (3.5) would be substituted by the following:
68
ce
req_in
ce
stage 1
ack_out
handshaking
protocol
start
start
done
done
req_out
stage 1
ce
req_in
ce
ack_in
ack_out
stage 2
handshaking
protocol
start
start
done
done
req_out
stage 2
ce
stage n
ack_in
req_in
ce
ack_out
handshaking
protocol
start
start
done
done
req_out
stage n
(a)
ack_in
(b)
Fig. 3.14 a Pipelined circuit. b Self-timed pipelined circuit
d maxftAV 1; tAV 2; . . .; tAV ng;
3:6
T tAV 1 tAV 2 tAV n:
3:7
Example 3.3
The following process describes a handshaking protocol component. As before,
sync is a synchronization procedure (Sect. 2.5):
3.1 Pipeline
69
reqi-1
cei
acki
starti
donei
reqi
acki
segment i exectution
Fig. 3.15 Handshaking protocol
The corresponding signals are shown in Fig. 3.15.

Example 3.4
Consider a self-timed version of the circuit of Example 3.1 (Fig. 3.6). In stages 1
and 3, the done signals are the corresponding done outputs of the multipliers. In
stages 2, 4 and 5 an additional delay must be added. Stage 2 is shown in Fig. 3.16:
a synchronous delay D, greater than the sum of the computation times
of c0 = a0 ? b0 and d0 = (c0 )2, has been added. A complete VHDL model
pipeline_ST.vhd is available at the Authors web page.
70
xP
ack1
req1
xA
xB zA
xB
req_in
ce
ack_out
*
handshaking
protocol
start
+
c
done
req_out
ack_in
req2
ack2
sq
xP
zA xA
Fig. 3.16 Self-timed pipelined circuit: stage 2

Table 3.1 Redundant
encoding
s1
s0
Reset or in transition
0
1
0
0
1
0
1
0
With regards to the generation of the done signal in the case of combinational
components, an interesting method consists of using a redundant encoding of the
binary signals (Sect. 10.4 of [3]: every signal s is represented by a pair (s1, s0)
according to the definition of Table 3.1.
The circuit will be designed in such a way that during the initialization (reset),
and as long as the value of s has not yet been computed, (s1, s0) = (0, 0). Once the
value of s is known s1 = s and s0 = not(s).
Assume that the circuit includes n signals s1, s2,, sn. Every signal si is
substituted by a pair (si1, si0). Then the done flag is computed as follows:
done s11 s10 s21 s20 . . .sn1 sn0 :
During the initialization (reset) and as long as at least one of the signals is in
transition, the corresponding pair is equal to (0, 0), so that done = 0. The done
flag will be raised only when all signals have a stable value.
In the following example, only the signals belonging to the critical path of the
circuit are encoded.
Example 3.5
Generate an n-bit ripple-carry adder (Chap. 7) with end of computation detection.
For this purpose, all signals belonging to the carry chain, that is c0, c1, c2,, cn - 1,
are represented by the form (c0, cb0), (c1, cb1), (c2, cb2),, (cn - 1, cbn - 1). During
the initialization, all ci and cbi are equal to 0. When reset goes down,
3.1 Pipeline
71
reset
xi
reset
yi
FA
ci+1
ci
cin
c0
zi
xi
yi
FA
cbi+1
eoci+1
cbi
cb0
eoci
eoc0
(a)
(b)
Fig. 3.17 a Iterative cell. b Initial cell
c0 cin ;
ci1 xi yi xi ci yi ci ;
8i 2 f0; 1; . . .n 1g:
cb0 cin ;
cbi1 xi yi xi cbi yi cbi ;
The end of computation is detected when

cbi ci ;
8i 2 f0; 1; . . .; ng:
The following VHDL process describes the circuit:
The corresponding circuit is shown in Fig. 3.17.

A complete model adder_ST2.vhd is available at the Authors web page.
In order to observe the carry chain delay, after clauses have been added (1 ns for
72
Fig. 3.18 Iterative algorithm

implementation
initial_values
iterative_operations
data_1
data_2
data_p-1
registers
initially: initial_values
control
final_results
final_results
(b)
(a)
ci ? 1 and cbi
deleted.
? 1,
0.2 ns for eoci
? 1).
For synthesis purpose, they must be
3.2 Loop Unrolling and Digit-Serial Processing

Consider an iterative algorithm whose main operation consists of executing a
procedure iterative_operations(a: in; b: out):
Assuming that a combinational component that implements iterative_operations

has been previously developed, two straightforward implementations of the
algorithm are shown in Fig. 3.18. The first one is an iterative combinational circuit
whose cost and delay are
Ccombinational p Ccomponent ; Tcombinational \p Tcomponent :
The second one is a sequential circuit whose main characteristics are
Csequential Ccomponent Cregisters Ccontrol ; Tclk [ Tcomponent ; Tsequentiall p Tclk :
An alternative option consists of a partial unroll of the for loop [1, 2]. Assume
that p = ks. Then, s successive iteration steps are executed at each clock cycle.
73
Fig. 3.19 Unrolled loop

implementation (s = 3)
registers
initially: initial_values
control
final_results
An example, with s = 3, is shown in Fig. 3.19. Obviously, the clock cycle, say
Tclk0 , must be longer than in the sequential implementation of Fig. 3.18b (Tclk).
Nevertheless, it will be generally shorter than sTclk. On the one hand, the critical
path length of s serially connected combinational circuits is generally shorter than
the critical path length of a single circuit, multiplied by s. For example, the delay
of an n-bit ripple-carry adder is proportional to n; nevertheless the delay of two
serially connected adders, that compute (a ? b) ? c, is proportional to n ? 1, and
not to 2n. On the other hand, the register delays are divided by s. Furthermore,
when interconnecting several circuits, some additional logical simplifications can
be performed by the synthesis tool. So,
Cunrolled s Ccomponent Cregisters Ccontrol ; Tunrolled
0
0
p=s Tclk
; where Tclk
\s Tclk :
Example 3.6
Given two naturals x and y, with x \ y, the following restoring division algorithm
computes two fractional numbers q = 0.q-1 q-2 q-p and r \ y2-p such that
x = qy ? r and, therefore, q B x/y \ q ? 2-p:
74
Fig. 3.20 Restoring

algorithm (combinational
component)
2r
q-i
sign
subtractor
next_r
Algorithm 3.1: Restoring division algorithm
The corresponding circuit is made up of a combinational component (Fig. 3.20), a

register that stores the successive remainders r0, r1,, rp, a shift register that
serially stores the quotient bits q-1, q-2,, q-p, and a control unit. The combinational component can be defined as follows:
A complete model restoring.vhd is available at the Authors web page.

An unrolled version, with s = 2, is made up of a combinational component
consisting of two serially connected copies of the component of Fig. 3.20, a
register that stores the successive remainders r0, r2, r4,, a shift register that stores
the successive quotient bits q-1, q-3, q-5,, another shift register that stores q-2,
q-4, q-6,, and a control unit. The combinational component can be defined as
follows:

y
4r
sign1
subtractor
z1
sign1,sign2,sign3
q-2i+1
q-2i
1--
75
01-
2y
sign2
subtractor
z2
001
3y
sign3
subtractor
z3
000
comb.
circ.
next_r
Fig. 3.21 Digit serial restoring divider (D = 2)
A complete model unrolled_divider.vhd is available at the Authors web page.

The first implementation (restoring.vhd) of Example 3.6 generates one quotient
bit at each step, while the second one (unrolled_divider.vhd) generates two quotient bits at each step. So, as regards to the quotient generation, the first implementation could be considered as bit-serial and the second one as digit-serial,
defining in this case a digit as a 2-bit number. This is a common situation in
arithmetic function implementation: an algorithm processes data, or part of them,
in a bit-serial manner; a modified version of this initial algorithm permits the
processing of several bits, described as D, concurrently. The second implementation is called digital-serial and D is the digit size. This technique is used in many
examples during the course of this book, in order to explore costperformance
tradeoffs: small values of D generate cost-effective circuits, while high values of
D yield fast circuits.
76
Example 3.7
Consider again a restoring divider (Example 3.6). Algorithm 3.1 is modified in
order to generate two quotient bits (D = 2) at each step.
Algorithm 3.2: Base-4 restoring division algorithm (p even)
The corresponding circuit is made up of a combinational component (Fig. 3.21), a

register that stores the successive remainders r0, r2, r4,, a shift register that stores
the successive quotient bits q-1, q-3, q-5,, another shift register that stores q-2,
q-4, q-6,, a circuit that computes 3y, and a control unit. The combinational
component can be defined as follows:
A complete model restoringDS.vhd is available at the Authors web page.
77
The components of Fig. 3.20 (initial restoring algorithm) and Fig. 3.21 (digitserial restoring algorithm, with D = 2) have practically the same delay, namely
the computation time of an n-bit subtractor, so that the minimum clock period of
the corresponding dividers are practically the same. Nevertheless, the first divider
needs p clock periods to perform a division while the second only needs p/2. So, in
this example, the digit-serial approach practically divides by 2 the divider latency.
On the other hand the second divider includes three subtractors instead of one.
Loop unrolling and digit-serial processing are techniques that allow the
exploration of costperformance tradeoffs, in searching for intermediate options
between completely combinational (maximum cost, minimum latency) and completely sequential (minimum cost, maximum latency) circuits. Loop unrolling can
be directly performed at circuit level, whatever the implemented algorithm, while
digit-serial processing looks more like an algorithm transformation. Nevertheless it
is not always so clear that they are different techniques.
3.3 Data Path Connectivity

In the data paths described in Chap. 2, multiplexers are associated with all the
computation resource and register data inputs. With this structure, sets of operations such as Ri := CRj() using different resources CRj can be executed in
parallel. In other words, this type of data path has maximum connectivity.
Assuming that the computation resources have at most p data inputs, another
option is to add p - 1 registers acc1, acc2,, accp - 1, and to realize all the data
transfers with two multiplexers: the first one connects all the register outputs and
external signals to the first data input of every resource as well as to every register
acci; the second one connects the resource data outputs and the first multiplexer
output to the register data inputs. With this structure, an operation such as

Ri : CRj R0 ; R1 ; . . .; Rp1 ;
must be decomposed as follow:

acc1 : R1 ; acc2 : R2 ; . . .accp1 : Rp1 ; Ri : CRj R0 ; acc1 ; . . .; accp1 ;
Obviously, it is no longer possible to execute several operations in parallel. On the
contrary, every operation is divided up into (at most) p steps. So, this option only
makes sense if one resource of each type is used.
Example 3.8
Figure 3.22 shows a minimum connectivity data path for the example of Sect. 2.5.
The operations of the first branch (km - i = 0) of Algorithm 2.4 are executed as
follows:
78
This 2-multiplexer structure includes ten multiplexer inputs instead of thirty-two

in Figs. 2.18 and 2.19, but does not allow the concurrent execution of compatible
operations. Nevertheless, in this case, the total computation time is defined by the
only time consuming operations, which are the five products R := xAacc, xB
:= xBacc, xA := xAacc, xB := Racc and xA := xPacc, so that the latency of the
circuit is still of the order of 5tm, tm being the delay of a multiplier. In conclusion,
the circuit of Fig. 3.22 has practically the same computation time as that of
Figs. 2.18 and 2.19, and uses less multiplexing resources.
3.4 Exercises
1. Generate VHDL models of different pipelined 128-bit adders.
2. Design different digit-serial restoring dividers (D = 3, D = 4, etc.).
3. The following algorithm computes the product of two natural numbers x and y:
3.4 Exercises
79
xA
xB
zA
zB
xP
acc
mult
XOR
square
register
initially: 1
register
initially: xP
register
initially: 0
register
initially: 1
register
xA
xB
zA
zB
Fig. 3.22 Example of minimum connectivity data path
A computation resource mult_and_add that computes 2r ? xiy, is available.

a. Define a combinational circuit using mult_and_add as a computation
resource.
b. Define and compare several pipelined versions of the circuit.
c. Unroll the loop in several ways and synthesize the corresponding circuits.
4. The following algorithm divides an integer x by a natural y, where -y B x \ y,
and generates a quotient q = q0q1 q2qp % x/y (Chap. 9).
80
An adder-subtractor that computes 2r y, under the control of an add/sub

variable, is available.
a. Define a combinational circuit.
b. Define and compare several pipelined versions of the circuit.
c. Unroll the loop in several ways and synthesize the corresponding circuits.
5. In the following combinational circuits, the delays of every cell and of every
connection are equal to 5 ns and 2 ns, respectively.
For each circuit:

a.
b.
c.
d.
e.
Compute the combinational delay.

Segment the circuit in two stages. How many registers must be added?
Segment the circuit in three stages. How many registers must be added?
What is the maximum number of segmentation stages?
Assume that the cutting of a connection generates two new connections
whose delays are still equal to 2 ns, and that the registers have a propagation delay of 1 ns and a setup time of 0.5 ns. Which is the maximum
frequency of circuits b. and c. ?
6. The following pipelined floating-point components are available: fpmul computes the product in 2 cycles, fpadd computes the sum in 3 cycles, and fpsqrt
computes the square root in 5 cycles, all of them with a rate r = 1.
a. Define the schedule of a circuit that computes the distance d between two
points (x1, y1) and (x2, y2) of the (x, y)-plane.
b. Define the schedule of a circuit that computes the distance d between two
points (x1, y1, z1) and (x2, y2, z2) of the three-dimensional space.
c. Define the schedule of a circuit that computes d = a ? ((a - b)c)0.5.
d. In every case, how many registers must be added?
References
81
References
1. Parhami B (2000) Computer arithmetic: algorithms and hardware design. Oxford University
Press, New York
2. De Micheli G (1994) Synthesis and optimization of digital circuits. McGraw-Hill, New York
3. Rabaey JM, Chandrakasan A, Nikolic B (2003) Digital integrated circuits: a design
perspective. Prentice Hall, Upper Saddle River
Chapter 4
Control Unit Synthesis
Modern Electronic Design Automation tools have the capacity to synthesize the
control unit from a finite state machine description, or even to extract and synthesize the control unit from a functional description of the complete circuit
(Chap. 5). Nevertheless, in some cases the digital circuit designer can himself be
interested in performing part of the control unit synthesis. Two specific synthesis
techniques are presented in this chapter: command encoding and hierarchical
decomposition [1]. Both of them pursue a double objective. On the one hand they
aim at reducing the circuit cost. On the other hand they can make the circuit easier
to understand and to debug. The latter is probably the most important aspect.
The use of components whose latency is data-dependent has been implicitly
dealt with in Sect. 2.5. Some additional comments about variable-latency operations are made in the last section of this chapter.
4.1 Command Encoding

Consider the control unit of Fig. 2.6 and assume that commands is an m-bit vector,
conditions a p-bit vector and internal_state an n-bit vector. Thus, the command
generation block generates m ? 1 binary function of p ? n binary variables.
Nevertheless, the number s of different commands is generally much smaller than
2m. An alternative option is to encode the s commands with a t-bit vector, with
2t C s. The command generation block of Fig. 2.6 can be decomposed into two
blocks as shown in Fig. 4.1: the first one generates t ? 1 binary functions of
p ? n variables, and the second one (the command decoder) m binary functions of
t binary variables.
A generic circuit-complexity measure is the number of bits that a memory
(ROM) must store in order to implement the same functions. Thus, the complexity
of a circuit implementing m ? 1 functions of p ? n variables is

83
84
4 Control Unit Synthesis

done
commands
command
decoder
generation of
encoded
commands
p
conditions
internal state
Fig. 4.1 Command encoding
m 1 2pn bits;
4:1
and the total complexity of two circuits implementing t ? 1 function of

p ? n variables and m functions of t variables, respectively, is
t 1 2pn m 2t bits:
4:2
Obviously, this complexity measure only takes into account the numbers of outputs and inputs of the combinational blocks, and not the functions they actually
implement.
Another generic complexity measure is the minimum number of LUTs (Chap. 1)
necessary to implement the functions, assuming that no LUT is shared by two or
more functions. If k-input LUTs are used, the minimum number of LUTs for
implementing a function of r variables is
dr 1=k 1eLUTs;
and the minimum delay of the circuit is
dlogk r e TLUT
being TLUT the delay of a k-input LUT.
The complexities corresponding to the two previously described options are
m 1 dp n 1=k 1eLUTs
4:3
t 1 dp n 1=k 1e m dt 1=k 1eLUTs;
4:4
and
and the delays

dlogk p ne TLUT and dlogk p ne dlogk te T
4:5
Example 4.1
Consider the circuit of Sect. 2.5 (scalar_product.vhd, available at the Authors
web page). The commands consist of 26 bits: eight one-bit signals
4.1 Command Encoding
85
and nine two-bit signals
There are four binary conditions:
and the finite-state machine has 40 states. Thus, m = 26, p = 4 and n = 6.

Nevertheless, there are only 31 226 different commands, namely
that can be encoded with t = 5 bits.

Thus, the complexities in numbers of stored bits (4.1 and 4.2) to be compared
are
m 1 2pn 27 210 27; 648 bits;
4:6
t 1 2pn m 2t 6 210 26 25 6; 976 bits;
4:7
and the complexities in numbers of LUTs (4.3 and 4.4), assuming that 4-input
LUTs are used, are
m 1 dp n 1=3e 27 d9=3e 81 LUTS;
4:8
t 1 dp n 1=3e m dt 1=3e 6 d9=3e 26 d4=3e 70 LUTs:

4:9
The corresponding minimum delays (4.7) are
dlogk p ne dlog4 10e 2TLUT ;
dlogk p ne dlogk te TLUT dlog4 10e dlog4 5e 4TLUT :
4:10
4:11
The second complexity measure (number of LUTs) is surely more accurate than
the first one. Thus, according to (4.84.11), the encoding of the commands
hardly reduces the cost and increases the delay. So, in this particular case, the
main advantage is clarity, flexibility and ease of debugging, and not cost
reduction.
86
Fig. 4.2 Hierarchical circuit
data path
(interleaved
multiplier)
control
(interleaved
multiplier)
control
other components
4.2 Hierarchical Control Unit

Complex circuits are generally designed in a hierarchical way. As an example, the
data path of the scalar product circuit of Sect. 2.5 (Fig. 2.18) includes a polynomial adder (XOR gates), a classic squarer and an interleaved multiplier, and the
latter in turn consists of a data path and a control unit (Fig. 4.2). This is a common
strategy in many fields of system engineering: hierarchy improves clarity, security,
ease of debugging and maintenance, thus reducing development times.
Nevertheless, this type of hierarchy based on the use of previously defined
components does not allow for the sharing of computation resources between
several components. As an example, one of the components of the circuit of
Sect. 2.5 is a polynomial adder, and the interleaved multiplier also includes a
polynomial adder. A slight modification of the operation scheduling, avoiding
executing field multiplications and additions at the same time, would allow to use
the same polynomial adder for both operations. Then, instead of the architecture of
Fig. 4.2, a conventional (flat) structure with a data path including only a polynomial adder could be considered. In order to maintain some type of hierarchy, the
corresponding control unit could be divided up into a main control unit, in charge
of controlling the execution of the main algorithm (scalar product) and a secondary control unit, in charge of controlling the execution of the interleaved
multiplication.
Consider another, simpler, example.
Example 4.2
Design a circuit that computes
z
p
x2 y2 :
87
The following algorithm computes z:
A first solution is to use three components: a squaring circuit, an adder and a square
rooting circuit, for example that of Sect. 2.1. The corresponding circuit would include
two adders, one for computing c, and the other within the square_root component
(Fig. 2.3). Another option is to substitute, in the preceding algorithm, the call to
square_root with the corresponding sequence of operations. After scheduling the
operations and assigning registers to variables, the following algorithm is obtained:
This algorithm can be executed by the data path of Fig. 4.3.

In order to distinguish between the main algorithm and the square root
computation, the control unit can be divided up as shown in Fig. 4.4.
A command decoder (Sect. 4.1) is used. There are eight different commands,
namely
that are encoded with three bits. The following process describes the command
decoder:
88

r
x
sel_ sq
squaring
sel_a1
cy_ out
x2
adder
square
sel_a2
cy_ in
sum
sel_r
sel_s
en_r
load
en_ c
en_s
load
c
en_ signb
cy_ out
signb
Fig. 4.3 Data path
start
done
conditions
commands
start_ root
control unit 1
(main algorithm)
Fig. 4.4 Hierarchical control unit
root _done
control unit 2
(square root )
89
The two control units communicate through the start_root and root_done signals. The first control unit has six states corresponding to a wait for start loop,
four steps of the main algorithm (operations 1, 2, 3, and the set of operations 48),
and an end of computation detection. It can be described by the following
process:
The second control unit has five states corresponding to operations 4, 5, 6, and
7, and end of root computation detection:
90
The code corresponding to nop is 000, so that the actual command can be
generated by ORing the commands generated by both control units:
A complete VHDL model example4_1.vhd is available at the Authors web

page.
Comments 4.1
This technique is similar to the use of procedures and functions in software
generation.
In the former example, the dividing up of the control unit was not necessary. It
was done only for didactic purposes. As in the case of software development,
this method is useful when there are several calls to the same procedure or
function.
91
This type of approach to control unit synthesis is more a question of clarity (well
structured control unit) and ease of debugging and maintenance, than of cost
reduction (control units are not expensive).

In Sect. 2.3, operation scheduling was performed assuming that the computation
times tJM of all operations were constant values. Nevertheless, in some cases the
computation time is not a constant but a data-dependent value. As an example, the
latency tm of the field multiplier interleaved_mult.vhd of Sect. 2.5 is dependent on
the particular operand values. In this case, the scheduling of the operations was
done using an upper bound of tm. So, an implementation based on this schedule
should include wait for tm cycles loops. Nevertheless, the proposed implementations (scalar_product.vhd and scalar_product_DF2.vhd) are slightly different: they use the mult_done flag generated by the multiplier. For example, in
scalar_product_DF2.vhd (Sect. 2.5), there are several sentences, thus:
In an implementation that strictly respects the schedule of Fig. 2.14, these particular sentences should be substituted by constructions equivalent to
In fact, the pipelined circuit of Fig. 3.6 (pipeline_DF2.vhd) has been designed
using such an upper bound of tm. For that, a generic parameter delta was defined
and a signal time_out generated by the control unit every delta cycles. On the other
hand, the self-timed version of this same circuit (Example 3.4) used the mult_done
flags generated by the multipliers.
Thus, in the case of variable-latency components, two options could be considered: a first one is to previously compute an upper bound of their computation
times, if such a bound exists; another option is to use a start-done protocol: done is
lowered on the start positive edge, and raised when the results are available. The
second option is more general and generates circuits whose average latency is
shorter. Nevertheless, in some cases, for example for pipelining purpose, the first
option is better.
Comment 4.2
A typical case of data-dependent computation time corresponds to algorithms that
include while loops: some iteration is executed as long as some condition holds
true. Nevertheless, for unrolling purpose, the algorithm should be modified and the
while loop substituted by a for loop including a fixed number of steps, such as for i
92
in 0 to n - 1 loop. Thus, in some cases it may be worthwhile to substitute a

variable-latency slow component by a constant-latency fast one.
An example of a circuit including variable-latency components is presented.
Example 4.3
Consider again Algorithm 2.3, with the schedule of Fig. 2.17, so that two finite
field multipliers are necessary. Assume that they generate output flags done1 and
done2 when they complete their respective operations. The part of the algorithm
corresponding to km-i = 0 can be executed as follows:
The following VHDL model describes the circuit:
93
A complete model unbounded_DF.vhd is available at the Authors web page.

Other implementations, using latency upper bounds and/or pipelining or selftiming, are proposed as exercises.
94
4.4 Exercises
1. Design a circuit that computes z = (x1 - x2)1/2 ? (y1 - y2)1/2 with a hierarchical control unit (separate square rooter control units, see Example 4.2).
2. Design a 2-step self-timed circuit that computes z = (x1 - x2)1/4 using two
square rooters controlled by a start/done protocol.
3. Design a 2-step pipelined circuit that computes z = (x1 - x2)1/4 using two
square rooters, with a start input, whose maximum latencies are known.
4. Consider several implementations of the scalar product circuit of Sect. 2.5,
taking into account Comment 2.2. The following options could be considered:
hierarchical control unit;

with an upper bound of the multiplier latency;
pipelined version;
self-timed version.
Reference
1. De Micheli G (1994) Synthesis and optimization of digital circuits. McGraw-Hill, New York
Chapter 5
Electronic Aspects of Digital Design
This chapter is devoted to those electronics aspects important for digital circuit
design. The digital devices are built with analog components, and then some
considerations should be taken into account in order to obtain good and reliable
designs.
Some important electronics aspects, related to the circuit design, the timing and
synchronization aspects are discussed in this chapter. Most of those details are
hidden in reprogrammable logic for simplicity, but these do not eliminate the
consequences.
5.1 Basic Electronic Aspects of Digital Design

Digital devices represent signals by discrete bands of analog levels, rather than by
a continuous range. All levels within a range represent the same signal state.
Typically the number of these states is two, and they are represented by two
voltage bands: one near zero volts (referenced also as ground or earth) and a higher
level near the supply voltage, corresponding to the false (0) and true (1)
values of the Boolean domain, respectively (Fig. 5.4).
5.1.1 Basic Concepts

Before we start looking at more specific ideas, we need to remember a few basic
concepts.

95
96
5 Electronic Aspects of Digital Design
(a)
(b)
p-type
Vdd
n-type
Vdd
Vdd
Vdd
0
i0
i0
i1
out
out
i1
i0
1
out
0
i0
i1
i1
gnd
switching action
gnd
i
out
i0
i1
gnd
out
NOR
gnd
i0
i1
out
NAND
Fig. 5.1 CMOS transistors. a Switching action for p-type and n-type. b Some basic gates
5.1.1.1 CMOS Circuits

Complementary metaloxidesemiconductor (CMOS) is the dominant technology
for constructing digital integrated circuits since mid 1970s. The word complementary refers to the fact that the typical digital design style with CMOS uses
complementary and symmetrical pairs of p-type and n-type metal oxide semiconductor field effect transistors (MOSFETs) for logic functions. The key characteristics of CMOS devices are high noise immunity and low static power
consumption. Neglecting other technological details, the power is only drawn
while the transistors in the CMOS device are switching between on and off states.
The CMOS transistors allow creating the logic gates of Fig. 1.1 that are the basics
of the digital design. The key idea is the use of transistors as switches as described
in Fig. 5.1a. The p transistor conducts better the supply voltage meanwhile the
n-type conducts to the ground. With this, the main principle behind CMOS circuits
is the use of p-type and n-type transistors to create paths to the output from either
the voltage source or ground. Figure 5.1b shows the complementary-symmetric
interconnecttion of p-type and n-type transistors to build some basic gates. More
technological details of these constructions are out of the scope of this book. One
can find excellent surveys and material in [1, 2].
The FPGA technology is a CMOS (re)programmable Application Specific
Integrated Circuit (ASIC). Hence, all the characteristics, peculiarities and consequences of an ASIC design are present.
5.1.1.2 Fan-in and Fan-out

In ASICs technologies and others the fan-in and fan-out are measured in capacitance units (in actual technologies in picofarad or femtofarad). So, the fan-in is the
97
Table 5.1 Fan-in, fan-out, internal and external delay of typical gates
fan-in
fan-out
t_int_hl
t_int_lh
t_ext_hl
(pf)
(pf)
(ns)
(ns)
(ns/pf)
t_ext_lh
(ns/pf)
INV
BUF
AND2
AND3
NAND2
NAND3
NOR2
XOR2
CKBUF
CKBUF_N
DFF
4.891
2.399
4.271
2.376
4.271
7.212
8.035
2.560
1.782
1.881
2.360
0.003
0.004
0.003
0.007
0.004
0.003
0.004
0.008
0.006
0.006
0.003
0.337
0.425
0.334
0.673
0.197
0.140
0.205
0.645
1.160
1.460
0.703
0.079
0.265
0.105
0.211
0.105
0.071
0.091
0.279
0.355
0.183
0.402
0.151
0.056
0.144
0.131
0.144
0.192
0.162
0.331
0.350
0.537
0.354
2.710
1.334
4.470
1.362
4.470
6.088
3.101
1.435
0.782
0.628
1.256
capacitance that an input of a gate has. Thus, the fan-out is the maximum
capacitance controllable by a gate, while providing voltage levels in the guaranteed range. The fan-out really depends on the amount of electric current a gate can
source or sink while driving other gates. Table 5.1 shows examples of fan-in and
fan-out for gates in 0.25 lm technologies (a ten years old technology). Observe,
for example, that a NAND2 gate can drive up 49 similar gates (197pf/4pf) if we
neglect the capacitance of interconnection.
In the case of FPGA the concepts of fan-in and fan-out are simplified and
measured in number of connections. Remember that most of the logic is implemented in look-up tables (LUTs). Then the fan-in is the number of inputs a
computing block has, like a two input AND gate implemented in a LUT has a fanin of two, and a three input NAND gate a fan-in of three. The concept of fan-out is
used to express the number of gates that each gate has connected at the output. In
some cases there appears the concept of maximum fan-out as the maximum
number of connections that we can control by a gate.
5.1.1.3 Drive Strength or Drive Capabilities

This means the amount of power a gate or circuit can output (i.e. same concept of
fan-out). The larger drive strength a circuit has, the more power it can deliver and
thus the more capable it is to drive a higher number of other gates. In the particular
case of ASIC libraries, a typical gate has different drive strengths (x1, x2, x4, etc.)
to be able to drive more gates. This allows the designer or the synthesizer to
choose the right one at each time. In FPGAs the internal drive strength is not
controlled by programmers. Nevertheless, the output strength of the output buffers
can be controlled.
98

input pad
vcc
pull-up
resistor
Internal value
pull-down
resistor
input pad
Internal value
Fig. 5.2 Pull-up and pull-down resistors connected to input pads
(a)
(b)
bus
in
out
In G out
0 1 0
1 1 1
0
Z
in
resistor
out
G
Fig. 5.3 Tri-state buffer and bus keeper
5.1.1.4 Pull-up and Pull-down Resistors

The purpose of these resistors is to force an input (or output) to a defined state. The
pull-up and pull-down resistors are mainly used to avoid a circuit input from being
floating if the corresponding external device is disconnected. Figure 5.2 shows a
pull-up and a pull-down resistor. The pull-up resistor weakly pulls the internal
voltage towards Vcc (logical 1) when the other components are inactive, and the
pull-down resistor weakly pulls the internal voltage towards GND (logical 0).
Pull-up/down resistors may be discrete devices mounted on the same circuit
board as the logic devices, but many microcontrollers and FGPA have internal,
programmable pull-up/pull-down resistors, so that less external components are
needed.
5.1.1.5 Tri-States Buffers and Bus-Keeper

Tri-state (also named as three-state or 3-state) logic allows an output port to
assume a high impedance state (Z, or Hi-Z) in addition to the 0 and 1 logic levels.
The tri-state buffers are used to connect multiple devices to a common bus (or
line).

Logic values
Voltage (arbitrary units)
1.0
0.8
transition time
high-to-low
99
transition time
low-to-high
1.0
Logic 1
noise margin
0.9
0.6
invalid
0.4
0.2
0.0
noise margin
Logic 0
0.1
Rise time
Fall time
Fig. 5.4 Logic values and transition times (rise and fall time)
As is suggested in Fig. 5.3a, a tri-state buffer can be thought of as a switch. If

G is on, the switch is closed, transmitting either 0 or 1 logic levels. If G is off, the
switch is open (high impedance).
A related concept is the bus-keeper (or bus-holder). The bus-keeper is a weak
latch circuit which holds the last value on a tri-state bus. The circuit is basically a
delay element with the output connected back to the input through a comparatively
high impedance. This is usually achieved with two inverters connected back to
back. The resistor drives the bus weakly; therefore, other circuits can override the
value of the bus when they are not in tri-state mode (Fig. 5.3b).
Many microcontroller and FGPA have internal, programmable tri-states buffers
and the bus-keeper logic at outputs pins, so that minimal external components are
needed to interface to other components.
5.1.2 Propagation DelayTransition Time

It is the time that an electronic circuit needs to switch between two different stable
states. In a logic circuit performing a change of state, it identifies the rise-time and
the fall-time of the output voltage. These times are related to the times to charge
and discharge capacitances. Figure 5.4 shows typical charge and discharge ramps
of a CMOS gate.
5.1.2.1 Rise Time (Transition Time Low-to-High)

It refers to the time required for a signal to change from a specified low value to a
specified high value. Typically, these values are 10% and 90% of the step height.
100

Derating Curve for Operating Temperatures
1.2
Derating Curve for Supply Voltage

2
1.1
1.5
1.0
Typcial Case.
Derating Factoris 1
0.9
0.8
-60 -40 -20 0
Typcial Case.
Derating Factoris 1
0.5
20
40
60
80 100 120 140 160
1.8
2.2
Operating Temperature (C)
2.4
2.6
2.8
3.2
3.4
3.6
Core Supply Voltage (V)
Fig. 5.5 Typical derating curves for temperature and supply voltage
5.1.2.2 Fall Time (Transition Time High-to-Low)

It is the time required for the amplitude of a pulse to decrease (fall) from a
specified value, typically 90% of the peak value, to another specified value, usually
10% of the minimum value.
5.1.2.3 Slew Rate

This concept is used in linear amplifiers but also in digital circuits. In the second
case, it is an approximation of the time necessary to change a logic value, without
distinguishing between high-to-low and low-to-high transitions. In the FPGA
arena, the output pins can be configured as slow slew rate and fast slew rate. The
second one is faster, but consumes more power and is prone to transmit internal
glitches to the outputs.
5.1.2.4 Propagation Delay, Intrinsic and Extrinsic Delays

The propagation delay of a gate or digital component depends on two factors, the
intrinsic and the extrinsic delays. The intrinsic delay is the delay internal to the
gate (also known as gate delay or internal delay). In other words, it is the time
taken by the gate to produce an output after the change of an input (in ns, ps, etc.).
On the other hand, the extrinsic delay (also called, differential, load-dependent
or fan-out delay) depends on the load attached to the gate. This value is expressed
in ns/pf. Table 5.1 shows some intrinsic and extrinsic delays for 0.25 lm gates.
The intrinsic delay in fact depends on which input is changing, but for simplicity
this is not taken into account in the table. Observe that the high to low (HL) and
low to high (LH) values are different. As an example, consider an inverter (INV)
whose output is connected to 10 inverters. Assume that the capacitance of every
101
Table 5.2 A derating table for temperature and supply voltage

Vcc
Temperature (C)
2.6
2.8
3.0
3.2
3.3
3.4
120
100
80
60
40
32
20
-20
1.39
1.29
1.23
1.18
1.16
1.14
1.36
1.25
1.20
1.15
1.13
1.11
1.32
1.22
1.17
1.12
1.10
1.08
1.27
1.18
1.12
1.08
1.06
1.04
1.24
1.14
1.09
1.05
1.03
1.01
1.20
1.11
1.06
1.02
1.00
0.98
1.19
1.10
1.05
1.01
0.99
0.97
1.15
1.07
1.02
0.98
0.96
0.94
1.12
1.03
0.99
0.95
0.93
0.91
connection is equal to 0.1 pf. What is the time required for the INV gate to
propagate the transitions 0 to 1 and the 1 to 0?
Thl t inthl t exthl capacity
0:079ns 2:710ns=pf 10 0:003pf 0:1pf 0:4313ns
Tlh t intlh t extlh capacity
0:151ns 4:891ns=pf 10 0:003pf 0:1pf 0:7868ns
In FPGA technologies these concepts are hidden, given only a propagation
delay for internal nodes (LUTs, multiplexers, nets, etc.), and do not distinguish
between high to low and low to high transitions.
5.1.2.5 Timing Derating Factors

The propagation delays depend also on the temperature and the supply voltage.
Variations in those parameters affect the circuit delay. The circuit manufacturers
even give tables, graphics or embed the information in the static timing analyzer to
adapt the timing to the operating conditions.
The derating factors are coefficients that the timing data are multiplied by, in
order to estimate the delays corresponding to the operating conditions. Figure 5.5
shows typical derating curves for different operating temperature and supply
voltage (the example is based on a 180 nm process). Table 5.2 is another form to
show the same information.
5.1.3 Glitches in Digital Circuits

An electronics glitch is an undesired transition that occurs before the signal settles
to its proposed final value. In other words, a glitch is an electrical pulse of short
duration, usually unwanted, and typically produced by an imbalance in internal
interconnection delays. A simple example is depicted in Fig. 5.6 where different
102

a
A
B
a < b
C
A
B
a > b
C
`1
`0
`1
`0
`0
`1
`0
`1
`1
`1
`0
`0
`1
`0
`1
`1
`1
`0
a
t0 t1 t2
`0
`0
a
t0 t1 t2
Fig. 5.6 Glitches of a simple gate due to unbalanced interconnection delay
delays in interconnections produce unnecessary changes at the output. For

simplicity, in the preceding figure the gates delays are assumed to be equal to 0.
The glitch effect grows with the logic depth. As a simple example, consider two
levels of XOR gates (Fig. 5.7). The four inputs change at the same time (t0) but the
net delays produce different times of arrival to the gates. In this case, changing
from 1010 to 0101 should not produce any output change. Nevertheless, in the
example, four transitions in output G are generated.
5.1.3.1 Runt Pulse and Spikes

A related concept to glitches is the runt pulse. It is a pulse whose amplitude is
smaller than the minimum level specified for correct operation. It is a narrow pulse
that, due to rise and fall times of the signal, does not reach a valid high or low level
value. Typically, it has no influence on the operations, but it increases the power
consumption (see Sect. 5.4).
Some authors define the concept of spike as being a short pulse similar to a
glitch, but caused by ringing (an unwanted oscillation of a signal) or crosstalk
(undesired capacitive or inductive coupling).
5.2 Synchronous Design Issues
103
c
Fig. 5.7 The avalanche

effect of glitches as the logic
depth grows
b
d
A
`1
`0
`0
`1
`1
`0
`0
`1
c > a > b> d

e
E
B
C
D
E
F
G
`1
`0
`1
`1
`0
`1
`0
`0
`0
b B
c C
d D
c > a > b > d

e = f = XOR = 0
t0
t1 t 2 t 3 t 4

Digital circuits are mainly synchronous. The Register Transfer Level (RTL)
model recognizes register and combinational logic connected between them.
The typical form, adopted to synchronize these components, is the True Single
Phase Clock (TSPC). In TSPC a single clock signal is supposed to be distributed at different register levels. All the registers (flip-flops) are capturing on
the rising (or falling) edge. Figure 5.8 illustrates the classical situation. A clock
edge (rising in Fig. 5.8) captures the information into the first level of registers
(ffi). The contents of registers ffi (B) are propagated into the combinational
logic and generate new results. The output of the combinational circuitry (C)
should be stable before the next clock arrives to the next register stage (ffj).
The following clock edge captures the content of C in ffj and makes D
available.
In an ideal synchronous circuit, every change in the logical levels of its registers
(flip-flops) is simultaneous. These transitions follow the level change of the clock
(clk) signal (positive or negative clock edge). In normal function, the input to each
storage element has reached its final value before the next edge clock occurs, so
the behavior of the whole circuit is exactly predictable.
104
Fig. 5.8 Typical single
phase synchronization

dff
m A
dff
Q
Combinational
circuit
ffi
D n
ff j
clk
clk period
clk
B
C
D
Combinational delay
5.2.1 Edge Sensitive and Level Sensitive Registers

The edge sensitive and level sensitive registers (flip-flop and latch) are electronic
circuits that have two stable states and can be used to store binary information. The
circuits are made to change state by signals applied to one or more control inputs
and will have one or two outputs. The latches are level sensitive circuits that pass
their input D to their output Q when the clock is high (or low) (transparent mode),
and the input sampled on the falling edge of the clock is held stable when the clock
is low (or high)-hold mode. The flip-flops (also called edge-triggered latches) are
edge sensitive circuits that sample the inputs on a clock transition (positive edgetriggered: 0 ? 1 or negative edge-triggered: 1 ? 0). They are built using latches
(e.g., masterslave flip-flops). Historically, there have been different classes of flipflops (FF) depending on the way they work (SR set-reset, D data, T
toggle, and JK). Todays FPGA have registers that can be configured as latches
or flip-flop, but only as a D-FF.
5.2.2 Temporal Parameters of Flip-Flops

The synchronous flip-flop requires that the data to be sampled is stable some time
before and after (setup and hold times) the clock edge Fig. 5.9. If the input data
changes in the setup-hold windows, metastability could occur (next section).
To summarize: the setup time (tsu) is the minimum amount of time the data signal
should be held steady before the clock event; and the hold time (th) is the minimum
amount of time the data signal should be held steady after the clock event, so that
105
dff
IN
OUT
clk
t su
th
data
stable
IN
clk
t pr
t h = hold time
t su = setup time
OUT
output
stable
t pr = propagation delay
output
stable
time
Fig. 5.9 Setup time, hold time and propagation delay of a register
the data are reliably sampled. These times are specified for any device or technology, and are typically between a few tens of picoseconds and a few nanoseconds for modern devices.
The data stored in the flip-flop is visible at its output as a propagation delay (tpr)
after the clock edge. This timing value is also known as clock-to-output delay.
Another related concept is the minimum clock pulse of a flip-flop. It is the
minimum width of the clock pulse necessary to control the register.
5.2.3 Metastability
Whenever there is a setup or a hold time violation, the flip-flop could enter in a
metastable state (a quasi-stable state). In this state the flip-flop output is unpredictable and it is considered as a failure of the logic design. At the end of a
metastable state, when the output could reveal an in between value, the flip-flop
settles down to either 1 or 0. This whole process is known as metastability.
Figure 5.10 illustrates this situation. The duration of the metastable state is a
random variable that depends on the technology of the flip-flop. The circuits
vendors provide information about the metastability in their devices. As an
example, the main vendors of FPGA provide this kind of information [35].
Comments 5.1
1. Not all the setup-hold window violations imply a metastable state. It is a
probabilistic event.
2. Not all the metastability states cause design failures. In fact, if the data output
signal resolves to a valid state before the next register captures the data, then
the metastable signal does not impact negatively in the system operation.
106
clk
dff
IN
t su t h
OUT
t su t h
t su t h
IN
t pr
clk
t pr
t pr
OUT
Metastable
State
Fig. 5.10 Metastability capturing data in a flip-flop
5.2.3.1 Main Causes of Metastability

As previously mentioned, a setup or hold time violation, could produce metastability, so we have to see when signals violate this timing requirement:
When the input signal is an asynchronous signal.

When interfacing two domains operating at two different clock frequencies.
When the clock skew is too high. (Sect. 5.3.1)
When the clock slew rate is too high (rise and fall times are longer than the
tolerable values).
When interfacing two domains operating at the same frequency but with different phase.
When the combinational delays are such that flip-flop data input change within
the critical window. (setup-hold window)
Observe that that only the first two cases are real metastability problems and we
will discuss possible solutions in the next sections. The other cases are related with
the clock frequency and simple increases of the clock period could solve the
problem. Even a static timing analysis detects these kinds of problems.
5.2.3.2 Mean Time Between Failures (MTBF) in Metastability

The Mean Time Between Failures (MTBF) is a general concept used in reliability.
This concept, applied to the metastability, gives the average time interval between
two successive failures due to this phenomenon. The expression that computes
MTBF is:
MTBF
eTrec K2
K1 fclk fdata
Where:
fclk is the frequency of the clock receiving the asynchronous signal.
fdata is the toggling frequency of the asynchronous input data signal.
5:1
107
1e18
1 Billon Years
1e15
MTBF (log seconds)
1 Millon Years
1e12
1000 Years
1e9
1 Years
1e6
1e3
1 Day
1 Hr
1 min
1e0
1e-3
1e-6
0.5
1.0
1.5
2.0
2.5
3.0
Trec (ns)
Fig. 5.11 MTBF for 300 MHz clock and 50 MHZ data in a 130 nm technology
Trec is the available metastability settling time (recovery time), or the timing
until the potentially metastable signal goes to a known value 0 or 1.
K1 (in ns) and K2 (in 1/ns) are constants that depend on the device process and
on the operating conditions.
If we can tolerate more time to recover from the metastability (Trec) the MTBF
is reduced exponentially. Faster clock frequencies (fclk) and faster-toggling (fdata)
data worsen (reduce) the MTBF since the probability to have a setup-hold window
violation augments. Figure 5.11 shows a typical MTBF graphic. These graphics
are average data for 300 MHz clock and 50 MHz data in a 130 nm technology.
Comments 5.2
1. Observe the sensitivity to the recovery time. Following Fig. 5.11, if you are
able to wait 2 ns to recover from metastability, the MTBF is less than two
weeks. But if you can wait for 2.5 ns the MTBF is more than three thousand
years.
2. Suppose for the previous data (fclk = 300 MHz, fdata = 50 MHz, Trec =
2.5 ns) that give a MTBF of around 3200 years. But, if we have a system with
256 input bits the MTBF is reduced to 12.5 years. If we additionally produce
100,000 systems we have MTBF of 1 h!
3. Measuring the time between metastability events using real designs under real
operating conditions is impractical, because it is in the order of years. FPGA
vendors determine the constant parameters in the MTBF equation by characterizing the FPGA for metastability using a detector circuit designed to have a
short, measurable MTBF [3, 5].
108
syncronization chain
asynch
input
clk
dff
dff
Q
INT
t su th
Q
OUT
asynch
input
t pr
clk
INT
th = hold time
Metastable
tsu = setup time

tpr = propagation delay
OUT
tpr
t pr
Fig. 5.12 Using a two flip-flops synchronizer to reduce metastability
5.2.3.3 How to Avoid or Mitigate Metastability

In reality, one cannot avoid metastability, without the use of tricky self-timed
circuits. So a more appropriate question might be How to mitigate the effect of
metastability?
The simplest solution is by making sure the clock period is long enough to
allow the resolution of metastable states and for the delay of whatever logic may
be in the path to the next flip-flop. This approach, while simple, is unpractical
given the performance requirements of modern designs.
The classical solution is the use of a sequence of registers (a synchronization
registers chain or synchronizer) in the destination clock domain (Fig. 5.12). The
idea is to cascade one or more successive synchronizing flip-flops to the flip-flop
that is interfacing with the asynchronous input.
This approach cannot guarantee that metastability cannot pass through the synchronizer; they simply reduce the probability to practical levels. In order to evaluate
the effect, consider the three cases of Fig. 5.13, assuming that tcomb ? tsu % tclk
(combinational delay plus setup time similar to clock period). In the first case,
without registering the asynchronous input, Trec is near zero, raising the MTBF to an
unreliable system. In the second case, with a simple flip-flop capturing the
asynchronous input, Trec % tclk (tcomb ? tsu). In the final case, using two FFs to
interface the asynchronous input, the recovery time Trec % tclk tsu. For the data of
Fig. 5.11 (i.e. tclk = 3.33 ns) and assuming a tsu = 350 ps, the MTBF of the
synchronization chain is more than 1 billon years.
5.3 Clock Distribution Network

The clock distribution network (or clock tree) distributes the clock signal(s) from a
common input (an input pin) to the entire synchronous element (registers). Since
this function is vital in an synchronous system, special attention is given to the
109
(a)
(b)
asynch
input
comb
n
D
asynch
input
clk
comb
clk
(c)
asynch
input
D
n
D
clk
Fig. 5.13 Synchronizing asynchronous inputs in a design due to the metastability problem. a No
registers. b One synchronization register. c Two register in synchronization chain
(a)
D Q
...
(b)
.
.
.
ck2
D Q
clk
ck1
.
.
.
ck2
...
ck3
.
.
.
clk
D Q
ck3
...
.
.
.
.
.
.
D Q
Fig. 5.14 Typical clock tree distribution. a Buffers tree. b Ideal H-routing distribution
design of this component. In ASIC design a special clock synthesizer is used

(a step in physical synthesis) during the design. In FPGA several dedicated prefabricated clock distribution networks are present.
Clock signals are typically loaded with the greatest amount of interconnections
and operate at the highest speeds of any signal in the system. Todays clock signals
are easily distributed to tens of thousands of points. Due to the driving capability
of the buffers, the clock distribution is designed as trees (Fig. 5.14a).
As a technological example, suppose use of the data of Table 5.1, and use the
buffer BUF to distribute a clock signal. If we neglect the interconnection load (an
unreal case), a simple BUF can drive 141 flip-flops (DFF) (0.425 pf/0.003 pf) and
also 106 similar buffers (0.425 pf/0.004 pf). If we want to distribute the signal to
110
(a)
Launching
flip-flop(s)
m A
cki
Capturing
flip-flop(s)
B
Combinational
circuit
ffi
ckj
D n
ffj
clk
(b)
(c)
clk period (T)
clk
clk
ti
ti
tj
cki
ck i
ckj
ck j
tprop
B
C
clk period (T)
old_c
new_c
ERROR
Combinational delay
tj
t prop
old_b
old_c
new_b
new_c
new_c
Combinational
delay
Fig. 5.15 Setup and hold violations due to clock skew. a Example circuit. b Setup violation or
long-path fault. B. Hold violation or race-through
64,000 FF we need at least three levels of those buffers. Real clock trees use
special buffers with more driving capability and with other special characteristics,
and have additional problems related with the interconnections. Figure 5.14 shows
a classic clock tree that present multiple levels of buffers, and the typical H-routing
used to reduce the skew.
5.3.1 Clock Skew

The variation in the arrival times of a clock transition at two different locations on
a chip is commonly known as the clock skew (also timing skew). Being the cause
of the unbalanced delays in the clock distribution network, clock edges could
arrive at clock pins cki and ckj in Fig. 5.15 at different times. This spatial variation
can be caused by many different things, such as wire-interconnect length, temperature variations, material imperfections, capacitance coupling, etc. The skew
between two registers stages i and j could be defined as
skewi;j delaycki delaycki ti tj
5:2
That leads to two types of clock skew: positive skew (skewi,j [ 0) and negative
skew (skewi,j \ 0). Positive skew occurs when the transmitting register (launching flip-flop, ffi) receives the clock later than the receiving register (capturing
111
flip-flop, ffj). Negative skew is the opposite; the sending register gets the clock
earlier than the receiving register. Two types of time violation (synchronization
failures) can be caused by clock skew: setup and hold violations. They are
described in what follows.
5.3.1.1 Setup Violation due to Clock Skew

If the destination flip-flop receives the clock edge earlier than the source flip-flop
(positive skew), the data signal has that much less time to reach the destination
flip-flop before the next clock edge. If it fails to reach the destination timeously, a
setup violation occurs, so-called because the new data was not stable before the
Tsu (setup time) of the next clock edge. This error is drawn in Fig. 5.15b and is
also called zero-clocking (because nothing is captured properly) or also long-path
fault.
In this case, at time tmax
(assuming several FF at level i, the worst case), the
i
clock edge arrives at flip-flops ffi. In the worst case, to propagate through the flipmax
max
flops and the combinational logic it takestprop
tcomb
: The signal must have settled
max
down for a duration of tsu before the next clock edge arrives at the next flip-flop
stage (ffj) at time tmin
? Tin order to be captured properly. That transform in the
j
following general setup time constrain:
max
max
tjmin T [ timax tprop
tcomb
5:3
Observe that this inequation could be always satisfied if the clock period (T) could
be incremented. In other words, a positive clock skew, as shown in Fig 5.15b,
places a lower bound on the allowable clock period (or an upper bound on the
operating frequency) as follows:
max
max
max
T [ tprop
tcomb
tsu
skewmax
i;j
5:4
While a positive skewi,j decreases the maximum achievable clock frequency, a

negative clock skew (skewi,j \ 0) actually increases the effective clock period.
Effectively, we may have a combinational block between two adjacent flip-flops
that has a propagation delay longer than the given clock period.
5.3.1.2 Hold Violation due to Clock Skew

The hold violation is caused when the clock travels more slowly than the path from
one register to another (negative skew), allowing data to penetrate two registers in
the same clock tick, or maybe destroying the integrity of the latched data. This is
called a hold violation because the previous data is not held long enough at the
destination flip-flop to be properly clocked through. This situation is also known as
112
Fig. 5.16 Clock jitter

example
Ideal clock
edges
actual clock
edges
jitter
double-clocking (because two FFs could capture with the same clock tick) or
race-through and is common in shift registers.
To describe this situation, consider the example in Fig. 5.15c. At time tmin
i , the
clock edge triggers flip-flops ffi, and the signals propagates through the flip-flop
min
and the combinational logic tmin
prop + tcomb (we use the shortest propagation delay,
because we are considering the possibility of data racing). The input signal at ffj
has to remain stable for tmax
hold after the clock edge of the same clock cycle arrives
(tmax
). We are using max value since we are considering the worst case. To
j
summarize the constraint to avoid race condition (hold violation) is:
max
min
min
tjmax thold
\timin tprop
tcomb
5:5
This can be rewritten as a constraint for the skew:

max
min
min
skewi;j [ thold
tprop
tcomb
5:6
Observe that a hold violation is more serious than a setup violation because it
cannot be fixed by increasing the clock period.
In the previous analysis we consider several launching and capturing flip-flops.
If the entire FFs have the same characteristics we do not need to consider maximum or minimum setup and hold time in equations 5.3, 5.4, 5.5 and 5.6.
5.3.2 Clock Jitter

The clock edges at a flip-flop may sometimes arrive earlier and sometimes later
with respect to an ideal reference clock, depending on the operating condition of
the circuit. Such temporal variation in the clock period is referred to as clock jitter.
The concept of absolute jitter refers to the worst case deviation (absolute value)
of the arrival time with respect to an ideal reference clock edge (Fig. 5.16).
The clock period may be shortened or lengthened by the clock jitter. For timing
purposes the worst case is considered, then jitter impacts negatively in the maximum frequency of operation.
113
There are several sources of clock jitter. The analog component of the clock
generation circuitry and the clock buffer tree in the distribution network are both
significant contributors to clock jitter. There are also environmental variations that
cause jitter, such as power supply noise, temperature gradients, coupling of
adjacent signals, etc.
5.3.3 Clock Gating

Since the clock distribution network could take a significant fraction of the power
consumed by a chip and, moreover, a large amount of power could be wasted
inside computations blocks even when their output are not used (unnecessary
switching activity), a typical power saving technique applied is the clock gating,
which selectively switches off part of the clock tree.
Clock gating works by taking the enable conditions attached to registers, and
uses them to gate the clocks (Figs. 5.17a and b). The clock gating logic can be
added into a design mainly in two ways. Coding into the RTL code as enable
conditions and using automatic clock gating by synthesis tools, or manually
inserting library specific modules that implement this functionality.
Using clock gating in FPGA deserves some especial considerations. The FPGA
have several low skew and low jitter dedicated clock trees; the use of this network
is vital to distribute the clock and achieve high clock frequencies. In order to
disable safely the clock tree, special clock buffers should be used that allow
switching off the clock without glitches. Observe the potential problem to use
general logic to implement gated clock in the timing diagram of Fig. 5.17c.
Todays FPGA have several clock trees and also clock buffers to disable part of
it safely at different granularities (more or less logic elements). Moreover, the
synthesis tools allow applying this technique automatically.
5.3.4 Clock Managers

A clock manager is a general name to describe a component that can manipulate
some characteristics of the input clock. The most typical actions performed by a
clock manager are:
Delay Locked Loop (DLL), synchronizes the input clock signal with the internal
clock (clock deskew, Fig. 5.18).
Frequency Synthesis (FS): Multiply and divide an incoming clock.
Phase Shifter (PS): a phase shift (skew) with respect to the rising edge of the
input clock may be configured.
Recondition clock signal: Reduce jitter, duty cycle correction, etc.
114
(a)
(b)
(c)
dff
dff
IN
en
D
en
IN
OUT
clk
OUT
en
clk enable
clk enable
clk
clk
en
OUT
delay
...
delay
delay
clk_in
delay
Fig. 5.17 Clock gating. a Register with enable. b Register with clock gating. c Clock gating risk
clock
network
clk_out
...
phase
controller
clk_fb
Fig. 5.18 A typical delay locked loop used to synchronize the internal clock
Todays FPGA have several of those components, for example the Xilinx
Digital Clock Managers (DCM) and the Altera Phase-Locked Loops (PLL). One of
the main actions performed by those components is to act as a Delay Locked Loop
(DLL).
In order to illustrate the necessity of a DLL, consider the clock tree described in
Sect. 5.3.1, Fig. 5.14, based on the cells of Table 5.1, that is, 64,000 FF that uses
three levels of buffers. The interconnection load is neglected (an unreal case).
Assume, additionally, that we decide to connect 40 buffers to the first buffer, 40
buffers at the output of each buffer of the second level, and finally 40 FFs at the
output of the last 1600 buffers. We will calculate the transition time of a rising
edge at input clk using the concepts of Sect. 5.1.2.4. (Tlh = t_intlh ? t_extlh
capacity). In this scenario the signal ck1 will see the rising edge 0.44 ns later
(0.056 ns ? 2.399 ns/pf (40 0.004pf ? interconnection)). The ck2 rising edge
will be 0.44 ns after the ck1, and finally the ck3 will be propagated 0.344 ns after
the ck1. That is the clk signal will be at the FF 1.22 ns later (remember that for
simplicity we do not consider the interconnection load). If we want to work with a
clock frequency of 400 MHz (2.5 ns period) the clock will arrive half a cycle later.
(a)
115
(b)
System 1
100 MHZ
System 2
75 MHz
asynch
input
System 1
D
synch
input
75 MHz
clk1
clk 2
clk
clk1
clk 2
System 2
Fig. 5.19 Interfacing different clock domains using synchronization chain
5.3.4.1 Delay-Locked Loop (DLL)

A delay-locked loop (DLL) is conceptually similar to a phase-locked loop (PLL)
but using a different operating principle. The main objective is to maintain the
same phase of the input clock inside a device.
The main component of a DLL is a delay chain composed of many delay
components. The input of the chain, and thus of the DLL, is connected to the input
clock (clk_in). A multiplexer is connected to each stage of the delay chain; the
selector of this multiplexer is a phase controller that compares the clock feedback
(clk_fb) from the clock network and the input clock.
This circuit, after several clock cycles, ensures that the input clock rising edge
is in phase with the clock feedback rising edge (offset = 3608).
5.3.5 Interfacing Different Clock Domains

Several times, a digital design needs to interface two different clock domains. This
interfacing is difficult in the sense that design becomes asynchronous at the
interface boundary, which could fail in setup and hold violations with the
consequent metastability (see Sect. 5.2.3). Hence, particular design and interfacing
techniques are necessary.
Two systems become asynchronous to each other when they operate at two
different frequencies or when they operate at the same frequency, but using
different and independent clock sources (Fig. 5.19a). Observe that if we use the
same clock source and a multiple of the same clock (for example, divided or
multiplied by 2) or a phase shifting of the clock (1808) they do not necessarily
become asynchronous.
Synchronization failure is lethal and difficult to debug, so it is important to take
care about this issue. If we have two systems, asynchronous to each other, and we
need to transfer data between them, several methods can be considered:
116

clk period (T)
sig1[0]
sig0[0]
D
Q
clk
sig0[0]
sig0[1]
sig1[1]
D
sig0[1]
sig1[0]
sig1[1]
clk
t prop
Fig. 5.20 Synchronization problems with buses
Using synchronizer.
Handshake signaling method.
Asynchronous FIFO.
Open loop communication.
5.3.5.1 Using Synchronizer

It is the simplest method, useful only for very low speed communication. As
described in Sect. 5.2.3, a simple method to mitigate metastability is the use a
cascade of flip-flops. Fig. 5.19b shows the simple solution using a synchronization
chain (synchronizer).
This method could lead to errors when it is used to synchronize a bus. As a
simple example consider Fig. 5.20. If the input bus changes at the capturing edge
of the clock, different paths and different flip-flops could react differently, making
a misalignment of the data in the bus. The answer to this problem is the use of a
handshaking protocol.
Comments 5.3
A related synchronization pitfall could occur if a single bit is synchronized in
different places. The same problem described for buses in Fig. 5.20 could appear.
5.3.5.2 Handshake Signaling

In this method system 1 sends data to system 2 based on the handshake signals req
(request) and ack (acknowledge). A simple handshaking protocol (known as
4-phase) will work as follow:
Sender outputs data and asserts req.

Receiver captures data and asserts ack.
Sender, after ack is detected, deasserts req.
Receiver sees req deasserted, deasserts ack when ready to continue.
117
clk1
clk2
new_value
data_s
req
ack
data_r
D Q
en
new_value
data_r
data_s
D Q
en
req
sending
hand
shaking
FSM
D Q
D Q
clk 2
Q D
Q D
receiving
hand
shaking
FSM
ack
clk2
clk 1
clk1
System 1
System 2
Fig. 5.21 Handshaking protocol using a two stage synchronizer
This method is straightforward, but the metastability problems could be present. In

fact, when system 2 samples system 1s req and system 1 samples system 2s ack
line, they use their internal clock, so setup and hold time violations could arise. To
avoid this, we can use double or triple stage synchronizers, which increase the
MTBF and thus reduce the metastability to a good extent. Figure 5.21 shows a
handshaking protocol using a two stage synchronizer in req and ack lines.
As shown in the example, the use of double or triple stage synchronizing
reduces significantly the transfer rate, due to the fact that a lot of clock cycles are
wasted for handshaking.
However, a simpler handshaking is possible (2-phase or edge based). In this, the
sender outputs data and changes the state of req; it will not change the state of req
again until after the state of ack changes. The receiver latches data; once the
receiver is ready for more, it changes the state of ack. This method (2-phase)
requires one bit of state to be kept on each side of the transaction to know the ack
state. Additionally, a reliable reset is necessary to start the synchronization.
Handshaking works great, but reduces the bandwidth at the clock crossing
interface because many cycles of handshaking are wasted. The high bandwidth
solution that maintains reliable communication is the use of asyncronous FIFOs.
5.3.5.3 Asynchronous FIFO

An asynchronous FIFO (First In First Out) has two interfaces, one for writing the
data into the FIFO and the other for reading the data out (Fig. 5.22a). Ideal dual
118
(a)
data_in
data_out
we
full
empty
read
clk 1
clk 2
(b)
WR_ ptr
data_in
FIFO
Write
logic
we
full
WR_data
data_out
FIFO
Read
logic
RD_ptr
we
clk1
clk2
RD_data
clk1
Port 1
Port 2
empty
read
clk2
Dual Port RAM
Fig. 5.22 Asynchronous FIFO. a Abstract FIFO design. b Detailed view
port FIFOs write with one clock, and read with another. The FIFO storage provides
buffering to help rate match between different frequencies. Flow control is needed
in case the FIFO gets totally full or totally empty. These signals are generated with
respect to the corresponding clock. The full signal is used by system 1 (when the
FIFO is full, we do not want system 1 to write data because this data will be lost or
will overwrite an existing data), so it will be driven by the write clock. Similarly,
the empty signal will be driven by the read clock.
FIFOs of any significant size are implemented using an on-chip dual port RAM
(it has two independent ports). The FIFO is managed as a circular buffer using
pointers. A write pointer to determine the write address and a read pointer to
determine the read address are used (Fig. 5.22b). To generate full/empty conditions, the write logic needs to see the read pointer and the read logic needs to see
the write pointer. That leads to more synchronization problems (in certain cases,
they can produce metastability) that are solved using synchronizers and Gray
encoding.
Comments 5.4
1. Asynchronous FIFOs are used at places where the performance matters, that is,
when one does not want to waste clock cycles in handshake signals.
2. Most of todays FPGAs vendors offer blocks of on-chip RAMs that can also be
configured as asynchronous FIFOs.
3. A FIFO is the hardware implementation of a data stream used in some computation models.
119
clk1
data_s
a_value
new_value
req
D Q
en
data_r
System 2
data_s
D Q
en
D Q
D Q
receiving
hand
shaking
FSM
sending
hand
shaking
FSM
clk2
clk 1
System 1
Assuming:
clk 1@clk 2 5%
Fig. 5.23 Open loop communication (mesosynchronous)
5.3.5.4 Open Loop Communication

When two systems of bounded frequency need to communicate, open loop
synchronization circuits can be used (also known as mesosynchronous designs). In
this approach no ack signal is used)
The benefits of mesosynchronous designs are in less synchronization based
circuitry and are of a lower latency in contrast to a 4-phase handshaking. The main
drawback is that this method only works at certain frequency ratios
(clk1 % clk2 5%). Figure 5.23 shows this synchronization method assuming
that the sender holds two cycles of the req signal.
5.4 Power Consumption

Power dissipation is one of the main concerns in todays digital design. For
portable devices the battery life is essential, the amount of current and energy
available in a battery is nearly constant and the power dissipation of a circuit or
system defines the battery life. On the other hand, the heat generated is proportional to the power dissipated by the chip or by the system; an excessive heat
dissipation may increase the operating temperature and thus degrades the speed
(see Sect. 5.1.2.5, derating factor) and causes circuitry malfunctions. The necessary cooling systems (fans, heatsinks, etc.) when excessive power is used, increase
the total system cost. Additionally, the life of the circuitry is typically shortened
when working at higher temperatures (electromigration and others).
120
5.4.1 Sources of Power Consumption

CMOS gates are considered very power efficient because they dissipate nearly zero
power when idle. As the CMOS technology moved below sub-micron size this
assumption became relative since the static power is more important. The power
dissipation in CMOS circuits occurs because of two components:
Static power dissipation. Is the power consumed when the output or input are
not changing. The main contributions are the subthreshold and leakage current.
Dynamic power dissipation. It is the power consumed during state transitions.
The two main components are charging and discharging of load capacitances
and the short circuit dissipation.
Thus the total power dissipation is the addition of the static power dissipation plus
the dynamic power dissipation.
5.4.1.1 Static Power Consumption

In the past, the subthreshold conduction of transistors has been very small, but as
transistors have been scaled down, leakages from all sources have increased.
Historically, CMOS designs operated at supply voltages much larger than their
threshold voltages (Vdd 5 V, and Vth around 700 mV) and the static power consumption was negligible. As transistor size is reduced, i.e. below 90 nm, the static
current could be as high as 40% of the total power dissipation. The static power
consumption reduction is one of the main concerns in todays technological
development of new CMOS processes.
5.4.1.2 Dynamic Power Consumption

CMOS circuits dissipate power mainly by charging the different load capacitances
(gates, wire capacitance, etc.) whenever they are switching (Fig. 5.24). In one
complete cycle of CMOS logic, current flows from power supply Vdd to the load
capacitance (0 to 1 transition) to charge it and then flows from the charged load
capacitance to ground during discharge (1 to 0 transition). Therefore in one
complete charge/discharge cycle, a total load Q = C Vdd is thus transferred from
Vdd to ground. If we multiply by the switching frequency to get the current, and
multiply again to supply voltage we obtain the power dissipated for a gate as:
P = f C V2dd Some authors recognize the transition power (the power per
transition) as P = 1/2 f C V2dd (half power charging, half the power discharging). Since most gates do not operate/switch at the clock frequency, they are
often accompanied by a factor a, called the activity factor (also switching activity).
Now, the dynamic power dissipation may be re-written as P = a f C V2dd.
(a)
121
(b)
Vdd
Vdd
Q 01= CL . Vdd
pMOS
network
i0
i1
out
out
in
gnd
nMOS
network
CL
gnd
gnd
gnd
Fig. 5.24 Dynamic power consumption. a An inverter. b A general CMOS gate charging a
capacitance
A clock in a system has an activity factor a = 1, since it rises and falls every
cycle. Most data has an activity factor lower than 0.5, i.e. switches less than one
time per clock cycle. But real systems could have internal nodes with activity
factors grater then one due to the glitches.
If correct load capacitance is calculated on each node together with its activity
factor, the dynamic power dissipation at the total system could be calculated as:
X
2
P
f ai ci Vdd
5:7
i
Another component in the dynamic component is the short circuit power dissipation. During transition, due to the rise/fall time both pMOS and nMOS transistors will be on for a small period of time in which current will find a path
directly from Vdd to gnd, hence creating a short circuit current. In power estimation tools this current also can be modeled as an extra capacitance at the output
of the gates.
5.4.1.3 Power and Energy

Power consumption is expressed in Watts and determines the design of the power
supply, voltage regulators and the dimensions of the interconnections and eventually short time cooling. Moreover, the energy consumed is expressed in Joules
and indicates the potency consumed over time, as shown in Equation 5.8.
Energy power delay joules watts seconds
5:8
The energy is associated with the battery life. Thus, less energy indicates less
power to perform a calculation at the same frequency. Energy is thus independent
122
of the clock frequency. Reducing clock speed alone will degrade performance, but
will not achieve savings in battery life (unless you change the voltage).
That is why, typically, the consumption is expressed in mW/Mhz when comparing circuits and algorithms that produce the same amount of data results per
clock cycle, and normally nJoules (nanoJoules) are used when comparing the total
consumption of different alternatives that require different numbers of clock cycles
for computing.
5.4.2 Power Reduction Techniques

The power reduction could be tackled at different abstraction levels. At higher
abstraction levels, bigger possibilities to reduce the power exist.
The static power consumption is reduced mainly at a technological level and is
beyond the scope of this book. Excellent surveys and trends are in [6] and [7].
To a designer of digital systems, Equation 5.7 gives information about the main
sources of reduction. The quadratic influence of the power supply is an interesting
source of power reduction. The lower supply voltage the device employs, the
lower power the systems will use. Todays systems are supplied at different
voltages for cores and pads in order to reduce, when possible, the power budget.
On the other hand, the operation frequency (f), the activity (a), and the
capacitance (c) are linear sources of power reduction.
At circuit level, the reduction of glitches, data reordering to reduce activity, or
the fan-out reduction are typical techniques. Parts of these ideas are present in the
automatic synthesis for low power provided by EDA tool vendors.
At algorithm level, the complexity and the regularity of algorithms, the data
representation, and other techniques, offer a big scope for optimizations. The
concept of energy and power is useful in this case (Sect. 5.4.1.3); a faster algorithm consuming the same power will consume less energy.
At system level, the partition of the system allows to apply dynamic power
management, that is, applying sleep modes (shut down) to part of or to the
complete system. These power management techniques could be applied manually
by the designer or partially automated by the synthesis tools.
The power consumption is an important issue for todays FPGA vendors.
Compared to ASICs, FPGAs are not power-efficient components because they use
a larger amount of transistors to provide programmability on the chip (typically
an order of one magnitude more power). That leads to some specific FPGAs
called low power FPGA families, optimized for power dissipation at the cost of
performance. FPGA vendors use different transistors, trading off speed versus
power in their devices, and also the availability to shut off the unused components.
They also offer power estimation tools and some optimization tools at synthesis
level.
123
5.4.3 Power Measurement and Estimation

The power measurement is based on the measure of the supply voltage and the
instant current (P = I * V). Some systems have supply sources that measure the
current and allow measurement of the power. There also are other methods to
measure the total energy consumed, but they are out of scope of the present
discussion.
For power estimation, there are specific tools either for ASIC and FPGA. Some
tools use estimated activities for different blocks; others are based on post place
and route simulation to obtain the exact activity. The power estimation tool uses
this activity information in Equation 5.7 to calculate the dynamic power consumption (more details are given in Chap. 6). The main drawback is the possibility
to simulate a real situation to obtain the real activity of a circuit.
5.5 Exercises
1. Redraw Fig. 5.5 (glitches for unbalanced paths) considering: (a) NAND gate,
(b) NOR gate, (c) XOR gate.
2. Redraw Fig. 5.6 (cascade effect of glitches paths) considering the transition
(a) from 0000 to 1111; (b) from 0100 to 1101; (c) from 0101 to 1101;
3. Supposing one has a simple ripple carry adder (Fig. 7.1) of four bits. Analyze
the glitch propagation when changing to add 0000 ? 0000 to 1111 ? 1111.
4. What is the rise time and fall time of an AND2 gate that connects at his output
four XOR2 gates? Assume a total interconnection load of 12 pf (use data of
Table 5.1).
5. How many DFF can drive a CKBUF assuming the unreal case of no interconnection capacitance? Which is the propagation delay of the rising edge?
6. Assuming the derating factors of Table 5.2, what is the delay of exercise 4 and
5 for 80C and a supply voltage of 3.0 V.
7. Determine the MTBF for K1 = 0.1 ns, K2 = 2 ns-1; with clock frequency of
100 MHz and data arrival at 1 MHz, for recovery time of 1, 5, 10 and 20 ns.
8. What is the MTBF expected for an asynchronous input that uses two synchronization flip-flops working at 100 MHz and using the data of the previous
problem? The FFs have a setup time of one ns.
9. What is the delay of the falling edge with the data used in Sect. 5.3.4. i.e.
64,000 FF that uses three levels of BUF, neglecting the interconnection load?
10. Calculate the level of clock buffers (CKBUF) necessary to control 128,000
registers (DFF) of Table 5.1. Suppose additionally that any interconnection
has a load of 0.006 pf.
11. For the previous clock tree. Determine the propagation delay from the input
clock signal to the clk input of a FF.
124
12. Draw a timing diagram of a two stage handshaking protocol. Assume that the
sending clock is faster than the receiving clock.
13. For the circuit of the figure, suppose that the FF has a propagation delay
between 0.9 and 1.2 ns, a setup time between 0.4 and 0.5 ns and a hold time
between 0.2 and 0.3 ns.
Launching
flip-flop(s)
m
Capturing
flip -flop(s)
Combinational
circuit
Q
ff i
ck i
D
ckj
Q
ffj
clk
The clock arrives to the different FF of level i with a delay between 2.1 ns and
3.3 ns, and to level j with delays between 2.5 ns and 3.9 ns. What is the
maximum combinational delay acceptable to work at 100 MHz?
14. Using the data of the previous exercise, what is the minimum combinational
delay necessary to ensure a correct functionality?
15. A system A works with a supply voltage of 1.2 V and needs 1.3 mA during
10 s to perform a computation. A second system B powered at 1.0 V consumes an average of 1.2 mA and needs 40 s to perform the same task. Which
consumes less power and energy?
16. In the shift register of the figure, assuming that all the flip-flops have a
propagation delay of 0.9 ns, a setup time of 0.3 ns and a hold time of 0.2 ns,
what is the maximum skew tolerated if the interconnection has a delay (d1 and
d2) of 0.1 ns?
D
ck i
d1
ff i
Q
ffj
ckj
d2
D
ckk
Q
ffk
clk
17. For the previous problem. What is the maximum frequency of operation?
18. The following FF (with same temporal parameters as in exercise 16) is used to
divide the clock frequency. What is the minimum delay d of the inverter and
interconnection necessary for it to work properly?
d
ffi
clk
References
125
References
1. Rabaey JM, Chandrakasan A, Nikolic B (2003) Digital integrated circuits, 2nd edn. PrenticeHall, Englewood Cliffs
2. Wakerly JF (2005) Digital design principles and practices, 4th edn. Prentice-Hall, Englewood
Cliffs. ISBN 0-13-186389-4
3. Xilinx Corp (2005) XAPP094 (v3.0) Metastable Recovery in Virtex-II Pro FPGAs, XAPP094
(v3.0). http://www.xilinx.com/support/documentation/application_notes/xapp094.pdf
4. Actel Corp (2007) Metastability characterization report for Actel antifuse FPGAs; http://
www.actel.com/documents/Antifuse_MetaReport_AN.pdf
5. Altera corp (2009) White paper: Understanding metastability in FPGAs. http://www.altera.
com/literature/wp/wp-01082-quartus-ii-metastability.pdf
6. Pedram M (1996) Tutorial and survey paperPower minimization in IC design: principles and
applications. ACM Trans Design Autom Electron Syst 1(1):356
7. Rabaey JM (1996) Low power design methodologies. Kluwer Academic Publishers, Dordrecht
Chapter 6
EDA Tools
The Electronic Design Automation (EDA), is a group of software tools for

designing electronic systems such as integrated circuit (ASICs), printed circuit
boards (PCBs), or reprogrammable hardware as FPGA, etc. The general ideas of
EDA tools and the particular for FPGA designs will be discussed in this section.
Typically these tools work in a design flow that hardware and system designers use
to design and analyze entire system behavior. This chapter explains the main
concepts related to the EDA tools and presents an example using Xilinx ISE and
Altera Quartus tools.
6.1 Design Flow in FPGA EDA Tools

The design flows are the combination of EDA tools to carry out a circuit or system
design. Current digital flows are very modular and are the consequence of the
evolution of the standalone tools for synthesis, placement, and routing.
Naturally, the tools are evolving, driven by Moores law from standalone tools
to integrated construction and analysis flows for design closure.
The FPGA design flow is a kind of simplification of the ASIC design flow and
we will concentrate on it. The FPGA vendors groups the EDA tools in design
suites such as Libero from Actel [5, 7] or Quartus II from Altera [2]. Figure 6.1
shows a typical design flow; in what follows we will describe the different main
stages.
The design flow can be used as command-line executables or scripting, or by
using the GUI (graphical user interface). Figures 6.5 and 6.14 shows the GUI
aspects for Xilinx ISE and Altera Quartus II, respectively. The graphical interface
is the preferred design entry for newcomer designers and for small projects.

127
128
6 EDA Tools
system specification
Design Verification
Design Entry
Behavioural
Simulation
RTL (VHDL, Verilog , Schematic)

Synthesis constraints
Synthesis
Functional
Simulation
Netlist (edif, propietary )

Implementation
constraints
(pads and Timming )
Implementation
Post Implem.
Simulation
Timing
Simulation
Mapping
Place & Route
Back
annotation
Netlist (propietary)
Generate
program. File
In-circuit
Testing
Program.
Tool
Bitstream
Fig. 6.1 A typical design flow for FPGA design
6.1.1 Design Entry

The design entry is the way you describe the hardware that you want to implement
into an FPGA (or in an electronic device) based on a system specification. There
are different methods:
Using
Using
Using
Using
a Hardware Description Language (HDL).

schematics.
Intellectual Property (IP) blocks.
Electronic System Level (ESL) languages.
Todays EDA tools allow the mixing of different design entries in a hierarchical
structure. It is common to see a schematic top level, with several predesigned IPs,
some specific components developed in VHDL and or Verilog and subsystems
designed in an ESL language.
6.1.1.1 HDL Design Entry

There are now two industry standard hardware description languages (HDL),
Verilog and VHDL. Some vendors used to have their own proprietary HDL languages but these are displaced by the use of the standard languages.
129
HDL is the de facto design entry in most digital designs. In ASIC designs
Verilog is much more used but, FPGA designers use either VHDL and/or Verilog.
All FPGA tools support both languages and even projects mixing the use of booth
languages.
The Verilog hardware description language has been used far longer than
VHDL and has been used extensively since it was launched by Gateway Design
Automation in 1983. Cadence bought Gateway and opened Verilog to the public
domain in 1990. It became IEEE standard 1364 in 1995. It was updated in 2001
(IEEE 1364-2001) and had a minor update in 2005 (IEEE 1364-2001).
On the other hand, VHDL was originally developed at the request of the U.S
Department of Defense in order to document the behavior of existing and future
ASICs. VHDL stand for VHSIC-HDL (Very high speed integrated circuit Hardware Description Language) became IEEE standard 1076 in 1987. It was updated
in 1993 (IEEE standard 1076-1993), and the last update in 2008 (IEEE 1076-2008,
published in January 2009).
Xilinx ISE and Altera Quartus have text editors with syntax highlighting and
language templates for VHDL and Verilog to help in the edition.
6.1.1.2 Schematic Design Entry

With schematic capture or schematic entry you draw on your computer using a
schematic capture tool. The main advantage in the use of schematics is that it
documents the design in an easily readable format. However big designs rapidly
become difficult to maintain and the file formats are incompatible between vendors. A HDL code is easier to be parameterized and regular structures are easily
replicated. ASIC and FPGA users rarely use schematics.
6.1.1.3 Intellectual Property (IP) Blocks

In order to make simpler the design of complex systems, there exist libraries of
predefined complex functions and circuits that have been tested and optimized to
speed up the design process. These predefined circuits are commonly called IP
cores (or IP blocks) and are available from FPGA vendors and third-party IP
suppliers. The IP Cores can be distributed as a compiled netlist or as an HDL
source code. Moreover, the FPGA vendors have tools to generate most typical IP
cores (Xilinx CORE generator in Xilinx, Altera megafunctions). The simplest IP
cores are typically free, but the complex one rarely is without charge, and typically
released under proprietary licenses.
In the FPGA arena a related concept is the idea of hard IP core and soft IP
core. In todays FPGA several heterogeneous block are built in the FGPA such as
multipliers, blocks of memories, clock managers, transceivers, memory controllers, etc. These components are called hard IP cores to differentiate from the
ones implemented using general purpose logic (soft IP blocks).
130
6 EDA Tools
6.1.1.4 Electronic System Level (ESL) Languages

As FPGA applications have grown in complexity and FPGA chips have become
more complex, there is an increasing need for more efficient less time-consuming
development methodologies. In order to address this requirement, many chip and
tool providers have developed high-level development tools. In this environment, high-level means that the input to the tool is a target independent
language as C, C++or Java instead an HDL language.
To refer to these tools and methodologies, several names are used such as Highlevel synthesis (HLS), High Level Languages (HLL), C synthesis, C to hardware,
electronic system level (ESL) synthesis, algorithmic synthesis, or behavioral
synthesis and others.
The high-level synthesis seems to be the future design entry for numerous fields
of applications. There are several successful products from the big EDA vendors
and from small companies fighting in this field. Most of those products generate, as
a result, an HDL (Verilog or VHDL) description of the circuit in order to use it in a
traditional design flow.
Related to this high-level concept and in the DSP (Digital Signal Processing)
field, the design productivity of DSP system is increased using MatLab/Simulink
design entry. The big EDA companies have tools for DSP design based on
MatLab/Simulink, even booth leaders in the FPGA field: Altera (DSP builder) and
Xilinx (System Generator).
6.1.2 Synthesis
The synthesis (or logic synthesis) is the process by which an abstract form (an
HDL description) of the circuit behavior is transformed into a design implementation in terms of logic gates and interconnections. The output is typically a netlist
and various reports. In this context, a netlist describes the connectivity of the
electronic design, using instances, nets and, perhaps, some attributes.
There are several proprietary netlist formats but, most synthesizers can generate
EDIF (Electronic Design Interchange Format) that is a vendor-neutral format to
store electronic netlist.
FPGA vendors have their own synthesizers (Xilinx XST, Altera Quartus II Integrated
Synthesis) but, main EDA vendors have syntheses for FPGA (Precision by Mentor
Graphics and Synplify by Synplicity) that can be integrated in the FPGA EDA tools.
6.1.2.1 Synthesis Optimizations

The synthesis process performs several target independent optimizations (logic
simplification, state assignment, etc.) but, also the synthesis tools take into account
the target technologies and make target dependent optimizations.
131
The optimizations available depend on the synthesizer but, most typical optimizations are present in all of them. Typical optimizations are for the area
reduction, the speed optimization, the low power consumption, and the target
frequency of the whole design.
Nevertheless, much more details can be controlled, such as:
Hierarchy Preservation: control if the synthesizer flattens the design to get better
results by optimizing entities and module boundaries or maintaining the hierarchy during Synthesis.
Add I/O buffers: enables or disables the automatic input/output buffer insertion:
this option is useful to synthesize a part of a design to be instantiated, later on.
FSM Encoding: selects the Finite State Machine (FSM) coding technique.
Automatic selection, one-hot, gray, Johnson, user defined, etc.
Use of embedded components: use embedded memory or multipliers blocks or
use general purpose LUTs to implement these functionalities.
Maximum fan-out: limits the fan-out (maximum number of connections) of nets
or signals.
Register duplication: allows or limit the register duplication to reduce fan-out
and to improve timing.
Retiming or Register Balancing: automatically moves registers across combinatorial gates or LUTs to improve timing while maintaining the original behavior.
The complete description of synthesis optimization can be found at the
synthesis tool documentation (for example for FPGA [3, 4, 6, 8]).
The synthesis behavior and optimization can be controlled using synthesis
constraints. The constraints could be introduced using the integrated environment,
a constraint file or embedding in the HDL code.
6.1.2.2 Synthesis Constraints

The synthesis optimizations are controlled globally (for the complete design) or
locally for each part of the HDL code. The global optimization can be introduced
in the integrated environment (then translated as switches in command line or to an
additional file), or in a specific constrain file. The local optimization can be
specified in a constraint file or embedded in the HDL code.
The synthesis constraint file is, typically, a plain text file having different
syntax, depending on the tool. Xilinx use the xcf (Xilinx Constraint File) [8],
Altera for timing use SDC files (TimeQuest Timing Constrains) and QSF (Quartus
Settings File) for I/O and others [2, 3], and simplify uses SDC (Synopsys Design
Compiler) constraint files [6].
The main advantage of using constraint files (generated by a graphical interface
or entered in a text editor) is that it makes your source code more portable and
separates the functionality described in the HDL of the synthesis detail.
The HDLs (VHDL and Verilog) allow you to embed constraints in the design
code. This method is recommended for constraints that define the desired result of
132
6 EDA Tools
the HDL (encoding of FSM, type of memories or multipliers) or specific optimization (duplicate registers, registers balancing, maximum fan-out, etc.). In
VHDL this is done using attributes defined in the declaration. The following
lines of code define the maximum fan-out of signal a_signal to 20 connections
in XST [8].
6.1.2.3 Synthesis Reports

The synthesis reports show the results of the netlist generation synthesis process.
The synthesizer gives, typically, a text report, where you can see a summary of
your synthesis options, and a summary and analysis of the netlist generation. Some
tools generate an HTML, XML or other proprietary formats which are easier to
navigate into the information.
This report is important in order to see if the hardware generated by the synthesizer agrees with the described HDL. The main parts in a synthesis report are:
Synthesis Options: a summary of selected optimizations used. Important to
check if differences in results appear.
HDL Compilation and Analysis: syntax errors and hierarchy are analyzed and
information reported.
HDL Synthesis: is the main part, where the inferred hardware is reported. The
tool informs what is generated in each part of the code.
Optimization reports: the advanced optimization and low level optimization are
also informed.
Final resource and timing: a summary of the device resource utilization and a
preliminary timing analysis with the worst path is informed. Take into account
that this information is previous to the implementation (placement and routing)
and the interconnection delay is only estimated.
Additionally, most synthesis tools can generate graphical information (a
schematic) of the resulted synthesis. Typically, you can see graphically the RTL
view or a technological view (using the target low level components).
6.1.3 Implementation (Mapping, Placement and Routing)

The implementation step is a vendor specific tool that mainly places and routes the
design in the target device. In Altera Quartus II this tool is known as a fitter [2,
3], meanwhile, in Xilinx ISE [7, 8] is composed of three processes (translate, map,
and place & route).
133
The inputs to the implementation are the netlist(s) generated in synthesis and
the design implementation constraints. The output is a proprietary placed and
routed netlist (ncd file in Xilinx, an internal database representation in Altera) and
several reports summarizing the results. Typically the design implementation uses
timing and area constraints; a general description is in Sect. 6.2.
6.1.3.1 Implementation Reports

The implementation tools generate different information about that process.
Typically, a text report includes detailed information about the used resources, the
clock distribution (including skew), the final timing obtained with respect to the
constraints and other information.
Nevertheless, more information can be obtained of the implementation, such as
internal delay of the interconnections, a description of the used pads, a graphical
view of the placed and routed design in the target device, a simulatable HDL file
including timing information, and a power consumption estimation.
6.1.4 Programming File Generation and Programming

The programming file generation creates a bitstream file (.bit in Xilinx,.rbf in Altera)
that can be downloaded to the device or can be recorded in an external EPROM. This
tool in Altera Quartus II is call assembler, meanwhile, in Xilinx ISE Bitgen.
The generated bitstream can be converted in a standard format for EPROM
programming or directly downloaded to an FPGA device using a JTAG connection
(iMPACT in Xilinx, Programmer in Altera).
6.2 Implementation Constraints

The complexity of todays FPGA designs and the demand for higher performance
makes necessary the use of complex timing and placement constraints to meet the
performance requirements. Implementation constraints are instructions given to
the FPGA implementation tools to direct the mapping, placement, routing, timing
or other guidelines for the implementation tools.
Implementation constraints are placed in constraint file(s) but, may exist in the
HDL code, or in a synthesis constraints file and propagated for implementation.
Xilinx uses the User Constraint File (.ucf), meanwhile Altera uses the Quartus II
Settings File (.qsf) or, in the case of timing constraints, the Synopsys Design
Constraints file (.sdc). The constraint files are plain text but, there are several
graphical tools that help in the edition of such constraints avoiding the necessity to
know the exact syntax.
134
6 EDA Tools
input speed
clk speed
X
D
cki
clk
Q
ff i
clk
output speed
d2
d1
Combinational
path
Logic
Location
D
ckk
Comb
path
Z
ffk
Pin
Location
d3
Combinational
path
pin to pin speed
Fig. 6.2 What is possible to be constrained
Figure 6.2 shows the basic constraints of a digital synchronous design:
Internal clock speed for one or several clocks.

I/O speed.
Pin to Pin timing.
Pin Locations and Logic Locations (floor-planning).
The first three are timing constraints, meanwhile, the last are area constraints.
6.2.1 Timing Constrains

As suggested previously (Fig. 6.2) the timing constrain includes the clock definition, input and output timing requirements and the combinatorial path requirements. Creating global constraints for a design is the easiest way to provide
coverage of the constrainable connections in a design, and to guide the tools to
meet timing requirements for all paths. For example given a constraint of frequency of 100 MHz, we are constraining each combinational path in the design.
Nevertheless sometimes the designer needs to relax some global constraint and
inform the implementation tools that some path can take more time. The typical
cases are:
False Paths: if there is a paths that is static in nature, or is not of much significance for timing.
Multi-cycle paths: paths between registers (combinational path) that intentionally take more than one clock cycle to become stable.
Fast or Slow path: combinational path that can work at slower speed than the
global constrain.
In these cases you can use directives to eliminate the paths from timing consideration in implementation.
6.2 Implementation Constraints
135
6.2.2 Placement and Other Constrains

The placement constraints instruct the implementation tools, where to locate the
logic element into the FPGA (I/O pads, Flip-flops, ROMs, RAMs, LUTs, etc.).
Since every component in the design carries a unique name, you can use these
name to assign to a region in the FPGA where put this components. A representative placement constraint, always used in a design, is the position of the input/
output pins. Nevertheless, there are others commonly used constraints:
Relative placement constraints: allow to place logic blocks relative to each other
to increase speed and use die resources efficiently.
Routing constraints: mechanism of locking the routing in order to maintain
timing.
Input/output characteristics: specifies the i/o standard, the use of internal pullups or pull-downs, slew rate, asynchronous delay, etc.
Dedicated block configuration: how the multipliers are configured, digital clock
managers, memory blocks, SERDES, etc.
Mapping Directives: allow eliminating the simplification of internal nets for
observability, or how the logic is grouped.
Maximun Skew: allow to control the maximum skew in a line (Sect. 5.3.1).
Derating Factors: specifies different supply voltage and temperature. Used to derate the timing factors (Sect. 5.1.2.5).
As previously mentioned, implementation constraints (timing and placement) are
placed in separated constraint file(s) but, may be directly written in the HDL code.
6.3 System Verification

Figure 6.3 shows the typical system verification flow for FPGA. The logic simulation is the most used technique in early stage of development. Nevertheless, the
reprogrammable nature of FPGA gives other alternatives as in-circuit simulation,
testing and debugging.
6.3.1 Simulation
Logic simulation is the primary tool used for verifying the logical correctness of a
hardware design. In many cases, logic simulation is the first activity performed in
the process of design. There are different simulation points were you can simulate
your design, the three more relevant are:
RTL-level (behavioral) simulation. No timing information is used.
Post-Synthesis simulation. In order to verify synthesis result.
136
6 EDA Tools
Design Verification
HDL Desing
HDL RTL
(behavioual)
simulation
Tesbench
Stimulus
Synthesis
Implementation
(Place & Route)
Back
anotation
Generate
program. File
post-synthesis
(gate level)
Simulation
Timming
Simulation
Vendor
Libraries
Timing
Libraries
In-circuit
Simulation
Bitstream
Program.
Tool
In-circuit
Testing
Fig. 6.3 The design verification flow in FPGAs
Post implementation (post place & route) simulation. Also known as timing
simulation because it includes blocks and nets delays.
The behavioral (RTL-level) simulation enables you to verify or simulate a
description at the system level. This first pass simulation is typically performed to
verify code syntax, and to confirm that the code is functioning as intended. At this
step, no timing information is available, and simulation is performed in unit-delay
mode to avoid the possibility of a race condition. The RTL simulation is not
architecture-specific, but can contain instantiations of architecture-specific components, but in this case additional libraries are necessary to simulate.
Post-synthesis simulation, allows you to verify that your design has been
synthesized correctly, and you can be aware of any differences due to the lower
level of abstraction. Most synthesis tools can write out a post-synthesis HDL
netlist in order to use a simulator. If the synthesizer uses architecture-specific
components, additional libraries should be provided.
The timing simulation (post implementation or post Place & Route full timing)
is performed after the circuit is implemented. The general functionality of the
design was defined at the beginning but, timing information cannot be accurately
calculated until the design has been placed and routed.
After the implementation tools, a timing simulation netlist can be created and
this process is known as back annotation. The result of a back annotation is an
HDL file describing the implemented design in terms of low level components and
additional SDF (Standard Delay Format) file with the internal delays that allows
you to see how your design behaves in the actual circuit.
Xilinx ISE has his own simulator (ISE simulator, ISim) but, can operate with
Mentor Modelsim or Questa and operate with external third-party simulation tools
6.3 System Verification
137
(Synopsys VCS-MX, Cadence NCSim, Aldec Active-HDL). Altera Quartus II uses

Mentor Modelsim and can use other third party simulators.
6.3.2 Formal Verification

The formal verification is the act of proving the correctness of a system with
respect to a certain formal specification, using formal mathematical methods.
Since hardware complexity growth continues to follow Moores Law, the verification complexity is even more challenging and is impossible to simulate all
possible states in a design. In order to implement the formal verification, Hardware
Verification Language (HVL) can be used. A HVL is a programming language
used to verify the designs of electronic circuits written in a hardware description
language (HDL). System-Verilog, OpenVera, and SystemC are the most commonly used HVLs. The formal verification is widely used by the big companies in
the ASIC world but, is relatively new in the FPGA arena. The adoption of formal
verification in FPGA flow is still poor but, increasingly important.
6.3.3 In-Circuit Co-Simulation

The idea is to simulate the system at hardware speed but, maintaining a part of the
flexibility of a traditional simulation. The simulation executes the RTL code
serially while a hardware implementation executes it fully in parallel. This leads to
differences not only in execution time but, also in debugging. In simulation, the
user can stop simulation to inspect the design state (signals and memory contents),
interact with the design, and resume simulation. Downloading the design to an
FPGA, the visibility and observability is greatly reduced.
EDA vendors offer products to simulate the complete or a part of the design in
circuits but, controlling externally the execution. In this line, Xilinx added their
simulator (Isim), the so called Hardware co-simulation, as a complementary
flow to the software-based HDL simulation. This feature allows the simulation of a
design or a sub-module of the design to be offloaded to hardware (a Xilinx FPGA
regular board). It can accelerate the simulation of a complex design and verify that
the design actually works in hardware.
6.3.4 In-Circuit Testing and Debugging

In order to test the functionality, a simple configuration of the FPGA allows testing
of the circuit. In the typical in-circuit debugging, the user employs an external
138
6 EDA Tools
logic analyzer for visibility but, can see only a limited number of signals which
they determined ahead of time.
The reprogramability of FPGA opens new ways to debug the designs. It is
possible to add an internal logic analyzer within the programmed logic. Xilinx
ChipScope Pro Analyzer and Altera SignalTap II Logic Analyzer are tools that
allow performing in-circuit verification, also known as on-chip debugging. They
use the internal RAM to store values of internal signals and communicate to the
external world using the JTAG connection.
Another intermediate solution includes inserting probes of internal nets
anywhere in the design and connecting of the selected signals to unused pins
(using Xilinx FPGA Editor, Altera Signal Probe, Actel Designer Probe Insertion).
6.3.5 Design for Test

A recurring topic in digital design is the design for test (DFT). The DFT (or also
Design for Testability) is the general name for design techniques that add certain
testability features to a hardware design.
In the ASIC world, tests are applied at several stages in the manufacturing flow.
The tests usually are accomplished by test programs that execute in Automatic
Test Equipment (ATE). For test program generation, several automatic algorithms
are used as the Stuck-at fault model and other algorithmic methods.
One of the main issues in a test is to gain control (controllability) and observe
(observability) internal nodes in order to check functionality and this leads to the
concept of scan chain.
In scan chain, registers (flip-flops or latches) in the design are connected in a
shift register chain in order to set the vales or read the values when a test mode is
selected. Figure 6.4 shows the addition of a multiplexer to a simple register to
support the scan chain mode. Observe that only using four signals (clk, Sin, Sout,
test) and an option reset it is possible to gain access to any register inside a device.
The FPGA are pretested ASIC circuits that have their own circuitry to test the
manufacturing procedure. Nevertheless, some techniques of DFT are useful for
FPGA design debugging and testing. The main useful characteristic is the access to
the internal state of the FPGA using the internal scan chain through the JTAG port.
JTAG (Joint Test Action Group) is the name for standard test access port and
boundary-scan architecture. It was initially devised for testing printed circuit boards
using boundary but, today it is also widely used for integrated circuit debug ports.
The FPGA devices allow the accessing of the internal state (read-back) of any
internal register, even the configuration ones using the JTAG port. The possibility
to read and write the internal state of the device allows several standard and ad-hoc
techniques for testing purposes. The access to the entire registers content allows,
for example, the reconfiguration of the device, the partial reconfiguration, read
internal states for in-circuit debugging (Sect. 6.3.4), communicate internal values
to implement an internal logic analyzer, and several other techniques.
6.4 Timing Analysis
139
IN
D
dff
IN
SIN
test
OUT
IN
OUT
IN
OUT
IN
...
Q
IN
OUT
...
...
OUT
...
OUT
clk
reset
IN
D
IN
IN
OUT
IN
IN
OUT
...
SIN (Scan Input )
IN
OUT
OUT
IN
OUT
OUT
OUT
SOUT (Scan output )
Fig. 6.4 Scan chain register and organization of a scan chain
6.4 Timing Analysis

Static timing analysis (STA or simply timing analysis) is the method of computing
the expected timing of a digital circuit without requiring simulation. The word
static refers to the fact that this timing analysis is carried out in an input independent manner. The objective is to find the worst case delay of the circuit over all
possible input combinations.
The computational task is to review the interconnection graph that represents
the final netlist and determine the worst case delay. The methods used are specific
optimization of typical graph algorithm such as a depth-first search. Modern static
timing analyzers bear similarities with project management models such as PERT
(Program Evaluation and Review Technique) or CPM (Critical Path Method).
There are some common terms used in timing analysis:
Critical path: is the path with the maximum delay between an input and an
output (or two synchronous points).
Arrival time: is the time elapsed for a signal to arrive at a certain point.
Required time: is the latest time at which a signal can arrive without making
the clock cycle longer than desired (or a malfunction).
Slack: is the difference between the required time and the arrival time. A negative
slack implies that a path is too slow. A positive slack at a node implies that the arrival
time at that node may be increased without affecting the overall delay of the circuit.
In a synchronous digital system, data is supposed to move on each tick of the
clock signal (Sect. 5.2). In this context, only two kinds of timing errors are possible (see Sect. 5.2.2):
Hold time violation: when an input signal changes too fast, after the clocks
active transition (race conditions).
Setup time violation: when a signal arrives too late to a synchronous element
and misses the corresponding clocks edge (long path fault).
140
6 EDA Tools
Since the timing analysis is capable of verifying every path, it can detect the
problems in consequence of the clock skew (Sect. 5.3) or due to glitches
(Sect. 5.1.3).
The timing analysis can be performed interactively, asking for different paths
but, typically is used to report the slack upon the specified timing requirements
expressed in the timing constraint (Sect. 6.2.1).
The FPGA tools, after implementing the design, make a default timing analysis
to determine the design system performance. The analysis is based on the basic
types of timing paths: Clock period; input pads to first level of registers; last level
of registers to output pads, and pad to pad (in asynchronous paths).
Each of these paths goes through a sequence of routing and logic. In Xilinx ISE
the tool calls Timing Analyzer and in Altera Quartus II Altera TimeQuest.
More advanced options can be analyzed upon using the specific tool.
6.5 Power Consumption Estimation

Power dissipated in a design can be divided into static and dynamic power
(Sect. 5.3). In FPGA designs, due to the reprogramability nature, the total power is
also divided into three components for each power supply:
Total Power Device Static Design Static Dynamic Power
Where the components are:
Device Static: depends on manufacturing, process properties, applied voltage,
and temperature.
Design Static: blocks in an FPGA (I/O termination, transceivers, block RAM,
etc.) are disabled by default and enabled depending on the design requirements.
When these blocks are enabled they consume power, regardless of user design
activity.
Dynamic Power: depends on the capacitance and activity of the resources used,
and also scales with the applied voltage level.
FPGA vendors offer early power estimators spreadsheet (Altera PowerPlay
Early Power Estimator, Xilinx XPower Estimator) typically used the pre-design
and pre-implementation phases of a project. These spreadsheets are used for
architecture evaluation; device and power supply selection and helps to choose the
thermal management components which may be required for the application.
For a more accurate estimation, the power analyzer tools (Xilinx XPower,
Altera PowerPlay) perform power estimation, post implementation. They are more
precise tools since they can read from the implemented design netlist the exact
logic and routing resources used. In order to obtain good and reliable results the
activity of each node should be provided. If you do not supply the activity, the
software can predict the activity of each node but, the accuracy is degraded.
6.5 Power Consumption Estimation
141
The power analyzer takes the design netlist, the activity of the circuit, the
supply voltage and the ambient temperature and reports the consumed current
(power) and the junction temperature. Nevertheless, the junction temperature itself
depends on the ambient temperature, voltage level, and total current supplied. But
the total current supplied includes the static current as a component that depends
on temperature and voltage, so a clear circular dependency exists. The tools use a
series of iterations to converge on an approximation of the static power for given
operating conditions.
A significant test bench that models the real operation of the circuit provides the
necessary activity of each node. The simulation result of the activity is saved in a
file (SAIF or VCD file) that is later used in conjunction with the capacitance
information by the power analyzer.
The Value Change Dump (VCD) file is an ASCII file containing the value
change details for each step of the simulation and can be generated by most
simulators. The computation time of this file can be very long, and the resulting
file size is typically huge. On the other hand, the SAIF (Switching Activity
Interchange format) file contains toggle counts (number of changes) on the signals
of the design and is supported also for most simulators. The SAIF file is smaller
than the VCD file, and recommended for power analysis.
6.5.1 Reducing the Power Consumption

Using the results of a power analyzer, having a complete system-level understanding and the accurate power model will permit the designer to make the
decisions necessary to reduce the power budget (Sect. 5.3), including:
Selecting the best device.

Reducing the device operating voltage.
Optimizing the clock frequencies.
Reducing long routes in the design.
Optimizing encodings.
With regards to the EDA automatic improvements for low power, as mentioned
previously, in synthesis it is possible to have, as a target, the power reduction. In
implementation, it is possible to specify optimal routing to reduce power
consumption. In this case, it allows to specify an activity file (VCD or SAIF) to
guide place & route when it optimizes for power reduction.
6.6 Example of EDA Tool Usage

In order to review the concepts of EDA tools we will deploy a simple example
using Xilinx ISE 13.1 and Altera Quartus 11.0. The example is the combinational
142
6 EDA Tools
Fig. 6.5 A simple project in Xilinx ISE
floating point adder of Chap. 12 that we will implement in both vendor tools. The
simple example uses three VHDL files: FP_add.vhd, FP_right_shifter, and
FP_leading_zeros_and_shift.vhd.
If you want to begin with no previous experience, it is highly recommendable to
start with the tutorial included in the tools. In the case of Quartus, follow the
Quartus II Introduction for VHDL/verilog Users accessible form help menu. If
you will start using ISE we recommend using the ISE In-Depth Tutorial,
accessible using help ? Xilinx on the web ? Tutorials.
6.6.1 Simple Example Using Xilinx ISE

We create a new project (file ? new project) with name fp_add. We select a
Virtex 5, XC5VLX30 ff324 -1. For synthesis XST, simulation ISIM and preferred
language VHDL. Then we add to the project the three VHDL source files (project ? add source) (Fig. 6.5).
6.6.1.1 Design Entry and Behavioral Simulation

The design entry in this simple project is not used, since the VHDL code is
provided. Nevertheless to create a new source file you can use project ? new
source, and then choose the design entry type of source code. In the case of using
VHDL or Verilog module you have a graphical wizard to describe the input and
output of the circuit.
143
Fig. 6.6 Behavioral simulation in Xilinx ISE using ISIM simulator
Fig. 6.7 Synthesis and synthesis report
In order to make a behavioral simulation we add to the project a test bench

(fp_add_tb.vhd) selecting project ? add source. The simple test bench reads
the stimuli text file (contains two encoded operands to add and the corresponding
result) and verifies the results (Fig. 6.6).
In order to launch the simulation, select the test bench (fp_add_tb.vhd), and
double click Simulate Behavioral Model in the processes window. Before that,
ensure that you have selected the simulation view radio button, in order to be
able to see the simulation process.
6.6.1.2 Synthesis and Synthesis Report

To run the synthesis, double click the synthesisXST process in processes
view (Fig. 6.7). The global options of the synthesis are available using the right
button at the option process option. The more specific constrain can be either
embedded in the HDL code or in the Xilinx constraint file (.xcf). The xcf is a
plain text file that needs to be linked in the process option. The details are
available at [8].
144
6 EDA Tools
Fig. 6.8 Assigning timing constraints, and the generated ucf text file
After the synthesis you can review the graphical representation of the synthesized circuit (View RTL schematic) or review the synthesis report (.syr file). This
report gives relevant information. The HDL synthesis part of report describes
the inferred component as a function of the HDL and the final report summarizes the resource utilization and performs a preliminary timing report.
6.6.1.3 Implementation: Constraints and Reports

In order to implement a design, we need to generate the implementation
constraints. We will firstly assign the position of the pads. We can edit manually
the ucf (user constraint file) text file or double click the I/O pin planning (plan
ahead) and use the graphical interface.
Then we can assign timing constraints using the create timing constraint
option. We will assign a period constraints (clock domain in constraint type
view) of 80 MHz (12.5 ns) and the same restriction for inputs and outputs
(inputs and outputs in constraint type view, respectively). You can check the
generated restriction editing the ucf file (Fig. 6.8).
We can assign global options to the implementation using right click over the
implementation icon in processes window and selecting process properties
We implement the design double clicking in implementation. The three step
in the Xilinx implementation flow will be executed (translate, map, place & route).
In ISE the design implementation, comprises the following steps:
Translate: merges the incoming netlists from synthesis and the imple-mentation
constraints into a proprietary Xilinx native generic database (NGD) file.
Map: fits the design into the available resources on the target device. The output
is a native circuit description (NCD) file.
Place and Route: takes a mapped NCD file, places and routes the design, and
produces another NCD file.
145
Fig. 6.9 Implemented design. Access to the reports
The result of the implementation is a proprietary placed and routed netlist (ncd
file) and several reports. There are three main reports generated: the map report
(.mrp), the place and route report (.par) and the Post-PAR static timing report
(.twr) (Fig. 6.9).
The map report shows detailed information about the used resources (LUTs, slices,
IOBs, etc.), the place and route report gives clock network report including the skew
and informs which constraints were met and which were not. Finally, Post-PAR static
timing report gives the worst case delay with respect to the specified constraints.
You can obtain additional graphical information about the place and route
results using FPGA editor and PlanAhead. Use the FPGA Editor to view the actual
design layout of the FPGA (in the Processes pane, expand Place & Route, and
double-click View/Edit Routed Design (FPGA Editor)). The PlanAhead software can be used to perform post place & route design analysis. You can observe,
graphically, the timing path onto the layout, and also perform floorplaning of the
design. In order to open PlanAhead in the Processes pane, expand Place & Route,
and double-click Analyze Timing/Floorplan Design (PlanAhead).
6.6.1.4 Post Place and Route Simulation

The post place and route (or timing) simulation is accessed by selecting simulation in the view panel and selecting post-route from the drop-down list. Then
select the test bench and double click Simulate Post-Place & Route Model. This
will execute the back annotation process (netgen). Then the simulation is
performed. Observe that the resulting simulation gives some errors. Why?
The answer is simple, the test bench generates a clock of 100 MHz (10 ns period)
146
6 EDA Tools
Fig. 6.10 Post place and route simulation
and the implementation that we perform is slower. Then, some input pattern
combination violates setup time and gives, consequently, errors. You can modify
the test bench in order to operate at a lower frequency (Fig. 6.10).
6.6.1.5 Running a Static Timing Analysis

We can review, more carefully, the timing aspects, by opening the analyze postPlace & Route Static Timing (expand Implement Design ? Place &
Route ? Generate Post-Place & Route Static Timing to access this process). By
default, this runs an analysis of the worst case delays with respect to the specified
constraints giving the three worst paths (Fig. 6.11).
In order to control more aspects of the timing analysis you can run an analysis
(Timing ? Run Analysis), you can control the type of analysis (Analyze against),
control the amount of path reported, control the timing derating factors
(Sect. 5.1.2.5), filter nets and paths in the analysis, etc.
6.6.1.6 Generating Programming File and Programming the FPGA

After implementing the design and performed the corresponding timing analysis,
you need to create configuration data. A configuration bitstream (.bit) is created for
downloading to a target device or for formatting into a PROM programming file
(Fig. 6.12).
6.6.1.7 Using Command Line Implementation

All the previous step could be executed in command line and group it in order to
create scripts. The ISE software allows you extract the command line arguments
147
Fig. 6.11 Post place and route static timing analysis
Fig. 6.12 Generating programming file and programming the FPGA
for the various steps of the implementation process (Design Utilities ? View
Command Line Log File). This allows you to verify the options being used or to
create a command batch file to replicate the design flow.
148
6 EDA Tools
Fig. 6.13 Xpower analyzer result
6.6.1.8 Estimating the Power Consumption

After place and route of the design it is possible to obtain power consumption
estimation. The accuracy of the estimation relies upon the activity given to the
xpower tool. The activity file is generated with a post place & route simulation
(Sect. 6.6.1.4) but instructing the ISim simulator to generate the SAIF (Switching
Activity Interchange Format) file. In order to do that, right click simulate post place
and route model, click on process properties then select Generate SAIF File
for Power Optimization/Estimation, it is possible to assign different file names.
In order to check the power estimation results, based on the activity computed,
you can even generate a simple text report by double clicking on Generate Text
Power Report, or by opening the interactive tool Xpower double clicking
Analyze Power Distribution (Xpower Analyzer) (Fig. 6.13). Observe that by
browsing into the details you have access to the activity of each node, the fan-out
of nets and the corresponding power.
6.6.2 Simple Example Using Altera Quartus II

Designing for Altera devices is very similar, in concept and practice, to designing
for Xilinx devices. In most cases, the same RTL code can be compiled by Alteras
149
Fig. 6.14 Quartus II project
Quartus II as was explained for Xilinx ISE (Sect. 6.6.1). Altera has an application
note Altera Design Flow for Xilinx Users [1], where this process is carefully
explained.
We can start creating a new project (file ? New Project Wizard) with name
fp_add. We add to the project the three VHDL source files (step 2 of 5).We select a
Stratix III device, EP3SL50F484C3. Leave the default synthesis and simulation
options (Fig. 6.14).
For the rest of the steps in the implementation and simulation flow, Table 6.1
summarizes the GUI (graphical user interface) names for similar task in Xilinx ISE
and Altera Quartus II.
6.7 Exercises
1. Implement the floating adder of Chap. 12 in Altera Quartus in a Stratix III
device. What are the area results? Add implementation constraints in order to
add FP numbers at 100 MHz. Constraints are met?
2. Implement the floating point multiplier of Chap. 12 in Xilinx ISE in a Virtex 5
device. What are the area results? It is possible to multiply at 100 MHz?
Remember to add implementation constraints. Multiplying at 20 MHZ what is the
expected power consumption? Use the XPower tool and the provided test bench.
150
6 EDA Tools
Table 6.1 GUI names for similar tasks in Xilinx ISE and Altera Quartus II
GUI feature
Xilinx ISE
Altera quartus II
HDL design entry
Schematic entry
IP entry
Synthesis
Synthesis constraints
Implementation
constraint
Timing constraint
wizard
Pin constrain wizard
Implementation
Static timing analysis
Generate programming
file
Power estimator
Power analysis
Simulation
Co-simulation
In-chip verification
View and editing
placement
Configure device
HDL editor
Schematic editor
CoreGen and architecture
wizard
Xilinx synthesis technology
(XST)
Third-party EDA synthesis
XCF (xilinx contraint file)
UCF (user constraint file)
HDL editor
Schematic editor
Megawizard plug-in manager
PlanAhead (PinPlanning)
Translate, map, place and
route
Xilinx timing analyzer
BitGen
Quartus II integrated
synthesis (QIS)
Third-party EDA synthesis
Same as implementation
QSF (quartus II settings file) and
SDC (synopsys design
constraints file)
Quartus II timequest timing analyzer
SDC editor
Pin planner
Quartus II integrated synthesis (QIS),
fitter
Timequest timing analyzer
Assembler
XPower estimator
XPower analyzer
ISE simulator (ISim)
Third-party simulation tools
ISim co-simulation
Chipscope pro
PlanAhead, FPGA editor
Powerplay early power estimator

Powerplay power analyzer
ModelsimAltera starter edition
Third-party simulation tools
SignalTap II logic analyzer

Chip planner
iMPACT
Programmer
Create timing constraints
3. Implement the pipelined adder of Chap. 3. Analyze the area-time-power trade off
for different logic depths in the circuit. Use, for the experiment, a Virtex 5 device
and compare the result with respect to the new generation Virtex 7. What happened in Altera, in comparing stratix III devices with respect to Stratix V?
4. Implement the adders of Chap. 7 (use the provided VHDL models). Add to the
model input and output registers in order to estimate the maximum frequency of
operation.
5. Compare the results of implementation of radix 2 k adders of Sect. 7.3 in Xilinx
devices using muxcy and a behavioural description of the multiplexer (use the
provided VHDL models of Chap. 7).
6. Implement the restoring, non-resorting, radix-2 SRT and radix 2 k SRT divider
of Chap. 9 in Xilinx and Altera devices (use the provided VHDL models of
Chap. 9).
7. Compare the results of square root methods of Chap. 10 in Altera and Xilinx
design flow.
References
151
References
1. Altera corp (2009) AN 307: Altera design flow for Xilinx users. http://www.altera.com/
literature/an/an307.pdf
2. Altera corp (2011a) Altera quartus II software environment. http://www.altera.com/
3. Altera corp (2011b) Design and synthesis. In: Quartus II integrated synthesis, quartus II
handbook version 11.0, vol 1. http://www.altera.com/
4. Mentor Graphics (2011) Mentor precision synthesis reference manual 2011a. http://
www.mentor.com/
5. Microsemi SoC Products Group (2010) Libero integrated design environment (IDE) v9.1.
http://www.actel.com/
6. Synopsys (2011) Synopsys FPGA synthesis user guide (Synplify, Synplify Pro, or Synplify
Premier). http://www.synopsys.com/
7. Xilinx inc (2011a) Xilinx ISE (Integrated Integrated software environment) design suite. http://
www.xilinx.com/
8. Xilinx inc (2011b) XST (Xilinx Synthesis Technology) user guide, UG687 (v 13.1). http://
www.xilinx.com/
Chapter 7
Adders
Addition is a primitive operation for most arithmetic functions, so that FPGA

vendors have dedicated a particular attention to the design of optimized adders. As
a consequence, in many cases the synthesis tools are able to generate fast and costeffective adders from simple VHDL expressions. Only in the case of relatively
long operands can it be worthwhile to consider more complex structures such as
carry-skip, carry-select and logarithmic adders.
Another important topic is the design of multi-operand adders. In this case, the
key concept is that of carry-save adder or, more generally, of parallel counter.
Obviously, the general design methods presented in Chap. 3 (pipelining, digitserial processing, self-timing) can be applied in order to optimize the proposed
circuits. Numerous examples of practical FPGA implementations are reported in
Sect. 7.9.
7.1 Addition of Natural Numbers

Consider two radix-B numbers
x xn1 Bn1 xn2 Bn2 x1 B x0
and
y yn1 Bn1 yn2 Bn2 y1 B y0 ;
where all digits xi and yi belong to {0, 1,, B-1}, and an input carry c0 belonging
to {0, 1}. An n-digit adder generates a radix-B number
z zn1 Bn1 zn2 Bn2 z1 B z0 ;

153
154
7 Adders
Fig. 7.1 n-digit adder
xn-1 yn-1
cn
1-digit
adder
x1
cn-1
....
y1
x0 y0
c 2 1-digit
adder
c 1 1-digit
adder
zn-1
z1
c0
z0
and an output carry cn, such that

x y c0 cn Bn z:
Observe that x ? y ? c0 B 2(Bn-1) ? 1 = 2Bn-1, so that cn belongs to {0, 1}.
The common way to implement an n-digit adder consists of connecting in series
n 1-digit adders (Fig. 7.1). For each of them
xi y i ci ci 1 B zi ;
where ci and ci
? 1
belong to {0, 1}. In other words
zi xi yi ci mod B; ci 1 bxi yi ci =Bc:

The critical path is
x0 ; y0 ; c0 ! c1 ! c2 ! ! cn1 ! zn1 ; cn ;
so that the total computation time is approximately equal to nTcarry where Tcarry is
the computation time of ci ? 1 in function of xi, yi and ci.
In order to reduce Tcarry, it is convenient to compute two binary functions
p (propagate) and g (generate) of xi and yi:
pxi ; yi 1 if xi yi B 1; pxi ; yi 0 otherwise;
gxi ; yi 1 if xi yi B; gxi ; yi 0 if xi yi B 2; otherwise; any value:
So, ci ? 1 can be expressed under the following way:
ci 1 pxi ; yi ci pxi ; yi gxi ; yi
The last relation expresses that if xi ? yi = B-1, then ci+1 is equal to ci; if
xi ? yi C B, then ci+1 = 1; if xi ? yi B B-2, then ci+1 = 0. The corresponding
implementation is shown in Fig. 7.2. It is made up of two 2-operand combinational
circuits that compute p(xi, yi) and g(xi, yi), and a multiplexer. In an n-digit adder
(Fig. 7.1), all functions p(xi, yi) and g(xi, yi) are computed in parallel, so that the
value of Tcarry is practically equal to the multiplexer delay Tmux.
7.2 Binary Adder
155
Fig. 7.2 Carry computation
xi
yi
xi
c i+1
yi
0
1
ci
7.2 Binary Adder

If B = 2, then p(xi, yi) = xi XOR yi, and g(xi, yi) can be chosen equal to xi (or yi).
A complete n-bit adder is shown in Fig. 7.3. Its computation time is equal to
Tadder n Txor n 1 Tmux maxfTmux ; Txor g;
7:1
and the delay from the input carry to the output carry is equal to
Tcarrytocarry n n Tmux :
7:2
Comment 7.1
Most FPGAs include the basic components to implement the structure of Fig. 7.3,
and the synthesis tools automatically generate this optimized adder from a simple
VHDL expression such as
7.3 Radix-2k Adder

If B = 2k, then p(xi, yi) = 1 if xi ? yi = 2k-1, that is, if the k less significant bits
of si = xi ? yi are equal to 1, and g(xi, yi) = 1 if xi ? yi C 2k, that is, if the most
significant bit of si is equal to 1. The iterative cell of a radix-2k adder is shown
in Fig. 7.4. The critical path of the part of the circuit that computes g(xi, yi) and
p(xi, yi) has been shaded. Its computation time is equal to Tadder(k) +Tand.
An m-digit radix-2k adder is equivalent to an n-bit adder with n = m k. The total
computation time is
Tadder n Tadder k Tand m 1 Tmux Thalf adder k;
7:3
and the delay from the input carry to the output carry to
Tcarrytocarry n m Tmux :
7:4
156
7 Adders
xn-1 yn-1
zn
0
1
z n-1
x1 y1
c n-1
....
x0 y0
c1
z1
c0
z0
Fig. 7.3 n-bit adder
The following VHDL model describes the basic cell of Fig. 7.4.
An alternative way of computing p = s0s1 sk-1 is
The corresponding circuit is a k-bit half adder that computes t = (s mod 2k) + 1.
The most significant bit tk of t is equal to 1 if, and only if, all the bits of (s mod 2k)
are equal to 1. As mentioned above (Comment 7.1) most FPGAs include the basic
components to implement the structure of Fig. 7.3. In the particular case where
x = 0, y = s and c0 = 1, the circuit of Fig. 7.4 is obtained. The apparently
unnecessary XOR gates are included because there is generally no direct connection between the adder inputs and the multiplexer control inputs. Actually, the
7.3 Radix-2k Adder
157
xi (k-1) yi (k-1)
si (k)
1-bit
adder
xi (1) yi (1)
xi (0) yi (0)
1-bit
adder
1-bit
adder
....
si(k-1)
pi 2-input
AND
......
si(1)
si(0)
2-input
AND
....
2-input
AND
0
1
ci+1
si(k-1)
1-bit
half adder
....
zi(k-1)
ci
si(1)
si(0)
1-bit
half adder
1-bit
half adder
zi(1)
zi(0)
......
Fig. 7.4 Radix 2k adder
Fig. 7.5 FPGA

implementation of a k-input
AND
0 si (k-1)
pi
0
1
0 si (1)
....
0
1
0 si (0)
0
1
XOR gates are LUTs whose outputs are permanently connected to the carry-logic
multiplexers.
A complete generic model base_2 k_adder.vhd is available at the Authors web
page and examples of FPGA implementations are given in Sect. 7.9.
According to (7.3), the non-constant terms of Tadder(n) are:
mTmux,
kTmux included in Tadder(k) according to (7.1),
kTmux included in Thalf-adder(k) according to (7.1).
Thus, the sum of the non-constant terms of Tadder(n) is equal to (2k ? m)Tmux.
The value of 2k ? m, with mk = n, is minimum when 2k % m, that is, when
k % (n/2)1/2. With this value of k, the sum of the non-constant terms of Tadder(n) is
equal to (8n)Tmux. Thus, the computation time is O(n) instead of O(n).
158
7 Adders
xi (k-1) yi (k-1)
ci0
1-bit
adder
xi (1) yi (1)
xi (0) yi (0)
1-bit
adder
1-bit
adder
....
zi0(k-1)
1-bit half
adder
1-bit half
adder
zi0(1)
....
zi0(0)
1-bit half
adder
1-bit half
adder
zi1(1)
zi1(0)
ci1
zi1(k-1)
0
1
c i+1
......
ci
......
zi 0
zi1
zi
Fig. 7.6 Carry select adder
Comments 7.2
1. The circuit of Fig. 7.4 is an example of carry-skip adder. For every group of
k bits, both the carry-propagate and carry-generate functions are computed.
If the carry-propagate function is equal to 1, the input carry is directly propagated to the carry output of the k-bit group, thus skipping k bits.
2. A mixed-radix numeration system could be used. Assume that n = k1 ?
k2 ? ? km; then a radix
2k1 ; 2k2 ; ; 2km
representation can be considered. The corresponding adder consists of m
blocks, similar to that of Fig. 7.3, whose sizes are k1, k2,, and km, respectively. Nevertheless, within an FPGA it is generally better to use adders that fit
within a single column. Assuming that the chosen device has r carry-logic cells
per column, a good option could be a fixed-radix adder with k B r. In order to
minimize the computation time, k must be approximately equal to (n/2)1/2, so
that n must be smaller than 2r2, which is a very large number.
7.4 Carry Select Adders
159
Fig. 7.7 Carry-select adder

(second version)
xi(k-10)
yi(k-10)
k-bit adder
ci0
zi0(k-10)
k-bit adder
ci1
zi1(k-10)
7.4 Carry Select Adders

Another way of reducing the computation time of a radix-2k adder consists in
computing, at each step, the next carry and the output digit for both values of the
input carry. The corresponding circuit is shown in Fig. 7.6.
The critical path of the part of the circuits that computes the two possible values
of the next carry and of the output digit has been shaded. Its computation time is
equal to Tadder(k) ? Tadder(2). The total computation time is (n = mk)
Tadder n Tadder k Thalf
adder 2
m 1 Tmux Tmux ;
7:5
and the delay from the input carry to the output carry to
Tcarrytocarry m k m Tmux :
7:6
The following VHDL model describes the basic cell of Fig. 7.6.
A complete generic model carry_select_adder.vhd is available at the Authors web

page and examples of FPGA implementations are given in Sect. 7.9.
The non-constant term of Tadder(n) is equal to (k ? m)Tmux. The minimum
value is obtained when k % m, that is k % (n). With this value of k, the nonconstant term of Tadder(n) is equal to (4n)Tmux. Thus, the computation time is
O(n) instead of O(n).
160
7 Adders
Comment 7.3
As before (Comments 7.2) a mixed-radix numeration system could be
considered.
As a matter of fact, the FPGA implementation of a half-adder is generally not
more cost-effective than the implementation of a full adder. So, the circuit of
Fig. 7.6 could be slightly modified: instead of computing ci0 and ci1 with a full
adder and a half adder, two independent full adders of any type can be used
(Fig. 7.7).
The following VHDL model describes the modified cell:
The computation time of the modified circuit is

Tadder n Tadder k m 1 Tmux Tmux Tadder k m Tmux :
7:7
A complete generic model carry_select_adder2.vhd is available at the Authors

web page and examples of FPGA implementations are given in Sect. 7.9.
7.5 Logarithmic Adders

Several types of adders whose computation time are proportional to the logarithm
of n have been proposed. For example: carry-lookahead adders ([1], Chap. 6), Ling
adders [2], Brent-Kung prefixed adders [3], Ladner-Fischer prefixed adders [4].
Nevertheless, their FPGA implementations are generally not as fast as what could
be theoretically expected. There are two reasons for that. On the one hand, the
special purpose carry-logic included in most FPGAs is very fast, so that ripplecarry adders are fast. Their computation time is approximately equal to a ? bn,
where a and b are very small constants: a is the delay of a LUT and b is the delay
of a multiplexer belonging to the carry logic. On the other hand, the structure of
most logarithmic adders is not so regular as the structure of ripple-carry adders, so
that they include long connections which in turn introduce long additional delays.
The practical result is that, except for very great values of n, the adders described
in Sects. 7.27.4 are faster and more cost-effective.
Obviously, any optimization method that allows the dividing up of an n-bit
adder into smaller k-bit and m-bit adders, with km = n, in such a way that
7.5 Logarithmic Adders
161
Tadder n Tadder k Tadder m;

can be recursively used in order to generate a logarithmic adder. As an example,
consider again a carry-select adder. According to (7.7)
Tadder n Tadder k m:Tmux :
Assume that k = rs. Then each k-bit adder (Fig. 7.7) can in turn be decomposed
in such a way that
Tadder k Tadder r s:Tmux ;
so that the computation time of the corresponding 2-level carry-select adder is
Tadder n Tadder r s m Tmux ;
where n = rsm. Finally, if n = n1n2 nt, then a (t-1)-level carry-select adder,
whose computation time is equal to
Tadder n1 n2 . . . nt Tadder n1
n2 . . . nt Tmux On1 n2 . . . nt ;
can be generated.
Example 7.1
The following VHDL model describes an n-bit 2-level carry-select adder with
n = n1n2n3. First, define the basic cell carry_select_step3, in which two 1-level
carry-select adders, with k = n1 and m = n2, are used:
The complete circuit is made up of n3 basic cells:
162
7 Adders
xn-1..n-s
x2s-1..s
xs-1..0 yn-1..n-s
.....
k-1
y2s-1..s
ys-1..0
.....
cout
k-1
s-digit adder
sel
cin
DFF
init.:cin
enk-1
zn
zn-1..n-s
en1
.....
z2s-1..s
en0
zs-1..0
Fig. 7.8 Long-operand adder
A complete generic model carry_select_adder3.vhd is available at the Authors

web page and examples of FPGA implementations are given in Sect. 7.9.
7.6 Long-Operand Adder

In the case of long-operand additions, the n-digit operands can be broken down
into s-digit groups and the addition computed according to the following algorithm
in which natural_addition is a procedure that computes
zi xi yi ci mod Bs and ci1 bxi yi ci =Bs c;
where xi, yi and zi are s-digit numbers, and ci and ci+1 are bits.
7.6 Long-Operand Adder
163
Algorithm 7.1: Long-operand addition
The complete circuit (Fig. 7.8, with k = n/s) is made up of an s digit adder,
connection resources (k-to-1 multiplexers) giving access to the s-digit groups,
a D-flip-flop which stores the carries (ci in Algorithm 7.1), an output register
storing z, and a control unit whose main component is a k-state counter.
The following VHDL model describes the circuit of Fig. 7.8 (B = 2).
A complete generic model long_operand_adder.vhd is available at the Authors

web page.
164
7 Adders
Fig. 7.9 Multioperand

addition
x0
x1 x2
x m-1
.....
m -1
sel
n-digit adder
n-digit register
initially: 0
en_acc
load
7.7 Multioperand Adders

Consider m natural numbers x0, x1,, xm-1. A multioperand adder computes
z x0 x1 xm1 :
7:8
7.7.1 Sequential Multioperand Adders

In order to compute (7.8), the following (obvious) algorithm can be used.
Algorithm 7.2: Basic multioperand addition
The corresponding sequential circuit (Fig. 7.9) is made up of an n-digit adder, an

n-digit register, an m-to-1 n-digit multiplexer, and a control unit whose main
component is an m-state counter.
The following VHDL model describes the circuit of Fig. 7.9 (B = 2). The nmbit vector x is the concatenation of x0, x1,, xm-1.

Fig. 7.10 Carry-save adder
165
xn-1 yn-1 cn-1

1-digit
adder
zn-1
x1 y-1
....
dn-1
c1
1-digit
adder
z1
x0 y0
c0
1-digit
adder
d1
z0
0
d0
A complete generic model multioperand_adder.vhd is available at the Authors

web page.
The computation time of the preceding m-operand n-digit sequential adder is
approximately equal to
Tsequential m; n m Tadder n:
7:9
In order to reduce the computation time, a carry-save adder can be used. The basic
component is shown in Fig. 7.10: it consists of n 1-digit adders working in parallel.
Given two n-digit numbers x and y, and an n-bit number c, it expresses the sum
(x ? y ? c) mod Bn under the form z ? d, where z is an n-digit number and d an
n-bit number. In other words, the carries are stored within the output binary vector
d instead of being propagated (stored-carry encoding). As all cells work in parallel
the computation time is independent of n.Let CSA be the function implemented by
the circuit of Fig. 7.10, that is
CSAx; y; c z; d;
where
zi xi yi ci mod B; di bxi yi ci =Bc; 8i 2 f0; 1; . . .; n 1g:
Assume that at every step of Algorithm 7.2 the value of accumulator is represented
under the form u ? v, where u is an n-digit number and v an n-bit number. Then, at
step j, the following operation must be executed:

u; v : CSA u; xj ; v :
The following formal algorithm computes z.
166
Fig. 7.11 Carry save adder
7 Adders
x0
x1 x2
xm-1
.....
m-1
sel
n-digit carry-save adder
n-bit register
initially: 0
n-digit register
initially: 0
en_acc
load
n-digit adder
Algorithm 7.3: Multioperand addition with stored-carry encoding
The sequential circuit corresponding to Algorithm 7.3 (Fig. 7.11) is made up of an

n-digit carry-save adder (Fig. 7.10), an n-digit register, an n-bit register, an m-to-1
n-digit multiplexer, a conventional n-digit adder implementing the last step of
Algorithm 7.3, and a control unit whose main component is an m-state counter.The
following VHDL model describes the circuit of Fig. 7.10 (B = 2). As before, x is
the concatenation of x0, x1,, xm-1.

Fig. 7.12 Multioperand
addition array
167
x0,n-1 x1,n-1
x0,n-2 x1,n-2
x0,n-3 x1,n-3
x0,0 x1,0
x2,n-1
x2,n-2
x2,n-3
x2,0
x3,n-1
x3,n-2
x3,n-3
x3,0
...
....
xm-1,n-1
zn-1
xm-1,n-2
zn-2
xm-1,n-3
zn-3
xm-1,0
z0
A complete generic model CSA_multioperand_adder.vhd is available at the

Authors web page.
Taking into account that the computation time of the circuit of Fig. 7.10 is
independent of the number n of digits, the computation time of the circuit of
Fig. 7.10 is approximately equal to
Tsequential
csa m; n
m Tadder 1 Tadder n:
7:10
168
7 Adders
x0,2 x1,2 x0,1 x1,1 x0,0 x1,0
x2,2 x3,2 x2,1 x3,1 x2,0 x3,0
x4,2 x5,2 x4,1 x5,1 x4,0 x5,0
z2
z1
x6,2 x7,2 x6,1 x7,1 x6,0 x7,0
z0
Fig. 7.13 Multioperand addition tree
7.7.2 Combinational Multioperand Adders

The combinational circuit that corresponds to Algorithm 7.2 is an iterative circuit
made up of m-1 2-operand n-digit adders. If every adder is a simple ripple-carry
adder, then the complete circuit is a 2-dimensional array made up of (m-1)n onedigit adders, as shown in Fig. 7.12 in which one of the critical paths has been
shaded. The corresponding computation time is equal to
Tcombinational m; n m n 2 Tadder 1:
7:11
The following VHDL model describes the circuit of Fig. 7.12 (B = 2). As before,
x is the concatenation of x0, x1,, xm-1.

x0,n-1
x1,n-1
169
x2,n-1
x0,1
x1,1
x2,1
x0,0
x1,0
x2,0
...
x3,n-1
x3,1
x3,0
x4,1
x4,0
xm-1,0
...
x4,n-1
...
xm-1,n-1
xm-1,1
...
un-1
vn-1
u1
v1
u0
v0
Fig. 7.14 Combinational carry-save adder
A complete generic model comb_multioperand_adder.vhd is available at the

Authors web page.
A most time-effective solution is a binary tree of 2-operand n-digit adders
instead of an iterative circuit. An example, with n = 3 and m = 8, is shown in
Fig. 7.13:
x0 x0;2 x0;1 x0;0 ; x1 x1;2 x1;1 x1;0 ; . . .; x7 x7;2 x7;1 x7;0 :
170
7 Adders
The depth of the tree is equal to dlog2me and its computation time (one of the
critical paths has been shaded) is approximately equal to
Taddertree m; n n log2 m 1 Tadder 1:
7:12
A complete generic model eight_operand_adder.vhd is available at the Authors

web page.
Another way to reduce the computation time, with an iterative architecture
similar to that of Fig. 7.12, is to use the carry-save principle. An m-operand
carry-save array (Algorithm 7.3) is shown in Fig. 7.14 (if B [ 2, x2 must be an
n-bit number or an initial file that computes x0 ? x1 ? 0 must be added). The
result is the sum of two n-digit numbers u and v. In order to get the actual
result, an additional 2-operand n-digit adder is necessary for computing
u ? v (last instruction of Algorithm 7.3). The corresponding computation time
is equal to
Tcombinational
csa m; n
m 2 Tadder 1 Tadder n:
7:13
The following VHDL model describes a 2-operand carry-save adder, also called 3to-2 counter (Sect. 7.7.3). It corresponds to a file of the circuit of Fig. 7.14.
The complete circuit is made up of m-2 3-to-2 counters:
171
A complete generic model comb_CSA_mutioperand_adder.vhd is available at the

Authors web page and examples of FPGA implementations are given in Sect. 7.9.
Comment 7.4
In all of the previously described multioperand adders, the operands, as well as the
result, were assumed to be n-digit numbers. If all of the operands belong to the
same range, and the result is known to be an n-digit number, whatever the value of
the operands, then the operands can be represented with (nk) digits where
k % logBm, and the previously described circuits can be pruned.
7.7.3 Parallel Counters

Given two n-digit numbers x and y, and an n-bit number c, the carry-save adder of
Fig. 7.10 allows the expression of the sum (x ? y ? c) mod Bn under the form
z ? d, where z is an n-digit numbers and d an n-bit number. In other words, it
reduces the sum of three digits x, y and c to the sum of two digits z and d. For that
reason, it is also called a 3-to-2 counter.
This 3-to-2 counter can be used as a computation resource for reducing the sum
of m digits x0, x1,, xm-1 to the sum of two digits u and v as shown in Fig. 7.14.
Thus, the circuit of Fig. 7.14 could be considered as an m-to-2 counter.
This type of construction can be generalized. As an example, consider an adder
that computes the sum of 6 bits x0, x1,, x5. The result, smaller than or equal to 6,
is a 3-bit number. Thus, this 6-operand 1-bit adder computes
x0 x1 x5 4z2 2z1 z0
7:14
and can be implemented by three 6-input Look Up Tables (LUT6) working in

parallel:
172
Fig. 7.15 6-to-3 counter
7 Adders
x0,n-1x1,n-1 x5,n-1
6-operand
1-bit adder
x0,1x1,1 x5,1
x0,0x1,0 x5,0
6-operand
1-bit adder
6-operand
1-bit adder
...
un-1
vn-1
u1
u0
v0
v1
wn-1
w0
w1
Fig. 7.16 24-to-3 counter
x0x1 x5
6-to-3
counter
x6x7 x11
x18x19 x23
x12x13 x17
6-to-3
counter
6-to-3
counter
6-to-3
counter
6-to-3
counter
6-to-3
counter
6-to-3
counter

Fig. 7.17 24-operand adder
173
x0x1 x23
24-to-3
counter
3-to-2
counter
2-operand adder
z
Then, by connecting in parallel n circuits of this type, a binary 6-to-3 counter is
obtained (Fig. 7.15):
The counter of Fig. 7.15 can in turn be used as a building block for generating
174
7 Adders
Fig. 7.18 Radix-B Bs

complement adder
xn yn
xn-10
yn-10
cin
n-digit adder
zn-10
zn
Fig. 7.19 Radix-B Bs
complement subractor
xn yn
xn-10
yn-10
(B-1)s
compl.
yn-10
n-digit
adder
zn
zn-10
more complex counters. As an example, the circuit of Fig. 7.16 is a 24-to-3

counter.
The computation time of the circuit of Fig. 7.16 is equal to 3TLUT6. More
generally, a tree made up of 2k-1 6-to-3 counters generates a 62k-1-to-3 counter,
with a computation time equal to kTLUT6. In the case of Fig. 7.16, k = 3 and
62k-1 = 24.
Finally, with an additional 3-to-2 counter and an n-bit adder a 24-operand adder
is obtained (Fig. 7.17). Complete VHDL models six_to_three_counter.vhd and
twenty_four_operand_adder.vhd are available at the Authors web page and
examples of FPGA implementations are given in Sect. 7.9.
To summarize, an m-operand adder, with m = 62k-1, can be synthesized with
k
2 -1 6-to-3 counters plus a 3-to-2 counter and an n-bit adder. Its computation time is
T m; n k TLUT6 TFA Tadder n;
where k = log2m ? 1-log26 \ log2m.
Comment 7.5
More complex types of counters have been proposed (see, for example, Chap. 8 of
[1], Chap. 3 of [5], Chap. 11 of [6]). Nevertheless, they do not necessarily give
high performance FPGA implementations. As a matter of fact, in many cases the
best FPGA implementations are based on relatively simple algorithms, to which
correspond regular circuits that allow taking advantage of the special purpose carry

Fig. 7.20 20 s complement
adder and subractor
175
yn0
xn0
xn
yn0
yn
(n+2)-bit
adder
xn0
xn
cin
zn+10
yn yn0
(n+2)-bit
adder
zn+10
logic circuitry, and permit the use of efficient design techniques such as pipelining
and digit-serial processing.
7.8 Subtractors and AdderSubtractors

Given two radix-B naturals x and y, the difference z = x-y could be negative. So,
the subtraction operation must be considered over the set of integers. A convenient
way to represent integers is Bs complement: the vector
xn xn1 xn2 . . .x1 x0 ; with xn 2 f0; 1g and xi 2 f0; 1; . . .; B 1g8i\n;
represents
x xn Bn xn1 Bn1 xn2 Bn2 x1 B x0 :
Thus, xn is a sign bit: if xn = 0, x is a non-negative integer (a natural), and if
xn = 1, x is a negative integer. The range of represented integers is
Bn x\Bn :
Let xn xn-1 xn-2 x1 x0 and yn yn-1 yn-2 y1 y0 be the Bs complement representations of x and y. If the sum z = x ? y ? cin, being cin an initial carry, belongs
to the interval -Bn B z \ Bn, then z is represented by the vector zn zn-1 zn-2 z1 z0
generated by the mixed-radix adder of Fig. 7.18 (all radix-B digits but the most
significant binary digits xn, yn and zn).
If the difference z = x-y belongs to the interval -Bn B z \ Bn, then z is
represented by the vector zn zn-1 zn-2 z1 z0, generated by the circuit of Fig. 7.19 in
which yi is the (B-1)s complement of yi, Vi \ n.
The sum z = x ? y or the difference z = x-y could lie outside the interval Bn B z \ Bn (an overflow situation). In order to avoid overflows, both x and
y should be represented with an additional digit. In the case of Bs complement
representations, digit extension is performed as follows:
xn xn1 xn2 . . .x1 x0 ! xn wn xn1 xn2 . . .x1 x0 ; with wn xn B 1:
For example, if B = 10 and x = 249, then x is represented by
176
Table 7.1 Binary adders
Table 7.2 Radix-2k n-bit

adders
7 Adders
n
LUTs
Delay
32
64
128
256
512
1024
32
64
128
256
512
1024
2.25
2.98
4.44
7.35
13.1
24.82
LUTs
Delay
32
64
64
64
64
128
128
128
256
256
256
512
1024
1024
1024
1024
4
4
5
8
16
8
12
16
16
12
13
16
16
22
23
32
88
176
143
152
140
304
286
280
560
572
551
1120
2240
2303
2295
2304
2.92
3.11
3.05
3.64
4.95
3.85
4.73
5.04
5.22
4.98
4.99
5.59
6.31
6.15
6.13
6.41
0249; 00249; 000249; etc:

If B = 10 and x = -249, then x is represented by
1751; 19751; 199751; etc:
Observe that if B = 2, then the bit extension operation amounts to repeating the
most significant bit. In Fig. 7.20 a 20 s complement adder and a 20 s complement
subtractor are shown. In both cases the comparison of bits zn+1 and zn allows the
detection of overflows: if zn+1 = zn then the result does not belong to the interval
- Bn B z \ Bn.
The following VHDL models describe the circuits of Fig. 7.20.
Generic models two_s_comp_adder.vhd and two_s_comp_subtractor.vhd are

7.9 FPGA Implementations
177
Table 7.3 n-bit carry-select

adders
LUTs
Delay
32
32
32
64
64
64
128
128
256
256
256
512
512
512
1024
1024
1024
6
8
4
8
16
4
12
16
16
32
13
16
32
23
16
64
32
84
72
60
144
199
120
417
399
799
783
819
1599
1567
1561
3199
3103
3135
4.83
3.99
3.64
6.06
4.17
4.03
5.37
4.87
5.69
5.26
5.64
6.10
6.09
6.16
6.69
6.74
6.52
Table 7.4 n-bit carry-select

adders (version 2)
Delay
32
256
512
1024
8
16
32
32
3.99
5.69
6.09
6.52
Table 7.5 n-bit adders with

n = n1n2n3
n1
n2
n3
LUTs
Delay
256
256
512
512
512
1024
1024
16
4
8
4
16
16
16
4
16
8
16
4
4
16
4
4
8
8
8
16
4
1452
684
2120
1364
2904
5808
6242
6.32
6.81
10.20
7.40
7.33
10.33
7.79

Several adders have been implemented within a Virtex 5-2 device. All along this
section, the times are expressed in ns and the costs in numbers of Look Up Tables
(LUTs) and flip-flops (FFs). All VHDL models as well as several test benches are
178
7 Adders
Table 7.6 Long-operand adders

n
s
k
FF
LUTs
Period
Total time
128
128
128
128
256
256
256
512
512
512
1024
1024
1024
2048
2048
135
134
133
132
263
262
261
520
519
518
1033
1034
1031
2063
2056
107
97
132
137
187
177
234
381
347
337
757
717
667
1427
1389
3.21
3.14
3.18
3.45
3.40
3.51
3.87
3.92
3.78
4.26
4.20
4.32
4.55
4.45
5.04
51.36
25.12
12.72
6.90
54.40
28.08
15.48
125.44
60.48
34.08
268.80
138.24
72.80
284.80
161.28
8
16
32
64
16
32
64
16
32
64
16
32
64
32
64
16
8
4
2
16
8
4
32
16
8
64
32
16
64
32
Table 7.7 Sequential

multioperand adders
FF
LUTs
Period
Total time
8
8
16
16
32
32
32
64
64
64
4
8
16
8
32
16
64
64
32
16
12
13
22
21
39
38
40
72
71
70
23
32
90
56
363
170
684
1356
715
330
2.25
2.37
2.71
2.57
3.72
3.09
3.89
4.62
4.48
4.41
9.00
18.96
43.36
20.56
119.04
49.44
248.96
295.68
143.36
70.56
7.9.1 Binary Adder

The circuit is shown in Fig. 7.3. The synthesis results for several numbers n of bits
are given in Table 7.1.
7.9.2 Radix 2k Adders

The circuit is shown in Fig. 7.4. The synthesis results for several numbers n = 2k
of bits are given in the Table 7.2. In these implementations, the carry propagation multiplexer muxcy has been explicitly instantiated within the VHDL
description.
179
Table 7.8 Sequential carrysave adders
FFs
LUTs
Period
Total time
8
8
16
16
32
32
32
64
64
64
4
8
16
8
32
16
64
64
32
16
19
20
37
36
70
69
71
135
134
133
37
46
120
86
425
232
746
1482
841
456
1.81
1.81
1.87
1.84
2.57
1.88
2.68
2.69
2.61
1.9
7.24
14.48
29.92
14.72
82.24
30.08
171.52
172.16
83.52
30.40
Table 7.9 Multioperand

addition array
LUTs
Delay
8
8
16
16
32
32
32
32
64
64
64
4
8
16
8
32
8
16
64
64
32
16
21
47
219
103
947
215
459
1923
3939
1939
939
2.82
5.82
11.98
6.00
24.32
6.36
12.35
47.11
49.98
25.04
13.07
7.9.3 Carry Select Adder

The circuit is shown in Fig. 7.6. The synthesis results for several numbers n = mk
of bits are given in Table 7.3.
The alternative circuit of Fig. 7.7 has also been implemented for several values
of n. The results are given in Table 7.4.
7.9.4 Logarithmic Adders

The synthesis results for several numbers n = n1n2n3 of bits are given in Table 7.5.
7.9.5 Long Operand Adder

The circuit is shown in Fig. 7.8. The synthesis results for several numbers n = ks
of bits are given in Table 7.6. Both the clock period Tclk and the total delay (kTclk)
are given.
180
7 Adders
Table 7.10 Combinational

carry-save adder
LUTs
Delay
8
8
16
16
32
32
32
32
64
64
64
4
8
16
8
32
8
16
64
64
32
16
22
68
314
135
1388
279
649
2868
5844
2828
1321
2.93
5.49
10.26
5.59
20.03
5.95
10.65
37.75
39.09
20.31
11.35
Table 7.11 8-operand

addition trees
LUTs
Delay
8
16
32
64
50
106
218
442
3.78
3.97
4.33
5.06
Table 7.12 24-operand

adders based on 6-to-3
counters
LUTs
Delay
8
16
24
32
64
157
341
525
709
1445
4.59
4.77
4.95
5.13
5.86
7.9.6 Sequential Multioperand Adders

The circuit is shown in Fig. 7.9. The synthesis results for several numbers n of bits
and m of operands are given in Table 7.7. Both the clock period Tclk and the total
delay (mTclk) are given.
The carry-save adder of Fig. 7.10 has also been implemented. The results are
given in Table 7.8.
7.9.7 Combinational Multioperand Adders

The circuit of Fig. 7.12 has been implemented. The synthesis results for several
numbers n of bits and m of operands are given in Table 7.9.
The carry-save adder of Fig. 7.14 has also been implemented. The results are
given in Table 7.10.
181
30
25
20
normal
Base_2k
15
Carry_sel
10
log_add
5
0
0
200
400
600
800
1000
1200
Fig. 7.21 Delay in function of the number of bits for several 2-operand adders
As an example of multioperand adddition trees (Fig. 7.13), several 8-bit adders

have been implemented, with the results given in Table 7.11.
Examples of implementation results for 24-operand adders based on 6-to-3
counters (Fig. 7.17) are given in Table 7.12.
7.9.8 Comparison
A comparison between four types of 2-operand adders, namely binary (normal),
radix-2k, carry-select and logarithmic adders, has been done: Fig. 7.21 gives the
corresponding adder delays (ns) in function of the number n of bits.
7.10 Exercises
1. Generate a generic model of a 20 s complement addersubtractor with overflow
detection.
2. An integer x can be represented under the form (1)s m where s is the sign of
x and m its magnitude (absolute value). Design an n-bit sign-magnitude adder
subtractor.
3. Design several n-bit counters, for example
7-to-3,
31-to-3,
5-to-2,
26-to-2.
4. Design a self-timed 64-bit adder with end of computation detection (done
signal).
182
7 Adders
5. Generate several generic models of an incrementer/decrementer, that is, a

circuit that computes x 1 mod m under the control of an upb/down binary
signal.
References
Press, New York
2. Ling H (1981) High-speed binary adder. IBM J Res Dev 25(3):156166
3. Brent R, Kung HT (1982) A regular layout for parallel adders. IEEE Trans Comput C-31:260
264
4. Ladner RE, Fischer MJ (1980) Parallel prefix computation. J ACM 27:831838
5. Ercegovac MD, Lang T (2004) Digital arithmetic. Morgan Kaufmann, San Francisco
6. Deschamps JP, Bioul G, Sutter G (2006) Synthesis of arithmetic circuits. Wiley, New York
Chapter 8
Multipliers
Multiplication is a basic arithmetic operation whose execution is based on 1-digit

by 1-digit multipliers and multi-operand adders. Most FPGA families include the
basic components for implementing fast and cost-effective multipliers. Furthermore, they also include optimized fixed-size multipliers which, in turn, can be used
for implementing larger-size multipliers.
The basic multiplication algorithm is described in Sect. 8.1. Several combinational implementations are proposed in Sect. 8.2. They correspond to different
types of multi-operand adders: iterative ripple-carry adders, carry-save adders,
multi-operand adders based on counters, radix-2k and mixed-radix adders.
Sequential implementations are proposed in Sect. 8.3. They used the shift and add
method implemented with either a ripple-carry adder or a carry-save adder. If
integer operands are considered, several options are proposed in Sect. 8.4. A first
method consists of multiplying Bs complement integers as they are naturals; the
drawback of this conceptually simple method is that the operands must be represented, and multiplied, with as many digits as the final result. Better options are a
modification of the shift and add algorithm, multiplication of naturals followed
by a post-correction, and the Booth algorithms. The last section describes a
LUT-based method for implementing a constant multiplier, that is to say, circuits
that compute c y ? u, where c is a constant.
8.1 Basic Algorithm

Consider two radix-B numbers
x xn1 Bn1 xn2 Bn2 . . . x1 B x0 and
y ym1 Bm1 ym2 Bm2 . . . y1 B y0 ;

183
184
8 Multipliers
Fig. 8.1 1-digit by 1-digit

multiplier. a Symbol,
b internal structure (B = 2)
ab c
b
c
x
d
(a)
FA
(b)
where xi and yi belong to {0, 1,, B -1}. An n-digit by m-digit multiplier

generates a radix-B number
z znm1 Bnm1 znm2 Bnm2 . . . z1 B z0
such that
z x y:
A somewhat more general definition considers the addition of two additional
numbers
u un1 Bn1 un2 Bn2 . . . u1 B u0 and
v vm1 Bm1 vm2 Bm2 . . . v1 B v0 ;
so that
z x y u v:
8:1
Observe that the maximum value of z is

Bn 1Bm 1 Bn 1 Bm 1 Bnm 1:
In order to compute (8.1), first define a 1-digit by 1-digit multiplier: given four
B-ary digits a, b, c and d, it generates two B-ary digits e and f such that
a bcd e Bf
8:2
(Fig. 8.1a).
If B = 2, it amounts to a 2-input AND gate and a 1-digit adder (Fig. 8.1b).
An n-digit by 1-digit multiplier made up of n 1-digit by 1-digit multipliers is
shown in Fig. 8.2. It computes as
zx bud
8:3
where x and u are n-digit numbers, b and d are 1-digit numbers, and z is an
(n ? 1)-digit number. Observe that the maximum value of z is
Bn 1B 1 Bn 1 B 1 Bn1 1:
8.1 Basic Algorithm

x n-1
185
u n-1 xn-2
un-2
x1
.....
u 1 x0
u0
d
zn
z n-1
z n-2
z1
z0
Fig. 8.2 n-digit by 1-digit multiplier
Using the iterative circuit of Fig. 8.2 as a computation resource, the computation
of (8.1) amounts to computing the m n-digit by 1-digit products
z0 x y0 u v0 ;
z1 x y1 v1 B;
z2 x y2 v2 B2 ;
8:4
...
z
m1
x ym1 vm1 Bm1 ;
and to adding them, that is

z z0 z1 z2 . . . zm1 x y u v:
8:5
For that, one of the multioperand adders of Sect. 7.7 can be used. As an example, if
Algorithm 7.2 is used, then z is computed as follows.
Algorithm 8.1: Multiplication, right to left algorithm
8.2 Combinational Multipliers

8.2.1 Ripple-Carry Parallel Multiplier
The combinational circuit of Fig. 8.3 implements Algorithm 8.1 (with n = 4 and
m = 3). One of its critical paths has been shaded. Its computation time is equal to
Tmultiplier n; m n 2m 2 Tmultiplier 1; 1:
8:6
186
8 Multipliers
x3
y0
u3 x2
y0
u 2 x1
y0
u 1 x0
y0
u0
x
v0
x3
y1
x2
y1
x1
y1
x0
y1
z0
x
v1
x3
y2
x2
y2
x1
y2
x0
y2
z1
x
v2
z6
z5
z4
z3
z2
Fig. 8.3 Combinational multiplier
187
A complete generic model parallel_multiplier.vhd is available at the Authors

web page.
8.2.2 Carry-Save Parallel Multiplier

A straightforward modification of the multiplier of Fig. 8.3, similar to the
carry-save principle, is shown in Fig. 8.4. The circuit is made up of an n-bym array of 1-by-1 multipliers, whose computation time is equal to n T(1,1),
plus an m-digit output adder. Its critical path has been shaded. Its computation
time is equal to
Tmultiplier n; m n Tmultiplier 1; 1 m Tadder 1 n m Tmultiplier 1; 1:
8:7
188
8 Multipliers
x3
y0
u3
x3
x2
y0
y1
x2
u2
x1
y0
v2
y1
x1
u1
x0
y0
u0
v1
v0
y1
x0
y1
z0
0
x
x3
y2
x2
y2
x1
y2
x0
y2
z1
0
x
z6
z5
z4
z3
z2
Fig. 8.4 Carry-save combinational multiplier
A complete generic model parallel_csa_multiplier.vhd is available at the Authors

web page.
189
8.2.3 Multipliers Based on Multioperand Adders

A straightforward implementation of Eqs. (8.4) and (8.5) can also be considered
(Fig. 8.5). For that, any type of multioperand adder can be used.
Example 8.1
Consider an n-bit by 7-bit multiplier. The 7-operand adder can be divided up into a 7to-3 counter, a 3-to-2 counter and a ripple-carry adder. The complete structure is shown
in Fig. 8.6 and is described by the following VHDL model:
A complete generic model N_by_7_multiplier.vhd is available at the Authors

web page.
Numerous multipliers, based on trees of counters, have been proposed and
reported, among others the Wallace and Dadda multipliers (Wallace [4]; Dadda
[3]). Nevertheless, as already mentioned before (Comment 7.5), in many cases the
best FPGA implementations are based on relatively simple algorithms, to which
correspond regular circuits that allow taking advantage of the special purpose carry
logic circuitry. To follow, an example of efficient FPGA implementation is
described.
Consider the set of equations (8.4). If two successive steps are merged within an
only step (loop unrolling), the new set of equations is:
z1;0 x y1 v1 B x y0 u v0 ;
z3;2 x y3 v3 B x y2 v2 B2 ;
...
zm1;m2 x ym1 vm1 B x ym2 vm2 Bm2 ;
8:8
190
8 Multipliers
x
ym-1
u
d
z=xb+u+d
xB
0
v m-1
y1
z=xb+u+d
m-1
xB
z (m-1)
u
d
0
v1
y0
z=xb+u+d
xB
z(1)
u
d
u
v0
z (0)
multioperand adder
Fig. 8.5 Multiplier with a multioperand adder
y6
z=xb+u+d
u
d
0
v6
y1
z=xb+u+d
0
v1
y0
z=xb+u+d
u
d
w (1)
w (1)
z (1)
z (0)
w (6)
x2
u
d
x2
(6)
7-to-3 counter
x (1)
x (2)
x (3)
3-to-2 counter
x (1)
x (2)
ripple adder
Fig. 8.6 An n-bit by 7-bit multiplier
u
v0

x3
yj
191
x2
yj
x1
yj
0 x0
yj
x
vj
x3
yj+1
x2
y j+1
x1
y j+1
x0
y j+1
z0
(j+1,j)
x
v j+1
z 5(j+1,j) z4(j+1,j)
z3(j+1,j)
z 2(j+1,j)
z1(j+1,j)
Fig. 8.7 4-digit by 2-digit multiplier
and the product is equal to

z z1;0 z3;2 . . . zm1;m2 :
Assuming that u = 0, the basic operation to implement (8.8) is

zi1;i x yj1 vj1 B x yj vj
to which corresponds the circuit of Fig. 8.7 (with n = 4).
The circuit of Fig. 8.7 can be decomposed into n ? 1 vertical slices of the type
shown in Fig. 8.8a (with obvious simplifications regarding the first and last slices).
Finally, if B = 2 and vj = 0, the carries of the first line are equal to 0, so that the
circuit of Fig. 8.8a can be implemented as shown in Fig. 8.8b.
Comment 8.1
Most FPGAs include the basic components for implementing the structure of
Fig. 8.8b, and the synthesis tools have the capability to generate optimized multipliers from a simple VHDL expression, such as
z \ x y;
Furthermore, many FPGAs also include fixed-size multiplier blocks.
8.2.4 Radix-2k and Mixed-Radix Parallel Multipliers

The basic multiplication algorithm (Sect. 8.1) and the corresponding ripple-carry
and carry-save multipliers (Sects. 8.2.1 and 8.2.2) have been defined for any radixB. In particular, radix-2k multipliers can be defined. This allows the synthesis of
n k-bit by m k-bit multipliers using k-bit by k-bit multipliers as building blocks.
192
8 Multipliers
xi
yj
c j,i+1
xi
c ji
cj+1,i
zi
x i-1 yj+1
p = xiyj XOR xi-1yj+1
x i-1 y j+1
cj+1,i
yj
0
1
c j+1,i-1
c j+1,i-1
(j+1,j)
z i(j+1,j)
(a)
(b)
Fig. 8.8 Iterative cell of a parallel multiplier
The following VHDL model defines a radix-2k ripple-carry parallel multiplier.

The main iteration consists of m n instantiations of any type of k-bit by k-bit
combinational multiplier that computes z = a b ? c ? d and represents z under
the form zH 2k zL ; where zH and zL are k-bit numbers:
193
A complete generic model base_2k_parallel_multiplier.vhd is available at the

Authors web page.
The stored-carry encoding can also be applied. Once again, the main iteration
consists of m n instantiations of any type of k-bit by k-bit multiplier, and the
connections are similar to those of the carry-save multiplier of Sect. 8.2.2. A
complete generic model base_2k_csa_multiplier.vhd is available at the Authors
web page.
A straightforward generalization of relations (8.2) to (8.5) allows defining
mixed-radix combinational multipliers. First consider the circuit of Fig. 8.1a,
assuming that
a; c 2 f0; 1; . . . ; B1 1g; and b; d 2 f0; 1; . . . ; B2 1g:
Then
z a b c d B1 1 B2 1 B1 1 B2 1 B1 B2 1;
so that z can be expressed under the form
z e B1 f ; with e 2 f0; 1; . . . ; B2 1g; f 2 f0; 1; . . . ; B1 1g:
Then, consider the circuit of Fig. 8.2, assuming that x and u are n-digit radix-B1
numbers, and b and d are 1-digit radix-B2 numbers. Thus,
x b u d zn Bn1 zn1 Bn1
. . . z 1 B1 z 0 ;
1
8:9
with
zn 2 f0; 1; . . . ; B2 1g and zi 2 f0; 1; . . . ; B1 1g; 8i in f0; 1; . . . ; n 1g:
Finally, given two n-digit radix-B1 numbers x and u, and two m-digit radix-B2
numbers y and v, compute
z0 x y0 u v0 ;
z1 x y1 v1 B2 ;
z2 x y2 v2 B22 ;
...
8:10
zm1 x ym1 vm1 Bm1

:
2
Then
z z0 z1 z2 . . . zm1 x y u v:
Consider the case where
B 1 2 k 1 ; B 2 2 k2 :
8:11
194
8 Multipliers
An easy way to define a VHDL model of the corresponding multiplier consists

in first modelling a circuit that implements (8.9). The main iteration consists of
n instantiations of any type of k1-bit by k2-bit combinational multiplier that
computes
a b c d z H 2 k1 z L ;
where zH is a k2-bit number and zL a k1-bit number:
Then, it remains to instantiate m rows:
A complete generic model MR_parallel_multiplier.vhd is available at the Authors

web page.
The circuit defined by the preceding VHDL model is a bidirectional array
similar to that of Fig. 8.3, but with more complex connections. As an example,
with k1 = 4 and k2 = 2, the connections corresponding to cell (j, i) are shown in
195
x 4i+74i+4 y 2j-12j-2
x 4i+34i
y 2j-12j-2
b
c
j-1,i+1
j-1,i
d
x 4i+34i
y2j+12j
x 4i-14i-4
y2j+12j
j,i
j,i-1
d
e
f
e
f
Fig. 8.9 Part of a 4n-bit by 2n-bit multiplier using 4-bit by 2-bit multiplication blocks
Fig. 8.9. As before, a stored-carry encoding circuit could also be designed, but
with an even more complex connection pattern. It is left as an exercise.
8.3 Sequential Multipliers

8.3.1 Shift and Add Multiplier
In order to synthesize sequential multipliers, the basic algorithm of Sect. 8.1 can be
modified. For that, Eq. (8.4) are substituted by the following:
196
8 Multipliers
x
shift register
initially: y
x b
Fig. 8.2
zn..1 z 0 d
x b
Fig. 8.2
z n..1 z 0 d
register
initially: u
shift register
initially: v
register
initially: u
zm-1..0
zn+m-1..m
zn+m-1..m
(a)
shift register
initially: y
zm-1..0
(b)
Fig. 8.10 Shift and add multipliers
z0 u x y0 v0 =B;

z1 z0 x y1 v1 =B;

z2 z1 x y2 v2 =B;
zm1
8:12
...

zm2 x ym1 vm1 =B:
Multiply the first equation by B, the second by B2, and so on, and add the so
obtained equations. The result is
zm1 Bm u x y0 v0 x y1 v1 B . . . x ym1 vm1 Bm1
xy u v:
Algorithm 8.2: Shift and add multiplication
A data path for executing Algorithm 8.2 is shown in Fig. 8.10a. The following
VHDL model describes the circuit of Fig. 8.10a (B = 2).
197
The complete circuit also includes an m-state counter and a control unit.
A complete generic model shift_and_add_multiplier.vhd is available at the
Authors web page.
If v = 0, the same shift register can be used for storing both y and the least
significant bits of z. The modified circuit is shown in Fig. 8.10b. A complete
generic model shift_and_add_multiplier2 is also available.
198
8 Multipliers
x
Fig. 8.11 Sequential carrysave multiplier
y1
y2
y3
carry-save adder
sn..1
c n..1
register
initially: u
register
initially: v
s0
adder
shift register
initially: y
zm-1..0
z n+m-1..m
The computation time of the circuits of Fig. 8.10 is approximately equal to

Tmultiplier n; m m Tmultiplier n; 1 m n Tmultiplier 1; 1:
8:13
8.3.2 Shift and Add Multiplier with CSA

The shift and add algorithm can also be executed with stored-carry encoding. After
m steps the result is obtained under the form
sn1 Bnm1 cn2 sn2 Bnm2 . . . c0 s0 Bm zm1 Bm1 . . .
z1 B z0 ;
and an additional n-digit adder computes
sn1 Bn1 cn2 sn2 Bn2 . . . c0 s0
zmn1 Bn1 . . . zm1 B zm :
The corresponding data path is shown in Fig. 8.11. The carry-save adder computes
y1 y2 y3 s c;
where y1, y2 and y3 are n-bit numbers, and s and c are (n ? 1)-bit numbers. At the
end of step i, the less significant bit of s is zi, and the n most significant bits of s and
c are transmitted to the next step:
199
The complete circuit also includes an m-state counter and a control unit. A complete generic model sequential_CSA_multiplier.vhd is available at the Authors
web page. The minimum clock period is equal to the delay of a 1-bit by 1-bit
multiplier. Thus, the total computation time is equal to
Tmultiplier n; m m Tmultiplier 1; 1 Tadder n n m Tmultiplier 1; 1:
8:14
Comment 8.2
In sequential_CSA_multiplier.vhd the done flag is raised as soon as the final values
of the adder inputs are available. A more correct control unit should raise the flag
k cycles later, being k Tclk an upper bound of the n-bit adder delay. The value of
k could be defined as a generic parameter (Exercise 8.3).
8.4 Integers
Given four Bs complement integers
x xn xn1 xn2 . . .x0 ; y ym ym1 ym2 . . .y0 ; u un un1 un2 . . .u0 ;
v vm vm1 vm2 . . . v0 ;
belonging to the ranges
Bn x\Bn ; Bm y\Bm ; Bn u\Bn ; Bm v\Bm ;
then z = x y ? u ? v belongs to the interval
200
8 Multipliers
Bnm1 z\Bnm1 :
Thus, z is a Bs complement number of the form
z znm1 znm znm1 . . . z1 z0 :
8.4.1 Mod 2Bn+m Multiplication

The integer represented by a vector xn xn1 xn2 . . . x1 x0 is
x xn Bn xn1 Bn1 xn2 Bn2 . . . x1 B x0 ;
while the natural natural(x) represented by this same vector is
natural x xn Bn xn1 Bn1 xn2 Bn2 . . . x1 B x0 :
As xn [ {0, 1}, either natural(x) = x or natural(x) = x +2Bn. So,
natural x x mod 2Bn :
The following method can be used to compute z = x y ? u ? v. First, represent
the operands x, y, u and v with the same number of digits (n ? m ? 2) as the result
z (digit extension, Sect. 7.8). Then, compute z = x y ? u ? v as if x, y, u and
v were naturals:
z natural x natural y naturalu naturalv naturalx y u v:
Finally, reduce z modulo 2Bnm1 : Assume that before the mod 2Bnm1 reduction
z . . . znm1 Bnm1 znm Bnm znm1 Bnm1 . . . z1 B z0 ;
then
z mod 2Bnm1 . . . znm1 mod 2Bnm1 znm Bnm znm1 Bnm1
. . . z1 B z0 :
In particular, if B is even,
z mod 2Bnm1 znm1 2Bnm1 znm Bnm znm1 Bnm1 . . . z1 B z0 :
Example 8.2
Assume that B = 10, n = 4, m = 3, x = 7918, y = -541, u =
-7017, v = 742, and compute z = 7918(-541) ? (-7017) ? 742. In 10s
complement: x = 07918, y = 1459, u = 12983, v = 0742.
1. Express all operands with 9 digits: x = 000007918, y = 199999459, u =
199992983, v = 000000742.
8.4 Integers
v2
201
u3 v 2
u3 v2
u 3 v2
u 3 v2
u2 v 1
u1 v0
u0
3,0
3,0
3,0
3,0
2,0
1,0
0,0
3,1
3,1
3,1
2,1
1,1
0,1
3,2
3,2
2,2
1,2
0,2
3,2
2,2
1,2
0,2
2,2
1,2
0,2
1,2
0,2
z0
z1
z2
j,i
z3
xjyi + c + d = 2e + f
0,2
z4
z5
z6
Fig. 8.12 Carry-save multiplier for integers (n = 3, m = 2)
2. Compute x y ? u ? v: 000007918199999459 ? 199992983 ? 000000742 =

1583795710087.
3. Reduce 1583795710087 modulo 2108: (1583795710087) mod 2108 = (7 mod 2)
108 ? 95710087 = 195710087.
The result 195710087 is the 10s complement representation of -4289913.
Thus, any multiplier for natural numbers can be used. As an example, an
(n ? m ? 2)-digit by (n ? m ? 2)-digit carry-save multiplier could be used
(Fig. 8.4). As the result is reduced modulo 2Bn+m+1, only the rightmost part of the
circuit is used (if B is even), so that there is no output adder, and the most
significant digit is reduced mod 2. An example with n = 3 and m = 2 is shown in
Fig. 8.12. The corresponding computation time is equal to
n m 2 Tmultiplier 1; 1:
8:15
This delay is practically the same as that of a carry-save combinational multiplier

(8.7). Nevertheless, the number of 1-digit by 1-digit multiplication cells is equal to
1 2 3 . . . n m 2 n m 2n m 3=2 instead of n m.
202
8 Multipliers
A very simple way to generate a VHDL model consists of defining

(n ? m ? 2)-bit representations of all operands and instantiating an (n ? m ? 2)bit by (n ? m ? 2)-bit carry-save multiplier:
Only n ? m ? 2 output bits of the carry-save multiplier are connected to output

ports, and the synthesis program will prune the circuit accordingly.
A complete generic model integer_CSA_multiplier.vhd is available at the
Authors web page.
To conclude, this approach is conceptually attractive because any type of
multiplier for natural numbers can be used. Nevertheless, the cost of the corresponding circuits is very high.
8.4.2 Modified Shift and Add Algorithm

Consider again four Bs complement integers
x xn xn1 xn2 . . . x0 ; y ym ym1 ym2 . . . y0 ; u un un1 un2 . . . u0 ;
v vm vm1 vm2 . . . v0 :
A set of equations similar to (8.12) can be defined:
z0 u x y0 v0 =B;

z1 z0 x y1 v1 =B;

z2 z1 x y2 v2 =B;
...

zm1 zm2 x ym1 vm1 =B;

zm zm1 x ym vm =B:
8:16
8.4 Integers
203
Multiply the first equation by B, the second by B2, and so on, and add the m ? 1
so obtained equations. The result is
zm Bm1 u x y0 v0 x y1 v1 B . . . x ym1 vm1 Bm1
x ym vm Bm
xy u v:
Algorithm 8.3: Modified shift and add multiplication
In what follows it is assumed that vm = 0, that is to say v C 0; so, in order to

implement Algorithm 8.3, the two following computation primitives must be
defined:
zux bd
8:17
z u x b;
8:18
and
where
Bn x\Bn ; Bn u\Bn ; 0 b\B; 0 d\B:
Thus, in the first case,
Bn1 z\Bn1 ;
and in the second case
Bn1 B 1 z\Bn1 ;
so that in both cases z is an (n ? 2)-digit Bs complement integer and natural(z) = z mod 2Bn+1.
The first primitive (8.17) is implemented by the circuit of Fig. 8.13 and the
second (8.18) by the circuit of Fig. 8.14. In both, circuit zn+1 is computed
modulo 2.
As an example, the combinational circuit of Fig. 8.15 implements Algorithm
8.3 (with n = m = 2). Its cost and computation time are practically the same as in
the case of a ripple-carry multiplier for natural numbers. It can be described by the
following VHDL model.
204
8 Multipliers
xn
u n xn
un x n-1
u n-1
u1 x0
u0
.....
d
z n+1
zn
z n-1
z0
Fig. 8.13 First computation primitive
xn
x n-1
x0
(B-1)'s
compl.
(B-1)'s
compl.
(B-1)'s
compl.
un
un
un-1
.....
u0
x
b
z n+1
zn
Fig. 8.14 Second computation primitive
z n-1
z0
8.4 Integers
Fig. 8.15 Combinational
multiplier for integers
(B = 2, m = n = 2)
205
x2
y0
u 3 x2
y0
x2
x2
x1
x2
x1
x2
y3
x1
y1
x0
y2
x0
z6
z5
z4
z3
z1
v2
y3
z0
v1
y3
u0
v0
y2
y3
x0
y2
y0
y1
y2
u1 x0
y1
x2
y0
y1
x2
u2 x 1
z2
y3
206
8 Multipliers
A complete generic model modified_parallel_multiplier.vhd is available at the

Authors web page.
The design of a sequential multiplier based on Algorithm 8.3 is left as an
exercise.
8.4.3 Post Correction Multiplication

Given four Bs complement integers
x xn xn1 xn2 . . . x0 ; y ym ym1 ym2 . . . y0 ; u un un1 un2 . . . u0 ;
v vm vm1 vm2 . . . v0 ;
then z = x y ? u ? v, belonging to the interval Bnm1 z\Bnm1 ; can be
expressed under the form
z X0 Y0 U0 V0 xn ym Bnm xn Y0 un Bn
y m X0 v m Bn ;
where X0, Y0, U0 and V0 are four naturals
X0 xn1 xn2 . . . x0 ; Y0 ym1 ym2 . . . y0 ; U0 un1 un2 . . . u0 ;
X0 vm1 vm2 . . . v1 v0
deduced from x, y, u and v by eliminating the sign bits. Thus, the computation of
z amounts to the computation of
Z0 X0 Y0 U0 V0 ;
that can be executed by any type of multiplier for naturals, plus a post correction
that consists of several additions and left shifts.
If B = 2 and u = v = 0, then
8.4 Integers
207
x3
y0
x2
x
nand
x3
y1
y2
z6
z5
x1
y0
x0
y0
x
0
x2
x
nand
x3
y0
y1
x1
y1
x0
y1
z0
x
0
x2
y2
x
nand
z4
x1
y2
x
nand
z3
x0
y2
x
nand
z1
z2
Fig. 8.16 Multiplier with post correction
z x y X0 Y0 xn ym 2nm xn Y0 2n ym X0 2n :
The (n ? m ? 2)-bit 2s complement representations of xn Y0 2n and
ym X0 2m are
2m1 2m xn ym1 2m1 . . . xn y0 20 1 2n mod 2nm2 ;
and
2n1 2n ym xn1 2n1 . . . ym x0 20 1 2m mod 2nm2 ;
so that the representation of xn ym 2nm xn Y0 2n ym X0 2n is
2nm1 xn ym 2nm xn ym1 2nm1 . . . xn y0 2n
2n ym xn1 2nm1 . . . ym x0 2m 2m mod 2nm2 :
A simple modification of the combinational multipliers of Fig. 8.3 and 8.4
allows computing x y, where x is an (n ? 1)-bit 2s complement integer and y an
(m ? 1)-bit 2s complement integer. An example is shown in Fig. 8.16 (n = 3,
208
8 Multipliers
m = 2). The nand multiplication cells are similar to that of Fig. 8.1b, but for the
substitution of the AND gate by a NAND gate [1].
The following VHDL model describes the circuit of Fig. 8.16.
A complete generic model postcorrection_multiplier.vhd is available at the

Authors web page.
8.4.4 Booth Multiplier

Given an (m ? 1)-bit 2s complement integer y ym 2m ym1 2m1
. . . y1 2 y0 ; define
y00 y0 and y0j yj yj1 ; 8i in f1; 2; . . .; mg;
8.4 Integers
209
so that all coefficients yi belong to {-1, 0, 1}. Then y can be represented under the
form
y y0m 2m y0m1 2m1 . . . y01 2 y00 ;
the so-called Booths encoding of y (Booth [2]. Unlike the 2s complement representation in which ym has a specific function, all coefficients yi have the same
function. Formally, the Booths representation of an integer is the same as the
binary representation of a natural. The basic multiplication algorithm (Algorithm
8.1), with v = 0, can be used.
Algorithm 8.4: Booth multiplication, z 5 xy 1 u
The following VHDL model describes a combinational circuit based on Algorithm

8.4.
A complete generic model Booth1_multiplier.vhd is available at the Authors web

page.
Higher radix Booth multipliers can be defined. Given an (m ? 1)-bit 2s
complement integer y ym 2m ym1 2m1 . . . y1 2 y0 ; where m is
odd, define
y00 2 y1 y0 ; y0i 2 y2 i1 y2 i y2 i1 ; 8i inf1; 2; . . .; m 1=2g;
so that all coefficients yi belong to {-2, -1, 0, 1, 2}. Then y can be represented
under the form
210
8 Multipliers
Fig. 8.17 Sequential radix-4

Booth multiplier
011
010,001
111,000
110,101
100
adder
n+1
2x
x
0
-x
-2x
y
2
register
initially: u
shift register
initially: y.0
zn+m+1..m+1
z m..0
y y0m1=2 4m1=2 y0m1=21 4m1=21 . . . y01 4 y00 ;

the so-called Booth-2 encoding of y.
Example 8.3
Consider the case where m = 9 and thus (m-1)/2 = 4. The 2s complement
representation of -137 is 1101110111. The corresponding Booth-2 encoding is 1 2 -1 2 -1 and, indeed, -44 ? 2 43 - 42 +2 4 - 1 = -137.The basic radix4 multiplication algorithm, with v = 0, can be used.
Algorithm 8.5: Radix-4 Booth multiplication, z 5 x y 1 u
A sequential implementation is shown in Fig. 8.17. It includes a shift register

whose content is shifted two positions at each step, a parallel register and an adder
whose second operand is -2x, -x, 0, x or 2x depending on the three least significant bits (y2i+1, y2i, y2i-1) of the shift register. At each step, two output bits are
generated. Hence, the total computation time is equal to (m ? 1)/2Tclk, where Tclk
must be greater than the computation time of an (n ? 3)-bit adder. Thus,
Tn; m
m1
Tadder n 3:
2
With respect to a radix-2 shift and add multiplier (Sect. 8.2.1), the computation
time has been divided by 2.
8.4 Integers
211
The complete circuit also includes an (m ? 1)/2-state counter and a control

unit. A complete generic model Booth2_sequential_multiplier.vhd is available at
8.5 Constant Multipliers

Given an n-bit constant natural c and an m-bit natural y, the computation of
c y can be performed with any n-bit by m-bit multiplier whose first operand is
connected to the constant value c. Then, the synthesis tool will eliminate useless
components. In the case of FPGA implementations, an alternative method is to
store the constant c within the LUTs.
Assume that the technology at hand includes k-input LUTs. The basic component is a circuit that computes w = c b, where b is a k-bit natural. The
maximum value of w is

2n 1 2k 1 2nk 2k 2n 1;
212
Fig. 8.18 LUT
implementation of a k-bit by
n-bit constant multiplier
8 Multipliers
b 5..0
w5..0
w11..6
wn+5..n
so w is an (n ? k)-bit number. The circuit is shown in Fig. 8.18, with k = 6. It is

made up of n ? 6 LUT-6, each of them being programmed in such a way that
w6j5 ... 6j b c1 b6j5 ... 6j :
Its computation time is equal to TLUT6.
The function LUT_definition defines the LUT contents.
8.5 Constant Multipliers
213
Fig. 8.19 Computation of

w=cb?u
u
w n+5..0
b 5..0
Fig.8.18
(n+6)-bit
adder
The circuit of Fig. 8.18 can be used as a component for generating constant
multipliers. As an example, a sequential n-bit by m-bit constant multiplier is
synthesized. First define a component similar to that of Fig. 8.2, with x constant. It
computes z = c b ? u, where c is an n-bit constant natural, b a k-bit natural, and
u an n-bit natural. The maximum value of z is

2n 1 2k 1 2n 1 2nk 2k ;
so it is an (n ? k)-bit number. It consists of a k-bit by n-bit multiplier (Fig. 8.18)
and an (n ? k)-bit adder (Fig. 8.19).
Finally, the circuit of Fig. 8.19 can be used to generate a radix-2k shift and add
multiplier that computes z = c y ? u, where c is an n-bit constant natural, y an
m-bit natural, and u an n-bit natural. The maximum value of z is
2n 12m 1 2n 1 2nm 2m ;
so z is an (n ? m)-bit number. Assume that the radix-2k representation of y is
Ym/k-1 Ym/k-2 Y0, where each Yi is a k-bit number. The circuit implements the
following set of equations:
z0 u c Y0 =2k ;

z1 z0 c Y1 =2k ;

z2 z1 c Y2 =2k ;
zm=k1
8:19
...

zm=k2 c Ym=k1 =2k :
Thus,
zm=k1 2k
m=k
u c Y0 c Y1 2k . . . c Ym=k1 2k
that is to say
zm=k1 2m c y u:
The circuit is shown in Fig. 8.20.
The computation time is approximately equal to
T m=k TLUTk Tadder n k:
m=k1
214
8 Multipliers
Fig. 8.20 n-bit by m-bit

constant multiplier
u
Fig. 8.19
Yi
z n+k-1..k z k-1..0
register
initially: u
shift register
initially: y
zn+m-1..m
z m-1..0
The following VHDL model describes the circuit of Fig. 8.20 (k = 6).
A complete model sequential_constant_multiplier.vhd is available at the

Authors web page.
The synthesis of constant multipliers for integers is left as an exercise.
215

multiplier
LUTS
Delay
8
16
32
32
64
64
8
16
16
32
32
64
96
384
771
1536
3073
6181
13.29
28.26
36.91
57.46
74.12
119.33
Table 8.2 Carry-save

combinational multiplier
LUTS
Delay
8
16
32
32
64
64
8
16
16
32
32
64
102
399
788
1580
3165
6354
8.05
15.42
16.99
29.50
32.08
60.90

Several multipliers have been implemented within a Virtex 5-2 device. Those
devices include Digital Signal Processing (DSP) slices that efficiently perform
multiplications (25 bits by 18 bits), additions and accumulations. Apart from
multiplier implementations based on LUTs and FFs, more efficient implementations, taking advantage of the availability of DSP slices, are also reported. As
before, the times are expressed in ns and the costs in numbers of Look Up Tables
(LUTs), flip-flops (FFs) and DSP slices. All VHDL models as well as several test
benches are available at the Authors web page.
8.6.1 Combinational Multipliers

The circuit is shown in Fig. 8.3. The synthesis results for several numbers n and
m of bits are given in Table 8.1.
A faster implementation is obtained by using the carry-save method (Fig. 8.4;
Table 8.2).
If multipliers based on the cell of Fig. 8.8b are considered, more efficient
circuits can be generated. It is the by default option of the synthesizer
(Table 8.3).
Finally, if DSP slices are used, better implementations are obtained (Table 8.4).
216
8 Multipliers
Table 8.3 Optimized

combinational multiplier
LUTs
Delay
8
16
32
32
64
64
32
8
16
16
32
64
32
64
113
435
835
1668
6460
3236
3236
5.343
6.897
7.281
7.901
11.41
9.535
9.535

multiplier with DSP slices
LUTs
DSPs
Delay
8
16
32
32
64
64
32
8
16
16
32
64
32
64
0
0
77
93
346
211
211
2
2
2
4
12
6
6
4.926
4.926
6.773
9.866
12.86
11.76
11.76
Table 8.5 Radix-2k parallel

multipliers
mk
nk
LUTs
Delay
2
4
2
4
4
4
2
4
2
4
2
4
8
4
16
8
16
16
16
16
32
32
64
64
16
16
32
32
32
64
452
448
1740
1808
3480
6960
10.23
17.40
12.11
20.29
15.96
22.91
8.6.2 Radix-2k Parallel Multipliers

Several m k bits by n k bits multipliers (Sect. 8.2.4) have been implemented
(Table 8.5).
A faster implementation is obtained by using the carry-save method
(Table 8.6).
The same circuits have been implemented with DSP slices. The implementation
results are given in Tables 8.7, 8.8
8.6.3 Sequential Multipliers

Several shift and add multipliers have been implemented. The implementation
results are given in Tables 8.9, 8.10. Both the clock period Tclk and the total delay
(m Tclk) are given.

Table 8.6 Carry-save radix2k parallel multipliers
Table 8.7 Radix-2k parallel

multipliers with DSPs
Table 8.8 Carry-save radix2k parallel multipliers with

DSPs
Table 8.9 Shift and add

multipliers
217
m
mk
nk
LUTs
Delay
2
4
2
4
4
4
2
4
2
4
2
4
8
4
16
8
16
16
16
16
32
32
64
64
16
16
32
32
32
64
461
457
1757
3501
1821
6981
8.48
10.09
10.36
11.10
12.32
14.93
mk
nk
DSPs
LUTs
Delay
2
4
2
4
4
4
2
4
2
4
2
4
8
4
16
8
16
16
16
16
32
32
64
64
16
16
32
32
32
64
8
32
8
32
16
32
0
0
0
0
0
0
12.58
27.89
12.58
27.89
16.70
27.89
mk
nk
DSPs
LUTs
Delay
2
4
2
4
4
4
2
4
2
4
2
4
8
4
16
8
16
16
16
16
32
32
64
64
16
16
32
32
32
64
8
32
8
32
16
32
15
15
31
31
63
63
9.90
17.00
10.26
17.36
11.07
18.09
FFs
LUTs
Period
Total time
8
8
16
16
32
32
64
64
8
16
8
16
16
32
32
64
29
46
38
55
71
104
136
201
43
61
72
90
112
161
203
306
2.87
2.87
4.19
4.19
7.50
7.50
15.55
15.55
23.0
45.9
33.5
67.0
120.0
240.0
497.6
995.2
8.6.4 Combinational Multipliers for Integers

A carry-save multiplier for integers is shown in Fig. 8.12. The synthesis results for
several numbers n and m of bits are given in Table 8.11.
218
8 Multipliers
Table 8.10 Sequential carrysave multipliers
FFs
LUTs
Period
Total time
8
16
16
32
32
64
64
8
8
16
16
32
32
64
29
47
56
88
106
170
203
43
64
74
122
139
235
268
1.87
1.88
1.92
1.93
1.84
1.84
1.84
15.0
15.0
30.7
30.9
58.9
58.9
117.8
Table 8.11 Carry-save mod

2n+m+1 multipliers
LUTs
Delay
8
8
16
16
32
32
8
16
8
16
16
32
179
420
421
677
1662
2488
12.49
18.00
20.41
25.86
42.95
55.69
Table 8.12 Modified shift

and add algorithm
LUTs
Delay
8
8
16
16
32
32
8
16
8
16
16
32
122
230
231
435
844
1635
15.90
27.51
20.20
31.81
39.96
62.91
Table 8.13 Multipliers with

post correction
LUTs
Delay
8
8
16
16
32
32
8
16
8
16
16
32
106
209
204
407
794
1586
14.18
24.60
18.91
30.60
39.43
62.91
Another option is the modified shift and add algorithm of Sect. 8.4.2 (Fig. 8.15;
Table 8.12).
In Table 8.13, examples of post correction implementations are reported.
As a last option, several Booth multipliers have been implemented
(Table 8.14).
219

Booth multipliers
LUTs
Delay
8
8
16
16
32
32
8
16
8
16
16
32
188
356
332
628
1172
2276
13.49
25.12
13.68
25.31
25.67
49.09
Table 8.15 Sequential radix4 Booth multipliers
FFs
LUTs
Period
Total time
8
8
16
16
32
32
64
64
9
17
9
17
17
33
33
65
25
34
33
42
58
75
107
140
58
68
125
135
231
248
440
473
2.90
2.90
3.12
3.12
3.48
3.48
4.22
4.22
26.1
49.3
28.1
53.0
59.2
114.8
139.1
274.0
8.6.5 Sequential Multipliers for Integers

Several radix-4 Booth multipliers have been implemented (Fig. 8.17). Both the
clock period Tclk and the total delay (m Tclk) are given (Table 8.15).
8.7 Exercises
1. Generate the VHDL model of a mixed-radix parallel multiplier (Sect. 8.2.4).
2. Synthesize a 2n-bit by 2n-bit parallel multiplier using n-bit by n-bit multipliers
as building blocks.
3. Modify the VHDL model sequential_CSA_multiplier.vhd so that the done flag
is raised when the final result is available (Comment 8.2).
4. Generate the VHDL model of a carry-save multiplier with post correction
(Sect. 8.4.3).
5. Synthesize a sequential multiplier based on Algorithm 8.3.
6. Synthesize a parallel constant multiplier (Sect. 8.5).
7. Generate models of constant multipliers for integers.
8. Synthesize a constant multiplier that computes z = c1y1 ? c2y2 ? ?
csys ? u.
220
8 Multipliers
References
1. Baugh CR, Wooley BA (1973) A twos complement parallel array multiplication algorithm.
IEEE Trans Comput C 31:10451047
2. Booth AD (1951) A signed binary multiplication technique. Q J Mech Appl Mech 4:126140
3. Dadda L (1965) Some schemes for parallel multipliers. Alta Frequenza 34:349356
4. Wallace CS (1964) A suggestion for fast multipliers. IEEE Trans Electron Comput EC13:1417
Chapter 9
Dividers
Division is a basic arithmetic operation whose execution is based on 1-digit by

m-digit multiplications and subtractions. Nevertheless, unlike addition and multiplication, division is generally not included as a predefined block within FPGA
families. So, in many cases, the circuit designer will have to generate dividers by
choosing some division algorithm and implementing it with adders and
multipliers.
The basic digit-recurrence algorithm is described in Sect. 9.1. Among others, it
introduces the Robertson and PD diagrams [1, 2]. Several radix-2 dividers are
proposed in Sect. 9.2: non-restoring and restoring dividers, binary SRT dividers
[2, 3]), and radix-2k dividers. Section 9.3 gives some information about radixB division. As a matter of fact, unless B = 2k, the main and perhaps unique
application of radix-B operations is decimal arithmetic (Chap. 11). The last section
describes two convergence algorithms, namely the Newton-Raphson and Goldschmidt algorithms, which could be considered in the case of real number operations (Chap. 12).
9.1 Basic Digit-Recurrence Algorithm

Consider an integer x and a natural y [ 0 such that -y B x \ y. The quotient
q and the remainder r, with an accuracy of p fractional radix-B digits, are defined
by the following relation:
x Bp q y r; with y r\y:
9:1
In fact, Eq. (9.1) has two solutions: one with r C 0 and another with r \ 0.

221
222
9 Dividers
ri+1
qi+1 =
-B -(B-1) -(B-2)
-2
-2y
-By
-(B-1)y
-1
-y
2y
B-3 B-2
(B-3)y
(B-2)y
(B-1)y
B-1
By
-(B-2)y
Bri
Fig. 9.1 Robertson diagram
The following set of equations generates q and r:

r0 x;
B r0 q1 y r1 ;
B r1 q2 y r2 ;
...
9:2
B rp1 qp y rp :
From (9.2) the following relation is obtained:

x Bp q1 Bp1 q2 Bp2 qp1 B1 qp B0 y
rp ; with y rp \y:
9:3
Thus,
q q1 Bp1 q2 Bp2 qp1 B1 qp B0 and r rp :
At each step of (9.2) ri
? 1
and qi
? 1
are computed in function of y and ri:
ri 1 B ri qi 1 y; where y ri 1 \y:
9:4
The Robertson diagram of Fig. 9.1 defines the set of possible solutions: the dotted
lines define the domain fB ri ; ri 1 j B y B ri \B y and y ri 1 \yg;
and the diagonals correspond to the equations ri 1 B ri k y with k 2
fB; B 1; . . .; 1; 0; 1; . . .; B 1; Bg: If k y B ri \k 1 y; there are
two possible solutions for qi ? 1, namely k and k ? 1. To the first one corresponds
a non-negative value of ri ? 1, and to the second one a negative value.
If the values qi ? 1 = -B and qi ? 1 = B are discarded, the solutions of (9.4)
are the following:
if B ri B 1 y then qi 1 B 1;
if k y B ri \k 1 y then qi 1 k or k 1; 8k in fB 1; . . .; 1; 0;
1; . . .; B 2g;
if B ri \ B 1 y then qi 1 B 1;
so all qis are radix-B digits or negative radix-B digits. The quotient q is obtained
under the form
9.1 Basic Digit-Recurrence Algorithm
223
Bri
(B-1)ymax
qi+1 =
B-1
B-1 or B-2
Bri = (B-1)y
Bri = (B-2)y
3ymax
2ymax
ymax
Bri = 2y
1 or 2
Bri = y
0 or 1
0 or -1
-ymax
-1 or -2
ymax
Bri = -y
Bri = -2y
-2ymax
-3ymax
-(B-1) or -(B-2)
Bri = -(B-2)y
-(B-1)
-(B-1)ymax
Bri = -(B-1)y
Fig. 9.2 P-D diagram
q q1 Bp1 q2 Bp2 qp1 B1 qp B0 ; with qi in

fB 1; . . .; 1; 0; 1; . . .; B 1g;
that is to say a signed-digit radix-B number. According to (9.3)
x=y 0 q1 q2 . . .qp rp Bp =y; with Bp rp Bp =y\Bp :
9:5
In other words, 0: q1 q2 . . .qp is an approximation of x/y with an absolute error

smaller than or equal to B-p.
Another useful diagram is the Partial remainder-Divisor diagram (P-D diagram) of Fig. 9.2. It defines zones of the (Bri, y) plane and the corresponding
solutions for qi ? 1:
224
9 Dividers
beyond the line defined by B ri B 1 y; qi 1 B 1;

betweenthelinesdefinedby B ri k y and B ri k 1 y; qi 1 k or k 1;
below the line defined by B ri B 1 y; qi 1 B 1:
Assume that a function quotient_selection has been defined, compatible with
Figs. 9.1 or 9.2. The following formal algorithm computes q and r.
Algorithm 9.1: Division, digit-recurrence algorithm
9.2 Radix-2 Division

The Robertson and P-D diagram corresponding to B = 2 are shown in Fig. 9.3.
9.2.1 Non-Restoring Divider

According to the diagrams of Fig. 9.3, a possible choice of the quotient_selection
function of Algorithm 1 is qi ? 1 = -1 if ri \ 0 and qi ? 1 = 1 if ri C 0. The
corresponding algorithm is
Algorithm 9.2: Division, non-restoring algorithm, version 1
In this case, a slight modification allows avoiding the conversion of a signed-digit

number to an unsigned-digit one. Let q q1 2p1 q2 2p2 qp1 21
qp 20 ; with qi in {-1, 1}, be the quotient in signed-digit form computed with the
previously defined quotient_selection function. Define
q0i 1 qi 1 =2; 8i in f0; 1; . . . ; p 1g:
225
2ri
ymax
ri+1
qi+1=
-2y
-2
-y
qi+1 = 0 or 1
2y
ymax
2ri
qi+1= -1
qi+1 = 0 or -1
(a)
2ri = y
qi+1 = 1
-ymax
qi+1 = -1
2ri = -y
(b)
Fig. 9.3 Robertson diagram (a) and P-D diagram (b), when B = 2

Then
q0i 2 f0; 1g and q 1 q00 2p q01 2p1 q02 2p2 q0p1
21 20 : In other words, the vector

1 q00 q01 q02 . . .q0p1 1
is the 2s complement representation of q. The corresponding modified algorithm
is the following.
Algorithm 9.3: Division, non-restoring algorithm, version 2

The following VHDL model describes the circuit of Fig. 9.4. The dividend
x xn xn1 . . .x0 and the remainder r rn rn1 . . .r0 are (n ? 1)-bit 2s complement integers, the divisor y yn1 yn2 . . .y0 is an n-bit natural, the quotient
q q0 q1 . . .qp is a (p ? 1)-bit 2s complement integer, and condition
y x\y
must hold. Then 2px = qy ? r, with -y B r \ y, and

x q0 q1 q2 . . .qp y r=y 2p ; with 2p r=y 2p \2p :
9:6
226
9 Dividers
ri-1
y
sign bit
update
qi-1
shift register
.....
addb /sub
adder/
subtractor
ri = 2r i-1
quotient
update
load
register
initially: x
ri-1
Fig. 9.4 Non-restoring divider
The complete circuit also includes a p-state counter and a control unit. A complete
generic model non_restoring.vhd is available at the Authors web page.
227

Tdivider n; p p Tadder n p n Tadder 1:
9:7
Comment 9.1
The least significant bit is always equal to 1 while the sign of the remainder could
be different from that of the dividend. For some applications a final correction step
could be necessary.
9.2.2 Restoring Divider

If x is a natural, so that r0 C 0, the quotient_selection function can be chosen as
follows (Fig. 9.3):
if 2 ri \y then qi 1 0 and ri 1 2 ri ;
if 2 ri y then qi 1 1 and ri 1 2 ri y
So, all along the algorithm execution the remainders ri are non-negative.
Algorithm 9.4: Division, restoring algorithm

x xn1 . . .x0 ; the divisor y yn1 yn2 . . .y0 and the remainder r rn1 . . .r0 are
n-bit naturals, the quotient q q1 q2 . . .qp is a p-bit natural, and the condition
x \ y must hold. Then 2p x q y r; with r \ y, so

x 0 q1 q2 . . . qp y r=y 2p ; with r=y 2p \2p :
228
9 Dividers
ri-1
update
shift register
qi
sign
subtractor
.....
2ri-1
quotient
2ri-1 - y
1
0
ri
update
load
register
initially: x
ri-1
Fig. 9.5 Restoring divider
generic model restoring.vhd is available at the Authors web page.

Fig. 9.6 P-D diagram,
B 2; y ymin
229
2ri
2ri = y
ymax
qi+1 = 1
ymin
qi+1 = 1
qi+1 = 0 or 1
qi+1 = 0
ymin
-ymin
ymax
qi+1 = 0 or -1
qi+1 = -1
qi+1 = -1
-ymax
2ri = -y
The computation time is approximately the same as that of a non-restoring

divider (9.7).
9.2.3 Binary SRT Divider

Assume that a lower bound ymin of y is known. Then, according to the P-D diagram
of Fig. 9.6, the quotient_selection function can be defined as follows:
if 2 ri ymin then qi 1 1;
if ymin 2 ri \ymin then qi 1 0;
if 2 ri \ymin then qi 1 1:
Algorithm 9.5: Division, binary SRT algorithm, version 1
Assume that x is an (n ? 1)-bit 2s complement integer and y a normalized n-bit

natural, that is 2n1 y\2n ; so ymin 2n1 : The condition 2 r ymin is equivalent
to r=2n2 1; and the condition 2 r\ymin to r=2n2 \1: If r rn rn1 rn2
rn3 . . .r0 ; then 2 r ymin if rn rn1 rn2 1; and 2 r\ymin if rn rn1 rn2 \ 1:
In order to convert the signed-bit representation of q to a 2s complement
representation, the following on-the-fly conversion algorithm can be used. Assume
that q q1 21 q2 22 qp 2p ; with qi in {-1, 0, 1}, and define
230
9 Dividers
Fig. 9.7 On-the-fly

conversion
qi
load
update
q (i)
qm (i)
0,1
-1
|qi |
qi
register
initially: 0
q
q (i)
qm (i)
0,-1
load
1-|qi |
register
initially: 1
update
(i)
(i)
qm
Fig. 9.8 SRT divider
r
rn..n-2
quotient
selection
function
operation
adder/
subtractor
2r-y or 2r or 2r+y
load
update
load
on the fly
conversion
update
register
initially: x
r
q
q0 0; qi q1 21 q2 22 qi 2i ; qm0 1; qmi qi 2i ;

Vi in f1; 2; . . .; pg: Then
if qi 1 : qi qi1 2i ; qmi qi1 ;
if qi 0 : qi qi1 ; qmi qi1 2i qi1 2i1 2i qmi1 2i ;
if qi 1 : qi qi1 2i qmi1 2i ; qmi qmi1 :
x xn xn1 . . .x0 and the remainder r rn rn1 . . .r0 are (n ? 1)-bit 2s complement integers, the divisor y 1yn2 . . . y0 is an n-bit normalized natural, the
quotient q q0 q1 . . .qp is a (p ? 1)-bit 2s complement integer, and condition
(9.6) must hold.
231
232
9 Dividers
generic model srt_divider.vhd is available at the Authors web page.
The computation time is approximately the same as that of a non-restoring
divider (9.7).
9.2.4 Binary SRT Divider with Carry-Save Adder

Assume again that a lower bound ymin of y is known and that an integer estimation
of vi 2 ri =ymin satisfying
2 ri =ymin 2\vi 2 ri =ymin
9:8
is computed at each step. The quotient_selection function can be defined as

follows:
if vi 2 then 2 ri =ymin \0; select qi 1 1;
if vi 1 then 1 2 ri =ymin \1; select qi 1 0;
if vi 0 then 2 ri =ymin 0; select qi 1 1:
In the following algorithm the function estimated_value generates an integer vi
belonging to (9.8).
Algorithm 9.6: Division, binary SRT algorithm, version 2
In what follows x is an (n ? 1)-bit 2s complement integer and y a normalized

n-bit natural, that is 2n1 y\2n ; so that ymin 2n1 : Assume that all remainders
ri are represented in stored-carry form, that is ri = ci ? si. Define

vi ci =2n2 si =2n2 :
233
cn+1..n-2
y
sn+1..n-2
adder
v
quotient
selection
function
operation
carry-save
adder/subtractor
2r-y or 2r or 2r+y
(stored-carry form)
load
update
on the fly
conversion
register
initially: 0
register
initially: x
load
update
q
Fig. 9.9 SRT divider with carry-save adder

Thus vi ci =2n2 si =2n2 ri =2n2 and vi [ ci =2n2 1 si =2n2

1 ri =2n2 2; so that (9.8), with ymin = 2n - 1, holds. Lower and upper
bounds of vi are computed as follows:

vi ri =2n2 \2n =2n2 4 and vi [ ri =2n2 2 2n =2n2 2 6:
Thus, vi belongs to the range -5 B vi B 3 and is a 4-bit 2s complement integer.
In order to compute vi, both ci and si must be represented with 4 ? (n - 2) = n ? 2
bits.
x = xn xn 1 x0 and the remainder r = rn rn - 1 r0 are (n ? 1)-bit 2s
complement integers, the divisor y = 1 yn - 2 y0 is an n-bit normalized natural,
the quotient q = q0 q1 qp is a (p ? 1)-bit 2s complement integer, and condition (9.6) must hold.
234
9 Dividers
generic model srt_csa_divider.vhd is available at the Authors web page.
Tdivider n; p p Tadder 1 Tadder n n p Tadder 1:
9:9
9.2.5 Radix-2k SRT Dividers

Another way to accelerate the quotient calculation consists of generating k quotient bits at each step. For that, the digit-recurrence algorithm can be executed in
base 2k. As an example, an SRT divider with k = 2 is defined.
235
4r
n
4r = 3y
32
3
52n-1
2 or 3
2n+1
4r = 2y
32n-1
1 or 2
2n
4r = y
2n-1
0 or 1
2n-1
32n-2
0 or -1
2n
n-1
-2
-2n
4r = -y
-1 or -2
-32n-1
4r = -2y
-2n+1
-2 or -3
-52n-1
-32n
-3
4r = -3y
Fig. 9.10 P-D diagram, radix-4 SRT, y ymin 2n1
As before x = xn xn - 1 x0 is an (n ? 1)-bit 2s complement integer and y

1yn2 . . .y0 is an n-bit normalized natural. The basic algorithm, based on Eq. (9.2)
with B = 4, is used, so that the quotient is obtained under the form q q1 4p1
q2 4p2 qp1 41 qp 40 with qi in {-3, -2, -1, 0, 1, 2, 3}.
According to the P-D diagram of Fig. 9.10, the quotient_selection function can
be defined as follows:
if
if
if
if
if
if
if
if
4 ri 2n1 then qi 1 3;
3 2n1 4 ri \2n1 and y\3 2n2 then qi 1 3;
3 2n1 4 ri \2n1 and y 3 2n2 then qi 1 2;
2n 4 ri \3 2n1 then qi 1 2;
2n1 4 ri \2n then qi 1 1;
2n1 4 ri \2n1 then qi 1 0;
2n 4 ri \ 3 2n1 then qi 1 1;
3 2n1 4 ri \ 2n then qi 1 2;
236
Table 9.1 Digit selection
9 Dividers
rn rn
- 1
rn
- 2
rn
- 3
r/2n
0000
0001
0010
0011
0
1
2
3
0100
0101
0110
0111
1000
1001
1010
1011
1100
4
5
6
7
-8
-7
-6
-5
-4
1101
1110
1111
-3
-2
-1
- 2
q-(i
? 1)
0
1
2
2 if yn - 2 = 1, 3
if yn - 2 = 0
3
3
3
3
-3
-3
-3
-3
-3 if yn - 2 = 0, -2
if yn - 2 = 1
-2
-1
0
if 2n1 4 ri \ 3 2n1 and y 3 2n2 then qi 1 2;

if 2n1 4 ri \ 3 2n1 and y\3 2n2 then qi 1 3;
if 4 ri \ 2n1 then qi 1 3:
Given a remainder r rn rn1 rn2 rn3 . . .r1 r0 ; then the next digit value is a
function of 4 r=2n1 r=2n3 and of y/2n - 2, that is rn rn1 rn2 rn3 and
yn1 yn2 1yn2 : Table 9.1 defines the value of the next digit in function of
rn ; rn1;1 ; rn1;0 ; rn2;1 and yn - 2.
In order to convert the signed-digit representation of q to a 2s complement
representation, the following on-the-fly conversion algorithm can be used. Assume
that q q1 41 q2 42 qp 4p ; with qi in {-3, -2, -1, 0, 1, 2, 3},
and define
q0 0; qi q1 41 q2 42 qi 4i ; qm0 3; qmi qi 4i ;
Vi in {1, 2, , p}. Then
if qi [ 0 : qi qi1 qi 4i ; qmi qi1 qi 1 4i ;
if qi 0 : qi qi1 ; qmi qi1 4i qi1 4i1 3 4i qmi1
3 4i ;
if qi \0 : qi qi1 qi 4i qi1 4 qi 4i 4i1 qmi1
4 qi 4i ;
qmi qi 4i qmi1 3 qi 4i :
237
qi
1,-3
2,-2
3,-1
1,-3
2,-2
3,-1
q(i-1) qm(i-1)
qi 0?
q(i-1) qm(i-1)
qi > 0?
register
initially: 0...00
register
initially: 0...03
q(i-1)
qm(i-1)
load
update
Fig. 9.11 Radix-4 on-the-fly conversion
The complete circuit is shown in Fig. 9.12. The dividend x

xn xn1 xn2 . . .x1 x0 and the remainder r rn rn1 rn2 . . .r1 r0 are (n ? 1)-bit 2s
complement integers, the divisor y 1 yn2 . . .y1 y0 is an n-bit natural, the quotient
q q0 q1 q2 . . .q2p1 q2p is a (2p ? 1)-bit 2s complement integer, and condition
(9.6) must hold.
238
9 Dividers
r
rn..n-3
yn-2
quotient
selection
function
load
update
on the fly
conversion
magnitude
sign
adder/subtractor
load
update
register
initially: x
r
Fig. 9.12 Radix-4 SRT divider
y
2y
3y
239
generic model radix_four_divider.vhd is available at the Authors web page.
9.3 Radix-B Dividers

When B is greater than 2, the quotient_selection function of Algorithm 9.1 is not easy
to implement. According to the Robertson diagram of Fig. 9.1, the definition of
qi ? 1 is based on the knowledge of the interval to which belongs ri: if
ky B Bri \ (k ? 1)y, then qi ? 1 can be chosen equal to k or k ? 1. A conceptually
simple method would consist of computing Bri - ky for all k in {B-1, , -1, 0,
1, , B-1} and choosing qi ? 1 in function of the signs of the results. Better
methods would use approximate values of ri and y, taking advantage of the fact that
for each interval there are two possible values of qi ? 1. In other words, larger
intervals can be considered: if ky B Bri \ (k ? 2)y, then qi ? 1 = k ? 1. Algorithms based on truncated values of ri and y have been proposed in the case where
B = 10 [4, 5], and the method proposed in the second reference can be generalized to
240
9 Dividers
the case of any radix B [6]. Nevertheless, the corresponding FPGA implementations
are generally very costly in terms of used cells.
Another method for executing radix-B division consists of converting the radixB operands to binary operands, executing the division with any type of binary
divider, and converting the binary results to radix-B results. Radix-B to binary and
binary to radix-B converters are described in Chap. 10. Decimal dividers (B = 10)
based on this method are reported in Chap. 12.
An alternative option, similar to the previous one, is to use a classical binary
algorithm (Sect. 9.2) and to execute all the operations in base B. All along this
section it is assumed that B is even.
Consider the following digit-recurrence algorithm (Algorithm 9.1 with B = 2):
Algorithm 9.7: Division, binary digit-recurrence algorithm
It generates q and rp such that

x=y q 2p rp =y 2p ; with 2p rp =y 2p \2p :
9:10
Algorithm 9.7 can be executed whatever the radix used for representing the
numbers. If Bs complement radix-B representation is used, then the following
operations must be available: radix-B doubling, adding, subtracting and halving.
Algorithm 9.8: Division, binary digit-recurrence algorithm, version 2
241
Fig. 9.13 Digit-recurrence

divider
ri-1
ri-1
x2
-y
0 -ulp ulp 0
two_r
quotient
selection
2
ulp
adder
adder
register
initially: x
register
initially: 0
register
initially: 2-1
ri-1
ulp
load
update
The corresponding data path is shown in Fig. 9.13.

Radix-B adders and subtractors can be generated using the general principles
presented in Sects. 7.1 and 7.8. The p and g blocks of Fig. 7.2 compute binary
functions of radix-B variables, namely p = 1 if xi ? yi = B-1, p = 0 otherwise,
g = 1 if xi ? yi C B, g = 0 if xi ? yi B B-2, otherwise, any value. The carry
chain is a binary circuit, whatever the radix. An output block computes
(xi ? yi ? ci) mod B. Furthermore, in the case of a subtractor, additional (B-1)s
complement cells, are necessary. Decimal adders and subtractors (B = 10) are
described and implemented in Chap. 12. In order to synthesize the circuit of
Fig. 9.13, it remains to generate circuits which compute 2x and x/2.
Consider a radix-B natural x xn1 Bn1 xn2 Bn2 x0 B0 : First define
a digit-doubling cell: given a radix-B digit xi, the cell computes the radix-B representation of 2xi. As xi B B-1, then 2xi 2B 2 B B 2; so that 2xi can
be expressed under the form
2xi di B ui ; where di f0; 1g and ui B 2if B is even; so is ui :
Then, 2x can be expressed under the form
2x dn1 Bn un1 dn2 Bn1 un2 dn3 Bn2 u1 d0 B1
u 0 B0 :
As all coefficients di belong to {0, 1} and all coefficients ui are even, then all
sums
ui ? di - 1 amount to a simple concatenation, that is ui di1 uj =2 &di1 ;
and no carry is generated. The corresponding circuit is shown in Fig. 9.14.
Assume that radix-B digits are represented as k-bit naturals and that k-input
LUTs are available. Then, every cell of Fig. 9.14 can be implemented with
k LUT-k. As an example, if B = 10, the following model defines the doubling cell
of Fig. 9.14.
242
9 Dividers
xn-1
x2
x1
x0
x2
x2
x2
x2
un-1/2
zn
dn-2
zn-1
....
u2/2
d1 u1/2
z2
cin=0
d0 u0/2
z1
z0
Fig. 9.14 Doubling circuit
The complete circuit is defined by instantiating n doubling cells. Its computation

time is equal to TLUT-k. A complete model doubling_circuit2.vhd is available at the
Authors web page (B = 10).
As B is even, halving is equivalent to multiplying by B/2 and dividing by B. Let
ak1 ak2 . . .a0 be the binary representation of a radix-B digit a, that is a
ak1 2k1 ak2 2k2 a0 20 : Then
aB=2 dB u; where d ak1 2k2 ak2 2k3 a1 20 and u
a0 B=2:
Observe that d B a/2 \ B/2 and u B B/2.
Consider now a radix-B natural x xn1 Bn1 xn2 Bn2 x0 B0 :
The product (B/2)xi can be represented under the form diB ? ui where di \ B/2
and ui B B/2, so that
B=2 x dn1 Bn un1 dn2 Bn1 un2 dn3 Bn2
u1 d0 B1 u0 B0 :
All sums ui ? di
- 1
are smaller than B, so that no carry is generated. Finally
243
xn-1 0
xn-1 k-1...xn-1 1
x10
xn-2 k-1...xn-2 1
conn.
conn.
xn-2 0B/2
zn
x00
x0 k-1...x02x01
conn.
x10B/2
x00B/2
(k-1)-bit
adder
(k-1)-bit
adder
zn-1
z1
z0
Fig. 9.15 Multiplication by B/2 (conn. only includes connections)
x=2 dn1 Bn1 un1 dn2 Bn2 un2 dn3 Bn3

u1 d0 B0 u0 B1 :
The circuit of Fig. 9.15 computes z = (B/2)x.
If k-input LUTs are available, then every cell of Fig. 9.15 can be implemented
with k LUT-k. As an example, if B = 10, the following model defines the multiply-by-five cell of Fig. 9.15. It computes b 5a0 4a3 2a2 a1 :
The complete circuit is defined by instantiating n - 1 multiply-by-five cells. Its

computation time is equal to TLUT-k. A complete model multiply_by_five.vhd is
available at the Authors web page (B = 10).
Example 9.1
A decimal divider (B = 10) based on the circuit of Fig. 9.13 can be synthesized. The
quotient selection is done according to the non-restoring algorithm method, that is to
say ri ? 1 = 2ri ? y if ri is negative, and ri ? 1 = 2ri - y if ri is non-negative. The
number p of steps is defined as follows: according to 9.10, the maximum value of the
error | x/y - q| is smaller than or equal to 2-p. If q is rounded in such a way that
the final result qrounded has m fractional digits, the additional rounding error is smaller
than 10-m, and the final error jx=y qrounded j is smaller than 2-p ? 10-m. By
choosing p so that 2-p % 10-m, that is p % 3.3 m, then jx=y qrounded j\2 10m :
In the following VHDL models, behavioral descriptions of the decimal adders have
been used. Real implementations are proposed in Chap. 12.
244
9 Dividers
245
A complete generic model decimal_divider.vhd is available at the Authors web

page.
9.4 Convergence Algorithms

Instead of using digit-recurrence algorithms, an alternative option is the use of
convergence algorithms. This type of algorithm is mainly attractive in the case of
real number operations (Chap. 12).
Given a real number x belonging to the interval 1 B x \ 2, the inverse y of
x belongs to the interval 1/2 \ y B 1. The following Newton-Raphson algorithm
generates successive approximations y0 ; y1 ; y2 ; . . . of y 1=x:
Algorithm 9.9: Newton-Raphson inversion algorithm
If all operations were executed with full precision, then the maximum error after
i steps would be given by the following relation:
i
y yi \1=22 :
9:11
Another option is the Goldschmidt algorithm. Given two real numbers x and
y belonging to the intervals 1 B x \ 2 and 1 B y \ 2, it generates two sequences
of real numbers a0, a1, a2, and b0, b1, b2, , and the second one constitutes a
sequence of successive approximations of x/y.
Algorithm 9.10: Goldschmidts algorithm
If all operations were executed with full precision, then the maximum error after
i steps would be defined by the following relation:
246
9 Dividers
x
i
bi \1=22 1 :
y
9:12
The following VHDL model describes the corresponding circuit; n is the number
of fractional bits of x and y, and p is the number of fractional bits of the internal
variables a, b and c.
A complete generic model goldschmidt.vhd is available at the Authors web page.

It includes an m-state counter, where m is the number of steps (a generic
parameter), and a control unit.

Several dividers have been implemented within a Virtex 5-2 device. As before, the
times are expressed in ns and the costs in numbers of Look Up Tables (LUTs) and
flip-flops (FFs). All VHDL models as well as several test benches are available at
247
Table 9.2 Non-restoring

dividers
FFs
LUTs
Period
Total time
8
16
24
32
32
53
64
64
8
16
27
16
32
56
32
64
22
39
59
55
72
118
104
137
28
46
64
78
79
123
143
144
2.21
2.43
2.63
2.82
2.82
3.30
3.55
3.55
19.9
41.3
73.6
47.9
93.1
188.1
117.2
230.8
Table 9.3 Restoring dividers
FFs
LUTs
Period
Total time
8
16
24
32
32
53
64
64
8
16
27
16
32
56
32
64
21
38
58
54
71
117
103
136
26
44
62
76
77
121
141
142
2.46
2.68
2.90
3.08
3.08
3.56
3.82
3.82
22.1
45.6
81.2
52.4
101.6
202.9
126.1
248.3
FFs
LUTs
Period
Total time
8
16
24
32
32
53
64
64
8
16
27
16
32
56
32
64
33
58
89
75
107
177
139
204
44
78
118
110
143
235
207
272
2.43
2.57
2.76
2.94
2.94
3.42
3.66
3.66
21.9
43.7
77.3
50.0
97.0
194.9
120.8
237.9
Table 9.4 Binary SRT

dividers
9.5.1 Digit-Recurrence Algorithms

Tables 9.2, 9.3, 9.4, 9.5 give implementation results for several
non-restoring dividers,
restoring dividers,
binary SRT dividers,
binary SRT dividers with carry-save adder.
Radix-4 SRT dividers have also been implemented. The quotient is a (2p ? 1)-bit
2s complement integer (Table 9.6).
Examples of decimal dividers are given in Chap. 11.
248
Table 9.5 Binary SRT
dividers with carry-save
adders
Table 9.6 Radix-4 SRT

dividers
Table 9.7 Dividers based on

the Goldschmidts algorithm
9 Dividers
n
FFs
LUTs
Period
Total time
8
16
24
32
32
53
64
64
8
16
27
16
32
56
32
64
40
73
112
105
138
229
202
267
86
143
215
239
272
448
464
529
2.81
2.86
2.86
2.86
2.86
2.87
2.87
2.87
25.3
48.6
80.1
48.6
94.4
163.6
94.7
186.6
2p
FFs
LUTs
Period
Total time
8
16
24
32
32
53
64
64
4
8
14
8
16
28
16
32
8
16
28
16
32
56
32
64
30
55
88
71
104
174
136
201
70
112
170
175
209
343
337
402
3.53
3.86
4.08
4.27
4.27
4.76
5.01
5.01
17.7
34.7
61.2
38.4
72.6
138.0
85.2
165.3
FFs
LUTs
DSPs
Period
Total time
8
8
16
16
24
53
10
16
18
32
27
56
4
5
5
7
6
8
25
38
43
71
62
118
38
42
119
76
89
441
2
2
2
8
8
24
5.76
5.68
7.3
10.5
10.5
14.42
23.0
28.4
36.5
73.5
63.0
115.4
9.5.2 Convergence Algorithms

Table 9.7 gives implementation results of several dividers based on the Goldschmidts algorithm. In this case, DSP slices have been used.
9.6 Exercises
1. Synthesize several types of combinational dividers: restoring, non-restoring,
SRT, radix-2k.
2. Generate a generic VHDL model of a Newton-Raphson inverter.
9.6 Exercises
249
3. Synthesize several dividers using a Newton-Raphson inverter and different

types of multipliers.
4. Generate generic models of combinational circuits with three n-bit inputs x,
y and m, and an n-bit output z = xy mod m. For that, use any type of combinational multiplier to compute p = xy and any type of combinational divider
to compute z = p mod m.
5. Generate generic models of sequential circuits that implement the same function as in exercise 4.
References
1. Freiman CV (1961) Statistical analysis of certain binary division algorithms. IRE Proc
49:91103
2. Robertson JE (1958) A new class of division methods. IRE Trans Electron Comput
EC-7:218222
3. Cocke J, Sweeney DW (1957) High speed arithmetic in a parallel device. IBM technical
report, Feb 1957
4. Deschamps JP, Sutter G (2010) Decimal division: algorithms and FPGA implementations.
In: 6th southern conference on programmable logic (SPL)
5. Lang T, Nannarelli A (2007) A radix-10 digit-recurrence division unit: algorithm and
architecture. IEEE Trans Comput 56(6):113
6. Deschamps JP (2010) A radix-B divider. Contact: [email protected]
Chapter 10
Other Operations
This chapter is devoted to arithmetic functions and operations other than the four
basic ones. The conversion of binary numbers to radix-B ones, and conversely, is
dealt with in Sects. 10.1 and 10.2. An important particular case is B = 10 as human
interfaces generally use decimal representations while internal computations are
performed with binary circuits. In Sect. 10.3, several square rooting circuits are
presented, based on digit-recurrence or convergence algorithms. Logarithms and
exponentials are the topics of Sects. 10.4 and 10.5. Finally, the computation of
trigonometric functions, based on the CORDIC algorithm [2, 3], is described in
Sect. 10.6.
10.1 Binary to Radix-B Conversion (B even)

Assume that B is even and greater than 2. Consider the binary representation
of x:
x xn1 2n1 xn2 2n2 x1 2 x0 :
10:1
Taking into account that B [ 2, the bits xi can be considered as radix-B digits,
and a simple conversion method consists of computing (10.1) in radix-B.
Algorithm 10.1: Binary to radix-B conversion

251
252
Fig. 10.1 Binary to
radix-B converter
10
zm-1..0
Fig.9.14
Other Operations
x n-1..0 (binary )
c in
x n-i
shift register (binary)
load
update
wm-1..0
register (radix B)
initially: 0
load
update
ym-1..0 (base B)
In order to compute z2 + xni the circuit of Fig. 9.14, with cin xni instead of 0,
can be used. A sequential binary to radix-B converter is shown in Fig. 10.1. It is
described by the following VHDL model.
The complete circuit also includes an n-state counter and a control unit.
A complete model BinaryToDecimal2.vhd is available at the Authors web page
(B = 10).
The computation time of the circuit of Fig. 10.1 is equal to nTclk where Tclk
must be greater than TLUT-k (Sect. 9.3). Thus
TbinaryradixB n TLUTk :
10:2
10.2 Radix-B to Binary Conversion (B even)

Given a natural z, smaller than 2n, its binary representation is deduced from the
following set of integer divisions
z q1 2 r 0 ;
q1 q2 2 r 1 ;
...
qn1 qn 2 rn1 ;
10.2
Radix-B to Binary Conversion (B even)
Fig. 10.2 Decimal to binary

converter
253
y m-1..0
Fig.9.15
wm..1
w0
shift register (binary)
register (radix -B)

initially: z
load
update
load
update
xn-1..0 (binary )
so
z qn 2n rn1 2n1 rn2 2n2 r1 2 r0 :
As z is smaller than 2n, then qn = 0, and the binary representation of z is
constituted by the set of remainders rn1 rn2 . . . r1 r0 :
Algorithm 10.2: Radix-B to binary conversion
Observe that if qi = qi+12 + ri, then qi(B/2) = qi+1B + ri(B/2) where ri(B/2) \
2(B/2) = B.
Algorithm 10.3: Radix-B to binary conversion, version 2
In order to compute (B/2)qi the circuit of Fig. 9.15 can be used. A sequential
radix-B to binary converter is shown in Fig. 10.2. It is described by the following
VHDL model.
254
10
Other Operations
The complete circuit also includes an n-state counter and a control unit.
A complete model DecimalToBinary2.vhd is available at the Authors web page
(B = 10).
The computation time of the circuit of Fig. 10.2 is equal to nTclk where Tclk
must be greater than TLUT-k (Sect. 9.3). Thus
TradixBbinary n TLUTk :
10:3
10.3 Square Rooters

Consider a 2n-bit natural X x2n1 22n1 x2n2 22n2 x1 2 x0 ;

and compute Q X 1=2 : Thus, Q2 B X \ (Q+1)2, and the difference R = X - Q2
belongs to the range
0 R 2Q:
10:4
10.3.1 Restoring Algorithm

A digit recurrence algorithm consisting of n steps is defined. At each step two
numbers are generated:
Qi qn1 2i1 qn2 2i2 qni1 2 qni and Ri
2
X Qi 2ni ;
10.3
Square Rooters
255
such that
0 Ri \ 1 2Qi 22ni :
10:5
After n steps, 0 B Rn \ 1 + 2Qn, that is (10.4) with Q = Qn and R = Rn.

Initially define Q0 = 0 and R0 = X, so condition (10.5) amounts to 0 B X \ 22n.
Then, at step i, compute Qi and Ri in function of Qi-1 and Ri-1:
Qi 2Qi1 qni where qni 2 f0; 1g;
2
2
Ri X Qi 2ni X 2Qi1 qni 2ni
2
X Qi1 2ni1 qni 4Qi1 qni 22ni Ri1 qni 4Qi1 qni 22ni :
The value of qn-i is chosen in such a way that condition (10.5) holds. Consider
two cases:
If Ri1 \1 4Qi1 22ni ; then qni 0; Qi 2Qi1 ; Ri Ri1 :
As Ri Ri1 \1 4Qi1 22ni 1 2Qi 22ni and Ri Ri1 0;
condition (10.5) holds.
If Ri1 1 4Qi1 22ni ; then qni 1; Qi 2Qi1 1; Ri Ri1
1 4Qi1 22ni ; so that Ri 0 and Ri \1 2Qi1 22ni1 1 4Qi1
22ni 3 4Qi1 22ni 1 2Qi 22ni :
Algorithm 10.4: Square root, restoring algorithm
Qi is an i-bit number, Ri \ 1 2Qi 22ni Qi & 1& 0 0 0 is a (2n-i+1)bit number, and Pi1 1 4Qi1 22ni Qi & 01& 00 0 a 2n 2 ibit number.
256
10
r 3n..2n-2
n-bit shift register

initially: 0
sign
Other Operations
Q&01
(n+3)-bit
subtractor
r3n..2n-2
difn..0
Q
1
r2n-3..0&00
(3n+1)-bit register
initially: 000.x 2n-1x0
Fig. 10.3 Square root computation: data path
An equivalent algorithm is obtained if Pi and Ri are replaced by pi1

Pi1 =22ni and ri Ri =22ni :
Algorithm 10.5: Square root, restoring algorithm, version 2
As before, Qi is an i-bit number, ri is a (2n-i+1)-bit fixed-point number

with 2(n-i) fractional bits and an (i+1)-bit integer part, and pi-1 a (2ni+2)-bit fixed-point number with 2(n-i) fractional bits and an (i+2)-bit integer
part.
A sequential implementation is shown in Fig. 10.3. It can be described by the
10.3
Square Rooters
257
The only computational resource is an (n+3)-bit subtractor so that the computation time is approximately equal to nTadder(n). The complete circuit includes
an n-bit counter and a control unit. A generic VHDL model SquareRoot.vhd is
Another equivalent algorithm is obtained if Pi, Qi and Ri are replaced by
pi Pi =22ni1 ; qi Qi =2i ; ri Ri =22ni :
Algorithm 10.6: Square root, restoring algorithm, version 3
Algorithm 10.6 is similar to the restoring algorithm defined in Chap. 21 of [1].

Its implementation is left as an exercise.
258
10
Other Operations
10.3.2 Non-Restoring Algorithm

Instead of computing Ri and Qi as in Algorithm 10.4, an alternative option is the
following. Define Ri Ri1 1 4Qi1 22ni ; whatever the sign of Ri. If Ri is
non-negative, then its value is the same as before. If Ri is negative then it is equal
to Ri restoring 1 4Qi1 22ni where Ri restoring is the value that would have
been computed with Algorithm 10.4. Then, at the next step, Qi and Ri+1 are
computed as follows:
if Ri is non-negative, then Qi 2Qi1 1 and Ri1 Ri 1 4Qi 22ni1 ;
if Ri is negative, then Qi 2Qi1 and Ri1 Ri restorig 1 4Qi 22ni1
Ri 1 4Qi1 22ni 1 4Qi 22ni1 Ri 1 2Qi 22ni
1 4Qi 22ni1 Ri 3 4Qi 22ni1 :
Algorithm 10.7: Square root, non-restoring algorithm
Qi is an i-bit number and Ri is an (i+2)-bit signed number.

An equivalent algorithm is obtained if Pi and Ri are replaced by
pi1 Pi1 =22ni ; ri Ri =22ni :
Algorithm 10.8: Square root, non-restoring algorithm, version 2
As before, Qi is an i-bit number and ri a (2n+1)-bit fixed-point number a2n

a2n1 a2n2 . . . a0 initially equal to 0:x2n1 x2n2 . . . x0 :
10.3
Square Rooters
259
r3n+1..2n-2
r3n+1
n-bit shift register
initially: 0
Q&r3n+11
(n+2)-bit
adder/subtractor
sign
r2n-3..0&00
sumdifn+1..0
(3n+2)-bit register
initially: 000.x 2n-1x0
Fig. 10.4 Square root computation: non-restoring algorithm
In this case rn2n is the remainder only if rn is non-negative. In fact, the

remainder is equal to (rn-i4i)2n where rn-i is the last non-negative remainder.
A sequential implementation is shown in Fig. 10.4. It can be described by the
The only computation resource is an (n+2)-bit adder/subtractor so that the

computation time is again approximately equal to nTadder(n). The complete circuit
includes an n-bit counter and a control unit. A generic VHDL model SquareRoot3.vhd is available at the Authors web page.
260
10
Other Operations
10.3.3 Fractional Numbers

Assume that X is a 2(n+p)-bit fractional number x2n-1 x2n-2 x1 x0. x-1 x-2 x-2p.
The square root Q of X, with an accuracy of p fractional bits, is defined as follows:
Q qn1 2n1 qn2 22n2 . . . q0 q1 21 q2 22 . . . qp 2p ;
Q2 X and Q 2p 2 [ X;
so that the remainder R = X - Q2 belongs to the range 0 B R\21-pQ + 2-2p, that
is to say
0 R Q 21p :
In order to compute Q, first substitute X by X0 = X22p, which is a natural, and
then compute the square root Q0 qnp1 qnp2 . . . q1 q0 of X0 , and the corresponding remainder R0 rnp rnp1 . . . r1 r0 ; using for that one of the previously
defined algorithms. Thus, X0 = (Q0 )2 + R0 , with 0 B R0 B 2Q0 , so that X
Q0 2p 2 R0 22p ; with 0 B R0 2-2p B 2Q0 2-2p. Finally, define Q = Q0 2-p
and R = R0 2-2p. Thus
X = Q2 + R, with 0 B R B Q21-p,
where Q qnp1 qnp2 . . . qp qp1 . . . q1 q0
r2p1 . . . r1 r0 is smaller than or equal to Q21-p.
and
R rnp rnp1 . . . r2p
Comment 10.1
The previous method can also be used for computing the square root of a natural
X x2n1 x2n2 . . . x1 x0 with an accuracy of p bits: represent X as an (n+p)-bit
fractional number x2n1 x2n2 . . . x1 x0 00 . . . 0 and use the preceding method.
10.3.4 Convergence Methods (NewtonRaphson)

Instead of a digit-recurrence algorithm, an alternative option is the NewtonRaphson
convergence method. The following iteration can be used for computing X1/2
xi1 1=2 xi X=xi :
It corresponds to the graphical construction of Fig. 10.5.
First check that (1/2)(x + X/x) is a function whose minimum value, within the
half plane x [ 0, is equal to X1/2, and is obtained when xi = X1/2. Thus, whatever
the initial value x0, xi is greater than or equal to X1/2 for all i[0. Furthermore, if xi[X1/2

then xi1 \1=2 xi X=X 1=2 1=2 xi X 1=2 \1=2 xi xi xi :
Thus, either X 1=2 \xi1 \xi or X 1=2 xi1 xi : For x0 choose a first rough
approximation of X1/2. As regards the computation of X/xi, observe that if xi C X1/2
and X \ 22n, then xi 2n [ X 1=2 X 1=2 X: So, compute q % X/(xi2n), with an
10.3
Square Rooters
261
Fig. 10.5 NewtonRaphson

method:
computation of X1/2
f (x ) = x 2-X
-X
1/2
x i+1 x i
accuracy of p+n fractional bits, using any division algorithm, so that X 2np
q xi 2n r; with r\xi2n, and X Q xi R; where Q = q2-p and R = (r/xi)
2-(n+p) \2-p.
An example of implementation SquareRootNR4.vhd is available at the Authors
web page. The corresponding data path is shown in Fig. 10.6. The initial value x0
must be defined in such a way that x02n [X. In the preceding example, initial_y =
x0 is defined so that initial_y(n+p n+p-4)2-2 is an approximation of the square
root of X2n-1 2n-4 and that initial_y2n is greater than X.
table_x0 is a constant array defined within a user package:
The end of computation is detected when xi+1 = xi.
262
10
Other Operations
table
divider
start_ div
div_ done
x0
adder
x i+1
(n+p+1)-bit register
initially:x 0
load
update
xi
Fig. 10.6 NewtonRaphson method: data path
Comment 10.2
Every iteration step includes a division, an operation whose complexity is similar
to that of a complete square root computation using a digit recurrence algorithm.
Thus, this type of circuit is generally not time effective.
Another method is to first compute X-1/2. A final multiplication computes
X1/2 = X-1/2 X. The following iteration can be used for computing X-1/2

xi1 xi =2 3 x2i X ;
where the initial value x0 belongs to the range 0 \ x0 B X-1/2. The corresponding
graphical construction is shown in Fig. 10.7.
The corresponding circuit does not include dividers, only multipliers and an adder.
The implementation of this second convergence algorithm is left as an exercise.
10.4 Logarithm
Given an n-bit normalized fractional number x = 1x-1 x-2 x-n, compute y =
log2x with an accuracy of p fractional bits. As x belongs to the interval 1 B x\2, its
base-2 logarithm is a non-negative number smaller than 1, so y = 0.y-1 y-2 y-p.
If y = log2x, then x 20 y1 y2 ...yp ... ; so that x2 2y1 y2 ... yp ... : Thus
if x2 2 : y1 1 and x2 =2 20:y2 yp ;
if x2 \2 : y1 0 and x2 20:y2 yp :
10.4
Logarithm
263
Fig. 10.7 NewtonRaphson

method: computation of X-1/2
f(x) = 1/x 2-X
X -1/2
xi
x i+1
-X
Fig. 10.8 Logarithm: data path

squaring
m.s.b. z 2
z 2/2
load
update
p-bit shift register
log
load
update
(n+1)-bit register
initially:x
z
The following algorithm computes y:

Algorithm 10.9: Base-2 logarithm
The preceding algorithm can be executed by the data path of Fig. 10.8 to which
corresponds the following VHDL model.
264
10
Other Operations
A complete VHDL model Logarithm.vhd is available at the Authors web page.

Comments 10.3
1. If x belongs to the interval 2n1 x\2n ; then it can be expressed under the
form x 2n1 y where 1 B y \ 2, so that log2x = n -1 + log2y.
2. If the logarithm in another base, say b, must be computed, then the following
relation can be used: logb x log2 x=log2 b:
10.5 Exponential
Given an n-bit fractional number x 0:x1 x2 . . . xn ; compute y = 2x with an
accuracy of p fractional bits. As x x1 21 x2 22 . . . xn 2n ; then
1 x1 2 x2
n xn
2x 22
22
. . . 22
:
i
If all the constant values ai 22 are computed in advance, then the following
algorithm computes 2x.
Algorithm 10.10: Exponential 2x
The preceding algorithm can be executed by the data path of Fig. 10.9.
The problem is accuracy. Assume that all ais are computed with m fractional
bits so that the actual operand ai is equal to ai - ei, where ei belongs to the range
0 ei \2m :
10:6
Consider the worst case, that is y = 20.11 1. Then the obtained value is
y a1 e1 a2 e2 . . . an en : If second or higher order products ei ej ek
are not taken into account, then y0 % y - (e1 a2 an + a1 e2 an + a1 an-1 en).
As all products p1 = a2 an, p2 = a1 a3 an, etc., belong to the range 1 \ pi \ 2,
and ei to (10.6), then
0
y y0 \2 n 2m :
10:7
10.5
Exponential
265
ai
table
multiplier
x -i
shift-register
load
load
update
parallel register
initially:1.000
update
Fig. 10.9 Exponential: data path
Relation (10.7) would define the maximum error if all products were computed
exactly, but it is not the case. At each step the obtained product is rounded. Thus
Algorithm 10.8 successively computes
0
z 2 [ a1 a2 2m ;
0

0
0
0
0
0
0
z3 [ a1 a2 2m a3 2m a1 a2 a3 2m 1 a3 ;
0

0
0
0
0
0
0
0
0
0
0
0
a4 2m a1 a2 a3 a4 2m 1 a4 a3 a4 ;
z4 [ a1 a2 a3 2m 1 a3
and so on. Finally

0
0
0
0
0
0
0
zn [ y 2m 1 an an1 an a3 a4 an
0
[ y 2m 1 2n 2 [ y 2 n 2m :
10:8
Thus, from (10.7) and (10.8), the maximum error y - zn is smaller than 4n2-m.
In order to obtain the result y with p fractional bits, the following relation must hold
true: 4n2-m- B 2-p, and thus
m p log2 n 2:
10:9
As an example, with n = 8 and p = 12, the internal data must be computed with
m = 17 fractional bits.
266
10
Other Operations
powers is a constant array defined within a user package; it stores the fractional
part of ai with 24 bits:
A complete VHDL model Exponential.vhd is available at the Authors web

page.
Instead of storing the constants ai, an alternative solution is to store an, and to
compute the other values on the fly:
i 2
i1
ai1 22
22
a2i .
Algorithm 10.11: Exponential 2 x, version 2
The preceding algorithm can be executed by the data path of Fig. 10.9 in which
the table is substituted by the circuit of Fig. 10.10.
Once again, the problem is accuracy. In this case there is an additional problem:
in order to get all coefficients ai with an accuracy of m fractional bits, they must be
computed with an accuracy of k [ m bits. Algorithm 10.11 successively computes
10.5
Exponential
267
Fig. 10.10 Computation of

ai on the fly
ai
multiplier
a i2
load
parallel register
initially:an
update
2
0
0
an [ an 2k ; an1 [ an 2k 2k a2n 2an 2k 2k an1 2k
2
0
1 2an ; an2 [ an1 2k 1 2an 2k a2n1 2an1 2k 1 2an
0
2k an2 2k 1 2an1 4an1 an ; an3 [ an2 2k
1 2an1 4an1 an 2 2k a2n2 2an2 2k 1 2an1 4an1 an 2k
an3 2k 1 2an2 4an2 an1 8an2 an1 an ;
and so on. Finally
0
a1 [ a1 2k 1 2a2 4a2 a3 2n2 a2 a3 . . . an

[ a1 2k 1 4 8 2n1
a1 2k 2n 3:
In conclusion, a1 - a10 \ 2-k(2n - 3) \ 2n-k. The maximum error is smaller
than 2-m if n-k B -m, that is k C n + m. Thus, according to (10.9)
k n p log2 n 2:
As an example, with n = 8 and p = 8, the coefficients ai (Fig. 10.10) are
computed with k = 21 fractional bits and z (Fig. 10.9) with 13 fractional bits.
A complete VHDL model Exponential2.vhd, in which an, expressed with
k fractional bits, is a generic parameter, is available at the Authors web page.
Comment 10.3
Given an n-bit fractional number x and a number b [2, the computation of y = bx,
with an accuracy of p fractional bits, can be performed with Algorithm 10.10 if the
constants ai are defined as follows:
i
ai b2 :
So, the circuit is the same, but for the definition of the table which stores the
constants ai. In particular, it can be used for computing ex or 10x.
268
10
Other Operations
Fig. 10.11 Pseudo-rotation

yi+1
/2
yi
i
xi+1
xi
10.6 Trigonometric Functions

A digit-recurrence algorithm for computing ejz = cos z + jsin z, with 0 B z B p/2,
similar to Algorithm 10.10 can be defined. In the modified version, the operations
are performed over the complex field.
Algorithm 10.12: Exponential ejz ; z z0 z1 z2 . . . zn
The constants aRi and aIi are equal to cos 2-i and sin 2-i, respectively. The
synthesis of the corresponding circuit is left as an exercise.
A more efficient algorithm, which does not include multiplications, is CORDIC
[2, 3]. It is a convergence method based on the graphical construction of
Fig. 10.11. Given a vector (xi, yi), then a pseudo-rotation by ai radians defines a
rotated vector (xi+1, yi+1) where
xi1 xi yi tan ai xi cos ai yi sin ai 1 tan2 ai 0:5 ;
yi1 yi xi tan ai yi cos ai xi sin ai 1 tan2 ai 0:5 :
In the previous relations, xicos ai - yisin ai and yicos ai + xisin ai define the
vector obtained after a (true) rotation by ai radians. Therefore, if an initial vector (x0, y0) is rotated by successive angles a0, a1, , an-1, then the final vector is
(xn, yn) where
10.6
Trigonometric Functions
269
shifter
+/-
dm
shifter
-1
+/-
dm
-i
tan (2 )
+/-
register
init.: x 0
register
init.: 0
register
init.: z
table
dm
Fig. 10.12 Data path executing CORDIC
xn x0 cosa y0 sina1 tan2 a0 0:5 1 tan2 a1 0:5 . . . 1 tan2 an1 0:5 ;

yn y0 cosa x0 sina1 tan2 a0 0:5 1 tan2 a1 0:5 . . .1 tan2 an1 0:5 ;
with a = a0 + a1 + + an-1. The CORDIC method consists in choosing angles ai
such that
xi - yitan ai and yi + xitan ai are easy to calculate,
a0 + a1 + a2 + tends toward z.
As regards the first condition, ai is chosen as follows: ai = tan-12-i or -tan-12-i, so
that tan ai = 2-i or -2-i. In order to satisfy the second condition, ai is chosen in
function of the difference di = z - (a0 + a1 + + ai-1): if di \0, then ai = tan-12-i;
else ai = -tan-12-i. The initial values are
x0 1=k; where k 1 10:5 1 22 0:5 . . . 1 22n1 0:5 ; y0 0; d0 z:
Thus, if z a0 a1 . . . an1 ; then xn % cos z and yn % sin z. It can be
shown that the error is less than 2-n. In the following algorithm, x0 has been
computed with n = 16 and 32 fractional bits.
270
10
Other Operations
Algorithm 10.13: CORDIC, x = cos z, y = sin z
A circuit for executing Algorithm 10.13 is shown in Fig. 10.12. It can be

described by the following VDL model.
angles is a constant array defined within a user package; it stores tan-12-i, for i up
to 15, with 32 bits:
10.6
Trigonometric Functions
271
shifter is a previously defined component that computes b = a2-shift. A complete

VHDL model cordic2.vhd is available at the Authors web page. It includes an nstate counter, which generates the index i of Algorithm 10.13, and a control unit.
CORDIC can be used for computing other functions. In fact, Algorithm 10.13 is
based on circular CORDIC rotations, defined in such a way that the difference
di z a0 a1 . . . ai1 tends to 0. Another CORDIC mode, called circular vectoring, can be used. As an example, assume that at each step the value of
ai is chosen in such a way that yi tends toward 0: if sign(xi) = sign(yi), then ai = tan-12-i; else, ai = tan-12-i. Thus, if yn % 0, then xn is the length of the initial
vector multiplied by k. The following algorithm computes (x2 + y2)0.5.
Algorithm 10.14: CORDIC, z = (x2 + y2)0.5
272
Table 10.1 Binary to
decimal converters
10
Other Operations
FFs
LUTs
Period
Total time
8
16
24
32
48
64
3
5
8
10
15
20
27
43
54
82
119
155
29
45
56
82
119
155
1.73
1.91
1.91
1.83
1.83
1.83
15.6
32.5
47.8
60.4
89.7
119.0
Table 10.2 Decimal-to-binary converters
FFs
LUTs
Period
Total time
8
16
24
32
48
64
3
5
8
10
15
20
26
43
65
81
118
154
22
30
43
51
72
92
1.80
1.84
1.87
1.87
1.87
1.87
16.2
31.3
46.8
61.7
91.6
121.6
A complete VHDL model norm_cordic.vhd corresponding to the previous

aslgorithm is available at the Authors web page.

Several circuits have been implemented within a Virtex 5-2 device. The times are
expressed in ns and the costs in numbers of Look Up Tables (LUTs) and flip-flops
(FFs). All VHDL models are available at the Authors web page.
10.7.1 Converters
Table 10.1 gives implementation results of several binary-to-decimal converters.
They convert n-bit numbers to m-digit numbers.
In the case of decimal-to-binary converters, the implementation results are
given in Table 10.2.
10.7
FPGA Implementations
273
Table 10.3 Square rooters:

restoring algorithm
FFs
LUTs
Period
Total time
8
16
24
32
38
71
104
136
45
79
113
144
2.57
2.79
3.00
3.18
20.6
44.6
72.0
101.8
Table 10.4 Square rooters:

non-restoring algorithm
FFs
LUTs
Period
Total time
8
16
24
32
39
72
105
137
39
62
88
111
2.61
2.80
2.98
3.16
20.9
44.8
71.5
101.1
Table 10.5 Square rooter:

NewtonRaphson method
FFs
LUTs
Period
8
8
8
16
16
32
32
0
4
8
8
16
16
32
42
51
59
92
108
173
205
67
78
90
135
160
249
301
2.94
3.50
3.57
3.78
3.92
4.35
4.67
Table 10.6 Base-2

logarithm
FFs
LUTs
DSPs
Period
Total time
8
16
24
32
10
18
27
36
16
25
59
44
20
29
109
46
1
1
2
4
4.59
4.59
7.80
9.60
45.9
82.6
210.5
345.6
10.7.2 Square Rooters

Three types of square rooters have been considered, based on the restoring
algorithm (Fig. 10.3), the non-restoring algorithm (Fig. 10.4) and the Newton
Raphson method (Fig. 10.6). The implementation results are given in Tables
10.3, 10.4.
In the case of the NewtonRaphson method, the total time is data dependent. In
fact, as was already indicated above, this type of circuit is generally not time
effective (Table 10.5).
274
Table 10.7 Exponential 2x
10
Other Operations
FFs
LUTs
DSPs
Period
Total time
8
16
8
16
13
23
27
46
29
48
1
2
4.79
6.42
38.3
102.7
Table 10.8 Exponential 2x,

version 2
FFs
LUTs
DSPs
Period
Total time
8
16
8
16
13
23
21
39
49
86
17
71
3
10
5.64
10.64
45.1
170.2
Table 10.9 CORDIC: sine

and cosine
FFs
LUTs
Period
Total time
16
32
32
8
16
24
16
32
32
57
106
106
134
299
309
3.58
4.21
4.21
57,28
134.72
134.72
Table 10.10 CORDIC:

0:5
z x2 y2
FFs
LUTs
DSPs
Period
Total time
8
16
48
8
16
24
16
32
48
43
76
210
136
297
664
1
2
5
3.39
4.44
4.68
27.12
71.04
224.64
10.7.3 Logarithm and Exponential

Table 10.6 gives implementation results of the circuit of Fig. 10.8. DSP slices
have been used.
The circuit of Fig. 10.9 and the alternative circuit using a multiplier instead of a
table (Fig. 10.10) have been implemented. In both cases DSP slices have been
used (Tables 10.7, 10.8)
10.7.4 Trigonometric Functions

Circuits corresponding to algorithms 10.13 and 10.14 have been implemented. The
results are summarized in Tables 10.9, 10.10.
10.8 Exercises
1. Generate VHDL models of combinational binary-to-decimal and decimal-tobinary converters.
2. Synthesize binary-to-radix-60 and radix-60-to-binary converters using LUT-6.
10.8
Exercises
275
3. Implement Algorithm 10.6.

4. Implement the second square rooting convergence algorithm (based on
Fig. 10.7).
5. Synthesize circuits for computing ln, log10, ex and10x.
6. Generate a circuit which computes ejx cos x j sin x:
References
Press, New York
2. Volder JE (1959) The CORDIC trigonometric computing technique. IRE Trans Electron
Comput EC8:330334
3. Volder JE (2000) The birth of CORDIC. J VLSI Signal Process Sys 25:101105
Chapter 11
Decimal Operations
In a number of computer arithmetic applications, decimal systems are preferred to

the binary ones. The reasons come, not only from the complexity of coding/
decoding interfaces but, mostly from the lack of precision in the results of the
binary systems.
For the following circuits the input and output operators are supposed to be
encoded in Binary Encoded Digits (BCD).
11.1 Addition
Addition is a primitive operation for most arithmetic functions, and then it
deserves special attention. The general principles for addition are in Chap. 7, and
in this section we examine special consideration for efficient implementation of
decimal addition targeting FPGAs.
11.1.1 Decimal Ripple-Carry Adders

Consider the base-B representations of two n-digit numbers
x xn1 Bn1 xn2 Bn2 x0 B0 ;
y yn1 Bn1 yn2 Bn2 y0 B0 :
The following (pencil and paper) Algorithm 11.1 computes the (n ? 1)-digit
representation of the sum z = x ? y ? cin where cin is the initial carry.

277
278
11
Fig. 11.1 Decimal ripplecarry adder cell
Decimal Operations
x 3(i) y 3(i) x 2(i) y 2(i) x1(i) y 1(i) x 0(i) y 0(i)
s4
s4
s 3.(s 2s 1)
FA
FA
FA
HA
s3
s2
s1
HA
FA
HA
c(i)
s0
c(i+1)
z3(i)
z 2(i)
z 1(i)
z 0(i)
Algorithm 11.1: Classic addition (ripple carry)
For B = 10, the classic ripple-carry for a BCD decimal adder cell can be
implemented as suggested in Fig. 11.1 The mod-10 addition is performed adding 6
to the binary sum of the digits, when a carry for the next digit is generated. The
VHDL model ripple_carry_adder_BCD.vhd is available at the Authors web page.
As described in Chap. 7, the na implementation of an adder (ripple-carry,
Fig. 7.1) has a meaningful critical path. In order to reduce the execution time of
each iteration step, Algorithm 11.1 can be modified as shown in the next section.
11.1.2 Base-B Carry-Chain Adders

In order to improve the ripple-carry adder, a better solution is the use of two binary
functions of two B-valued variables, namely the propagate (P) and generate
(G) functions.
pi pxi ;yi 1if xi yi B 1; pxi ;yi 0 otherwise;
gi gxi ;yi 1if xi yi B; gxi ;yi 0if xi yi B 2; otherwise; anyvalue:
11.1
Addition
x(n-1)
279
y(n-1)
x(n-2)
G-P
g(n-1)
z(n)=c(n)
Cy.Ch.
x(n-1)
y(n-2)
x(1)
y(n-1)
Cy.Ch.
x(n-2)
x(0)
G-P
G-P
p(n-1) g(n-2)
c(n-1)
y(1)
p(n-2)
c(n-2)
y(n-2)
G-P
g(1)
.....
c(2)
p(1)
Cy.Ch.
x(1)
y(0)
y(1)
g(0)
c(1)
p(0)
Cy.Ch.
x(0)
c(0)=c_in
y(0)
mod B sum
mod B sum
mod B sum
mod B sum
z(n-1)
z(n-2)
z(1)
z(0)
Fig. 11.2 n-digits carry-chain adder
So, ci
? 1
can be expressed under the following way:

ci1 pi ci not pi gi :
The corresponding modified Algorithm 11.2 is the following one.

Algorithm 11.2: Carry-chain addition
The use of propagate and generate functions allow generating a n-digit adder
carry-chain array of Fig. 11.1. It is based on the Algorithm 11.2. The GeneratePropagate (G-P) cell calculates the Generate and Propagate functions; and the
carry-chain (Cy.Ch) cell computes the next carry. Observe that the carry-chain
cells are binary circuits, whereas the generate-propagate and the mod B sum cells
are B-ary ones. As regards the computation time, the critical path is shaded in
Fig. 11.2. (It has been assumed that Tsum [ TCy. Ch)
280
11
(a) x3 y3 x2 y2 x1 y1 x0 y0
(b)
Decimal Operations
x 3 (i) y 3(i) x 2(i) y 2(i) x1(i) y 1(i) x 0(i) y 0(i)
...
FA
s4
FA
s3
FA
s2
s1
HA
s0
G-P
p(i)
c(i+1)
FA
FA
HA
c(i)
p = s4.s3.s2.s1.s0
g = s4 s3.(s2 s1)
p
FA
g(i)
0
Cy.Ch.
HA
z 3(i)
FA
z2(i)
FA
z 1(i)
HA
z 0(i)
Fig. 11.3 a Simple G-P cell for BCD adder. b Carry-chain BCD adder ith digit computation
11.1.3 Base-10 Carry-Chain Adders

If B = 10, the carry-chain circuit remains unchanged but the P and G functions as
well as the modulo-10 sums are somewhat more complex. In base 2 (Chap. 7), the
P and G cells are respectively synthesized by XOR and AND functions, while in
base 10, P and G are now defined as follows:
pi 1 if xi yi 9; pi 0 otherwise;
11:1
gi 1 if xi yi [ 9; gi 0 otherwise;
11:2
A straightforward way to synthesize P and G is shown at Fig. 11.3a. That is add

the BCD numbers and detects if the sum is equal to nine and greater than nine
respectively. Nevertheless, functions P and G may be directly computed from xi, yi
inputs. The following formulas (11.3) and (11.4) are Boolean expressions of
conditions (11.1) and (11.2).
pi P0 K1 P3 K2 _ K3 G2 _ G1 K3 P2
11:3
gi G0 P3 _ G2 _ P2 G1 _ G3 _ P3 P2 _ P3 P1 _ G2 P1 _ G2 G1 11:4
0
where Pj xj yj; Gj xj yj and Kj xj yj are the binary propagator, generator

and carry-kill for the jth components of the BCD digits x(i) and y(i).
The BCD carry-chain adder ith digit computation is shown at Fig. 11.3b and it
is made of a first binary mod 16 adder stage, a carry-chain cell driven by the GP functions, and an output adder stage performing a correction (adding 6)
whenever the carry-out is one. Actually, a zero carry-out c(i ? 1) identifies that
the mod 16 sum does not exceed 9, so no corrections are needed. Otherwise, the
add-6 correction applies. Naturally, the G-P functions may be computed according
to Fig. 11.3, using the outputs of the mod 16 stage, including the carry-out s4.
The VHDL model cych_adder_BCD_v1.vhd that implements a behavioral
model of the decimal carry-chain adder is available at the Authors web page.
11.1
Addition
281
cout
BCD carry chain

x3:0(N-1)
y3:0(N-1)
4-bit
Adder
s 4:0(N-1)
c(N)
p(N-1)
G-P
s3:0(N-1)
correc z3:0(N-1)
Adder
g(N-1)
c(N-1)
x3:0(N-2)
y3:0(N-2)
4-bit
Adder
p(N-2)
s 4:0(N-2)
G-P
s3:0(N-2)
correc z 3:0(N-2)
Adder
g(N-2)
...
x 3:0(1)
y3:0(1)
..
.
...
...
c(2)
4-bit
Adder
s4:0(1)
4-bit
Adder
s4:0(0)
p(1)
G-P g(1)
c(1)
p(0)
G-P g(0)
s3:0(1)
c(1)
x 3:0(0)
y3:0(0)
c(N-2)
correc
Adder
z 3:0(1)
correc
Adder
z 3:0(0)
s3:0(0)
c(0)
c in = c(0)
Fig. 11.4 FPGA implementation of an N-digit BCD Adder
11.1.4 FPGA Implementation of the Base-10 Carry-Chain Adders

The FPGA vendors provides dedicated resources to implement binary carry-chain
adders [1, 2]. As mentioned in Chap. 7 a simple HDL description of a binary adder
is implemented efficiently using the carry-logic. Nevertheless in order to use this
resources in other designs it is necessary to instantiate the components manually.
Figure 11.4 shows a Xilinx implementation of the decimal carry-chain adder.
The VHDL model cych_adder_BCD_v2.vhd that implements a low level component instantiation model of the decimal carry-chain adder is available at the
Authors web page.
Observe that the critical path includes a 4-bit adder, the G-P computation; the ndigits carry propagation and a final 3-bit correction adder. Xilinx 6-input/2-output
LUT is built as two 5-input functions while the sixth input controls a 2-1 multiplexor allowing to implement either two 5-input functions or a single 6-input one;
so G and P functions fit in a single LUT as shown at Fig. 11.5a.
Other alternatives to implement in FPGA the decimal carry-chain adders, include
computing the functions P and G directly from the decimal digits (x(i), y(i) inputs)
using the Boolean expressions, instead of the intermediate sum bits sk [3].
282
11
Decimal Operations
For this purpose one could use formulas (11.3) and (11.4), nevertheless, in order
to minimize time and hardware consumption the implementation of p(i) and g(i) is
revisited as follows. Remembering that p(i) = 1 whenever the arithmetic sum
x(i) ? y(i) = 9, one defines a 6-input function pp(i) set to be 1 whenever the
arithmetic sum of the first 3 bits of x(i) and y(i) is 4. Then p(i) may be computed as:
pi x0 i y0 i ppi:
11:5
On the other hand, gg(i) is defined as a 6-input function set to be 1 whenever the
arithmetic sum of the first 3 bits of x(i) and y(i) is 5 or more. So, remembering that
g(i) = 1, whenever the arithmetic sum x(i) ? y(i) [ 9, g(i), may be computed as:
gi ggi _ ppi x0 i y0 i:
11:6
As Xilinx LUTs may compute 6-variable functions, then gg(i) and pp(i) may be
synthesized using 2 LUTs in parallel while g(i) and p(i) are computed through an
additional single LUT as shown at Fig. 11.5b.
11.2 Base-10 Complement and Addition: Subtration

11.2.1 Tens Complement Numeration System
Bs complement representation general principles are available in the literature.
One restricts to 10s complement system to cope with the needs of this section.
A one-to-one function R(x), associating a natural number to x is defined as follows.
Every integer x belonging to the range 10n/2 B x \ 10n/2, is represented by
R(x) = x mod 10n, so that the integer represented in the form xn1 xn2 x1 x0 is
xn1 10n1 xn2 10n2 x0 if xn1 10n1 xn2 10n2
x0 \10n =2;
11:7
xn1 10n1 xn2 10n2 x0 10n if xn1 10n1 xn2 10n2

x0 10n =2:
11:8
The conditions (11.7) and (11.8) may be more simply expressed as
xn1 10n1 xn2 10n2 x0 if xn1 \5;
11:9
xn1 10n1 xn2 10n2 x0 10n if xn1 5:
11:10
Another way to express a 10s complement number is:

0
xn1 10n1 xn2 10n2 x0
11:11
11.2
Base-10 Complement and Addition: Subtration

0
283
where xn1 xn1 10 if xn1 5 and xn1 xn1 if xn1 \5; while the sign
definition rule is the following one: if x is negative then xn 1 C 5; otherwise
xn - 1 \ 5.
11.2.2 Tens Complement Sign Change

Given an n-digit 10s complement integer x, the inverse z = -x of x, is an n-digit
10s complement integer. Actually the only case -x cannot be represented with
n digits is when x 10n =2; so x 10n =2: The computation of the representation of -x is based on the following property. Assuming x to be represented as an
n-digit 10s complement number R(x), -x may be readily computed as
x 10n1 R x:
11:12
A straightforward inversion algorithm then consists of representing x with

n ? 1 digits, complementing every digit to 9, then adding 1. Observe that sign
extension is obtained by adding a digit 0 to the left of a positive number, or 9 for a
negative number, respectively.
11.2.3 10s Complement BCD Carry-Chain Adder-Subtractor

To compute X ? Y similar algorithm as in Algorithm 11.2 (Sect. 11.1.2) can be
used. In order to compute X - Y, 10s complement subtraction algorithm actually
adds (-Y) to X.
10s complement sign change algorithm may be implemented through a digitwise 9s complement stage followed by an add-1 operation. It can be shown that
the 9s complement binary components w3 ; w2 ; w1 ; w0 of a given BCD digit
y3 ; y2 ; y1 ; y0 are expressed as
w3 y03 y02 y01 ;
w2 y2 y1 ;
w 1 y1 ;
w0 y00
11:13
An improvement to the adder stage could be carried out by avoiding the delay
produced by the 9s complement step. Thus, this operation may be carried out
within the first binary adder stage, where p(i) and g(i) are computed as
0
p0 i x0 i y0 i A =S ; p1 i x1 i y1 i;
0
p2 i x2 i y2 i y1 i: A =S
0 0

0
0
0
0
p3 i x3 i y3 i y2 i y1 i A =S y3 i A =S
11:14
gk i xk i; 8k:
11:15
284
11
(a)
s 4(i)
s 3(i)
s2(i)
s1(i)
s0(i)
Decimal Operations
(b)
c(i+1)
LUT
p(i)
g(i)
P-G
c(i)
x 1(i)
x 2(i)
x 3(i)
y1(i)
y2(i)
y3(i)
x1(i)
x2(i)
x3(i)
y1(i)
y2(i)
y3(i)
LUT
c(i+1)
x0(i)
y0(i)
pp(i)
LUT
p(i)
g(i)
c(i)
LUT
gg(i)
P-G
Fig. 11.5 FPGA carry-chain for decimal addition. a. P-G calculation using an intermediate
addition. b. P-G calculation directly from BCD digits
Fig. 11.6 FPGA 9s

complement circuit for BCD
digit
A/S
y3(i)
y2(i)
y1(i)
w3(i)
LUT
A/S
y1(i)
y0 (i)
w2(i)
w1(i)
LUT
w0(i)
A third alternative is computing G and P directly from the input data. As far as
addition is concerned, the P and G functions may be implemented according to
formulas (11.4) and (11.5). The idea is computing the corresponding functions in
the subtract mode and then multiplexing according to the add/subtract control
signal A0 /S.
11.2.4 FPGA Implementations of Adder Subtractors

To compute X Y, 10s complement subtraction algorithm actually adds (-Y) to
X. So for a first implementation, Fig. 11.6 presents a 9s complement implementation using 6-input/2-output LUTs, available in the Xilinx (Virtex-5, Virtex-6,
spatan6, 7-series) technology. A0 /S is the add/subtract control signal; if A0 /S = 1
(subtract), formulas (11.13) apply, otherwise A0 /S = 0 and wj i yj i8i; j:
The complete circuit is similar to the circuit of Fig. 11.4, but instead of input y,
the input w as produced by the circuit of Fig. 11.5.
The better alternative intended to avoid the delay produced by the 9s complement step, embeds the 9s complementation within the first binary adder stage,
as depicted in Fig. 11.7a, where p(i) and g(i) are computed as explained in (11.14)
and (11.15).
The VHDL model addsub_1BDC.vhd that implements the circuit of Fig. 11.7a
is available at the Authors web page. The VHDL model has two architectures a
behavioral and a low level that instantiates components (Luts, muxcy, xorcy, etc.).
11.2
Base-10 Complement and Addition: Subtration
(a)
(b)
s 4(i)
A/S
y 3(i)
y2(i)
y1(i)
x 3(i)
285
A/S
LUT
p3(i)
1
s3(i)
x 1(i)
x 2(i)
x 3(i)
y 1(i)
y 2(i)
y 3(i)
LUT
A/S
y2(i)
y1(i)
x 2(i)
LUT
p2(i)
1
s2(i)
x1 (i)
x2 (i)
x3 (i)
y1 (i)
y2 (i)
y3 (i)
c(i+1)
pps(i)
pp(i)
x 0(i)
y 0(i)
LUT
p(i)
0
g(i)
LUT
c(i)
ppa(i)
A/S
LUT
p1(i)
y 1(i)
x 1(i)
s1(i)
A/S
LUT
p0(i)
y 0(i)
x 0(i)
s0(i)
x 1(i)
x 2(i)
x 3(i)
y 1(i)
y 2(i)
y 3(i)
x1 (i)
x2 (i)
x3 (i)
y1 (i)
y2 (i)
y3 (i)
LUT
ggs(i)
1
0
LUT
gg(i)
gga(i)
P-G
Fig. 11.7 FPGA implementation of adder-subtractor. a Adder-subtractor for one BCD digit.
b Direct computation of P-G function
Then we can use the circuit of Fig. 11.6a to compute the P-G function using the
previous computed addition of BDC digits. The VHDL model addsub_BDC_v1.vhd
that implements a complete decimal adder-subtractor and it is available at the
Authors web page. To complete the circuit, a final correction adder (correct_add.vhd) corrects the decimal digit as a function of the carries.
The third alternative is computing G and P directly from the input data. For this
reason, assuming that the operation at hand is X ? (Y), one defines on one hand
ppa(i) and gga(i) according to (11.4) and (11.5) (Sect. 11.1.4), i.e. using the
straight values of Ys BCD components. On the other hand, pps(i) and ggs(i) are
defined using wk(i) as computed by the 9s complement circuit (11.13). As
wk(i) are expressed from the yk(i) both pps(i) and ggs(i) may be computed directly
from xk(i) and yk(i). Then the correct pp(i) and gg(i) signal is selected according to
the add/subtract control signal A0 /S. Finally, the propagate and generate function
are computed as:
pi x0 i y0 i A0 =S ppi;
11:16
gi ggi _ ppi x0 i y0 i A0 =S:
11:17
Figure 11.7b shows the Xilinx LUT based implementation. The multiplexers
are implemented using dedicated muxF7 resources.
The VHDL model addsub_BDC_v2.vhd that implements the adder-subtractor
using for the P-G computation from the direct inputs (carry-chain_v2.vhd) is
286
11
Fig. 11.8 Binary to BCD

arithmetic reduction
Decimal Operations
80
40
p6
20
p5
p6
10
p4
p5
8
p3
d3
d2
d1
d0
c3
4
p2
p4
p6
c2
2
p1
p4
p5
c1
1
p0
c0
11.3 Decimal Multiplication

This section presents approaches for BCD multipliers. It starts by discussing algorithmic alternatives to implement a 1 9 1 BCD digit multiplier. Then, this section
proposes an implementation for an N 9 1 BCD multiplier. This operation can be used
to carried out an N 9 M multiplier, but by itself, the N 9 1 multiplication appears to
be a primitive for other algorithms, such as logarithm or exponential functions.
11.3.1 One-Digit by One-Digit BCD Multiplication

11.3.1.1 Binary Arithmetic with Correction
The decimal product can be obtained through a binary product and a post correction stage [4, 5]. Let A and B be two BCD digits (a3 a2 a1 a0) and (b3 b2 b1 b0)
respectively. The BCD coded product consists of two BCD digits D and C such
that:
AB D10 C
11:18
A * B is first computed as a 7-bit binary number P(6:0) such that

AB P p6 p5 p4 p3 p2 p1 p0
11:19
Although a classic binary-to-BCD decoding algorithm can be used, it can be

shown that the BCD code for P can be computed through binary sums of correcting
terms described in Fig. 11.8. The first row in Figure shows the BCD weights. The
weights of p3, p2, p1 and p0 are the same as those of the original binary number
p6 p5 p4 p3 p2 p1 p0 . But weights 16, 32 and 64 of p4, p5, and p6 have been
respectively decomposed as (10, 4, 2), (20, 10, 2) and (40, 20, 4). Observe that p3
p2 p1 p0 could violate the interval [0, 9], then an additional adjust could be
necessary.
First the additions of Row 1, 2, 3, and correction of p3 p2 p1 p0 are completed
(least significant bit p0 is not necessary in computations). Then the final correction
is computed.
One defines (Arithmetic I):
1. the binary product AB P p6 p5 p4 p3 p2 p1 p0
2. the Boolean expression: adj1 p3 ^ p2 _ p1 ;
11.3
Decimal Multiplication
287
3. the arithmetic sum: dcp p6 p5 p4 p3 p2 p1 p0 0 p6 p5 0 p4 p3 0 0 0 0p6 p5 0

0 0 0 0 adji adj1 0;
4. the Boolean expression: adj2 dcp3 ^ dcp2 _ dcp1 _ p5 ^ p4 ^ p3 :
One computes
dc dcp 0 0 0 0 adj2 adj2 0:
11:20
Then
D dc7 dc6 dc5 dc4 and C dc3 dc2 dc1 dc0
A better implementation can be achieved using the following relations
(Arithmetic II):
1. the product AB P p6 p5 p4 p3 p2 p1 p0
2. compute:
cc p3 p2 p1 p0 0 p4 p4 0 0 p6 p5 0;
dd p6 p5 p4 0 p6 p5
(cc is 5 bits, dd has 4 bits, computed in parallel)
3. define:
cy1 1 iff cc [ 19; cy0 1 iff 9\cc\20
(cy1 y cy2 are function of cc3 cc2 cc1 cc0, and can be computed in parallel)
4. compute:
c cc3 cc2 cc1 cc0 cy1cy1 or cy0 cy 0 0;
11:21
d dd3 dd2 dd1 dd0 0 0 cy1 cy0

(c and d calculated in parallel)
Compared with the first approach, the second one requires smaller adders (5 and
4-bit vs. 8-bit) and the adders can operate in parallel as well. The VHDL models
bcd_mul_arith1.vhd and bcd_mul_arith2.vhd that described the previous method
for digit by digit multiplication are available at the Authors web page.
11.3.1.2 Using ROM

Actually a (100 9 8)-bit ROM can fit to store all the A* B multiplications.
However, as A and B are two 4-bit operands (BCD digits) the product can be
mapped into a 28 9 8-bit ROM. In FPGA devices there mainly two main memory
recourses: Block and distributed RAMS. For the examples in Xilinx devices two
288
11
Decimal Operations
main possibilities are considered: Block RAMs BRAMs or distributed RAMs

(LUT based implementation of RAMs).
BRAM-based implementation: Xilinx Block RAMs are 18-kbit (or 36-kbits)
configurable and synchronous dual-port RAM (or ROM) blocks. They can be
configured into different layouts. The 211 9 8-bit configuration option has been
selected, though wasting some memory capacity. As BRAM is dual-port, two oneby-one digit multiplications can be implemented in a single BRAM and in a single
clock cycle. However the main characteristic to cope with is that BRAMs are
synchronous, so either the address or the output should be registered.
Distributed RAM (LUT-based): 6-input LUTs can be configured as 64 9
1-bit RAMs. Moreover the four 6-LUTs in a slice has additional multiplexers to
implement 256 9 1 bit RAMs. Then the 28 9 8-bit ROM can be implemented
using 8 slices (32 LUTs).
A trivial optimization reduces the area considering the computation of the BCD
final result components c0 and d3 straightforward. Actually c0 a0 ^ b0 and d3
a0 ^ b0 ^ a3 ^ b3 : That is, c0 is related to the parity while d3 emphasizes that most
significant bit is set for one in only one case (decimal 9 9 9 = 81). It is thus
possible to reduce the required memory size to a 28 9 6-bit ROM only plus two
LUTs to implement c0 and d3.
Comments
1. BRAM-based design is fast but synchronous. It is useless for combinational
implementations, but suitable for sequential and pipelined ones.
2. The existence of do-not-care conditions in the memory definition allows the
synthesizer to reduce the effective memory requirement.
The VHDL model bcd_mul_bram.vhd implements the BRAM based implementation for the digit by digit multiplication. Additionally, and bcd_mul_mem1.vhd and bcd_mul_mem2.vhd provides the LUT based implementation of
decimal BCD multiplication. These models with the corresponding test bench
(test_mul_1by1BCB.vhd) are available at the Authors web page.
11.3.2 N by One BCD Digit Multiplier

A N 9 1 BCD digit multiplier is readily achieved through N 1 9 1-digit multiplications followed by a BCD decimal addition. Fig. 11.9 shows how the partial
products are arranged to feed the BCD N-digit adder stage. The carry-chain adder
of Sect. 13.1 can be used.
The VHDL model mult_Nx1_BCD.vhd and mult_Nx1_BCD_bram.vhd that
describe the N by one decimal multiplier are available at the Authors web page.
11.3
Decimal Multiplication
b
a(n-1)
1x1 BCD
multiplier
fast carrychain adder
289
a(n-2)
1x1 BCD
multiplier
1x1 BCD
multiplier
...
C(n-1)
C(n-2)
D(n-1)
D(n-2)
D(n-3)
Z(n)
Z(n-1)
Z(n-2)
a(2)
...
...
a(1)
1x1 BCD
multiplier
a(0)
1x1 BCD
multiplier
C(2)
C(1)
D(2)
D(1)
D(0)
Z(3)
Z(2)
Z(1)
C(0) = Z(0)
Z(0)
Fig. 11.9 N by one digit circuit multiplication. It includes N one by one digit multipliers and a
fast carry-chain adder
A(n-1:0)
B(m-1:0)
N-digit shift register
M-digit shift register
capt
shift
b(0)
N+1-digit register
Nx1 BCD
multiplier
partialsum(n-1:0)
N+1 BCD
Carry chain Adder
mult(n:0)
start
reset
clk
capt
Control
State
Machine
shift
N-digit register
sum(n:1)
sum(n:0)
partialsum (n-1:0)
& shiftmul(m-1:0)
sum(0)
shift
M-digit register
end
N+M-digit register
done
P(n+m-1:0)
shiftmul(m-1:0)
Fig. 11.10 A sequential N 9 M multiplier
11.3.3 N by M Digits Multiplier

Using the previously designed N by one digit multiplier it is possible to perform
the N 9 M digits multiplication. The best area compromise is obtained using a
sequential implementation that uses an N 9 1 digits multiplier and an N ? 1 digit
adder. Figure 11.10 show the scheme of the Least Significant Digit (LSD) first
algorithm implemented. The B operand is shifted right at each clock cycle. The
LSD digit of B is multiplied by the A operand and accumulated in the next cycle to
shorten the critical path. After M ? 1 cycles, the N ? M digit result is available.
The VHDL models mult_BCD_seq.vhd and mult_BCD_bram_seq.vhd that
describes the N by M digits decimal multiplier and the matching test bench
(test_mult_BCB_seq.vhd) is available at the Authors web page.
290
11
A(n-1:0)
B(6)
B(7)
Nx1 digit
BCD Mult
Nx1 digit
BCD Mult
m(7)(n:0)
Nx1 digit
BCD Mult
m(6)(n:0)
m(6)(n:1)
B(4)
B(5)
N+1 BCD
Carry chain Adder
s(3)(n+1:1)
s(3)(0)
Nx1 digit
BCD Mult
m(2)(n:0)
m(2)(n:1)
s(2)(0)
Nx1 digit
BCD Mult
m(0)(n:0)
m(1)(n:0)
m(2)(0)
m(0)(0)
m(0)(n:1)
N+1 BCD
Carry chain Adder
s(2)(n+1:1)
B(0)
B(1)
Nx1 digit
BCD Mult
m(3)(n:0)
m(4)(0)
m(4)(n:1)
N+1 BCD
Carry chain Adder
Nx1 digit
BCD Mult
m(4)(n:0)
m(5)(n:0)
m(6)(0)
B(2)
B(3)
Nx1 digit
BCD Mult
Decimal Operations
N+1 BCD
Carry chain Adder
s(1)(n+1:1)
s(2)(n+1:2)
s(1)(0)
s(0)(n+1:1)
s(0)(0)
t(0)(1)
t(0)(0)
s(0)(n+1:2)
s(3)(n+1:0)
s(1)(n+1:0)
N+2 BCD
Carry chain Adder
t(1)(n+3:2)
N+2 BCD
Carry chain Adder
t(1)(1)
t(1)(0)
t(1)(n+3:0)
t(0)(n+3:2)
t(0)(n+3:4)
N+3 BCD
Carry chain Adder
p(n+7:4)
p(3:2)
p(1)
p(0)
Fig. 11.11 An example of a paralell N 9 M multiplier (N 9 8)
Fully combinational implementations of N 9 M digits multipliers are possible

also based on the N 9 1 digit multiplication cell. After the first multiplication, an
adder tree sums up the intermediate result and gives the N ? M digit result.
Figure 11.11 show an N 9 8 digit multiplication example.
The VHDL model mult_BCD_comb.vhd that describes the N by M digits
decimal multiplier and the test bench (test_mult_BCB_comb.vhd) are available at
11.4 Decimal Division

As described in Chap. 9, the basic algorithms are based on digit recurrence. Using
Eqs. (9.1), (9.2) and (9.3) at each step of (9.2) q-(i ? 1) and ri+1 are computed in
function of ri and y in such a way that 10 ri qi1 y ri1 ; with
y ri1 \y; that is
ri1 10 ri qi1 y; with y ri1 \y:
11:22
The Robertson diagram applied to radix-10 (B = 10) is depicted at Fig. 11.12. It

defines the set of possible solutions: the dotted lines define the domain
f10 ri ; ri1 j 10 y 10 ri \10 y and y ri1 \yg; and the diagonals
correspond to the equations ri1 10 ri ky with k 2 f10; 9; . . .;
1; 0; 1; . . .; 9; 10g: If ky 10 ri \k 1y; there are two possible solutions for
q-(i ? 1), namely k and k ? 1. To the first one corresponds a non-negative value of
ri ? 1, and to the second one a negative value.
11.4
Decimal Division
291
ri+1
qi+1 =
- 10
-9y
-9
-8
-7
-6
-5
-4
-3
-2
-8y
-7y
-6y
-5y
-4y
-3y
-2y
-y
-1
2y
3y
4y
5y
6y
7y
8y
9y
9
10
10.ri
Fig. 11.12 Robertson diagram for radix 10
11.4.1 Non-Restoring Division Algorithm

A slightly modified version of the base-10 non-restoring algorithm (Sect. 9.2.1) is
used. For that y must be a normalized n-digit natural, that is 10n1 y\10n : The
remainders ri satisfy the condition -y B ri \ y and belong to the range 10n \ ri \ 10n. Define wi 10 ri ; so that
10n1 \wi \10n1 :
11:23
Thus, wi is an (n ? 2)-digit 10s complement number. The selection of every

digit qi1 is based on a truncated value of wi, namely w0 bwi =10a c; for some a
that will be defined later, so that wi 10a \w0 10a wi and
w0 10a wi \w0 10a 10a :
11:24
According to (11.23) and (11.24), 10n1 10a \w0 10a \10n1 ; so that
0
10n1a w \10n1a :
11:25
Thus, w0 is an (n ? 2 - a)-digit 10s complement number. Assume that a set of

integer-valued functions mk(y), for k in {-10, -9,, -1, 0, 1,, 8, 9}, satisfying
k y mk y 10a \k 1 y 10a
11:26
has been defined. The interval [ky, (k ? 1)y - 10a] must include a multiple of
10a. Thus, y must be greater than or equal to 210a. Taking into account that
y C 10n - 1, the condition is satisfied if a B n - 2.
The following property is a straightforward consequence of (11.24) and (11.26):
0
Property 11.1 If mk y w \mk1 y; then k y wi \k 2 y:

According to the Robertson diagram of Fig. 11.12, a solution qi1 can be chosen
as follows:
0
if w \m9 y then qi1 9;

0
if w m8 y then qi1 9;
0
if mk y w \mk1 y for some k in f9; 8; . . .; 1; 0; 1; . . .; 7g; then qi1 k 1:
292
11
Decimal Operations
Thus, this non-restoring algorithm generates a p-digit decimal quotient 0.q-1

q-2q-p and a remainder rp satisfying

x 0 q1 q2 . . .qp y rp 10p ; with y 10p rp 10p \y 10p ;
11:27
where every q-i is a signed decimal digit. It can be converted to a decimal number
by computing the difference between two decimal numbers
pos q01 101 q02 102 q0p 10p and
neg q001 101 q002 102 q00p 10p ;
with q0i qi if qi [ 0; q0i 0 if qi \ 0; q001 qi if qi \0; q00i 0 if qi [ 0:
It remains to define a set of integer-valued functions mk(y) satisfying (11.26). In
order to simplify their computation, they should only depend on the most significant digits of y. The following definition satisfies (11.26):

mk y bk y0 =10c bias where y0 y=10a1 if k 0 and
mk y bk y0 =10c bias if k\0;
where bias is any natural belonging to the range 2 B bias B 6. With a = n-2, y0 as
a 3-digit natural, and w0 and mk(y) are 4-digit 10s complement numbers. In the
following Algorithm 1 mk(y) is computed without adding up bias to
bk y0 =10cor bk y0 =10c; and w0 is substituted by w0 bias.
Algorithm 11.3: Non-restoring algorithm for decimal numbers
11.4
Decimal Division
293
The structure of the data path corresponding to Algorithm 11.3 is shown in

Fig. 11.13. Additionally, two decimal shift registers storing pos and neg and an
output subtractor are necessary. Alternatively the conversion could be done on-the
fly. The computation time and complexity of the k_ys generation and range
detection components are independent of n. The execution time of the iteration
step is determined by the n-digit by 1-digit multiplier and by the (n ? 1)-digit
adder. The total computation time is O(pn).
The simplified VHDL model decimal_divider_nr.vhd that describes the N by
N digits divider that produces P digits of decimal result and the test bench
(test_dec_div_seq.vhd) are available at the Authors web page.
11.4.2 An SRT-Like Division Algorithm

In order to make the computation time linear, a classical method consists of using
carry-save adders and multipliers instead of ripple-carry ones. In this way, the
iteration step execution time can be made independent of the number n of digits.
Nevertheless, this poses the following problem: a solution qi1 of Eq. (9.2) must
be determined in function of a carry-stored representation (si, ci) of ri, without
actually computing ri = si ? ci. An algorithm similar to the SRT-algorithms for
radix-2K dividers can be defined for decimal numbers (Sect. 9.2.3). Once again the
divisor y must be assumed to be a normalized n-digit natural.
294
11
1_y
3
2_y
3
9_y
w=ri /10n-3
y=y / 10n-3
... 3
Decimal Operations
3
y
rangedetection
detection
range
q- (i+1)
q-(i +1)
generation of k_y = ky/10
n
3
1_y
2_y
...
B
a N by 1 digit multiplier
- aB
-q-(i +1)y
n+2
8_y
9_y
10ri
n+2
decimal adder
n+1
ri+ 1
register
n+1
ri
Fig. 11.13 Non-restoring radix 10 algorithm data path
All along the algorithm execution, ri will be encoded in the form wi si ci

(stored-carry encoding) where si and ci are 10s-complement numbers. Define wi
and w0 as in previous section, that is wi 10 ri ; and w0 bwi =10a c: Thus, wi
10 si 10 ci and w0 b10 si 10 ci =10a c: Define also truncated values st
and ct of 10si and 10ci, that is st b10 si =10a c and ct b10 ci =10a c; and let w00
be the result of adding st and ct. The difference between w0
b10 si 10 ci =10a c and w00 b10 si =10a c b10 ci =10a c is the possible
carry from the rightmost positions, so that w0 1 w00 w0 ; and thus
w00 w0 w00 1:
11:28
According to (11.25) and (11.28),

10n1a 1 w00 \10n1a ;
11:29
so that w00 is an (n ? 3-a)-digit 10s-complement number. The relation between

wi and the estimate w00 10a of wi is deduced from (11.24) and (11.28):
w00 10a wi \w00 10a 2 10a :
11:30
Assume that a set of integer-valued functions Mk(y), for k in {-10, -9,, -1,
0, 1,, 8, 9}, satisfying
k y Mk y 10a \k 1 y 2 10a;
11:31
have been defined. The interval [ky, (k ? 1)y - 210a[ must include a multiple of
10a. Thus y must be greater than or equal to 310a. Taking into account that
11.4
Decimal Division
295
y C 10n - 1, once again the condition is satisfied if a B n - 2. The following

property is a straightforward consequence of (11.30) and (11.31).
Property If Mk y w00 \Mk1 y; then k y wi \k 2 y:
Thus, according to the Robertson diagram of Fig. 11.12, a solution qi1 can
be chosen as follows:
if w00 \M9 y then qi1 9;
if w00 M8 y then qi1 9;
if Mk y w00 \Mk1 y for some k inf9; 8; . . .; 1; 0; 1; . . .; 7g; then qi1 k 1:
This SRT-like algorithm generates a p-digit decimal quotient 0.q-1 q-2q-p and
a remainder rp satisfying (11.27), and can be converted to a decimal number by
computing the difference between two decimal numbers as in non-restoring
algorithm.
It remains to define a set of integer-valued functions Mk(y) satisfying (11.31). In
order to simplify their computation, they should only depend on the most significant digits of y. Actually, the same definition as in Sect. 11.3.1 can be used, that is

Mk y bk y0 =10c bias where y0 y=10a1 if k 0 and
Mk y bk y0 =10c bias if k\0:
In this case the range of bias is 3 B bias B 6. With a = n - 2, y0 as a 3-digit
natural, w0 and mk(y) are 4-digit 10s complement numbers, and w00 is a 5-digit 10s
complement number. In the following Algorithm 2 Mk(y) is computed without
adding up bias to bk y0 =10c or bk y0 =10c; and w00 is substituted by w00 bias.
Algorithm 11.4: SRT like algorithm for decimal numbers
296
11
Decimal Operations
The structure of the data path corresponding to Algorithm 11.4 is shown in

Fig. 11.4. The carry-free multiplier is a set of 1-digit by 1-digit multipliers
working in parallel. Each of them generates two digits p1, j ? 1 and p0, j such that
qi1 yj 10 p1;j1 p0;j ; and the product qi1 y is obtained under the
form p1 ? p0. The 4-to-2 counter computes 10si ? 10ci - (p1 ? p0), that is
10ri - q-(i ? 1)y, under the form si ? 1 ? ci ? 1. Two decimal shift registers
storing pos and neg and an output ripple-carry subtractor are necessary. Another
output ripple-carry adder is necessary for computing the remainder rp = sp ? cp.
Thus, all the components, but the output ripple-carry components, have computation times independent of n and p. The total computation time is O(p ? n).
11.4
Decimal Division
ri-1
297
y
quotient
selection
-y
-ulp ulp
x2
0
two_ r
adder
ulp
adder
2
ri
register
initially: x
register
initially: 0
register
initially : 2 -1
ri-1
ulp
load
update
Fig. 11.14 Binary digit-recurrence data path for decimal division
The VHDL model decimal_divider_SRT_like.vhd that describes the N by

N digits divider that produces P digits decimal result with several modules
(decimal_shift_register.vhd, mult_Nx1_BCD_carrysave.vhd, special_5digit_
adder.vhd, range_detection3.vhd, bcd_csa_addsub_4to2.vhd) and the test bench
(test_dec_div_seq.vhd) are available at the Authors web page.
11.4.3 Other Methods for Decimal Division

Other methods of division could be considered such as the use digit recoding in
dividend and or divisor and also use extra degree of pre-normalization for the
operands. The idea behind these methods is to ease the digit selection process.
Another idea is the use of the binary digit recurrence algorithm of describe in
Chap. 9 but using decimal operands.
Observe that the Algorithm 9.1 can be executed whatever the representation of
the numbers. If Bs complement radix-B representation is used, then the following
operations must be available: radix-B doubling, adding, subtracting and halving.
298
11
Decimal Operations
Algorithm 11.5: Binary digit-recurrence algorithm for decimal division
In order to obtain m decimal digits results, you need to choose p so that

2-p % 10-m, that is p m log2 10 3:3 m: A possible data path is shown in
Fig. 11.14. The implementation of this architecture leads to a smaller circuit, but
slower than the decimal digit recurrence due to the difference in the amount of
cycles to be executed.
11.5 FPGA Implementation Results

The circuits have been implemented on Xilinx Virtex-5 family with speed grade -2
[6]. The Synthesis and implementation have been carried out on XST (Xilinx
Synthesis Technology) [7] and Xilinx Integrated System environment (ISE) version 13.1 [2]. The critical parts were designed using low level components
instantiation (lut6_2, muxcy, xorcy, etc.) in order to obtain the desired behavior.
11.5.1 Adder-Subtractor Implementations

The adder and adder-subtractor implementation results are presented in this section. Performances of different N-digit BCD adders have been compared to those
of an M-bit binary carry chain adder (implemented by XST) covering the same
range, i.e. such that M bN:log2 10c 3:322N:
11.5
FPGA Implementation Results
299
Table 11.1 Delays in ns for decimal and binary adders and adder-subtractor
N (digits) RpCy add CyCh add AddSub V1 AddSub V2 M (bits) Binary add-sub
8
16
32
48
64
96
12.4
24.4
48.5
72.3
95.9
3.5
3.8
4.5
5.1
5.2
5.9
Table 11.2 Area in 6-input

LUTs for different decimal
adders and adders-subtractors
3.5
3.8
4.6
5.2
5.5
6.1
3.4
3.7
4.8
5.3
5.5
6.1
27
54
107
160
213
319
2.1
2.6
3.8
5.2
6.6
8.8
Circuit
# LUTs
Ripple carry adder

Carry chain adder
Binary
Adder-subtractor V2 (PG from binary addition)
Adder-subtractor V2 (PG direct form inputs)
Binary adder and adder-subtractor
79N
99N
d3:32
Ne
99N
13 9 N
d3:32
Ne
Table 11.1 exhibits the post placement and routing delays in ns for the decimal
adder implementations Ad-I and Ad-II of Sect. 7.1; and the delays in ns
for the decimal adder-subtractor implementations AS-I and AS-II of Sect. 7.2.
Table 11.2 lists the consumed areas expressed in terms of 6-input look-up tables
(6-input LUTs). The estimated area presented in Table 11.2 was empirically
confirmed.
Comments
1. Observe that for large operands, the decimal operations are faster than the
binary ones.
2. The delay for the carry-chain adder and the adder-subtractor are similar in
theory. The small difference is due to the placement and routing algorithm.
3. The overall area with respect to binary computation is not negligible. In Xilinx
6-input LUT family an adder-subtractor is between 3 and 4 times bigger.
11.5.2 Multiplier Implementations

The decimal multipliers use the one by one digit multiplication described in
Sect. 11.2.1 and the decimal adders of Sect. 11.1. The results are for the same
Virtex 5 device speed grade -2.
300
11
Decimal Operations
Table 11.3 Results of BCD N 9 1 multipliers using LUTs cells and BRAM cells
N
Mem in LUTs cells
BRAM-based cells
4
8
16
32
T (ns)
# LUT
T (ns)
# LUT
# BRAM
5.0
5.1
5.3
6.1
118
242
490
986
5.0
5.1
5.4
6.1
41
81
169
345
1
2
4
8
Table 11.4 Results of sequential implementations of N 9 M multipliers using one by one digit
multiplication in LUTs
N
M
T (ns)
# FF
# LUT
# cycles
Delay (ns)
4
8
8
8
16
16
32
16
32
4
4
8
16
8
16
16
32
32
5.0
5.1
5.1
5.1
5.3
5.3
5.7
5.3
5.7
122
186
235
332
363
460
716
653
909
243
451
484
553
921
986
1,764
1,120
1,894
4
5
9
17
9
17
17
33
33
25.0
25.5
45.9
86.7
47.7
90.1
96.9
174.9
188.1
11.5.2.1 Decimal N 3 1 Digits Implementation Results

Comparative figures of merit (minimum clock period (T) in ns, and area in LUTs
and BRAMS) of the N 3 1 multiplier are shown in Table 11.3, for several values
of N; the result are shown for circuits using LUTs, and BRAMS to implement the
digit by digit multiplication respectively.
11.5.2.2 Sequential Implementations

The sequential circuit multiplies N digits by 1 digit per clock cycle, giving the
result M ? 1 cycles later. In order to speed up the computation, the addition of
partial results is processed in parallel, but one cycle later (Fig. 11.10). Results for
sequential implementation using LUT based cells for 1 3 1 BCD digit multiplication are given in Table 11.4. If the BRAM-based cell is used, similar periods (T)
can be achieved, by means of less LUTs but using BRAMs blocks.
11.5.2.3 Combinational Implementations of N by M Multipliers

For the implementation of the N 3 M-digit multiplier, the N 3 1 mux-based multiplication stage has been replicated M times: it is the best choice because BRAM-
11.5
FPGA Implementation Results
Table 11.5 Results of

combinational
implementations of
N 9 M multipliers
301
Delay (ns)
# LUT
4
8
8
8
16
16
4
4
8
16
8
16
10.2
10.7
13.4
15.7
13.6
16.3
719
1,368
2,911
6,020
5,924
12,165
based multipliers are synchronous. Partial products are inputs to an addition tree. For
all BCD additions the fast carry-chain adder of Sect. 11.1.4 has been used.
Input and output registers have been included in the design. Delays include FF
propagation and connections. The amount of FFs actually used is greater than
8*(M ? N) because the ISE tools [2] replicate the input register in order to reduce
fan-outs. The most useful data for area evaluation is the number of LUTs
(Table 11.5).
Comments
1. Observe that computation time and the required area are different for N by M
than M by N. That is due mainly by the use of carry logic in the adder tree.
11.5.3 Decimal Division Implementations

The algorithms have been implemented in the same Xilinx Virtex-5 device as
previous circuits in the chapter. The adders and multipliers used in division are the
ones described in previous sections. The non-restoring like division circuit correspond to Algorithm 11.4 (Fig. 11.13), meanwhile the SRT-like to Algorithm
11.5 (Fig. 11.15). In tables the number of decimal digits of dividend and divider is
expressed as N and the number of digits of quotient as P, meanwhile the period and
latency in ns (Table 11.6, 11.7).
Comments
1. Both types of dividers have approximately the same costs and similar delay.
They are also faster than equivalent binary dividers. As an example, the
computation time of an SRT-like 48-digit divider (n = p = 48) is about
5010.9 = 545 ns, while the computation time of an equivalent binary nonrestoring divider, that is a 160-bit one (48/log102 % 160), is more than 900 ns.
2. On the other hand, the area of the 48-digit divider (4,607 LUTs) is about five
times greater than that of the 160-bit binary divider (970 LUTs). Why so a great
difference? On the one hand it has been observed that decimal computation
resources (adders and subtractors) need about three times more slices than
binary resources (Sects. 11.4.1 and 11.4.2), mainly due to the more complex
definition of the carry propagate and carry generate functions, and to the final
302
11
st
5
1_y
3
9_y
2_y
3
... 3
ct
Decimal Operations
st=si /10n-3
ct=ci /10n-3
adder
5 w
y
n
range detection
q-(i+1)
carry-free multiplier
10si
p1
p0
n+2
n+2
10ci
n+2
n+2
p1 +p0= q-(i+1) y
4-to-2 counter
n+2
n+2
si+1
ci+1
register
n+2
n+2
si
ci
Fig. 11.15 SRT-like radix-10 algorithm data path

Table 11.6 Result for non-restoring division algorithm
N, P
FF
LUTS
Period
Latency
8
16
32
48
11.0
11.3
11.6
12.1
110.0
203.4
394.4
605.0
106
203
396
589
1,082
1,589
2,543
3,552
Table 11.7 Result for SRT-like division algorithm

N, P
FF
LUTS
Period
Latency
8
16
32
48
10.9
10.9
10.9
10.9
109.0
196.2
370.6
545.0
233
345
571
795
1,445
2,203
3,475
4,627
mod 10 reduction. On the other hand, the computation of the next quotient digit
is much more complex than in the binary case.
11.6
Exercises
303
11.6 Exercises
1. Implement a decimal comparator. Based on the decimal adder subtractor
architecture modify it to implement a comparator.
2. Implement a decimal greater than circuit that returns 1 if A C B else 0.
Tip: Base your design on the decimal adder subtractor.
3. Implement a N 9 2 digits circuit. In order to speed up computation analyze the
use of a 4 to 2 decimal reducer.
4. Implement a N 9 4 digits circuit using a 8 to 2 decimal reducer and only one
carry save adder.
5. Design a N by M digit multiplier using the N 9 2 or the N 9 4 digit multiplier.
Do you improve the multiplication time? What is the area penalty with respect
to the use of a N 9 1 multiplier?
6. Implement the binary digit-recurrence algorithm for decimal division (Algorithm 11.5). The key point is an efficient implementation of radix-B doubling,
adding, subtracting and halving.
References
1. Altera Corp (2011) Advanced synthesis cookbook. http://www.altera.com
2. Xilinx Inc (2011b) ISE Design Suite Software Manuals (v 13.1). http://www.xilinx.com
3. Bioul G, Vazquez M, Deschamps J-P, Sutter G (2010) High speed FPGA 10s complement
adders-subtractors. Int J Reconfigurable Comput 2010:14, Article ID 219764
4. Jaberipur G, Kaivani A (2007) Binary-coded decimal digit multiplier. IET Comput Digit Tech
1(4):377381
5. Sutter G, Todorovich E, Bioul G, Vazquez M, Deschamps J-P (2009) FPGA implementations
of BCD multipliers. V International Conference on ReConFigurable Computing and FPGAs
(ReConFig 09), Mexico
6. Xilinx Inc (2010) Virtex-5 FPGA Data Sheet, DS202 (v5.3). http://www.xilinx.com
7. Xilinx Inc (2011) XST User Guide UG687 (v 13.1). http://www.xilinx.com
Chapter 12
Floating Point Arithmetic
There are many data processing applications (e.g. image and voice processing),
which use a large range of values and that need a relatively high precision. In such
cases, instead of encoding the information in the form of integers or fixed-point
numbers, an alternative solution is a floating-point representation. In the first
section of this chapter, the IEEE standard for floating point is described. The next
section is devoted to the algorithms for executing the basic arithmetic operations.
The two following sections define the main rounding methods and introduce the
concept of guard digit. Finally, the last few sections propose basic implementations of the arithmetic operations, namely addition and subtraction, multiplication,
division and square root.
12.1 IEEE 754-2008 Standard

The IEEE-754 Standard is a technical standard established by the Institute of
Electrical and Electronics Engineers for floating-point operations. There are
numerous CPU, FPU and software implementations of this standard. The current
version is IEEE 754-2008 [1], which was published in August 2008. It includes
nearly all the definitions of the original IEEE 754-1985 and IEEE 854-1987
standards. The main enhancement in the new standard is the definition of decimal
floating point representations and operations. The standard defines the arithmetic
and interchange formats, rounding algorithms, arithmetic operations and exception
handling.

305
306
12 Floating Point Arithmetic
12.1.1 Formats
Formats in IEEE 754 describe sets of floating-point data and encodings for
interchanging them. This format allows representing a finite subset of real numbers. The floating-point numbers are represented using a triplet of natural numbers
(positive integers). The finite numbers may be expressed either in base 2 (binary)
or in base 10 (decimal). Each finite number is described by three integers: the sign
(zero or one), the significand s (also known as coefficient or mantissa), and the
exponent e. The numerical value of the represented number is (-1)sign 9 s 9 Be,
where B is the base (2 or 10).
For example, if sign = 1, s = 123456, e = -3 and B = 10, then the represented number is -123.456.
The format also allows the representation of infinite numbers (+? and -?),
and of special values, called Not a Number (NaN), to represent invalid values. In
fact there are two kinds of NaN: qNaN (quiet) and sNaN (signaling). The latter,
used for diagnostic purposes, indicates the source of the NaN.
The values that can be represented are determined by the base (B), the number
of digits of the significand (precision p), and the maximum and minimum values
emin and emax of e. Hence, s is an integer belonging to the range 0 to Bp-1, and e is
an integer such that emin e emax :
For example if B = 10 and p = 7 then s is included between 0 and 9999999. If
emin = -96 and emax = 96, then the smallest non-zero positive number that can be
represented is 1 9 10-101, the largest is 9999999 9 1090 (9.999999 9 1096), and
the full range of numbers is from -9.999999 9 1096 to 9.999999 9 1096. The
numbers closest to the inverse of these bounds (-1 9 10-95 and 1 9 10-95) are
considered to be the smallest (in magnitude) normal numbers. Non-zero numbers
between these smallest numbers are called subnormal (also denormalized)
numbers.
Zero values are finite values whose significand is 0. The sign bit specifies if a
zero is +0 (positive zero) or -0 (negative zero).
12.1.2 Arithmetic and Interchange Formats

The arithmetic format, based on the four parameters B, p, emin and emax, defines the
set of represented numbers, independently of the encoding that will be chosen for
storing and interchanging them (Table 12.1). The interchange formats define fixedlength bit-strings intended for the exchange of floating-point data. There are some
differences between binary and decimal interchange formats. Only the binary format
will be considered in this chapter. A complete description of both the binary and the
decimal format can be found in the document of the IEEE 754-2008 Standard [1].
For the interchange of binary floating-point numbers, formats of lengths equal
to 16, 32, 64, 128, and any multiple of 32 bits for lengths bigger than 128, are
12.1
IEEE 754-2008 Standard
307
Table 12.1 Binary and decimal floating point format in IEEE 754-2008
Binary formats (B = 2)
Decimal formats (B = 10)
Parameter
Binary
16
Binary
32
Binary
64
Binary
128
Decimal Decimal
132
l64
p, digits
emax
emin
Common
name
10 ? 1
+15
-14
Half
precision
23 ? 1
+127
-126
Single
precision
52 ? 1
+1023
-1022
Double
precision
112 ? 1
+16383
-16382
Quadruple
precision
7
+96
-95
16
+384
-383
Decimal
128
34
+16,383
-16,382
Table 12.2 Binary interchange format parameters

Parameter
Binary16 Binary32 Binary64 Binary128 Binary{k} (k C 128)
k, storage width in bits
p, precision in bits
emax
bias, E - e
w, exponent field width
t, trailing significand bits
16
11
15
15
5
10
32
24
127
127
8
23
64
53
1,023
1,023
11
52
128
113
16,383
16,383
15
112
Multiple of 32
k-w
2kp1 1
emax
Round(4log2 k) - 13
k-w-1
Fig. 12.1 Binary interchange format
defined (Table 12.2). The 16-bit format is only for the exchange or storage of small
numbers.
The binary interchange encoding scheme is the same as in the IEEE 754-1985
standard. The k-bit strings are made up of three fields (Fig. 12.1):
a 1-bit sign S,
a w-bit biased exponent E = e ? bias,
the p - 1 trailing bits of the significand; the missing bit is encoded in the
exponent (hidden first bit).
Each binary floating-point number has just one encoding. In the following
description, the significand s is expressed in scientific notation, with the radix point
immediately following the first digit. To make the encoding unique, the value of the
significand s is maximized by decreasing e until either e = emin or s C 1 (normalization). After normalization, there are two possibilities regarding the significand:
If s C 1 and e C emin then a normalized number of the form 1.d1 d2dp
obtained. The first 1 is not stored (implicit leading 1).
-1
is
308
If e emin and 0 \ s \ 1, the floating-point number is called subnormal. Subnormal numbers (and zero) are encoded with a reserved biased exponent value.
They have an implicit leading significand bit = 0 (Table 12.2).
The minimum exponent value is emin 1 emax : The range of the biased
exponent E is 1 to 2w - 2 to encode normal numbers. The reserved value 0 is used
to encode 0 and subnormal numbers. The value 2w - 1 is reserved to encode
? and NaNs.
The value of a binary floating-point data is inferred from the constituent fields
as follows:
If E 2w 1 (all 1s in E), then the data is an NaN or infinity. If T = 0, then it
is a qNaN or an sNaN. If the first bit of T is 1, it is an sNaN. If T = 0, then the
value is 1sign 1:
If E = 0 (all 0s in E), then the data is 0 or a subnormal number. If T = 0 it is a
signed 0. Otherwise (T = 0), the value of the corresponding floating-point
number is 1Sign 2emin 0 21p T :
If 1 B E B 2w - 2, then the data is 1sign 2Ebias 1 21 p T :
Remember that the significand of a normal number has an implicit leading 1.
Example 12.1
Convert the decimal number -9.6875 to its binary32 representation.
First convert the absolute value of the number to binary (Chap. 10):
9:687510 1001:10112 :
Normalize: 1001:1011 1:00110011 23 : Hence e 3; s 1:0011 0011:
Hide the first bit and complete with 0s up to 23 bits:
00110011000000000000000.
Add bias to the exponent. In this case, w = 8, bias = 28 - 1 = 127 and thus
E e bias 3 127 13010 100000102 :
Compose the final 32-bit representation:
1 10000010 001100110000000000000002 C119800016 :
Example 12.2
Convert the following binary32 numbers to their decimal representation.
7FC0000016: sig n 0; E FF16 ; T 6 0; hence it is an NaN. Since the first bit
of T is 1, it is a quiet NaN.
FF80000016: sig n 0; E FF16 ; T 0; hence it is -?.
6545AB7816 : sign 0; E CA16 ; 20210 ; e E bias 202 127 7510 ;
T 100010110101011011110002 ;
s 1:100010110101011011110002 1:544295310 :
The number is 1:100010110101011011112 275 5:8341827 1022 :
12.1
IEEE 754-2008 Standard
309
1234567816 : sign 0; E 2416 3610 ; e E bias 36 127 9110 ;

T 011010001010110011110002 ;
s 1:011010001010110011110002 1:40888810 :
The number is 1:01101000101011001111
291 5:69045661 1028 :8000000016: sign = 1, E = 0016, T = 0. The
number is -0.0.
8000000016: sign = 1, E = 0016, T = 0. The number is -0.0.
0001234516:sign = 0, E = 0016, T = 0, hence it is a subnormal number,
e E bias 127; T 00000100100011010001012 ;
s 0:00000100100011010001012 0:017777710 :
The number is 0:0000010010001101000101 2127 1:0448782 1040 :
12.2 Arithmetic Operations

First analyze the main arithmetic operations and generate the corresponding
computation algorithms. In what follows it will be assumed that the significand s is
represented in base B (in binary if B = 2, in decimal if B = 10) and that it belongs
to the interval 1 B s B B - ulp, where ulp stands for the unit in the last position
or unit of least precision. Thus s is expressed in the form (s0 s-1 s-2 s-p).
Be where emin e emax and 1 B s0 B B - 1.
Comment 12.1
The binary subnormals and the decimal floating point are not normalized numbers
and are not included in the following analysis. This situation deserves some special
treatment and is out of the scope of this section.
12.2.1 Addition of Positive Numbers

Given two positive floating-point numbers s1 Be1 and s2 Be2 their sum s Be is
computed as follows: assume that e1 is greater than or equal to e2; then (alignment)
the sum of s1 Be1 and s2 Be2 can be expressed in the form s Be where

s s1 s2 = Be1e2 and e e1 :
12:1
The value of s belongs to the interval
1 s 2 B 2 ulp;
12:2
so that s could be greater than or equal to B. If it is the case, that is if

B s 2 B 2 ulp;
12:3
310
then (normalization) substitute s by s/B, and e by e ? 1, so that the value of s Be

is the same as before, and the new value of s satisfies
1 s 2 2=B ulp B ulp:
12:4
The significands s1 and s2 of the operands are multiples of ulp. If e1 is greater than
e2, the value of s could no longer be a multiple of ulp and some rounding function
should be applied to s. Assume that s0 \ s \ s00 = s0 ? ulp, s0 and s00 being two
successive multiples of ulp. Then the rounding function associates to s either s0 or s00 ,
according to some rounding strategy. According to (12.4) and to the fact that 1 and
Bulp are multiples of ulp, it is obvious that 1 B s0 \ s00 B Bulp. Nevertheless, if
the condition (12.3) does not hold, that is if 1 B s \ B, s could belong to the interval
B ulp\s\B;
12:5
so that rounding(s) could be equal to B, then a new normalization step would be

necessary, i.e. substitution of s = B by s = 1 and e by e ? 1.
Algorithm 12.1: Sum of positive numbers
Examples 12.3
Assume that B = 2 and ulp = 2-5, so that the numbers are represented in the form
s 2e where 1 B s B 1.111112. For simplicity e is written in decimal (base 10).

1. Compute z 1:10101 23 1:00010 21 :
Alignment: z 1:10101 0:000100010 23 1:101110010 23 :
Rounding: s 1:10111:
Final result: z 1:10111 23 :
2. Compute z 1:11010 23 1:00110 22 :

Alignment:z 1:11010 0:100110 23 10:011010 23 :
Normalization : s 1:0011010; e 4:
12.2
Arithmetic Operations
311

3. Compute z 1:10101 23 1:10101 21 :
Alignment: z 1:10010 0:0110101 23 1:1111101 23 :
Normalization: s 1:00000; e 4:
Final result : z 1:00000 24 :
Comments 12.2
1. The addition of two positive numbers could produce an overflow as the final
value of e could be greater than emax.
2. Observe in the previous examples the lack of precision due to the small number
of bits (6 bits) used in the significand s.
12.2.2 Difference of Positive Numbers

Given two positive floating-point numbers s1 Be1 and s2 Be2 their difference s Be is
computed as follows: assume that e1 is greater than or equal to e2; then (for alignment)
the difference between s1 Be1 and s2 Be2 can be expressed in the form s Be where

s s1 s2 = Be1e2 and e e1 :
12:6
The value of s belongs to the interval s B ulp s B ulp: If s is negative, then it is substituted by and the sign of the final result will be modified
accordingly. If s is equal to 0, then an exception equal_zero could be raised. It
remains to consider the case where 0\s B ulp: The value of s could be smaller
than 1. In order to normalize the significand, s is substituted by s Bk and e by e - k,
where k is the minimum exponent k such that s Bk 1: Thus, the relation 1 s B
holds. It remains to round (up or down) the significand and to normalize it if
necessary.
In the following algorithm, the function leading_zeroes(s) computes the
smallest k such that s Bk C 1.
Algorithm 12.2: Difference of positive numbers
312
Examples 12.4
Assume again that B = 2 and ulp = 2-5, so that the numbers are represented in
the form s 2e where 1 B s B 1.111112. For computing the difference, the 2s
complement representation is used (one extra bit is used).

1. Compute z 1:10101 22 1:01010 21 :
Alignment: z 0:00110101 1:01010 21 :
20 s complement addition: 00:00110101 10:10101 00:00001 21
10:11100101 21 :
Change of sign: s 01:00011010 00:00000001 01:00011011:

2. Compute z 1:00010 23 1:10110 22 :
Alignment: z 1:00010 0:110110 23 :
20 s complement addition: 01:00010 11:001001 00:000001 23
00:001110 23 :
Leading zeroes: k 3; s 1:11000; e 0:

3. Compute z 1:01010 23 1:01001 21 :
Alignment: z 1:01010 0:0101001 23 0:1111111 23 :
Leading zeroes: k 1; s 1:111111; e 2:
Normalization: s 1:0000; e 3:
Final result:z 1:00000 23 :
Comment 12.3
The difference of two positive numbers could produce an underflow as the final
value of e could be smaller than emin.
12.2
Table 12.3 Effective

operation in floating point
addersubtractor
313
Operation
Sign1
Sign2
Actual
operation
0
0
0
0
1
1
1
1
0
0
1
1
0
0
1
1
0
1
0
1
0
1
0
1
s1 ?
s1 -(s1
-(s1
s1 s1 ?
-(s1
-(s1
s2
s2
s2
s2
?
-
s2)
s2)
s2)
s2)
12.2.3 Addition and Subtraction

Given two floating-point numbers 1sign1 s1 Be1 and 1sign2 s2 Be2 ; and a
control variable operation, an algorithm is defined for computing
z 1sign s Be 1sign1 s1 Be1 1sign2 s2 Be2 ;
z 1
sign
sign1
s B 1
e1
s1 B 1
sign2
e2
s2 B ;
if operation 0;
if operation 1:
Once the significands have been aligned, the actual operation (addition or
subtraction of the significands) depends on the values of operation, sign1 and sign2
(Table 12.3). The following algorithm computes z. The procedure swap (a, b)
interchanges a and b.
Algorithm 12.3: Addition and subtraction
314
12.2.4 Multiplication
Given two floating-point numbers 1sign1 s1 Be1 and 1sign2 s2 Be2 their
product (-1)sign s Be is computed as follows:
sign sign1 xor sign2 ; s s1 s2 ; e e1 e2 :
12:7
The value of s belongs to the interval 1 B s B (B - ulp) , and could be greater

than or equal to B. If it is the case, that is if B B s B (B - ulp)2, then (normalization) substitute s by s/B, and e by e ? 1. The new value of s satisfies
1 s B ulp2 =B B 2:ulp ulp2 =B\B ulp
12:8
(ulp \ B so that 2 - ulp/B [ 1). It remains to round the significand and to

normalize if necessary.
Algorithm 12.4: Multiplication
Examples 12.5
the form s 2e where 1 B s B 1.111112. The exponent e is represented in decimal.

1. Compute z 1:11101 22 1:00010 25 :
Multiplication : z 01:1100101100 23 :
Rounding : s 1:11001:

2. Compute z 1:11101 23 1:00011 21 :
Normalization : s 1:000010101112 ;
e 3:
12.2
315

3. Compute z 1:01000 x 21 1:10011 22 :
Normalization : s 1:11111;
e 3:
Normalization : s 1;
e 4:
Comment 12.4
The product of two real numbers could produce an overflow or an underflow as the
final value of e could be greater than emax or smaller than emin (addition of two
negative exponents).
12.2.5 Division
Given two floating-point numbers 1sign1 s1 Be1 and 1sign2 s2 Be2 their
quotient
(-1)sign s Be is computed as follows:
sign sign1 xor sign2 ;
s s1 =s2 ;
e e1 e2 :
12:9
The value of s belongs to the interval 1/B \ s B B - ulp, and could be smaller
than 1. If that is the case, that is if s s1 =s2 \1; then s1 \s2 ; s1 s2
ulp; s1 =s2 1 ulp=s2 \1 ulp=B; and 1=B\s\1 ulp=B:
Then (normalization) substitute s by s B, and e by e - 1. The new value of
s satisfies 1 \ s \ B - ulp. It remains to round the significand.
Algorithm 12.5: Division
Examples 12.6
the form s 2e where 1 B s B 1.111112. The exponent e is represented in
decimal.
316
1. Compute z 1:11101 23

1:00011 21 :
Division: z 1:1011111000 24 :
2. Compute z 1:01000 21

1:10011 22 :
Division: z 0:1100100011 21 :

Normalization: s 1:100100011;
e 2:
Comment 12.5
The quotient of two real numbers could produce an underflow or an overflow as the
final value of e could be smaller than emin or bigger than emax. Observe that a second
normalization is not necessary as in the case of addition, subtraction and multiplication.
12.2.6 Square Root

Given a positive floating-point number s1 Be1, its square root s Be is computed
as follows:
if e1 is even;
if e1 is odd;
s s1 1=2 ;
s s1 =B1=2 ;
e e1 =2;
e e1 1=2:
12:10
12:11
1/2
In the first case (12.10), 1 B s B (B - ulp) \ B - ulp. In the second case (1/
B)1/2 B s \ 1. Hence (normalization) s must be substituted by s B and e by e - 1,
so that 1 B s \ B. It remains to round the significand and to normalize if necessary.
Algorithm 12.6: Square root
An alternative is to replace (12.11) by:

if e1 is odd;
s s1 B1=2 ;
e e1 1=2:
12:12
12.2
317
In this case B1=2 s B2 ulp B1=2 \B; then the first normalization is not
necessary. Nevertheless, s could be B - ulp \ s \ B, and then depending on the
rounding strategy, normalization after rounding could be still necessary.
Algorithm 12.7: Square root, second version
Note that the round to nearest (default rounding in IEEE 754-2008) and the
truncation rounding schemes allow avoiding the second normalization.
Examples 12.7
Assume again that B = 2 and ulp = 2-5, so that the numbers are represented in the
form s 2e where 1 B s B 1.111112. The exponent e is represented in decimal form.
1=2
1. Compute z 1:11101 24
:
Square rooting : z 1:01100001 22 :
1=2
2. Compute z 1:00101 21 :
Even exponent : s 10:0101;
e 2:

Rounding : s 1:10000
3. Compute z 1:11111 23
1=2
Even exponent; s 11:1111;
:
e 2:

Rounding : s 1:11111round to nearest:
318
However, some rounding schemes (e.g. toward infinite) generate s 10:00000:

Then, the result after normalization is s 1:00000; e 2; and the final result
z 1:00000 22 :
Comment 12.6
The square rooting of a real number could produce an underflow as the final value
of e could be smaller than emin.
12.3 Rounding Schemes

Given a real number x and a floating-point representation system, the following
situations could happen:
j xj\smin Be min ; that is, an underflow situation,
j xj [ smax Be max ; that is, an overflow situation,
j xj s Be ; where emin e emax and smin s smax :
In the third case, either s is a multiple of ulp, in which case a rounding operation
is not necessary, or it is included between two multiples s0 and s00 of ulp:
s0 \ s \ s00 .
The rounding operation associates to s either s0 or s00 , according to some
rounding strategy. The most common are the following ones:
The truncation method (also called round toward 0 or chopping) is accomplished by dropping the extra digits, i.e. round(s) = s0 if s is positive, round(s) = s00 if s is negative.
The round toward plus infinity is defined by round(s) = s00 .
The round toward minus infinity is defined by round(s) = s0 .
The round to nearest method associates s with the closest value, that is, if
s \ s0 ? ulp/2, round(s) = s0 , and if s [ s0 ? ulp/2, round(s) = s00 .
If the distances to s0 and s00 are the same, that is, if s = s0 ? ulp/2, there are
several options. For instance:
round(s) = s0 ;
round(s) = s00 ;
round(s) = s0 if s is positive, round(s) = s00 if s is negative. It is the round to
nearest, ties to zero scheme.
round(s) = s00 if s is positive, round(s) = s0 if s is negative. It is the round to
nearest, ties away from zero scheme.
round(s) = s0 if s0 is an even multiple of ulp, round(s) = s00 if s00 is an even
multiple of ulp. It is the default scheme in the IEEE 754 standard.
round(s) = s0 if s0 is an odd multiple of ulp, round(s) = s00 if s00 is an odd
multiple of ulp.
12.3
Rounding Schemes
319
The preceding schemes (round to the nearest) produce the smallest absolute
error, and the two last (tie to even, tie to odd) also produce the smallest average
absolute error (unbiased or 0-bias representation systems).
Assume now that the exact result of an operation, after normalization, is
s 1:s1 s2 s3 . . . sp jsp1 sp2 sp3 . . .
where ulp is equal to B-p (the | symbol indicates the separation between the digit
which corresponds to the ulp and the following). Whatever the chosen rounding
scheme, it is not necessary to have previously computed all the digits s-(p+1) s(p+2); it is sufficient to know whether all the digits s-(p+1) s-(p+2) are equal to
0, or not. For example the following algorithm computes round(s) if the round to
the nearest, tie to even scheme is used.
Algorithm 12.8: Round to the nearest, tie to even
In order to execute the preceding algorithm it is sufficient to know the value of

s1 1 s1 s2 s3 . . . sp ; the value of s-(p+1), and whether s2
0:00 . . . 0j0 sp2 sp3 . . . is equal to 0, or not.
12.3.1 Rounding Schemes in IEEE 754

From the previous description, the IEEE 754-2008 standard defines five rounding
algorithms. The two first round to a nearest value; the others are called directed
roundings:
Round to nearest, ties to even; this is the default rounding for binary floatingpoint and the recommended default for decimal.
Round to nearest, ties away from zero.
Round toward 0directed rounding towards zero.
Round toward +?directed rounding towards positive infinity
Round toward -?directed rounding towards negative infinity.
320
12.4 Guard Digits

Consider the exact result r of an operation, before normalization. According to the
preceding paragraph:
r\B2 ; i:e: r r1 r0 r1 r2 r3 . . . rp jrp1 rp2 rp3 . . .
The normalization operation (if necessary) is accomplished by
dividing the result by B (sum of positive numbers, multiplication),
multiplying the result by B (division),
multiplying the result by Bk (difference of positive numbers).
Furthermore, if the operation is a difference of positive numbers (Algorithm
12.2), consider two cases:
if e1 e2 2; then r s1 s2 =Be1e2 [ 1 B=B2 1
1=B 1=Bas B 2; so that the number k of leading zeroes is equal to 0 or 1, and
the normalization operation (if necessary i.e. k = 1) is accomplished by multiplying the result by B;
if e1 - e2 B 1, then the result before normalization is either
r0 r1 r2 r3 . . . rp jrp1 0 0 . . . e1 e2 1; or
r0 r1 r2 r3 . . . rp j0 0 0 . . . e1 e2 0:
A consequence of the preceding analysis is that the result after normalization
can be either
r0 r1 r2 r3 . . . rp jrp1 rp2 rp3 . . . no normalization operation;
12:13
or
r1 r0 r1 r2 . . . rp1 jrp rp1 rp2 . . . divide by B;
12:14
r1 r2 r3 r4 . . . rp1 jrp2 rp3 rp4 . . . multiply by B;
12:15
or
or
rk rk1 rk2 . . . rp rp1 0 . . . 0j0 0 . . . multiply by Bk where k [ 1:
12:16
For executing a rounding operation, the worst case is (12.15). In particular, for
executing Algorithm 12.8, it is necessary to know
12.4
Guard Digits
321
the value of s1 r1 r2 r3 r4 . . . rp1 ;

the value of rp2 ;
whether s2 0:00 . . . 0j0 rp3 rp4 . . . is equal to 0, or not.
The conclusion is that the result r of an operation, before normalization, must
be computed in the form
r r1 r0 r1 r2 r3 . . . rp jrp1 rp2 T;
that is, with two guard digits r-(p+1) and r-(p+2)-, and
an additional sticky digit
T equal to 0 if all the other digits rp3 ; rp4 ; . . . are equal to 0, and equal to
any positive value otherwise.
After normalization, the significand will be obtained in the following general form:
s 1:s1 s2 s3 . . . sp jsp1 sp2 sp3 :
The new version of Algorithm 12.8 is the following:
Algorithm 12.9: Round to the nearest, tie to even, second version
Observe that in binary representation, the following algorithm is even simpler.

Algorithm 12.10: Round to the nearest, tie to even, third version
12.5 Arithmetic Circuits

This section proposes basic implementations of the arithmetic operations, namely
addition and subtraction, multiplication, division and square root. The implementation
is based on the previous section devoted to the algorithms, rounding and guard digit.
322
The circuits support normalized binary IEEE 754-2008 operands. Regarding the
binary subnormals the associated hardware to manage this situation is complex.
Some floating point implementations solve operations with subnormals via software routines. In the FPGA arena, most cores do not support denormalized
numbers. The dynamic range can be increased using fewer resources by increasing
the size of the exponent (a 1-bit increase in exponent, roughly doubles the dynamic
range) and is typically the solution adopted.
12.5.1 AdderSubtractor
An adder-subtractor based on Algorithm 12.3 will now be synthesized. The
operands are supposed to be in IEEE 754 binary encoding. It is made up of five
parts, namely unpacking, alignment, addition, normalization and rounding, and
packing. The following implementation does not support subnormal numbers; they
are interpreted as zero.
12.5.1.1 Unpacking
The unpacking separates the constitutive parts of the Floating Points and additionally detects the special numbers (infinite, zeros and NaNs). The special number
detection is implemented using simple comparators. The following VHDL process
defines the unpacking of a floating point operand FP; k is the number of bits of FP,
w is the number of bits of the exponent, and p is the significand precision.
The previous is implemented using two w bits comparators, one p bits comparator and some additional gates for the rest of the conditions.
12.5.1.2 Alignment
The alignment circuit implements the three first lines of the Algorithm 12.3, i.e.
12.5
Arithmetic Circuits
e1
323
e2
sign1 sign2 operation
s1
s2
actual_sign2
subtractor
subtractor
e1-e2
e2-e1
sign(e1-e2)
new_s2
0
s
000
dif
right shifter
(from 0 to p+3)
sign
new_sign2
aligned_s2
000
aligned_s1
Fig. 12.2 Alignment circuit in floating point addition/subtraction
An example of implementation is shown in Fig. 12.2. The principal and more

complex component is the right shifter.
Given a (p ? 3)-bit input vector 1 ak1 ak2 . . . a1 a0 0 0 0; the shifter generates a (p ? 3)-bit output vector. The lsb (least significand bit) of the output is the
sticky bit indicating if the shifted bits are equal to zero or not. If B = 2, the stickydigit circuit is an OR circuit.
Observe that if e1 e2 p 3; then the shifter output is equal to [0 0 0 1],
since the last bit is the sticky bit and the input number is a non-zero. The operation
with zero is treated differently.
12.5.1.3 Addition and Subtraction

Depending on the respective signs of the aligned operands, one of the following
operations must be executed: if they have the same sign, the sum aligned_s1 ? aligned_s2 must be computed; if they have different signs, the difference
324
aligned_s1aligned_s2 is computed. If the result of the significand difference is

negative, then aligned_s2aligned_s1 must be computed. Moreover, the only
situation in which the final result could be negative is when the e1 = e2.
In the circuit of Fig. 12.3 two additions are performed in parallel: result = aligned_s1 aligned_s2, where the actual operation is selected with the signs of the
operands, and alt_result = s2 - s1. At the end of this stage a multiplexer selects
the correct results. The operation selection is done as follows:
12.5.1.4 Normalization and Rounding

The normalization circuit executes the following part of Algorithm 12.3:
If the number of leading zeroes is greater than p 3; i:e: s1 s2 [ Bp2 ;

then s2 [ s1 Bp2 : If e1 were greater than e2 then s2 B ulp=B
1 Bp1 so that 1 Bp1 s2 \s1 Bp2 1 Bp2 ; which is
impossible. Thus, the only case where the number of leading zeroes can be greater
than p ? 3 is when e1 = e2 and s1 = s2. If more than p ? 3 leading 0s are
detected in the circuit of Fig. 12.4, a zero_flag is raised.
It remains to execute the following algorithm where operation is the internal
operation computed in Fig. 12.3:
12.5
Arithmetic Circuits
325
s1
s2
aligned_s1
aligned_s2
sign
sign2
(p+1)-bits subtractor
alt_result
alt_result >0
operation
e1
e2
significand
selection
iszero1
iszero2
(p+4)-bits adder /
subtractor
operation
result
s1 &000 alt_result & 000

s2 &000
result
3
signif
Fig. 12.3 Effective addition and subtraction
The rounding depends on the selected rounding strategy. An example of

rounding circuit implementation is shown in Fig. 12.4. If the round to the nearest,
tie to even method is used (Algorithm 12.10), the block named rounding decision
computes the following Boolean function decision:
The second rounding is avoided in the following way. The only possibility to need a
second rounding is when, as the result of an addition, the significand is 1.1111111xx.
This situation is detected in a combinational block that generates the signal isTwo
and adds one to the exponent. After rounding, the resulting number is 10.000000, but
the two most significand bits are discarded and the hidden 1 is appended.
12.5.1.5 Packing
The packing joins the constitutive parts of the floating point result. Additionally
depending on special cases (infinite, zeros and NaNs), generates the corresponding
codification.
Example 12.8 (Complete VHDL code available)
Generate the VHDL model of an IEEE decimal floating-point adder-subtractor. It
is made up of five previously described blocks. Fig. 12.5 summarizes the interconnections. For clearness and reusability the code is written using parameters,
326

signif
e
e
=
111...111
/B
+1
isTwo
signif(p+3)
0
1 0
leading
0's
zero_flag
sign
operation
x Bk
-k
alt_result > 0
Sign
computation
Rounding
new_e
p-bits
half adder
s3
s2
s1
s0
rounding
decision
new_sign
s(p+2:3)
1 (ulp)
new_s
Fig. 12.4 Normalization and rounding
where K is the size of the floating point numbers (sign, exponent, significand), E is
the size of the exponent and P is the size of the significand (including the hidden
1). The entity declaration of the circuit is:
For simplicity the code is written as a single VHDL code except additional files
to describe the right shifter of Fig. 12.2 and the leading zero detection and shifting
of Fig. 12.4. The code is available at the book home page.
12.5
Arithmetic Circuits
327
FP1
FP2
isInf1
add/sub
Unpack
e2
e1
sign1 sign2
isInf2
isZ1 isZ2 s1
s1
sign1 sign2
s2
e2
e1
sign1 sign2
s2 isNaN1 isNaN2
e2
e1
isZ1 isZ2 s1
s2
Alignment
sign
new_sign2
sign
sign2
aligned_s1 aligned_s2
aligned_s1 aligned_s2 s1 s2
Addition / Subtraction
operation
sign
operation
significand
significand
Normalization and Rounding

new_sign
sign
new_e
new_s
s
Pack
FP
overflow
undeflow
zero_flag
Conditions
Detection
isInf
isZero
isNaN
FP
Fig. 12.5 General structure of a floating point adder-subtractor
A two stage pipeline could be achieved dividing the data path between addtion/
subtraction and normalization and rounding stages (dotted line in Fig. 12.5).
12.5.2 Multiplier
A basic multiplier deduced from Algorithm 12.4 is shown in Fig. 12.6. The
unpacking and packing circuits are the same as in the case of the adder-subtractor
(Fig. 12.5, Sects. 12.5.1.1 and 12.5.1.5) and for simplicity, are not drawn. The
normalization and rounding is a simplified version of Fig. 12.4, where the part
related to the subtraction is not necessary.
328

s1
s2
e1
sign1 sign2
e2
adder
p-by-p-bits
multiplier
p-4 .. 0
2.p-1 .. 0
2.p-1 .. p-3
sticky digit
generation
Normalization
prod(p+3 .. 0)
e
/B
=
111...111
+1
isTwo
prod(p+3)
1
Rounding
s(p+2:3)
p-bits
half adder
s3
s2
s1
s0
rounding
decision
1 (ulp)
sign
Fig. 12.6 General structure of a floating point multiplier
The obvious method of computing the sticky bit is with a large fan-in OR gate
on the low order bits of the product. Observe, in this case, that the critical path
includes the p by p bits multiplication and the sticky digit generation.
An alternative method consists of determining the number of trailing zeros in
the two inputs of the multiplier. It is easy to demonstrate that the number of
trailing zeros in the product is equal to the sum of the number of trailing zeros in
each input operand. Notice that this method does not require the actual low order
product bits, just the input operands, so the computation can occur in parallel with
the actual multiply operation, removing the sticky computation from the critical
path.
The drawback of this method is that significant extra hardware is required. This
hardware includes two long length priority encoders to count the number of
12.5
Arithmetic Circuits
s1
329
s2
e1
sticky bit
generation
e2 sign1 sign2 isInf1 isNan1 isZero1

isInf2 isNan2 isZero2
conditions
detection
adder
p-by-p-bits
multiplier
isZero
2.p-1 .. 0
2.p-1 .. p-3
isInf
isNan
sticky bit
prod(p+3 .. 0)
e
Normalization and
Rounding
conditions
adjust
isZero
sign
isInf isNan
underflow overflow isZero
Fig. 12.7 A better general structure of a floating point multiplier
trailing zeros in the input operands, a small length adder, and a small length
comparator. On the other hand, some hardware is eliminated, since the actual low
order bits of the product are no longer needed.
A faster floating point multiplier architecture that computes the p by p multiplication and the sticky bit in parallel is presented in Fig. 12.7. The dotted lines
suggest a three stage pipeline implementation using a two stage p by p multiplication. The two extra blocks are shown to indicate the special conditions detections. In the second block, the range of the exponent is controlled to detect
overflow and underflow conditions. In this figure the packing and unpacking
process are omitted for simplicity.
Example 12.9 (complete VHDL code available)
Generate the VHDL model of a generic floating-point multiplier. It is made up of
the blocks depicted in Fig. 12.7 described in a single VHDL file. For clearness and
reusability the code is written using parameters, where K is the size of the floating
point numbers (sign, exponent, significand), E is the size of the exponent and P is
the size of the significand (including the hidden 1). The entity declaration of the
circuit is:
330

s1
s2
e1
sign1 sign2
e2
subtractor
p-by-p-bits
divider
q
e
r
p-1 .. 0
p+2 .. 0
sticky digit
generation
Normalization
div(p+3 .. 0)
e
*B
-1
div(p+2)
0
quotient < 0
Rounding
sign
Fig. 12.8 A general structure of a floating point divider
The code is available at the home page of this book. The combinational circuit
registers inputs and outputs to ease the synchronization. A two or three stage
pipeline is easily achievable adding the intermediate registers as suggested in
Fig. 12.7. In order to increase the clock frequency, more pipeline registers can be
inserted into the integer multiplier.
12.5
Arithmetic Circuits
331
s1
s2
e1
e2
sign1 sign2
subtractor
e
p-by-p-bits
divider
Sticky bit
div(p+3 .. 0)
Normalization and
Rounding
s
sign
Fig. 12.9 A pipelined structure of a floating point divider
12.5.3 Divider
A basic divider, deduced from Algorithm 12.6, is shown in Fig. 12.9. The
unpacking and packing circuits are similar to those of the adder-subtractor or
multiplier. The normalize and rounding is a simplified version of Fig. 12.4,
where the part related to the subtraction is not necessary.
The inputs of the p-bit divider are s1 and s2. The first operator s1 is internally
divided by two (s1/B, i.e. right shifted) so that the dividend is smaller than the
divisor. The precision is chosen equal to p ? 3 digits. Thus, the outputs quotient
and remainder satisfy the relation (s1/B).Bp+3 = s2.q ? r where r \ s2, that is,
s1 =s2 q:Bp2 r=s2 Bp2 where r=s2 Bp2 \Bp2 :
The sticky digit is equal to 1 if r \ 0 and to 0 if r = 0. The final approximation
of the exact result is
quotient q:Bp2 sticky digit: Bp3 :
332
If a non-restoring divider is used, a further optimization could be done in the

sticky bit computation. In the non-restoring divider the final remainder could be
negative. In this case, a final correction should be done. This final operation can be
avoided: the sticky bit is 0 if the remainder is equal to zero, otherwise it is 1.
A divider is a time consuming circuit. In order to obtain circuits, with frequencies similar to those of floating point multipliers or adders, more pipeline
stages are necessary. Figure 12.9 shows a possible pipeline for a floating point
divider where the integer divider is also pipelined.
Generate the VHDL model of a generic floating-point divider. It is made up of the
blocks depicted in Fig. 12.9. Most of the design is described in a single VHDL file,
but for the integer divider. The integer divider is a non-restoring divider
(div_nr_wsticky.vhd) that uses a basic non-restoring cell (a_s.vhd). For clearness
and reusability the code is written using parameters, where K is the size of the
floating point numbers (sign, exponent, significand), E is the size of the exponent
and P is the size of the significand (including the hidden 1). The entity declaration
of the circuit is:
12.5.4 Square Root

A basic square rooter deduced from Algorithm 12.7 is shown in Fig. 12.10. The
unpacking and packing circuits are the same as in previous operations and, for
simplicity, are not drawn. Remember that the first normalization is not necessary,
and for most rounding strategies the second normalization is not necessary either.
The exponent is calculated as follows:
E e1 =2 bias E1 bias=2 bias E1 =2 bias=2:
If e1 is even, then both E1 and bias are odd (bias is always odd). Thus, E
bE1 =2c bbias=2c 1 where bE1 =2c and bbias=2c amount to right shifts.
If e1 is odd, then E1 is even. The significand is multiplied by 2 and the exponent
reduced
by
one
unit.
The
biased
exponent
E E1 1=2 bias=2 E1 =2 bbias=2c:
To summarize, the biased exponent E bE1 =2c bbias=2c parityE1 :
12.5
Arithmetic Circuits
333
Fig. 12.10 Structure of a

floating point squarer
s1
e1
-1
mod 2
p+1 .. 0
/2
(p+2)-bits
square rooter
sticky
sticky bit
sqrt(p+1 .. 0)
p+2 .. 0
Normalization
s
Since the integer square root is a complex operation, a pipelined floating-point

square rooter should be based on a pipelined integer square root. The dotted line of
Fig. 12.10 shows a possible five stage pipeline circuit.
Generate the VHDL model of a generic floating-point square rooter. It is made up
of the blocks depicted in Fig. 12.10. The design is described in a single VHDL file,
but for the integer square root. The integer square root is based in a non-restoring
algorithm (sqrt_wsticky.vhd) that uses two basic non-restoring cells (sqrt_cell.vhd
and sqrt_cell_00.vhd). For clearness and reusability, the code is written using
parameters, where K is the size of the floating point numbers (sign, exponent,
significand), E is the size of the exponent and P is the size of the significand
(including the hidden 1). The entity declaration of the circuit is:
334
Table 12.4 Combinational floating point operators in binary32 format

FF
LUTs
DSP48
Slices
Delay
FP_add
FP_mult
FP_mult_luts
FP_div
FP_sqrt
11.7
8.4
9.7
46.6
38.0
96
96
98
119
64
699
189
802
789
409
275
105
234
262
123
Table 12.5 Combinational floating point operators in binary64 format

FF
LUTs
DSP48
Slices
Delay
FP_add
FP_mult
FP_mult_luts
FP_div
FP_sqrt
15.4
15.1
12.5
136.9
97.6
192
192
192
244
128
1372
495
3325
3291
1651
15
585
199
907
903
447
Table 12.6 Pipelined floating point operators in binary32 format

FF
LUTs
DSP48
Slices
Period
Latency
FP_add
FP_mult
FP_mult_luts
FP_mult
FP_mult_luts
FP_sqrt
FP_div
6.4
5.6
7.1
3.8
5.0
9.1
9.2
2
2
2
3
3
6
5
137
138
142
144
252
384
212
637
145
798
178
831
815
455
-0
247
76
235
89
272
266
141
12.5.5 Implementation Results

Several implementation results are now presented. The circuits were implemented
in a Virtex 5, speed grade 2, device, using ISE 13.1 and XST for synthesis.
Tables 12.4 and 12.5 show combinational circuit implementations for binary32
and binary64 floating point operators. The inputs and outputs are registered. When
the number of registers (FF) is greater than the number of inputs and outputs, this
is due to the register duplication made by synthesizer. The multiplier is implemented using the embedded multiplier (DSP48) and general purpose logic.
Table 12.6 shows the results for pipelined versions in the case of decimal32
data. The circuits include input and output registers. The adder is pipelined in two
stages (Fig. 12.5). The multiplier is segmented using three pipeline stages
(Fig. 12.7). The divider latency is equal to 6 cycles and the square root latency is
equal to five cycles (Figs. 12.9 and 12.10).
12.6
Exercises
335
12.6 Exercises
1. How many bits are there in the exponent and the significand of a 256-bit
binary floating point number? What are the ranges of the exponent and the
bias?
2. Convert the following decimal numbers to the binary32 and binary64 floatingpoint format. (a) 123.45; (b) -1.0; (c) 673.498e10; (d) qNAN; (e) -1.345e129; (f) ?; (g) 0.1; (h) 5.1e5
3. Convert the following binary32 number to the corresponding decimal number.
(a) 08F05352; (b) 7FC00000; (c) AAD2CBC4; (d) FF800000; (e) 484B0173;
(f) E9E55838; (g) E9E55838.
4. Add, subtract, multiply and divide the following binary floating point numbers
with B = 2 and ulp = 2-5, so that the numbers are represented in the form s
2e where 1 B s B 1.111112. For simplicity e is written in decimal (base 10).
(a)
(b)
(c)
(d)
1.10101 9 23 op 1.10101 9 21
1.00010 9 2-1 op 1.00010 9 2-1
1.00010 9 2-3 op 1.10110 9 22
1.10101 9 23 op 1.00000 9 24
5. Add, subtract, multiply and divide the following decimal floating point
numbers using B = 10 and ulp = 10-4, so that the numbers are represented in
the form s 10e where 1 B s B 9.9999 (normalized decimal numbers).
(a)
(b)
(c)
(d)
9.4375 9 103 op 8.6247 9 102

1.0014 9 103 op 9.9491 9 102
1.0714 9 104 op 7.1403 9 102
3.4518 9 10-1 op 7.2471 9 103
6. Analyze the consequences and implication to support denormalized (subnormal in IEEE 754-2008) numbers in the basic operations.
7. Analyze the hardware implication to deal with no-normalized significands
(s) instead of normalized as in the binary standard.
8. Generate VHDL models adding a pipeline stage to the binary floating point
adder of Sect. 12.5.1.
9. Add a pipeline stage to the binary floating point multiplier of Sect. 12.5.2.
10. Generate VHDL models adding two pipeline stages to the binary floating point
multiplier of Sect. 12.5.2.
11. Generate VHDL models adding several pipeline stages to the binary floating
point divider of Sect. 12.5.3.
12. Generate VHDL models adding several pipeline stages to the binary floating
point square root of Sect. 12.5.4.
336
Reference
1. IEEE (2008) IEEE standard for floating-point arithmetic, 29 Aug 2008
Chapter 13
Finite-Field Arithmetic
Finite fields are used in different types of computers and digital communication
systems. Two well-known examples are error-correction codes and cryptography.
The traditional way of implementing the corresponding algorithms is software,
running on general-purpose processors or on digital-signal processors. Nevertheless, in some cases the time constraints cannot be met with instruction-set processors, and specific hardware must be considered.
The operations over the finite ring Zm are described in Sect. 13.1. Two multiplication algorithms are considered: multiply and reduce (Sect. 13.1.2.1) and
interleaved multiplication (Sect. 13.1.2.2). The Montgomery multiplication,
and its application to exponentiation algorithms, are the topics of Sect. 13.1.2.3.
Section 13.2 is dedicated to the division over Zp, where p is a prime. The proposed
method is the binary algorithm, an extension of an algorithm that computes the
greatest common divider of two naturals. The operations over the polynomial ring
Z2[x]/f(x) are described in Sect. 13.3. Two multiplication algorithms are considered: multiply and reduce (Sect. 13.3.2.1) and interleaved multiplication
(Sect. 13.3.2.2). Squaring is the topic of Sect. 13.3.2.3. Finally, Sect. 13.4 is
dedicated to the division over GF(2n).
As a matter of fact, only some of the most important algorithms have been
considered. According to the Authors experience they generate efficient FPGA
implementations (Sect. 13.5). Furthermore, non-binary extension fields GF(pn) are
not considered. A much more complete presentation of finite field arithmetic can
be found in [1] and [2].
13.1 Operations Modulo m

Given a natural m, the set Zm f0; 1; . . .; m 1g is a ring whose operations are
defined as modulo m.

337
338
13
Fig. 13.1 Addersubtractor

modulo m
+/-
operation
sign1
s1
m
+/s2
0
sign 2
C.C.
13.1.1 Addition and Subtraction Mod m

Given two natural numbers x and y belonging to Zm, compute z = (x ? y) mod
m. Taking into account that 0 x y\2 m; z must be equal to either x ? y or
x ? y - m, the following algorithm computes z.
Algorithm 13.1: Mod m addition
As regards the computation of z = (x - y) mod m, take into account that

-m \ x - y \ m, so that z must be equal to either x y or x y m: The corresponding algorithm is the following.
Algorithm 13.2: Mod m subtraction
The circuit of Fig. 13.1 is an adder-subtractor, which computes z = (x ? y) mod

m if operation = 0 and z = (x - y) mod m if operation = 1. It is described by the
following VHDL model in which k is the number of bits of m.
13.1
Operations Modulo m
339
A complete VHDL model is mod_m_AS.vhd is available at the Authors web page.
13.1.2 Multiplication Mod m

Given x and y [ Zm, compute z = xy mod m, where m is a k-bit natural.
13.1.2.1 Multiply and Reduce
A straightforward method consists of multiplying x by y, so that a 2k-bit result
product is obtained, and then reducing product mod m. For that, any combination
of multiplier and mod m reducer can be used. For fixed values of m, specific
combinational mod m reducers can be considered.
As an example, synthesize a mod m multiplier with m = 2192 - 264 - 1. Any
192-by-192 multiplier can be used. A 384-bit to 192-bit mod m reducer can be
synthesized as follows: given x x383 2383 x382 2382 . . . x0 20 ; it can
be divided up under the form
x383 263 x382 262 ...x320 20 2320 x319 263 x318 262 ...x256 20 2256
x255 263 x254 262 ...x192 20 2192 x191 2191 x190 2190 ...x0 20 :
Then, substitute 2320 by 2128 ? 264 ? 1 : 2320 mod m,
2128 ? 264 : 2256 mod m, and 2192 by 264 ? 1 : 2192
So, x : x0 ? x00 ? x ? x0000 , where
x0 x383 2191 x382 2190 . . . x320 2128
x383 2127 x382 2126 . . . x320 264
x383 263 x382 262 . . . x320 20 ;
x00 x319 2191 x318 2190 . . . x256 2128
x319 2127 x318 2126 . . . x256 264 ;
x000 x255 2127 x254 2126 . . . x192 264
x255 263 x254 262 . . . x192 20 ;
x0000 x191 2191 x190 2190 . . . x0 20 :
2256
mod
by
m.
340
13
The sum s x0 x00 x000 x0000 is smaller than 4 2192 264 1 4 m; so

that x mod m is either s, s - m, s - 2m or s - 3m.
The corresponding circuit is shown in Fig. 13.2 and is described by the
A complete VHDL model is mod_p192_reducer2.vhd is available at the Authors

web page.
In order to complete the multiplier design, any 192-bit by 192-bit multiplier can
be used (Chap. 8). This is left as an exercise.
13.1.2.2 Interleaved Multiplier
Another option is to modify a classical left-to-right multiplication algorithm based
on the following computation scheme
x y . . .0 2 xn1 y 2 xn2 y 2 . . . x1 y 2 x0 y:
Algorithm 13.3: Mod m multiplication, left-to-right algorithm
The data path corresponding to Algorithm 13.3 is shown in Fig. 13.3. It is

described by the following VHDL model.
13.1
Operations Modulo m
341
x 383..320 &x 383..320 &x383..320
x 191..0
x 319..256 &x 319..256
192 -bit adder
x 255..192
x 255..192
128 -bit adder
s 1,192..0
s 2,192..64
s 2,63..0
193 -bit adder

s 193..0
m
subtractor
sign 1
sign 1..3
1--
z1
01-
2m
3m
subtractor
subtractor
sign 2
z2
001
sign 3
z3
000
Fig. 13.2 Mod 2192 264 1 reducer
A complete VHDL model mod-_m_multiplier.vhd-, including a k-state counter and

a control unit, is available at the Authors web page.
342
13
y
2acc
x n-i-1
adder
shif register
initially: x
load
update
s
m
2m
subtractor
sign1
sign 1..2
1-
subtractor
sign 2
z1
01
z2
00
register
initially: 0
load
update
acc
Fig. 13.3 Interleaved mod m multiplier
If ripple-carry adders are used, the minimum clock period is about kTFA, so that
the total computation time is approximately equal to k2TFA. In order to reduce the
computation time, the stored-carry encoding principle could be used [2]. For that,
Algorithm 13.3 is modified: accumulator is represented under the form
accs ? accc; the conditional sum (accs ? accc) 2 ? xn-k-iy is computed in storedcarry form, and every sum is followed by a division by m, also in stored-carry form
(Sect. 9.2.4), without on-the-fly conversion as only the remainder must be computed. The corresponding computation time is proportional to k instead of k2. The
design of the circuit is left as an exercise.
13.1
Operations Modulo m
343
13.1.2.3 Montgomery Multiplication

Assume that m is an odd k-bit number. As m is odd, then gcd(m, 2k) = 1, and there
exists an element 2-k of Zm such that 2k 2k 1 mod m. Define a one-to-one
and onto application T from Zm to Zm:
T x x:2k mod m and T 1 y y:2k mod m:
The following properties hold true: T((x ? y) mod m) = (T(x) ? T(y)) mod m,
T((x - y) mod m) = (T(x) - T(y)) mod m, T(xy mod m) = T(x)T(y).2-k mod
m. The latter suggests the definition of a new operation on Zm, the so-called
Montgomery product MP [3]:
MPx; y x y 2k mod m:
Assume that the value 22k mod m has been previously computed. Then

T x MP x; 22k mod m and T 1 y MPy; 1:
The main point is that the Montgomery product MP is easier to compute than the
mod m product. The following algorithm computes MP(x, y).
Algorithm 13.4: Montgomery product
The data path corresponding to Algorithm 13.4 (without the final correction) is
shown in Fig. 13.4. It is described by the following VHDL model.
If ripple-carry adders are used, the total computation time is approximately equal
to k2TFA. In order to reduce the computation time, the stored-carry encoding
principle could be used [2, 4].
344
13
xi y0
y
load
update
xi
shif register
initially: x
q
p0
adder
load
update
register
initially: 0
Fig. 13.4 Montgomery product
Algorithm 13.5: Montgomery product, carry-save addition
The corresponding computation time is proportional to k instead of k2. The design

of the circuit is left as an exercise.
In order to compute z = xy mod m, the Montgomery product concept should
be used in the following way: first (initial encoding) substitute x and y by
x0 T x MPx; 22k mod m and y0 T y MPy; 22k mod m; then compute z0 = MP(x0 , y); finally (result decoding) compute z T 1 z0 MPz; 1:
This method is not efficient, unless many operations involving the same initial data
are performed, in which case the initial encoding of those data is performed only
once. Consider the following modular exponentiation algorithm; it computes
z = yx mod m and is based on the following computation scheme:
z yx0 x1 2x2 2
...xk1 2k1
k1
yx0 y2 x1 y2 x1 . . . y2 xk1 mod m:
13.1
Operations Modulo m
345
Algorithm 13.6: Mod m exponentiation, LSB-first
Mod m operations can be substituted by Montgomery products. For that, 1 is

substituted by T 1 2k mod m, and y by T y MPy; 22k mod m: Thus,
assuming that 2k mod m and 22k mod m have been previously computed, the
following algorithm computes z = xy mod m.
Algorithm 13.7: Modulo m exponentiation, Montgomery algorithm,
LSB-first
The corresponding circuit is made up of two Montgomery multipliers working in

parallel, with some kind of synchronization mechanism. A data-flow VHDL
description mod_m_exponentiation.vhd is available at the Authors web page. At
each step both multipliers MP1 and MP2, with their corresponding done1 and done2
signals, are synchronized with a wait instruction:
An MSB-first exponentiation algorithm could also be considered. For that, use the
following computation scheme:
2
k1
z yx0 x1 2x2 2 ...xk1 2

. . .12 yxk1 2 yxk2 2 . . . yx1 2 yx0 mod m:
346
13
Algorithm 13.8: Mod m exponentiation, MSB-first
As before, mod m multiplications can be substituted by Montgomery products. For

that, 1 is substituted by T(1) = 2k mod m, and y by T(y) = MP(y, 22k mod m). There is
a precedence relation between the two main operations T(e) : = T(e2) = MP(T(e),
T(e)) and T(e) : = T(ey) = MP(T(e), T(y)). Thus, a direct implementation includes
only one Montgomery multiplier, but needs up to 2k cycles instead of k in the case of
the LSB-first algorithm. The design of the corresponding circuit is left as an exercise.
13.2 Division Modulo p

If p is prime, than all non-zero elements of Zp have a multiplicative inverse. Thus,
given x and y = 0 in Zp, there exists an element z of Zp such that z = xy-1 mod p.
There are several types of algorithms that compute z. Some of them are generalizations of algorithms that compute the greatest common divider: Euclidean
algorithm [5, 6], binary algorithm [7], plus-minus algorithm [810]. Another
option is to substitute division by multiplications: according to the Fermats little
theorem z = xyp-2 mod p. As an example, the following binary algorithm computes z = xy-1 mod p. It uses four variables a, b, c and d, initially equal to p, y, 0
and x, respectively. At each step, a and b are updated in such a way that their gcd is
unchanged and that b decreases. For that, observe that if b is even and a is odd,
then gcd(a, b) = gcd(a, b/2), and if both a and b are odd, then gcd(a, b) = gcd
(a, |b-a|) = gcd(b, |b-a|). As initially a = p and b = y, where p is a prime, after a
finite number of steps a is equal to 1. On the other hand, c and d are updated in
such a way that cy : ax mod p and dy : bx mod p. Initially, c = 0,
a = p : 0 mod p, d = x and b = y, so that the mentioned relations are satisfied.
It can be proven that if c and d are updated in the same way as a and b, both
relations remain true. In particular, if a = 1, then cy : x mod p, and z = c.
13.2
Division Modulo p
347
Algorithm 13.9: Mod p division, binary algorithm
The corresponding circuit is made up of adders, registers and connection resources. A data-flow VHDL description mod_p_division2.vhd is available at the
Authors web page.
An upper bound of the number of steps before a = 1 is 4k, k being the number
of bits of p. So, if ripple-carry adders are used, the computation time is shorter than
4k2TFA (a rather pessimistic estimation).
13.3 Operations Over Z2 x=fx

Given a polynomial f x xm fm1 xm1 . . . f1 x f0 ; whose coefficients fi
belong to the binary field Z2, the set of polynomials of degree smaller than m over
Z2 is a ring Z2[x]/f(x) whose operations are defined modulo f(x).
13.3.1 Addition and Subtraction of Polynomials

Given two polynomials a x am1 xm1 . . . a1 x a0 and b x bm1 xm1
. . . b1 x b0 ; then
a x b x a x b x cm1 xm1 . . . c1 x c0 ;
where ci ai bi mod 2; 8i inf0; 1; . . .; m 1g: In other words, the corresponding circuit is a set of m 2-input XOR gates working in parallel, which can be
described by the following VHDL sentence:
13.3.2 Multiplication Modulo f(x)

Given two polynomials a x am1 xm1 . . . a1 x a0 and b x bm1 xm1
. . . b1 x b0 of degree smaller than m, and a polynomial f x xm fm1 xm1
. . . f1 x f0 ; compute c(x) = a(x)b(x) mod f(x).
348
13
13.3.2.1 Multiply and Reduce

A straightforward method consists of multiplying a(x) by b(x), so that a polynomial d(x) of degree smaller than 2m-1 is obtained, and then reducing d(x) mod f(x).
The coefficients dk of d(x) are the following:
Xk
a bki ; k 0; 1; . . .; m 1;
dk
i0 i
X2m2
dk
akim1 bim1 ; k m; m 1; . . . ; 2m 2:
ik
The preceding equations can be implemented by a combinational circuit made up
of 2-input AND gates and XOR gates with up to m inputs.
It remains to reduce d(x) modulo f(x). Assume that all coefficients ri,j, such that
xmj mod f x rm1; j xm1 rm2; j xm2 r1;j x1 r0;j;
have been previously computed. Then the coefficients of c(x) = a(x)b(x) mod
f(x) are the following:
Xm2
r dmi ; j 0; 1; . . . ; m 1:
13:1
c j dj
i0 j;i
13.3
Operations Over Z2 x=fx
349
The preceding equations can be implemented by a combinational circuit made up

of m XOR gates with up to m inputs. The number of gate inputs is determined by
the maximum number of 1s within a column of the matrix [ri,j], and this depends
on the chosen polynomial f(x).
A complete VHDL model classic_multiplier.vhd, including both the polynomial multiplier and the polynomial reducer, is available at the Authors web
page.
Every coefficient dk is the sum of at most m products aibk-i, and every coefficient cj is the sum of, at most, m coefficients dk. Thus, if tree structures are used,
the computation time is proportional to log m. On the other hand, the cost is
proportional to m2. Hence, this type of multiplier is suitable for small values of
m. For great values of m, the cost could be excessive and sequential multipliers
should be considered.
13.3.2.2 Interleaved Multiplier

The following LSB-first algorithm computes c(x) = a(x)b(x) mod f(x) according
to the following computation scheme:

c x b0 a x b1 a x x b2 a x x2

. . . bm1 a x xm1 :
Algorithm 13.10: Interleaved multiplication, LSB-first
The first operation c(x) ? bia(x) is executed by the circuit of Fig. 13.5a. In order
to compute a(x)x mod f(x), use the fact that xm mod f x fm1 xm1
fm2 xm2 . . . f0 : Thus,
350
13
(a)
c m-1 a m-1 c m-2 a m-2
c0 a0
(b)
bi
am-1fm-1 a m-1fm-2
am-2
am-3
a m-1+
+
c m -1
+
c m -2
a m-1f1
am -1f0
am-2+
a0
a1+
a 0+
+
c0
Fig. 13.5 Interleaved multiplier, computation of c(x) ? bia(x) and a(x)x mod f(x)

a x x mod f x am1 fm1 xm1 fm2 xm2 . . . f0 am2 xm1
am3 xm2 . . . a0 x:
The corresponding circuit is shown in Fig. 13.5b.
For fixed

f(x), the products am-1fj
must not be computed: if fj 1; aj aj1 am1 mod 2, and if fj 0;

a
j aj1 :
The circuit also includes two parallel registers a and c, a shift register b, an
m-state counter, and a control unit. A complete VHDL model interleaved_mult.vhd
is available at the Authors web page.
13.3.2.3 Squaring
Given a x am1 xm1 . . . a1 x a0 ; the computation of c(x) = a2(x) mod
f(x) can be performed with the algorithm of Sect. 13.3.2.1. The first step (multiply)
is trivial:
d x a2 x am1 xm1 . . . a1 x a0 2
am1 x2m1 am2 x2m2 . . . a1 x2 a0 :
Thus, di = ai/2 if i is even, else di = 0. According to (13.1),
P
cj aj=2
rj;i ami=2 ; j 0; 2; 4; . . .
cj
0 i m2
mi even
rj;i ami=2 ; j 1; 3; 5; . . .
0 i m2
mi even
The cost of the circuit depends on the chosen polynomial f(x), which, in turn,
defines the matrix [ri,j]. If f(x) has few non-zero coefficients, as is the case of
trinomials and pentanomials, then the matrix [ri,j] also has few non-zero coefficients, and the corresponding circuit is very fast and cost-effective. Examples of
implementations are given in Sect. 13.5.
13.4
Division Over GF(2m)
351
13.4 Division Over GF(2m)

If f(x) is irreducible, then all non-zero polynomials in Z2[x]/f(x) have a multiplicative inverse. Thus, given g(x) and h x 6 0 in Z2 x=f x; there exists a polynomial z(x) in Z2[x]/f(x) such that z x g x h1 x mod f x:
There are several types of algorithms for computing z(x). Some of them are
generalizations of algorithms that compute the greatest common divider, like the
Euclidean algorithm and the binary algorithm. Another option is to substitute the
division by multiplications: according to the Fermats theorem z x g x
hq2 x mod f x where q = 2m. As an example, the following binary algorithm
computes z x g x h1 x mod f x: It uses four variables a(x), b(x),
u(x) and v(x), initially equal to f(x), h(x), 0 and g(x), respectively. At each step,
a(x) and b(x) are updated in such a way that their greatest common divider is
unchanged and that the degree of a(x) ? b(x) decreases. For that, observe that if
b(x) is divisible by x and a(x) is not, then gcd(a(x), b(x)) = gcd(a(x), b(x)/x), and if
neither a(x) nor b(x) are divisible by x, then gcd a x; b x gcda x; a x
b x=x gcdb x; a x b x=x: As initially a(x) = f(x) and b(x) = h(x),
where f(x) is irreducible, after a finite number of steps b(x) = 0 and
a(x) = gcd(f(x), h(x)) = 1. On the other hand, u(x) and v(x) are updated in such a
way that u(x)h(x) : a(x)g(x) mod f(x) and v(x)h(x) : b(x)g(x) mod f(x).
Initially, u(x) = 0, a(x) = f(x) : 0 mod f(x), v(x) = g(x) and b(x) = h(x), so that
both equivalence relations are satisfied. It can be proven that if u(x) and v(x) are
updated in the same way as a(x) and b(x), both relations remain true. In particular,
if a(x) = 1, then u(x)h(x) : g(x) mod f(x), and z(x) = u(x).
Algorithm 13.11: Mod f(x) division, binary algorithm
Given a polynomial w(x), then w x x1 mod f x w x w0 f x=x:

A data path for executing the preceding algorithm is shown in Fig. 13.6. The small
circles connected to f1, f2,, represent programmable connections: if fi = 1, the
rightmost input of the corresponding XOR gate is connected to w0, else it is
connected to 0.
352
13
vm-1 um-1
am
bm-1 am-1
...
bm-1
bm-2
...
...
am-2
w1
b0
b0
w0
fm-1
b0
f1
ce_ bv
initially:g(x)
ce_ au
vm-1
...
am-1
v0 u0
ce_bv
initially:f(x)
am
wm-1
initially: h(x)
0
v1 u1
b1 a1
0
0
vm-2
...
v0
ce_ au
initially: 0
a0
um-1
um-2
...
u0
Fig. 13.6 Binary algorithm for polynomials
A drawback of the proposed algorithm is that the degrees of a(x) and b(x) must be
computed at each step. A better option is to use upper bounds a and b of the
degrees of a(x) and b(x).
Algorithm 13.12: Mod f(x) division, binary algorithm, version 2
The data path is the same as before (Fig. 13.6). An upper bound of the number of
steps is 2m. As the operations are performed without carry propagation, the computation time is proportional to m. A data-flow VHDL description mod_f_division2.vhd is available at the Authors web page.
13.4
Division Over GF(2m)
353
In the preceding binary algorithms, the number of steps is not fixed; it depends
on the input data values. This is an inconvenience when optimization methods,
such as digit-serial processing (Chap. 3), are considered. In the following algorithm [11] a and b are substituted by count = |a-b-1|, and a binary variable state
equal to 0 if a [ b, and equal to 1 if a B b.
Algorithm 13.13: Mod f(x) division, binary algorithm, version 3
The data path is still the same as before (Fig. 13.6), and the number of steps is 2m,
independently of the input data values. As the operations are performed without
carry propagation, the computation time is proportional to m. A data-flow VHDL
description mod_f_division3.vhd is available at the Authors web page. Furthermore, several digit-serial implementations, with different digit definitions, are
given in Sect. 13.5.

Several circuits have been implemented within a Virtex 52 device. The times are
expressed in ns and the costs in numbers of Look Up Tables (LUTs) and flip-flops
(FFs). All VHDL models are available at the Authors web page.
Two combinational multipliers (multiply and reduce) have been implemented
(Table 13.1).
For greater values of the degree m of f(x), sequential implementations should be
considered. Several interleaved multipliers have been implemented (Table 13.2).
In the case of the squaring operation, combinational circuits can be used, even
for great valued of m (Table 13.3).
354
13
Table13.1 Classic mod f(x)

multipliers
LUTs
Delay
8
64
37
2,125
3.2
5.3
Table 13.2 Interleaved mod

f(x) multipliers
FFs
LUTs
Period
Total time
8
64
163
233
32
201
500
711
34
207
504
714
1.48
1.80
1.60
1.88
13.3
117.0
262.4
439.9
Table 13.3 Mod

f(x) squaring (square and
reduce)
LUTs
Delay
8
64
163
233
8
129
163
153
0.7
0.7
0.7
0.7
Table 13.4 Mod

f(x) dividers
FFs
LUTs
Period
Total time
8
64
163
233
43
273
673
953
34
157
370
510
2.41
2.63
2.96
2.98
41.0
339.3
967.9
1391.7
Several sequential mod f(x) dividers (Sect. 13.4) have been implemented
(Table 13.4).
13.6 Exercises
1.
2.
3.
4.
5.
Generate the VHDL model of a reducer modulo 239.

Generate the VHDL model of a multiplier modulo m 2192 264 1:
Design an interleaved modulo m multiplier using the stored-carry principle.
Design a modulo p divider based on Fermats little theorem.
Generate the VHDL model of a multiplier over Z2[x]/f(x) where f x x8
x4 x3 x 1:
13.6
Exercises
355
6. Generate the VHDL model of a divider over Z2[x]/f(x) where f x x8

x4 x3 x 1:
7. Design a squarer over Z2 x=f x where f x x8 x4 x3 x 1:
References
1. Rodrguez-Henrquez F, Saqib N, Daz-Prez A, Ko K (2006) Cryptographic algorithms
on reconfigurable hardware. Springer, Heidelberg
2. Deschamps JP, Imaa JL, Sutter G (2009) Hardware implementation of finite-field arithmetic.
McGraw-Hill, New York
3. Montgomery PL (1985) Modular multiplication without trial division. Math Comput 44:
519521
4. Sutter G, Deschamps JP, Imaa JL (2011) Modular multiplication and exponentiation
architectures for fast RSA cryptosystem based on digit serial computation. IEEE Trans
Industr Electron 58(7):31013109
5. Hankerson D, Menezes A, Vanstone S (2004) Guide to elliptic curve cryptography. Springer,
Heidelberg
6. Menezes A, van Oorschot PC, Vanstone S (1996) Handbook of applied cryptography. CRC
Press, Boca Raton
7. Knuth DE (1981) The art of computer programming, Seminumerical algorithmsvol, vol 2.
Addison-Wesley, Reading
8. Deschamps JP, Bioul G, Sutter G (2006) Synthesis of arithmetic circuits. Wiley, New York
9. Meurice de Dormale G, Bulens Ph, Quisquater JJ (2004) Efficient modular division
implementation. Lect Notes Comp Sci 3203:231240
10. Takagi N (1998) A VLSI algorithm for modular division based on the binary GCD algorithm.
IEICE Trans Fundam Electron Commun Comp Sci 81-A(5):724728
11. Kim C, Hong C (2002) High-speed division architecture for GF(2m). Electron Lett 38:
835836
Chapter 14
Systems on Chip
This chapter introduces the main concepts related to the implementation of

embedded systems on FPGA devices. Many of these concepts will appear during
the case studies that are exposed in the next chapter.
14.1 System on Chip

The concept System on Chip (SoC) indicates a technical direction where the
increasing capacity of transistors into a chip enables the integration of more complex
systems. A SoC integrates all the components of the system, such as processors,
memory, analog functions, etc., on a single chip substrate. Some systems may be too
complex to be built using a single microelectronic technological process; therefore a
systems subset is optimized to be implemented in a single chip. Typical SoC
applications are embedded systems that are specific-application computers.
SoCs can be implemented on several microelectronic technologies: full-custom,
standard-cells, FPGAs. Advances in FPGA devices permit the mapping of quite
complex embedded systems in a single programmable device, thus, leading a
growing technology.
14.2 Intellectual Property Cores

Most SoCs are developed from pre-designed hardware cores usually know as
Intellectual Property (IP) cores. The IP cores are reusable hardware units designed
and verified by third-party companies or open-source developers to perform a
specific task. They can be licensed to permit a faster SoC development, since they
can be integrated into the design flow as a qualified building block. Software

357
358
14
Systems on Chip
drivers are enclosed in some IP cores, in order to facilitate the development of

applications for embedded systems.
IP cores can be delivered in different abstraction levels. In general terms, a
higher level of abstraction provides more flexibility but poorer performances and
failure risks due to the following stages of the design flow.
Soft-core. They are described in files and can be represented in different levels.
Register Transfer Logic (RTL). The highest abstraction level where IP
cores are generally described in VHDL/Verilog files, therefore, they must be
synthesized and implemented on the target device. Open source IPs are
suitable on a large number of different devices and they may be modified to
match the required needs. By contrast many industrial IP cores, as the IP cores
deployed in the Xilinxs EDK suite, may be designed for a particular device
and protected in order to avoid their modification.
Gate-level. The next lower level are IPs described in netlist files that are obtained
after the synthesis for a particular family. They can not be (easily) modified or
ported to other devices. The IP is usually optimized to use the most appropriate
hardware resources, but the design flow must implement the layout for the target
device. The soft macros from Xilinx are pre-defined circuits, composed of library
components that have flexible mapping, placement and routing.
Layout-level. The next abstraction level is obtained after the IP implementation on a device. The layout is usually optimized to improve area and timing
constraints. The Relative Placed Macro (RPMs) from Xilinx are pre-defined
layouts that have flexible placement.
Hard-core. It is the physical IP implementation on the device, therefore, logic
and performances are fixed. FPGA devices, usually, provide hard IP cores that
are efficiently built in a devoted area. For instance, most FPGAs provide internal
memory blocks that can be employed by embedded processors. More advanced
devices, such as some of the Xilinx Virtex-4/5 family, integrate PowerPC
processors, Digital Signal Processing (DSP) blocks, Ethernet Media Access
Control (MAC), etc.
14.3 Embedded Systems

General-purpose computers are designed to execute a wide range of different
applications. Usually, they are built with a high-performance microprocessor, a large
memory capacity and many different peripherals (USB ports, video card, disk drives,
etc.). Applications rely on the Operating System (OS) which manages memory,
peripherals, file system, interrupts and task scheduling. By contrast, an embedded
system is a computer designed and programmed to meet the requirements of a
specific application, such as media players, printers, domotics, etc. Applications may
not require OS or may rely on a customized OS. Important aspects concerning these
14.3
Embedded Systems
359
systems are their reduced size, cost and power consumption when they are compared
against general-purpose computers. An embedded system is, usually, composed
of a low-cost microprocessor, memory and some peripherals. It may include a
customized peripheral or coprocessor to speed-up a specific task or computation.
14.3.1 Embedded Microprocessors

Microprocessors are the heart of any computer system since they manage the
rest of the system, performing the control tasks and many of the computations.
An embedded microprocessor is a general-purpose processor whose cost, size,
power consumption and computational performances are adequate for the specific
application it executes. The variety of general-purpose microprocessors is huge,
but they can be classified according to their architecture. Embedded microprocessors are usually classified according to:
Data-path width. The basic building blocks of a microprocessors data-path
are the registers and the Arithmetic Logic Unit (ALU) that are of n-bit width.
The data-path width indicates the maximum size carried out by a single computation. In general terms, a higher data-path width brings more computational
performance but higher cost and power consumption. Most embedded processors are 8-bit to 32-bit width since they offer a good trade-off.
Instruction set. General-purpose microprocessors can be divided into Reduced
Instruction Set Computing (RISC) and Complex Instruction Set Computing
(CISC). The machine instructions of a RISC processor are quite simple, such as
loading a register from memory, computing an arithmetic/logical operation on
registers, and storing a register to memory. A CISC processor improves the
efficiency of the instruction memory since a single instruction can perform
several of the previously described steps. However, RISC microprocessors are
predominant in embedded systems since they benefit from a simpler architecture
and increased clock frequency.
Memory architecture. The von Neumann architecture provides a single bus to
access memory which stores instructions and data. The memory architecture is
simplified, but it cannot simultaneously read the next instruction while it is
reading/writing data, due to the previous instruction. The Harvard architecture
provides separated busses (instruction-side and data-side) to memory, in order to
improve the execution performance. In the latter case, the system provides an
arbiter which multiplexes the accesses to memory, from both sides. The memory
devices that store instructions and data can be physically different or not.
Endianness. The minimum addressable space of memory that a microprocessor,
usually, can read/write is a byte (8-bit). A 32-bit processors can also access
memory to read/write 16/32-bit words that are stored in a set of 2/4 bytes
allocated in consecutive addresses. A little-endian processor stores the least
significant byte of a word into the byte of memory which is allocated at the
360
14
Systems on Chip
lower address. By contrast, a big-endian processor stores the most significant

byte of the word into the memory byte at the lower address.
Pipeline architecture. The pipeline architecture is a digital design technique
which divides a large combinational path into a set of simpler processing stages
that are serially connected through registers. During a time slot, a stage performs
a partial computation from the resulting data of the previous stage, which is
processing the next data. For instance, a typical 3-stage microprocessor can
simultaneously fetch, decode and execute three different instructions. Most
microprocessors provide several pipeline stages, since they can significantly
increase the instruction throughput and the clock frequency. Their main drawbacks are the increased number of registers and the conditions that flush the
pipeline. Some microprocessors can minimize the pipeline flushing, permitting
out-of-order execution or the predicting of conditional branches.
Superscalar architecture. They are microprocessors that have duplicated some of
the functional units. For instance, they may provide several ALUs, in order to
permit the parallel execution of computations that require several clock cycles.
Another possibility is to provide duplicated stages of the pipelined architecture,
in order to enhance the execution of conditional branches.
Single-core or Multi-core. Many embedded microprocessors are single-core
since they provide enough computing performance for the application.
However, other applications may require a faster microprocessor. Increasing the
clock frequency is not always feasible or reliable due to the physical limitations
of the technology. Multi-core processors can perform several tasks in parallel in
order to increase the performance at the same clock frequency. They are widely
applied on general-purpose computers and they are extending their applicability
on embedded systems. Usually, they are homogeneous cores since they are
composed of a set of microprocessors of the same type. A heterogeneous core is
composed of different processor types. For instance, the DM37xx [9] OMAP
processors are composed of a general-purpose ARM microprocessor and
a Digital Signal Processor (DSP) to accelerate processing of image, video
and audio.
FPGAs can implement embedded processors as hard- or soft-cores. Hard-core
microprocessors offer better performances. A fixed area of the FPGA implements
the microprocessor and the programmable logic fabric is available to implement
the rest of the embedded system (busses, peripherals, etc.). On the other hand,
soft-core microprocessors on FPGAs offer a greater flexibility since they can be
configured to match the required performances on a larger variety of applications.
However, they may occupy a large area of the programmable logic. The set of
machine instructions can be extended to support some of the optional features of
the microprocessor. This way a C/C ++ compiler can optimize the machine code
to take benefits from the enhanced architecture of the microprocessor.
Many of the soft-core microprocessors offer configurable features that can be
enabled or disabled to match the desired performance/cost trade-off.
14.3
Embedded Systems
361
Cache logic. The on-chip memory provides high bandwidth memory, but very
limited capacity. By contrast, the external memory is slower, but it provides a
higher capacity due to its lower cost per byte. When the on-chip memory does
not provide enough capacity for the application, it can be used as cache memory.
The cache improves the execution of applications from external memory,
storing transparently the most frequently accessed instructions and data. The
microprocessor architecture must provide cache logic which implements a
policy, in order to try to maximize the number of cache hits.
Memory Management Unit (MMU). Most sophisticated OS require virtual
memory support in order to provide memory protection, and to extend the logical
capacity of the physical memory. The address space of the memory is divided in
pages and the MMU translates the page numbers from the logical memory to the
physical memory. It relies on the Translation Lookaside Buffer (TLB) which stores
a page table, in order to provide a fast translation. Many embedded microprocessors
may not implement the MMU, since the OS does not support virtual memory.
Floating-Point Unit (FPU). The unit provides arithmetic operations on real
numbers that are usually formatted on the single or double-precision IEEE-754.
The FPU may not be adequate on many low-cost applications and it can be
emulated by software.
Parallel multiplier, divider, barrel-shifter, or other additional units. Some applications require frequent operation on integer numbers that can be accelerated using
a dedicated hardware unit. For instance, a barrel-shifter can perform an arbitrary
n-bit shift in a single clock cycle which is useful in some applications.
Debug module. Most embedded processors can add logic to support an external
debugger tool. This way the external debugger can control the microprocessor
which is running an application, by performing actions like stop it, inspect the
registers or set breakpoints.
There is a very large collection of embedded microprocessors from different
device vendors, third-parties or free sources. Some FPGAs embed hard-core
microprocessors that provide the highest performance, but with a lower flexibility
since they are physically implemented in the device. The FPGA vendor provides
the tool chain (compiler, linker, debugger, etc.) to easily develop the software.
Many FPGA vendors provide their own soft-core microprocessors, in order to
facilitate the development of embedded systems on their devices. They are highly
optimized for the target device, and they can be configured to match the required
specifications. However, the flexibility is quite limited since the microprocessor
cannot be (easily) modified or implemented in a device from other vendors. The
vendor also provides the design flow, which permits the easily development of the
hardware and software. The open-source microprocessors usually occupy a larger
area and cannot run at the same clock frequencies when they are compared with
the previous alternatives. However, they provide the greater flexibility since they
can be modified or implemented in any FPGA that provides enough logic capacity.
The design flow is also more complicated since the designer uses tools from
different parties and they may lack support and warranty.
362
14
Systems on Chip
The Harvard, RISC, 32-bit microprocessors are widely implemented on FPGAs

since they can provide a near to optimal performance/cost trade-off for the current
technology. Therefore, we introduce some of the most likely used.
Some of the Xilinx Virtex-4/5 FPGAs offer the hard-core PowerPC-405/440
32-bit processors. The PowerPC-405 [10] is a little-endian processor which features
a 5-stage pipeline, cache and MMU. The PowerPC-440 [11] is an enhanced version
which features superscalar architecture and 7-stage pipeline. The architecture of
both PowerPC processors supports an auxiliary FPU or other coprocessors. The
newer Zynq-7000 [12] FPGAs from Xilinx integrate a dual-core ARM Cortex-A9
processor. The ARM processors are the de facto standard for embedded 32-bit
microprocessors on Application Specific Integrated Circuit (ASICs).
Xilinx provides the MicroBlaze [13], a soft-core 32-bit microprocessor suitable for embedded systems implemented on their FPGA families. It can be
configured to a 3/5-stage pipeline, and it can enable cache logic, MMU, singleprecision FPU, divider unit, barrel-shifter, etc. Altera provides the Nios II [2]
32-bit microprocessor, a competing soft-core, licensable for their FPGAs or thirdparty standard-cells. The byte ordering is little-endian, with 1/5/6-stage pipeline,
and it provides many of the configurable features available as in MicroBlaze.
The LatticeMico32 [6] is an open source microprocessor which can be embedded
in any FPGA device or ASIC technology. It features a 6-stage pipeline and bigendian ordering. The core can be configured to enable cache logic, parallel
multiplier, divider or barrel shifter, but not MMU nor FPU. Actel FPGAs can
license the Cortex-M1 [1] microprocessor which is an implementation ARMv6
processor specially designed for major FPGA devices. The processor provides a
3-stage pipeline, but the configurable options are not as extended as in the previous microprocessors.
Leon-3/4 [4, 5] are synthesizable VHDL models of SPARC v8 microprocessors
that can be licensed for commercial applications. They feature a 7-stage pipeline,
cache, MMU, single or double-precision FPU, etc. The microprocessors can
be implemented on FPGAs, or attached to build a multi-core embedded system.
The OpenRISC-1200 [7] is delivered in Verilog files and it is suitable for many
FPGAs. It provides a 5-stage pipeline and optionally cache logic, MMU and
single-precision FPU.
14.3.2 Peripherals
A general-purpose microprocessor is attached to other hardware components to
add functionalities to the computer system. Peripherals extend the capabilities of
the computer and they are dependent on it. The peripheral provides a bus interface
which transforms signals and the handshake from the core to the system bus, in
order to permit the microprocessor to control/monitor it. Many peripherals can
generate an interrupt request which triggers the execution of a service routine on
the microprocessor when the peripheral detects an event.
14.3
Embedded Systems
363
There are many types of peripherals. Some examples of typical peripherals are:
General Purpose Input/Output (GPIO), a peripheral which can be used to get or
set the state of I/O pins that are externally connected to other device. Typical
applications are to drive led diodes or to read switches. However, they can control
more complex devices when the microprocessor is programmed accordingly.
Universal Asynchronous Receiver-Transmitter (UART). The data transported
on a serial communication interface is composed of character frames. The
frames on an asynchronous transmission are characterized by the start, stop and
parity bits, the character length, as well as the bit speed. The UART peripheral
generates and transmits a frame when it receives a new character from the
microprocessor. It also performs the reverse transformation when it receives a
frame from the serial interface. The transmitted and received characters are
temporally stored in First Input First Output (FIFOs) in order to permit the
microprocessor to not attend the UART immediately.
Counter/Timer. The peripheral implements a register which counts an event
from an external signal or from an internal clock. In the second case the elapsed
time is deducted from the number of counts and the clock period. Another
typical application is to generate a Programmable Interval Timer (PIT) which
periodically generates an interrupt request.
Memory controllers. They assign an address space to the memory and they
control efficiently the read/write accesses from the microprocessor, in order to
maximize the memory bandwidth. They can also match the data width between
the memory and system bus. Depending on the memory they may permit burst
accesses and refresh the dynamic memory.
Interrupt controller. Combines interrupt requests from several sources to the
microprocessors interrupt input. They can mask interrupts to enable/disable
them or assign priorities.
Direct Memory Access (DMA). Some peripherals, such as video and disk
controllers, require very frequent accesses to memory. The microprocessor can
move data from memory to the peripheral (or vice versa) through the system
bus, although, it is quite inefficient. A DMA is connected as a master of the
system bus in order to permit to copy data between memory and peripherals
while the microprocessor performs any other action. The microprocessor task
can be reduced to program the DMA with the source and destination addresses,
the data length and the burst size. These systems require a bus arbiter since they
are several bus masters that can concurrently initiate a new access. The DMA
can be a centralized peripheral or be a part of a peripheral.
Video controller. A peripheral which generates the video signals required to
display an image. They require a DMA in order to read pixels from the video
memory (frame buffer) to generate the images at the required frequency.
Peripherals implement a set of internal registers in order to control or monitor
them. Each register provides different functionalities that depend on the peripheral
type and designer. A basic classification of peripheral registers is:
364
14
Systems on Chip
Control register. A write-only register which configures and controls the

peripheral. For instance, it sets the communication parameters in an UART.
Status register. A read-only register used to get the status of the peripheral.
For instance, it gets the detected error conditions and the state of the FIFOs in a
UART.
Data register. A read/write register which sets or gets data through the peripheral. For instance, it is the register used in the UART to read/write the characters
stored in the reception/transmission FIFOs.
Each register is mapped on an address, in order to be accessible from the
microprocessor. Two registers may be assigned to the same address. This is a quite
common situation since the control and the status registers are, usually, mapped on
the same address. In this way, a read instruction on the assigned address will get
the status register and a write instruction will set the control register.
Peripherals attach to the system in order to permit the microprocessor to access
their internal registers. Depending on the attachment, peripherals are classified as:
Port-mapped peripherals are attached to a microprocessors dedicated ports that
are physically isolated from memory. The peripherals registers are accessed
through specific machine instructions that are different from the memory
instructions.
Memory-mapped peripherals share the system bus with memory. The peripherals registers are accessed with the same machine instructions used to read/
write memory. The bus provides handshake signals to introduce wait states, in
order to adapt it to the different latencies of the memory and peripherals.
Usually, the system reserves an address range for peripherals to easily differentiate them from the regular memory space, simplifying the address decoding
logic.
The address decoder enables the selected register according to the address
bus which attaches to the peripheral. In a memory-mapped peripheral, the address
bus is much larger than required, due to the number of registers. For instance, the
address decoder of a peripheral, which implements 4 registers, requires only 2 of
the bits. The address decoding is divided into:
Complete address decoding. The decoder checks all the arriving bits from the
address bus. Each register is mapped on a single address.
Incomplete address decoding. It does not check some bits of the address bus to
simplify the logic. It checks some of the Most Significant Bit (MSBs) to select
one of the peripherals and some of the Lower Significant Bit (LSBs) to enable
one of its registers. Therefore, a peripheral register can be accessed from different addresses because some bits of the address bus are not decoded.
A software application can directly access the peripherals registers in order to
get the desired behavior. This approach has several drawbacks. Firstly, the programmer must have a deep knowledge about the internal architecture of the
peripheral. Secondly, the application may not be portable since the peripheral
14.3
Embedded Systems
365
could change. Finally, it is not adequate on a multi-tasking OS, since several

processes could simultaneously control the peripheral. A better approach is to
program the application to use the peripherals driver which is composed of
high-level functions that rely on the OS.
14.3.3 Coprocessors
Some algorithms may require a set of specific computations that cannot be re-solved
in a general-purpose microprocessor in a reasonable execution time. A co-processor
is designed and optimized to quickly perform a specific set of computations to
accelerate the system performance. Coprocessors do not fetch instructions or data
from memory. The general-purpose microprocessor accesses to memory to read
data and enables the coprocessor to accelerate a computation. Coprocessors allow
the customization of computer systems, since they can improve computational
performances without upgrading the general-purpose microprocessor.
There is a wide range of coprocessors and some of the most typical are:
FPU coprocessors to accelerate the computations on real numbers. Many
embedded microprocessors do not integrate a FPU in order to reduce cost.
Others may integrate a single-precision FPU which may be not adequate for
some applications. The single and double-precision IEEE-754 may be emulated
through software libraries, but the system performance may be greatly affected.
Cryptoprocessors are designed to accelerate the most time spending computations related to cryptography, such as the encryption and decryption tasks. Their
architecture usually implements several parallel processing units to improve the
speed-up factor.
DSPs perform a digital signal processing from a data stream (voice, image,
etc.). The internal architecture may provide several pipeline stages to improve
the data throughput. The input and output streams are carried from/to the
microprocessor which accesses to memory or other components.
The coprocessors can attach to the microprocessor in several ways:
Data-path extension by custom instructions. For instance the ALU of the softcore NIOS II [3] is prepared to be connected to custom logic which accepts
users machine instructions that are similar to the native ones. Another example
is the APU (Auxiliary Processor Unit) of the PowerPC-405/440 [10, 11], which
permits the execution of new instructions on a tightly-coupled coprocessor.
Dedicated point-to-point attachment, such as the FSL (Fast Simplex Link) for
MicroBlaze [13]. The machine instructions set provides some instructions able
to transmit/receive data to/from the links attached to the coprocessor.
System bus. A coprocessor can be also attached to the system bus, as any
peripheral. This alternative is not usually implemented since it increases the
latency to communicate data, due to the more sophisticated bus handshake.
366
14
Systems on Chip
14.3.4 Memory
A memory circuit is a digital storage device which is commonly used to allocate
the program executed by a microprocessor. The memory on embedded systems
can be classified in several ways. A first logical division is:
Instruction memory. Microprocessors execute software applications when
reading machine instructions from the instruction memory.
Data memory. Some of the machine instructions command the microprocessor
to read or write data from/to the data memory.
The instruction and data memories may be physically isolated or implemented
on the same devices. Moreover, they can be composed of several banks and types
of memories. According to the physical placement, memory can be classified as
the following:
Internal memory. It is integrated in the same device as the embedded system.
The memory is usually SRAM technology which features fast access time but
small capacity (some KB). The internal memory can store a small executable
(like a boot loader) or serve as cache memory. Many FPGAs integrate internal
blocks of memory. For instance, Xilinx FPGAs implement Block RAM
(BRAMs), dual-port synchronous SRAM memories that permit simultaneous
access to instructions and data per clock cycle.
External memory. It is not integrated into the device which implements the
embedded system. The external memory can be semiconductor chips or storage
devices.
The semiconductor technologies that are, usually, applied as external memory
are DRAM and FLASH. They provide larger capacity (MB or GB) but slower
access time, when compared with SRAM memory. The DRAM can storage
instructions and data; however it is a volatile memory. By contrast, the
FLASH is non-volatile but it cannot be used as the data memory since it is
read-only memory. Many embedded systems make use of both technologies,
first copying the program from FLASH to DRAM during the boot-up, and
then executing the application from DRAM.
Storage devices, such as hard drives or FLASH cards, are more common in
general-purpose computers than in embedded systems. They are non-volatile
memories that feature very high capacity, but very slow access time.
Moreover, data is sequentially accessed in large blocks (some KB) making
it inefficient when randomly accessing data. The main purpose is to serve as
storage repository of programs and files, although some computer systems
can implement virtual memory on them if the microprocessor and OS
support it.
Another classification is based on the architecture of the bit cells. The most
common memories on embedded systems are as follows:
14.3
Embedded Systems
367
Static Random Access Memory (SRAM). The bit cell is larger than the other
alternatives; therefore, the memory density is lower and the cost per capacity is
increased. The main advantages are the lower access time and lower power
consumption when they are in standby mode. Many embedded systems and
FPGAs integrate some SRAM blocks, since they are compatible with conventional CMOS technologies to fabricate the integrated circuit.
Dynamic RAM (DRAM). It is a read/write and volatile memory as is the
SRAM. The main advantage is the larger memory density, since the bit cell is
very simple. However, the access time is, usually, slower and the bit cells must
be periodically refreshed, increasing the power consumption. Moreover, they
cannot be integrated efficiently in conventional CMOS technologies; therefore
they are usually connected as external memory.
FLASH. The bit cell is based on a floating-gate transistor which can be erased
and programmed providing read-only and non-volatile memory. There are two
important variants depending on the interconnections between cells: NAND and
NOR. The NAND FLASH memories are employed as storage devices, since
they provide higher capacity and bandwidth. However, the disadvantage is that
data is sequentially read on pages. By contrast, the NOR FLASHs are, usually,
employed as Read-Only Memory (ROMs), since they provide random access to
data and similar interface as in SRAM chips.
There are more memory technologies, but most of them are not currently
implemented or widely used on commercial applications. The Ferroelectric RAM
(FRAM) is an emerging technology on embedded systems. It provides randomaccess and non-volatile memory which features much better performances on
access time, power consumption, and write endurance when compared with
FLASH. However, the memory density is much lower. For instance, the
MSP430FR57xx family [8] of 16-bit microcontrollers can embed up to 16 KB of
FRAM.
Another classification is based in the synchronicity of accesses to data.
Asynchronous memory. Data is accessed by the address bus, independently of
any clock signal. Many external SRAM memories are asynchronous since they
provide a very simple interface and low latency when data is randomly accessed.
An interesting alternative to external SRAM is the Pseudo Static RAM
(PSRAM), which is DRAM surrounded with internal circuitry that permits to
control it as a SRAM chip. It offers external memory featured by a high capacity
and simple interface, since its internal circuitry hides the specific DRAM
operations.
Synchronous memory. Data is synchronously accessed according to a clock
signal. These memories improve the data bandwidth since they implement a
pipeline scheme, in order to increment clock frequency and data throughput.
However, the latency is increased due to the stage registers. Most of the present
Synchronous DRAM (SDRAMs), such as the DDR2 and DDR3, use Double
Data Rate (DDR) registers to provide very efficient burst accesses in order to
replace the cache pages. Another example of synchronous external memory is
368
14
Systems on Chip
the Zero-Bus Turn-around (ZBT) SRAM which improves the bandwidth of burst
accesses and eliminates the idle cycles between read and write operations.
Usually, the embedded memory blocks on FPGAs are double-port synchronous
SRAMs that can efficiently implement the data and instruction cache of a
Harvard microprocessor.
Embedded systems usually provide a memory controller to assign an address
space to memory and to adapt the handshake signals and the data width from the
system bus. The complexity of the external memory controller depends on the type
of memory. The SRAM controllers are quite simple, since data can be accessed
randomly. The DRAM controllers are more complex since the address bus is
multiplexed and they control the periodical refreshing operation. Moreover, data is
usually synchronously accessed on a burst basis; therefore the controller also
generates the clock signal and the logic to communicate data between clock
domains.
14.3.5 Busses
The interconnection between the microprocessor, peripherals and memory controllers in an embedded system is done through busses. The system bus permits the
transfer of data between the building components. It is physically composed of a
set of connecting lines that are shared by the components. The component must
follow a common communication protocol which is implemented on the bus
handshake.
The bus master starts a new access cycle in order to read/write data from/to a
slave. An embedded system can provide several masters, but only one of them can
take the bus control during a time slot. The bus arbiter assigns the control to one of
the masters that are requesting it concurrently. For instance, a typical embedded
system is composed of a Harvard microprocessor which connects to an external
memory controller through the system bus. The external memory stores the
instructions and data of an application program. The bus arbiter assigns the bus
control to the microprocessors instruction-side, or the data-side, according to the
arbitration policy, while executing the program from the external memory.
There is a wide variety of busses depending on the embedded processor.
Moreover many embedded systems provide several busses; depending on the
memory or peripherals that are connected. A typical bus can be divided into the
following:
The address bus is a unidirectional bus, from the masters to the slaves, that
carries the address of the memory (or a register in a memory-mapped peripheral)
that is going to be accessed.
The data bus is a bidirectional bus that carries the data which is read from
slaves, or written to them. A bidirectional bus requires tri-state buffers that are
not, usually, available in the internal FPGA fabric. In FPGAs, the data bus is
14.3
Embedded Systems
369
replaced by two unidirectional busses: the write data bus (from masters to
slaves) and the read data bus (from slaves to masters).
The control bus is the set of signals which implements the bus handshake, such
as the read/write signal, or signals that indicate the access start and to
acknowledge the completion. Some advanced busses permit burst accesses, in
order to improve the throughput of sequential data transfers, and error signals, to
permit the microprocessor to detect error conditions on the bus.
References
1.
2.
3.
4.
5.
6.
7.
9.
8.
10.
11.
12.
13.
Actel (2010) Cortex-M1 v3.1 Handbook

Altera (2011a) NiosII processor reference handbook (NII5V1-11.0)
Altera (2011b) Nios II custom instruction User Guide (UG-N2CSTNST-2-0)
Gaisler (2010a) LEON3 Multiprocessing CPU core
Gaisler (2010b) LEON4 32-bit processor core
Lattice (2011) LatticeMico32 processor reference manual
OpenCores (2011) OpenRISC 1000 architecture manual
Texas Instruments (2011a) DM3730, DM3725 Digital media processors (SPRS685D)
Texas Instruments (2011b) Mixed signal microcontroller MSP430FR573x MSP430572x
(SLAS639)
Xilinx (2008) PowerPC 405 Processor block reference guide (UG018)
Xilinx (2010a) PowerPC 440 Embedded processor block reference guide (UG200)
Xilinx (2010b) Extensible processing platform. Ideal solution for a wide range of embedded
systems (WP369)
Xilinx (2011) MicroBlaze processor reference guide (UG081)
Chapter 15
Embedded Systems Development:

Case Studies
Embedded systems are computers designed and programmed to meet the

requirements of a specific application. Applications may not require an OS
(Operating System) or may rely on a customized OS. The system architecture is
usually composed of a low-cost microprocessor, memory and peripherals interconnected through busses. It may also include a coprocessor to speed-up a specific
computation.
This chapter introduces the design of embedded systems on Xilinx FPGAs
through a set of case studies. The studies focus on the hardware development of a
peripheral and a coprocessor, as well as their software drivers. It assumes you are
familiarized with the VHDL necessary on the hardware implementation, and with
C/C++ which is required during the software development. In order to simplify the
contents, they only expose the most relevant steps and interesting portions of the
source files. They usually make references to specific documentation available in
the Xilinx software, to get more details. The full examples can be downloaded
from the authors website.
15.1 Introduction to Xilinx EDK

The FPGA implementation of embedded systems requires a set of tools which
permits the building of the hardware and the software (also named firmware).
It also provides utilities to simulate and debug the embedded system.
The Xilinx Embedded Development Kit (EDK) suite facilitates the development of those systems. The systems hardware is composed of a general-purpose
microprocessor connected to some peripherals through busses. The software
development provides the firmware executed by the microprocessor. The EDK
generates the bitstream of the system in order to program the FPGA, which can be
connected to the debugger tools, in order to test it on a real scenario. EDK also

371
372
15
Embedded Systems Development: Case Studies
permits the building of a simulation model in order to check the hardware design
through a Hardware Description Language (HDL) simulator.
The EDK is frequently updated and upgraded. We will focus this chapter on the
ISE Design Suite 13.1 for Windows since it is probably the most popular operating system for PCs, but there are no significant differences with versions for
other operating systems. Although the case studies included in this chapter can be
implemented on other upgraded EDK versions, they might require some small
changes.
The EDK is composed of a set of tools:
Xilinx Platform Studio (XPS). A graphical user interface that permits the
designing of the hardware of the embedded system from a set of interconnected
IP cores. It is the top-level tool which takes care of the necessary files and steps
needed to successfully complete the hardware design. The XPS implements the
design flow which runs other low-level tools in order to compute the hardware
synthesis and implementation (Platgen), the generation of the bitstream (BitGen) and the simulation model (Simgen).
Software Development Kit (SDK). An integrated environment based on the
Eclipse/CDT to manage the development of the software. It launches the C/C++
cross-compiler and linker to build the binary which is executed by the embedded
microprocessor. Moreover, it also provides a simple interface with the debugger
and profiler tools used by more advanced users. SDK also builds the BSP (Board
Support Package) through a low-level tool (Libgen). The BSP contains the set of
software drivers used to control the hardware from the executable.
IP cores. The library of configurable cores (microprocessors, peripherals, busses,
etc.) that are used as basic building blocks of the embedded system. Most of
these cores are licensed with the EDK, but there are also some cores that must
be licensed separately. Many cores include a set of programming functions and
drivers that can be used to facilitate the software development.
GNU tools chain. The set of tools that generate the software libraries and the
executable binaries. It includes the GNU C/C++ cross-compiler and linker. The
GNU tools are developed for the Linux operating system, but EDK includes
ported versions to the Windows OS.
Additionally, the EDK relies on other tools:
ISE tools. They synthesize and implement the hardware, generate the bitstream
and program the device. They also include other tools to generate the simulation
model, implement the internal memory, the timing analysis and others. Platform
Studio calls the required ISE tools, simplifying the design flow since the user
can be abstracted from many specific details.
Supported HDL simulator. It is recommended if the user designs a new IP, since
it permits the simulation of the hardware of the embedded system. Some IP
cores deployed with EDK are protected and encrypted, therefore, they can be
simulated only on supported simulators. The ISE tools provide the ISim, but the
user can choose a third-party simulator.
15.1
Introduction to Xilinx EDK
373
EDK and Xilinx
Simulation
libraries
Users Hardware
(VHDL files)
Xilinx Platform Studio (XPS)
XST
Microprocessor
Hardware
Specification
(MHS file)
Simgen
Platgen
ISim
BitGen
EDK IP cores
EDK IP drivers and

libraries
GNU Tools
Chain
(Compiler /
Linker )
Software Development Kit (SDK)
Libgen
Microprocessor
Software
Specification
(MSS file)
Users Drivers
(C files)
Data2MEM
Debugger
FPGA
development
board
GNU Tools
Chain
(Compiler /
Linker)
Users Software
(C/C++)
Fig. 15.1 Overview of the EDK design flow
A development board with a Xilinx FPGA. There is a wide range of boards with
different FPGAs (Spartan or Virtex series), memories, displays, communication
interfaces and other elements.
A Xilinx programmer, such as the Parallel III/IV or the Platform Cable USB
I/II. Some development boards provide an embedded USB programmer. The
programmers can be used to program the FPGA and to debug the executable
through the Xilinx Machine Debugger (XMD) low-level tool.
Two files play an important role in the design flow (see Fig. 15.1): the
Microprocessor Hardware Specification (MHS) and the Microprocessor Software
Specification (MSS). The XPS manages the hardware design flow using a Xilinx
Microprocessor Project (XMP) project file. The XPS project relies on the MHS file
which configures a set of instances to IP cores that are interconnected as building
blocks. The XPS can export the hardware design to SDK in order to generate the
Board Support Package (BSP) for the embedded system. The BSP generation
relies on the MSS file which configures the drivers and libraries that can be used by
the executable to control the hardware.
374
15
anode-3
anode-2
anode-1
anode-0
a
a
b
g
b
c
d
e
f
g
Fig. 15.2 The 4-digit, 7-segment led display
15.1.1 Case Study 1-A: A Basic Embedded System

This case study builds a basic embedded system, composed of the MicroBlaze
processor [16], internal memory, and some peripherals. The system controls a
4-digit, 7-segment led display to visualize the hexadecimal content of a variable.
The system attaches to two external switches that turn on/off the display and show/
hide the left-side zeros. The display can also be controlled by an external computer
connected through the serial interface.
The development FPGA board is the Xilinx Spartan-3 Starter Kit [2], a cheap
board composed of a XC3S200 FPGA interconnected to a 7-segment display, a
serial port and other elements. The four digits of the display share the segment inputs
and each digit is enabled with a dedicated anode input, as depicted in Fig. 15.2.
The system must continuously perform a time-multiplexed control to refresh the
display, driving a single digit during a short time slot (a few milliseconds).
Depending on the board, the inputs of the display are asserted to low or high logic
levels. The case study can be easily adapted to any FPGA board which provides a
similar display.
The display could be continuously controlled by the microprocessor in a loop to
enable one digit per refresh cycle. This system is quite inefficient, since most of the
execution time would be devoted to wait the refresh of the display, and to poll the
serial interface and the state of the switches. A better approach uses two interrupt
service routines (ISR) to attend the interrupts from two peripherals. A timer
peripheral can, periodically, request an interrupt which triggers an ISR in order to
read the switches and refresh the display. The Universal Asynchronous Receiver
Transmitter (UART) peripheral requests an interrupt when a new character is
received from the serial interface, and its ISR will process it.
The design of an embedded system involves the hardware and software development phases. The output of the hardware phase is a BIT (bitstream) file which
configures the hardware resources of the FPGA except the contents of the internal
BRAM (Block RAM) used as the microprocessors local memory. The output of
the software phase is the Executable and Linkable Format (ELF) file which must be
15.1
375
External
oscillator
50MHz
Reset
button
JTAG
programmer
system
clock
DLMB
Controller
reset
DLMB
MDM
MicroBlaze
BRAM
ILMB
Controller
Interrupt
Controller
ILMB
PLB
switches
GPIO
led7seg
GPIO
Timer
switches
4-digit,
7-segment
led display
rs232
UART
Serial
communication
Fig. 15.3 Overview of the system hardware
allocated into the microprocessors memory. In order to program the FPGA the
design flow executes the Data2MEM tool to generate a new bitstream file which
configures the FPGA including the BRAM contents to store the executable binary.
15.1.2 Hardware
The system is composed of the MicroBlaze, the internal BRAM and a set of
peripherals (see Fig. 15.3). The BRAM implements the microprocessors local
memory and it is connected through two Local Memory Bus (LMBs). The peripherals are connected through a Processor Local Bus (PLB). The MicroBlaze can
control the display and the two switches through General Purpose Input Output
(GPIO) peripherals. The UART peripheral permits the serial communication with
the external PC through the RS-232 interface. The timer and the UART will request
the microprocessor to interrupt, therefore, the system includes an interrupt controller. Finally, the Machine Debug Module (MDM) permits the debugging of the
application executed by the MicroBlaze through the FPGA programmer and the
XMD tool.
376
15
The hardware generation involves three main steps:

(1) Hardware specification in the MHS file
(2) Synthesis of the hardware
(3) Implementation of the FPGA layout and bitstream generation
15.1.2.1 Specification
The first step is to specify the hardware of the embedded system in the MHS file.
The easiest way to start is by using the Base System Builder (BSB) wizard.
It creates the project file and a valid MHS file [14] which describes the system
composed of the microprocessor attached to local memory and peripherals. Open
the Xilinx Platform Studio (XPS) and follow the steps:
(1) XPS opens a dialog window. Choose the BSB wizard
(2) The next dialog window configures the project file and directory. Change the
path to C:\edk13.1\led7seg and the project name to system.xmp
(3) Next, a new dialog configures the system bus. The AXI is currently supported
only on the newer FPGA families (Spartan-6, Virtex-6). Therefore, select the
PLB [4] since it is supported by all the FPGA families.
(4) Select the option Create a new design in the dialog Welcome.
(5) The dialog can configure a pre-built system for a supported FPGA board.
Choose Create a system for a custom board to setup the system from scratch.
Then select the Spartan-3 xc3s200-ft256-4 device and any polarity for the reset
button. These parameters can be easily modified later.
(6) Choose a single processor system since it simplifies the design.
(7) The next dialog configures the frequencies of the reference and system clocks
as well as the capacity of the microprocessors local memory. Leave the
default parameters since they are changed later.
(8) The BSB permits the connecting of a set of peripherals to the system. Just
continue until the end of the wizard since they are added later. The BSP
creates the hardware specification of a basic embedded system.
The XPS permits to display and modify the system in a graphical way using the
tab System Assembly View, as shown in Fig. 15.4. The view Bus Interfaces shows
the system composed of instances to IP components that are interconnected
through busses. The microprocessor (microblaze_0) is connected to the BRAM
(lmb_bram) through the data and instruction LMBs (dlmb, ilmb) [17] and their
memory controllers (dlmb_cntrl, ilmb_cntrl). The peripherals attach to the system
through the PLB as slaves. The instruction and data PLB sides of the microprocessor are the bus masters. The MDM (mdm_0) peripheral attaches to the
microprocessor through the PLB (mb_plb). The MicroBlaze is master of the
busses, meanwhile, the peripheral and memory controllers are the slaves. Finally
the last two instances are the generators of the internal clock (clock_generator_0)
and reset (proc_sys_reset_0) that are required by the system.
15.1
377
Fig. 15.4 Bus interfaces of the system in EDK
Fig. 15.5 Configuration parameters of the data LMB controller
The IPs provide a set of parameters that can be fixed, auto-computed or configurable. Select an instance and open the contextual menu (click the right button of
the mouse) to configure the IP graphically (see Fig. 15.5). The HDL name of a
parameter is the same as it appears in the MHS file. The configurable parameters
C_BASEADDR and C_HIGHADDR of the LMB controllers setup the address space
of the microprocessors local memory. Changing the C_HIGHADDR to 0 9 3FFF
of both LMB controllers increases the auto-computed parameter C_MEMSIZE of
the BRAM to 16 KB (0 9 4000). Every IP deployed by EDK provides a PDF file
which details the parameters, the input/output ports and internal architecture.
The MHS file is a text file which can be manually edited, as shown in Fig. 15.6.
It is synchronized with the System Assembly View. Therefore, they both update
378
15
Fig. 15.6 MHS file in the XPS project
when the other one is modified. The MHS specifies the systems hardware as a set
of interconnected instances and external FPGA ports. Each instance contains
configurable parameters, interface to busses and other ports. The parameters that
are not declared in the MHS take a default or the auto-computed value. The bus
interfaces or ports that are not declared in the MHS are disconnected.
At the beginning of the MHS file there are the declarations of the external
FPGA ports used to input the clock and the reset. The external ports are connected
to the internal signals CLK_S and sys_rst_s that are used by the clock and reset
generators. The parameter CLK_FREQ declares the frequency of the external
oscillator and the RST_POLARITY declares the logic level when the reset input
asserts. Both parameters must be modified according to the FPGA board.
The instance proc_sys_reset_0 generates the internal reset signals required by

the system. The configurable parameter C_EXT_RESET_HIGH must be modified
according to the reset polarity which was declared in the external port.
15.1
379
The instance clock_generator_0 infers a Digital Clock Manager (DCM) circuit to

generate the system clock from a reference clock (external oscillator). The parameter C_EXT_RESET_HIGH must be configured as in the reset generator.
The parameter C_CLKOUT0_FREQ configures the generated frequency from the
reference clock defined by the C_CLKIN_FREQ. The BSP generated a signal named
clk_66_6667 MHz (or similar) which carries out the generated clock, but it can be
renamed to sys_clk. Also change the name of the signal which drives the clock port
of the busses. An incorrect configuration of the clock frequencies will lead to errors
during the synthesis or implementation of the hardware.
In order to accommodate it to the desired application, the hardware must add

new peripherals. The first peripheral is a GPIO [5] which controls the 7 segments
and the 4 anodes of the display. As depicted in Fig. 15.7:
380
15
6
7
1
4
3
Fig. 15.7 Adding the new GPIO peripheral
(1) Drag the XPS General Purpose IO from the IP Catalog to the System
Assembly View.
(2) XPS opens a dialog to configure it. Set the data width of the first channel
(parameter C_GPIO_WIDTH) to 11 in order to control the display.
(3) Click the created instance and change the name to led7seg.
(4) Click the PLB interface to connect the peripherals SPLB (Slave PLB).
Go to the view Addresses and configure automatically the addresses of the
peripheral (see Fig. 15.8). The internal registers of the GPIO are accessible from
the microprocessor within this addresses range. By default, XPS assigns the
addresses range to 64 KB (0 9 10000) from an address above the 0 9 80000000.
15.1
381
Fig. 15.8 Automatic configuration of the addresses for the GPIO peripheral
The user can change the range, but they must not overlap to the other memorymapped peripherals.
Finally, the 11-bit width GPIOs output port must be connected to the external
FPGA ports that drive the display. Figure 15.9 shows the view Ports which permits
the external connection of the output port GPIO_IO_O. Change the default name to
fpga_0_led7seg_pin in a similar fashion as the rest of the external FPGA ports and
check that the direction is configured as output. Finally, set the range order to [10:0]
to declare them in descending order. The MSB and the LSB are indexed as 10 and 0,
respectively.
Open the MHS file to observe the previous changes. There is a new entry in the
section of the external FPGA ports. It also contains the new GPIO instance
including its configuration parameters and connections.
The next step adds a new GPIO to read the state of the two switches. The
hardware can also be modified by editing the MHS file. Copy the previous GPIO
instance and change the instance name to switches and the data width to 2. Set
the addresses range to 64 KB and do not overlap it to the other peripherals.
Connect the input port GPIO_IO_I to a signal which is connected to an external
FPGA port named fpga_0_switches_pin. Save the MHS file to update the
graphical view. The project will close if there is an error in the MHS file, which
must be manually corrected.
382
15
Fig. 15.9 External FPGA connection of the GPIOs output port
The next step adds the two peripherals to the PLB that will request interrupts.
The timer [6] will be programmed to request a periodic interrupt. The parameter
C_ ONE_TIMER_ONLY configures a single timer to minimize the size of the
peripheral.
15.1
383
The UART [7] will request an interrupt when it receives a new character from
the RS232. It will also transmit messages to the user through the serial communication. Therefore the UART ports transmission (TX) and reception (RX) are
connected to external FPGA ports. The parameters C_BAUDRATE and
C_USE_PARITY configure the speed and parity of the communication.
Some displays provide an extra input to turn on a dot placed beside the digit.
This case study just turns off the dot connecting the net_gnd or net_vcc to its
associated external port.
The system requires an interrupt controller [8] since MicroBlaze provides a

single input port for the interrupt requests. The interrupt controller attaches to the
PLB in order to permit the MicroBlaze to enable/disable interrupts or to check the
interrupt source. The interrupt controller (int_control_0) connects the interrupt
requests from the timer and UART peripherals to Microblaze. The controller
receives the concatenated signal from the two peripherals and drives the interrupt
port of MicroBlaze. The interrupt priority is higher when the interrupt source is
concatenated at the right side. The UART is assigned to the lower priority since it
384
15
provides a receiving First Input First Output (FIFO) memory which temporarily
stores the characters. Therefore, a new character received can be processed when
the display is not refreshing.
MicroBlaze provides configurable parameters to optimize area/performance and

to implement optional machine instructions. The parameter C_AREA_OPTIMIZED
configures a 3-stage pipeline architecture which optimizes the area of the implementation. The parameter C_USE_BARRELL implements a barrel shifter and its
related machine instructions. Therefore, the C/C++ compiler provides a set of flags
to build the BSP and the executable for the configured microprocessor. By default,
the MicroBlaze attaches to a MDM instance (mdm_0) which permits the debugging
of executables.
15.1.2.2 Synthesis
The XPS menu Hardware ? Generate Netlist synthesizes the design to generate
a set of NGC netlist files. It calls the Platgen tool [13] which starts performing a
Design Rule Check (DRC). Then it calls the Xilinx Synthesis Technology (XST)
[3] tool to synthesize the IP instances to get their NGC files. The embedded
system is finally synthesized and optimized to get the top-level netlist file.
A change in the MHS file will force Platgen to synthesise only the required
15.1
385
2
1
Fig. 15.10 Selection of the device in the XPS project
modules to speed-up the execution. If desired, the XPS can clean up the generated files to start the Platgen from scratch. Synthesis is dependent of the FPGA,
therefore, the user must select the correct device in the Project Options
(see Fig. 15.10) before proceeding.
Figure 15.11 shows the tab Design Summary which displays the report files
generated by the Platgen and XST in order to check details about the design, such
as the occupied FPGA resources or the estimated maximum frequency of the
clock.
15.1.2.3 Implementation
The implementation computes the FPGA layout which is stored in a Native Circuit
Description (NCD) file. The design flow executes three main tools: NGDBUILD,
MAP and PAR [15]. The NGDBUILD translates the NGC files and annotates
constraints from a User Constraints File (UCF). The following tools compute the
layout based on the annotated constraints. The design flow continues with
the MAP and PAR tools to map the netlist into the FPGA resources and to compute
their placements and routings.
The BSB wizard generates the UCF for the selected prototyping board. The
XPS project refers the UCF which must be edited to specify the attachment of the
display and switches to the FPGA board. The UCF also specifies the clock
frequency of the external oscillator.
386
15
3
1
Fig. 15.11 Report files from the Platgen and synthesis
The XPS menu Hardware ? Generate Bitstream launches the BitGen tool [15]
which generates the bitstream file from the FPGA layout. First, the XPS executes
the design flow to implement the FPGA layout, if necessary. Then it generates the
BIT (bitstream) file system.bit and the BlockRAM Memory Map (BMM) file
system_bd.bmm. The microprocessors local memory is implemented on BRAMs,
15.1
387
Fig. 15.12 Hardware export

from EDK to SDK
but the generated BIT file does not initialize them, since the executable binary is
not available at this stage. The file system_bd.bmm annotates the physical placement of the BRAMs with the microprocessors local memory. This file will be
required later to update the BRAM contents of the bitstream. The tab Design
Summary shows the reports generated by the implementation tools.
15.1.2.4 Software
The XPS menu Project ? Export Hardware opens a dialog window to export the
required files to SDK, as shown in Fig. 15.12. Select the option to export the BIT
and BMM files to permit SDK to program the FPGA. It creates a new directory
which is allocated in the XPS project folder.
SDK starts opening a dialog to set the workspace folder. Write the path
c:\edk13.1\led7seg\SDK\workspace to create it into the SDK folder which was generated
by XPS during the hardware exportation. The software development involves two stages:
(1) The BSP generation. Creates a set of headers and libraries to control the
hardware from the microprocessor.
(2) The executable ELF file. It builds the file executed by the embedded
microprocessor.
15.1.2.5 Board Support Package

The BSP provides the Application Programming Interface (API) for the target
hardware and Operating System (OS). The BSP generates a set of libraries and
header files that facilitates the development of software executables. The application C/C++ source files are compiled and linked using the API to build the
binary ELF which is executed by the microprocessor.
The SDK menu File ? New ? New Board Support Package Project launches a
wizard to create a BSP project (see Fig. 15.13). Choose the platform hw_platform_0
which is the hardware exported from XPS. Then set the standalone OS [18] since it
provides interrupts management and it does require a large memory capacity. Other
388
15
3
c:\edk13.1\led7seg\workspace\standalone_bsp_0
Fig. 15.13 Configuration of the BSP project
OS provide more advanced features, but they require external memory. The BSP
project standalone_bsp_0 is located by default in a folder contained in the SDK
workspace.
The wizard generates the BSP project which is linked to a MSS file. The MSS
[14] is a text file which list the drivers used by the peripherals and the OS for the
microprocessor. The Libgen [13] tool reads the MSS file to generate the BSP. As
with the MHS file, the MSS can be graphically or manually edited. Figure 15.14
shows the graphical view which configures the MSS. Change the standard input
(stdin) and output (stdout) to the instance rs232 in order to permit the console
functions to use the UART peripheral.
The MSS file can also be manually edited, and it reflects the configuration
changes done in the previous dialog.
The rest of the MSS file shows the drivers and peripherals. A peripheral driver
is a collection of declarations and functions that can be used to control it from the
executable. By default the BSP wizard sets a specific driver to every peripheral,
but the user can change it to set a generic driver or no driver. The generic driver
can control any peripheral, but the user must have a deeper knowledge about its
internal architecture. The systems hardware provides two GPIO peripherals: the
15.1
389
2
6
4
5
Fig. 15.14 Configuration on the BSP
switches and the led7seg instances. Select the generic driver for them in order to
understand better the role of internal registers of peripherals.
The SDK automatically calls the Libgen tool [13] when the MHS is changed to
build the BSP. The user may disable the Build Automatically behaviour in order to
clean or build the BSP using the commands under the menu Project. The Libgen
tool compiles the source files of the peripheral drivers and the OS, and it stores
them into the A (archive) library files. It also generates the H (header) files that
declare the functions contained in the libraries. The library and header files are
stored in the folders lib and include of the instance microblaze_0. The SDK can
display the contents of both folders and open the header files.
390
15
An important header file is the xparameters.h which declares a set of parameters about the hardware. Every peripheral has its own parameters that are obtained
from the exported hardware, as the addresses range of the GPIOs. The declarations
can be used by the C/C++ source files to control the peripherals.
15.1.2.6 Executable
The SDK will build the ELF executable from the source C++ files that are compiled
and linked with the functions stored in the BSP libraries. Click the menu
File ? New ? Xilinx New C++ Project which opens a wizard to create a C++
project for the BSP. Change the default project name to app1 and select the
previously generated BSP standalone_bsp_0, as depicted in Fig. 15.15. The wizard
creates a nested folder app1/src in the SDK workspace to store the source files that
will be compiled.
Using the Windows Explorer delete the template file main.cc which was
created by the wizard, and copy the new source files: ledseg7.cc, led7seg.h and
application.cc. Go to SDK and click the menu Refresh of the contextual menu
(right button of the mouse) of the project app1, in order to update the list of
source files. The SDK can open and display the source files in its integrated
editor (see Fig. 15.16).
The source files led7seg.h and led7seg.cc declare and implement a C++ class
named CLed7Seg which controls the display through the GPIO. The EDK
peripherals implement a set of 32-bit registers that are used to control them.
The peripherals registers are memory-mapped, therefore, MicroBlaze can access
them when it executes read/write instructions to the content of a C/C++ pointer.
The application can directly control the peripheral, although, it is necessary to
have a deeper knowledge about the internal architecture. The first register of a
GPIO [5] is the GPIO_DATA which is mapped at the base address of the
peripheral. The register retrieves/sets the state of the input/output ports depending
if the microprocessor reads/writes it.
The class constructor assigns the input argument to the integer (32-bit) pointer
GPIO_Data. Any pointer used to access a peripherals register should be declared
volatile. If not, the compiler may optimize a set of sequential memory accesses
through the pointer, changing the order or deleting some of them. The GPIO
15.1
391
Fig. 15.15 Creating a new C++ project in SDK
2
1
Fig. 15.16 Displaying the C++ source files
method concatenates the anodes and segments to write them into the GPIO_DATA
register through its pointer. The header file of the class declares the two parameters
that configure the active state of the anodes and segments of the display, therefore,
they can be easily changed to adapt it to another prototyping board.
392
15
The class declares two member variables: Data and Config. The Data is a
16-bit variable which stores the number which is displayed. The Config is an 8-bit
variable which uses the two LSBs to turn on/off the display and to show/hide the
left-side zeros of the number. The method Refresh is periodically executed since
the timers ISR calls it. It reads the member variables and calls the Digit method
to display one of the digits starting at the left side. The Digit method first computes the segments and anodes of a digit and then it calls the GPIO method to
display it.
The application C++ file is composed of 4 sections. The first section opens the
required header files. The application controls the GPIOs directly, but the rest of
the peripherals are controlled through their drivers. Therefore, it opens the header
files that declare the functions stored in the BSP libraries. The file xparameter.h
declares base addresses that are necessary to use the driver functions.
The second section initializes the object Display for the class CLed7Seg. The
objects constructor gets the base address of the GPIO which drives the display.
The section also declares the global variables data and endloop that are managed
by the ISR of the UART.
The third section is composed of the two ISRs. The timers ISR periodically
reads the state of the two external switches and refreshes the display. First, it reads
the register GPIO_DATA of the peripheral which attaches to the two external
switches. Then, it compares the state of the switches against the previous call.
A change in one of the switches will swap one of the two bits that configure the
display, using the bitwise XOR operators. Finally, it refreshes the display. The
other ISR is executed when the UART receives a new character which is read
15.1
393
using its driver function [9]. Depending on the received character, it changes the
data or the configuration of the display, or it quits the application.
The last section is the main function of the application. It configures and
enables the interrupt sources, and then it executes a loop until the application
quits. The loop can execute any computation without affecting the control of the
display.
The timer peripheral [6] implements two registers to periodically generate an
interrupt request: Timer Control/Status Register 0 (TCSR0) and Timer Load
Register 0 (TLR0). The configuration of both registers asserts the interrupt signal
every 5 ms (250,000 counts, 50 MHz clock). The constants and functions of the
timer driver are declared in the header file tmrctr_l.h [10] which was generated by
the BSP. They are low-level API functions since the programmer knows the
functionality of the registers. These functions compute the address of the registers
and write data into them through a pointer.
The driver of the interrupt controller provides functions [11] to register ISRs and
to enable the interrupts sources. Finally the application enables the MicroBlazes
interrupt input through an OS function [18].
394
15
By default, the wizard of the C++ project generates two targets: Debug and
Release. They differ in the flags of the GNU compiler [13]. The target Debug
compiles source files without optimizations and enabling debug symbols. The
target Release compiles source files with optimizations to build smaller and faster
code which is not suitable to debug. The targets configure other compiler flags that
are derived from the MicroBlazes configuration in order to use the optional
machine instructions. The menu Project ? Build Project builds the active target
which can be changed anytime using the Project ? Build Configurations ? Set
Active. Then, SDK runs the GNU tool chain which compiles the source files and it
links the resulting object code with the BSP libraries. The executable ELF file is
stored in a nested folder which is named as the target.
15.1.3 Programming and Debugging

The SDK menu Xilinx Tools ? Program FPGA opens a dialog window which
displays the BIT and BMM files that were imported from EDK, as shown in
Fig. 15.17. Select the ELF file app1.elf of any of the two targets and press the
button Program. It calls the Data2MEM tool [20] which generates a new bitstream
download.bit from an existing bitstream and its annotated BMM file. The new
bitstream configures the entire FPGA including the BRAM contents with the
selected ELF file. Then, the new bitstream is programmed into the FPGA and the
application begins to run.
The serial ports of the FPGA board and the PC are attached to test the
application. The PC should execute a terminal configured with the same
communication parameters as the embedded system. The terminal displays the
received characters from the FPGA and sends the characters that are pressed on
the keyboard. The SDK provides its own terminal which can be used for this
15.1
395
c:\edk13.1\led7seg\SDK\workspace\hw_platform_0\system.bit
c:\edk13.1\led7seg\SDK\workspace\hw_platform_0\system_bd.bmm
2
c:\edk13.1\led7seg\SDK\workspace\app1\Release\app1.elf
3
Fig. 15.17 Bitstream configuration to program the FPGA
3
1
Fig. 15.18 Configuration of the terminal in SDK
purpose (see Fig. 15.18). Press the reset button of the FPGA board and play with
the terminal.
The Xilinx Microprocessor Debugger (XMD) [13] is a low-level tool which
manages the programming and debugging of the embedded system through the
MDM peripheral and the JTAG programming cable. The user can interact with the
XMD clicking the SDK menu Xilinx Tools ? XMD console
396
15
1
2
3
Fig. 15.19 Configuration of the debugger
The debugging permits the programmer to inspect variables, insert breakpoints

or execute step-by-step, in order to correct or improve the executable. Click the
menu Run ? Debug Configurations to open the dialog shown in Fig. 15.19.
Ensure it selects the ELF under the target Debug. SDK will ask to switch to a new
perspective which facilitates the debugging tasks, therefore it is recommended to
confirm it. The debug perspective shows the source code, disassembly, variables,
microprocessors registers, memory, breakpoints, XMD console and more. The
user can debug the executable, manually launching XMD commands which is
quite uncomfortable. The SDK debugger relies on XMD to send or retrieve data
through a graphical view. The debugger starts uploading the ELF into the BRAM
and suspending it at the first executable line of the source file.
Set a breakpoint on a line of the timers ISR and resume the application to
observe the display refreshing (see Fig. 15.20). The tab Variables shows the local
variables of the ISR that are updated when the user plays with the switches and
resumes the application. The tab Expressions permits to display the object Display
and the global variables data and endloop.
15.1

2
397
3
Fig. 15.20 Debugging the application
In a similar way the user can place another breakpoint at the assignment of the
variable rs232_char in the ISR of the UART. The application will suspend when it
receives a new character from the PC. Then, the ISR updates the data or the
configuration of the display.
15.2 Case Study 1-B: Creating a Custom Peripheral

The previous embedded system devotes three peripherals (two GPIOs and a timer)
to drive the display and to read the switches. The executable uses the timers ISR
to periodically read the switches and refresh the display. A dedicated peripheral
can improve the design since it can replace the ISR and peripherals that are
devoted to a specific task. The main disadvantage is the greater design effort since
the designer must develop the source files of the hardware and the software driver.
The designer will surely require simulating the source VHDL files, in order to
verify and modify the hardware of the peripheral.
This case study starts from the previous one. It modifies the hardware design
and the executable to incorporate the new peripheral. Copy the previous project
folder and rename it to led7seg_ip. Then, open the XMP file to launch the XPS.
15.2.1 Design of a Custom Peripheral

A peripheral performs a specific task on hardware. Typically they are attached to
the PLB bus as a slave in order to permit the microprocessor to access their
398
15
internal registers to control them. More sophisticated peripherals can be attached

as PLB masters in order to access memory themselves. These kinds of peripherals
are quite harder to develop and they are much less common, therefore, they are not
covered in this example.
The hardware of a peripheral is implemented from a set of source VHDL files.
The driver of a peripheral is compiled from a set of source C and H files to build
the BSP. The source files must be organized in a set of named folders in order to
use the new peripheral in the EDK and SDK.
15.2.1.1 Hardware Design
The hardware is described in a set of VHDL files that are synthesized during the
design flow. EDK requires two files in order to permit the integration of the
peripheral to the embedded system: the MPD and PAO. The peripheral wizard from
XPS generates the folders and a set of template files that must be modified in order to
develop the desired IP. Open the wizard by clicking the menu Hardware ? Create
or Import Peripheral.
(1) Select the option Create templates for a new peripheral.
(2) Select the XPS project repository to create the template files into the local
repository which is stored in the XPS directory. The folders pcores and drivers
contain the local repository of the hardware and software.
(3) Set the name the peripheral to plb_led7seg, and the version to 1.00.a. The
wizard will create the folders plb_led7seg_v1_00_a in the local repositories.
(4) Select the PLB attachment.
(5) Unselect all the slave and master services of the IPIF (IP InterFace) to generate
the simplest template files. The IPIFs are IPs deployed with EDK to facilitate
the PLB interface.
(6) Continue with the default options of the next three dialogs.
(7) Select the option Generate template driver file. This option creates the template files for the driver in the software repository.
The MPD and PAO files are stored in the folder data which is nested in the
local hardware repository pcores.
The Microprocessor Peripheral Description (MPD) file [14] is composed of four
sections. The first section declares the description and implementation options.
The second section sets the bus interface to Slave PLB (SPLB). The last two
sections declare the parameters and ports of the SPLB. The data generated by the
wizard in the MPD template must not be modified. However, the designer can add
more parameters and ports to extend the peripheral. The parameters will be passed
as VHDL generics and ports during the synthesis.
The new peripheral adds three parameters in order to configure the peripheral
from XPS. The refresh period is an integer parameter which contains the number of
microseconds, and its default value is set to 5,000 ls. The boolean data configure the
active state of the ports that drive the display. The peripheral adds two output ports to
15.2
Case Study 1-B: Creating a Custom Peripheral
399
PLB
PLB_*
plb_led7seg.vhd
Sl_*
IPIF
Bus2IP_*
IP2Bus_*
user_logic.vhd
counter
reg_data
reg_control
2
16
led7seg.vhd
core
4
GLUE LOGIC
anodes
segments switch_off
switch_zeros
Fig. 15.21 Hierarchical schematic of the peripheral
drive the 7 segments and the 4 anodes of the display. Finally, it declares two input
ports that attaches to the external switches that configure the display.
400
15
Figure 15.21 shows a hierarchical schematic of the peripheral and related files.
The wizard created two VHDL files that are stored in the folder hdl\vhdl which is
nested in the hardware repository. The plb_led7seg.vhd is the top-level file which
connects an instance of the user logic to the PLB. The user_logic.vhd is a dummy
peripheral, therefore, the file must be modified to perform the desired functionality. The tasks related to the timers ISR are now described as hardware in this
file. The computation related to the class CLed7Seg class is now described in the
new hardware file led7seg.vhd.
The Peripheral Analyze Order (PAO) file [14] is the ordered list (bottom to top
level) of libraries and files required to synthesize the IP. The first two entries refer
to EDK libraries due to the selected IPIF. Then, it continues with the list of VHDL
files that will be synthesized into the target library. The target library must be
named as the repository folder of the peripheral.
The top-level file declares the entity plb_dec7seg and its architecture.
The template of the entity leaves space to add new generics and ports, therefore,
the user must complete it. The VHDL generics and ports of the entity must be
declared in the same way as in the MPD file.
The architecture declares an instance to the user logic. It leaves space to map the
new generics and ports. The two switches are directly connected from the top-level
15.2
401
input ports to the user logic. The glue logic attaches the output ports from the user
logic to the top-level ports, in order to drive the display. The synthesizer will
infer inverters between these ports if the configuration parameters are set to true.
The user logic will provide its internal timer, therefore, the instance requires a new
generic which configures the number of clock cycles to refresh the display. It is
computed from the MPD parameters that define the refreshing period (microseconds) and the clock period (picoseconds). Finally, the user logic also requires a
generic which configures the number of internal registers.
The architecture also contains the IPIF instance which eases the PLB connection. It adapts the PLB handshake to/from IP signals named ipif_Bus2IP_*/
ipif_IP2Bus_*. The IPIF also decodes the PLB address bus to enable one of the
peripherals internal registers. The architecture declares the number of internal
registers in the user logic, which is two in this case. This number affects the width
of the signals Bus2IP_RdCE (Read Chip Enable) and Bus2IP_WrCE (Write Chip
Enable) that arrives to the user logic. When MicroBlaze executes a read/write
instruction to a peripherals address range, the IPIF decoder sets one of the enable
bits, in order to read/write one of the internal registers of the user logic. The IPIF
maps registers on 32-bit boundaries from the peripherals base address. The first
register is mapped on C_BASEADDR, the second register on C_BASEADDR+4,
and so on. In order to simplify the IPIF, the C_HIGHADDR parameter is usually
402
15
configured much higher than necessary, but it leads to an incomplete address

decoding. In this case, the peripherals registers can be accessed from different
addresses since the IFPF does not decode some bits of the address bus.
The template of the user logic is a dummy design which must be modified. The
entity template declares an important generic which is the number of registers. The
entity must be completed according to the top-level file. Therefore, the designer must
add the new generics to configure the refreshing counts and the new ports to drive the
display and to read the switches. The default values of the generics are overwritten
since they are passed from the instance at the top-level file.
The architecture of the user logic declares two registers to control the display
that are accessible from the PLB. The first one (reg_data) is a 16-bit data register
which sets the 4 digits to display. The second one (reg_control) is a 2-bit register
which controls the display configuration: one bit to turn on/off the display and the
other bit to show/hide the left-hand zeros. The input ports Bus2IP_RdCE and
Bus2IP_WrCE provide a single bit for each register to read or write them. The data
comes from the port Bus2IP_Data during a write access to one of the registers.
During a read access, one of the registers drives the port IP2Bus_Data.
The architecture uses the signals bus_data/ip_data to get/set the data bus, since
they are declared in descending bit order as the registers. The PLB must
acknowledge the access completion before MicroBlaze can continue with a new
access. The user logic asserts the ports IP2Bus_RdAck/IP2Bus_WrAck when it
completes the access to the registers. In this case, the user logic does not require
wait states to read/write the registers.
15.2
403
The rest of the user logic performs the tasks executed in the timers ISR in the
software application. It implements a counter which periodically asserts the signal
refresh. This signal is used to refresh the display and to capture the state of the
switches in order to modify the control register. There is an instance, named core,
of the IP led7seg, which drives the display from the registers contents.
404
15
The file led7seg.vhd implements the functionalities described in the C++ class
CLed7Seg in the software application. It generates the signals that drive the display, updating the displayed digit when the port refresh is asserted. The architecture is composed of two processes. The first process updates the index of the
digit to refresh, starting at the left side. The second process computes the displays
ports based in the input ports off, zeros and data that are driven from the
peripherals registers.
15.2.1.2 Driver Design

A peripherals driver is built from header and implementation files. The C files
implement functions that control the peripheral. The header H files contain the
declarations required to use the functions. The SDK requires two additional files,
the MDD and the TCL, and a structured folder organization in a repository in order
to build the BSP with the peripherals driver. The Microprocessor Driver Definition
(MDD) file [14] declares the supported peripherals, dependencies of the driver and
files to copy. The TCL file is a script used to compile the source files and build the
library. The XPS peripheral wizard generates the template files and folders.
The designer usually does not modify the MDD and TCL files, but he must rewrite
the H and C files to develop the driver.
The MSS file controls the peripheral drivers that are built in the BSP. The
Libgen tool creates a folder named as the microprocessor and copies the H files
into the folder include. Then it compiles the C files and adds the functions into the
library libxil.a which is stored in the folder lib of the BSP.
The header file contains the declarations that are used to build the BSP and the
executable. First, it defines the masks to control the display. Next, it declares C
macros to read or write the peripherals registers through a pointer. Finally, there
are the declarations of the visible functions.
15.2
External
oscillator
50MHz
405
Reset
button
JTAG
programmer
system
clock
DLMB
Controller
reset
DLMB
MDM
MicroBlaze
BRAM
ILMB
Controller
Interrupt
Controller
ILMB
PLB
led7seg
plb_led7seg
rs232
UART
switches
4-digit,
7-segment
led display
Serial
communication
Fig. 15.22 Overview of the modified system architecture
The C file of the driver is compiled during the BSP generation. The library file
stores the functions, but not the C macros. The library may store other internal
functions that are not visible for the programmer since they are not declared in the H
file. The function which swaps one of the bits of the control register executes two
accesses to the register. First, MicroBlaze reads the register and changes one bit
according to a mask. Next, it writes the resulting value into the same register.
406
15
Fig. 15.23 Selection of the ELF to generate the simulation model
15.2.2 Systems Hardware/Software Modification

The new peripheral replaces the timer and the GPIOs connected to the display and
switches, as well as the software code related to them. Therefore, the hardware and
software of the new system must be modified, accordingly.
15.2.2.1 Hardware Modification
The MHS file of the new system adds the instance led7seg of the PLB display
controller plb_led7seg, and it removes the timer and GPIOs, as depicted in
Fig. 15.22. The external ports connected to the display and switches are assigned
to the display controller. The parameters that configure the refresh period and
active state of the anodes and segments can be easily modified to adapt them to
other FPGA board. The MHS parameters overwrite the default values defined in
the peripherals MPD file. The modified system requires just one interrupt source
for the UART, therefore, the interrupt controller could be eliminated. However, the
system maintains it since it does not increase significantly the area and it provides
the possibility to add new peripherals to request interrupts.
15.2
407
The names of the external FPGA ports, connected to the display and switches,
have changed, therefore, the UCF must be updated.
At this point, the designer can easily test the peripherals hardware using the
XMD tool. Using the XPS, implement the new hardware. Then, click Device
Configuration ? Download Bitstream to create the download.bit file and program
the FPGA. If the XPS project has no ELF file, the bitstream configures the BRAM
to store the default executable bootloop which runs a dummy loop. In order to
program the FPGA, XPS calls the iMPACT tool with the commands of the batch
file download.cmd. Check the number of the p flag, which configures the FPGA
position in the JTAG chain.
408
15
Once the FPGA is successfully programmed, click the Debug ? Launch XMD to open
the command shell. The XMD shell initially shows the microprocessor configuration when
it connects to the MDM. Then, the user can test the peripheral when he runs the XMD
commands: mwr (memory write) and mrd (memory read). Write the peripherals base
address (0 9 84C00000) to change the displayed number. Write the next register
(0 9 84C00004) to change the configuration of the display. The peripherals registers are
accessible from different addresses due to the incomplete IPIF decoder.
15.2.2.2 Software Modification

At the software level, the executable has to be also modified in order to delete the
code related to the GPIOs and the timers ISR. Export the new hardware from XPS
and create a SDK workspace in the c:\edk13.1\led7seg_ip\SDK\workspace path.
Check that the imported hardware presents the new display controller.
Create a new BSP project and select the standalone OS. By default, the MSS
assigns the generic driver to the display controller led7seg. A more convenient
way is to assign the specific driver to the peripheral in order to build the BSP.
15.2
409
The Libgen tool generates the BSP according to the information described in
the MSS file. It gets the source files from the folder plb_led7seg_v1_00_a which
must be stored in the driver repositories. Then, it copies the header files and builds
the libraries. The file libgen.options must be modified to configure the local
repository to the XPS directory.
The C++ code related to the GPIOs and timers ISR is deleted. Therefore, the
files that declare and implement the class CLed7Seg are deleted in the project.
The modified file application.cc calls the drivers functions to control the
display.
The application can be programmed and debugged following the same steps as
in the previous case study. Check that the BIT and BMM files are the imported
ones from the current XPS project, before proceeding.
15.2.3 Functional Simulation

The source VHDL files should be simulated if the peripheral is not working as
expected. The easiest way to test the VHDL files is by performing a functional
simulation of the embedded system. The system can be simulated with the ISim or
other supported simulator. The XPS can select the simulator when you click the
Edit ? Preferences. The Xilinx ISim has some limitations when it is compared
with other third-party simulators. However, the ISim is easy to use, and it provides
the simulation libraries.
Instead of simulating the application executable, which depends on the serial
communication, it is preferable to develop a much simpler executable. Create a
410
15
new Xilinx C++ project in SDK named simulation. Modify the C++ template file
main.cc to add a sequence which just writes the peripherals registers, as desired.
The function delay is called to improve the observation on the simulation results.
Build the target Debug to get the file simulation.elf.
The XPS project must import the ELF file in order to generate the simulation
model. Click the menu Project ? Select Elf file and choose the file simulation.elf
which is stored in the SDK project (see Fig. 15.23)
XPS launches the Simgen tool [13] to generate the set of files required to simulate the embedded system running the selected executable. Click the XPS menu
Project ? Project Options which opens the dialog window shown in Fig. 15.24.
Select the options: VHDL, generate testbench template and behavioural model.
This is the recommended model since it simulates the peripherals source files. The
other models are more complicated, since they simulate the output files from the
synthesis or the implementation stages. Click the XPS menu Simulation ?
Generate Simulation HDL Files to call the Simgen tool. The Simgen does not
support some FPGA families, such as the Spartan-3, but there is a workaround.
Choose any supported device since it does not affect the behavioural simulation of
the peripheral, because it does not require the synthesizing of its source VHDL files.
Finally, click the menu Simulation ? Launch HDL Simulator which takes some
time to compile the simulation files before starting the ISim [19].
Simgen creates the testbench file system_tb.vhd. It declares the instance dut
(Device Under Test) of the system and the stimulus applied to its inputs. ISim
shows the hierarchical view of the instances to simulate. Open the contextual menu
of the top-level instance system_tb and click Go To Source Code in order to edit
the testbench template, as shown in Fig. 15.25.
The template file drives the clock and reset signals, and it provides a users
section to write more stimulus. In order to simulate the external switches, change
the state of the switch, which turns off/on the display, at 250 and 260 ls.
15.2
411
5
1
6
2
3
Fig. 15.24 Configuration of the simulation model
The simulation will take a huge number of clocks to refresh the display. The
displays ports are updated every 200,000 clock cycles when the refresh cycle is
set to 4,000 ls (50 MHz clock frequency). Each instance of a peripheral is simulated from a wrapper file which configures the IP according the MHS file. Go
through the hierarchical view, and select the instance led7seg. Open the wrapper
file led7seg_wrapper.vhd, and change the generic C_REFRESH_PERIOD_US to
shorten the simulation of the refresh. This change does not affect the synthesis or
the implementation since it does not modify the MHS file.
412
15
Fig. 15.25 Hierarchical

view of the instances to open
the testbench file
2
3
3
4
Fig. 15.26 Configuring the waveform window
15.2
413
Fig. 15.27 Simulation waveform
Any change in the VHDL files requires the compilation of the model before starting
the simulation. Therefore, click the ISim menu Simulation ? Relaunch. In order to
display the external ports of the system, select the instance system_tb and drag the desired
signals into the waveform window. Add a divider, named USER LOGIC, in order to
display a new group of waveforms separately from the previous ones. Go to the hierarchy
of the instances system_tb ? dut ? led7seg ? led7seg ? USER_LOGIC_I and
drag the IPIF signals and the peripherals registers to the waveform. Repeat the previous
steps to display the signals of the instance core, as shown in Fig. 15.26.
414
15
By default, signals are displayed in binary radix which is uncomfortable for

some signals. Select the peripherals registers and the IPIF busses to change the
radix to hexadecimal.
Write 100 ls on the simulation time, and run the simulation, as shown in
Fig. 15.27. The resulting waveform can be zoomed in/out to show the desired time
interval. Check that the signal refresh is asserted during a clock cycle every 1 ls, due
to the change done in the wrapper file. Display the peripherals registers at 35 ls. At
this point MicroBlaze has already executed the first four lines of the C++ file,
therefore, the peripherals registers (reg_data, reg_control) contain the expected
values. The ports anodes and segments that drive the display are refreshed according
to the value stored in those registers.
The simulation should continue another 100 ls to display how Microblaze
executes the driver function Led7Seg_SwapOff. The function executes two
consecutive accesses to the register control_reg: a read followed by a write access.
The registers are accessed when the microprocessor executes read/write instructions
to memory addresses within the peripherals address range. The IPIF decodes the
PLB address bus in order to assert one of the enabler bits of the signals
Bus2IP_RdCE/Bus2IP_WrCE that arrive to the user logic, in order to read/write a
peripherals registers. The register control_reg is associated with the index 1 of these
signals. The waveform window permits a search for the next/previous transition of a
signal. Search a rising edge of the bit Bus2IP_RdCE(1), as shown in Fig. 15.28.
At this time, the MicroBlaze is reading the control register. The peripheral carries the
registers content on the signal IP2Bus_Data and it asserts the IP2Bus_RdAck to
complete the read access. The next access, MicroBlaze writes the control register.
The peripheral gets the new content from the Bus2IP_Data and asserts the
IP2Bus_WrAck to complete the write access.
Figure 15.29 shows the simulation of the user action on a switch at 250 ls and
260 ls to turn off/on the display, as described in the testbench file. The peripheral
updates the control register when the signal refresh is asserted, after the user action.
If the C/C++ code of the simulation executable is changed it is necessary to
close the ISim before executing Simgen. Save the waveform file before exiting
ISim. Then, build the SDK project to get the new ELF file and generate the new
simulation model with XPS. Finally, launch ISim, open the saved waveform file,
and run a new simulation.
Once the simulation is finished, select the target FPGA in the project options
and generate the bitstream. The system will be synthesized and implemented for
the selected device.
15.3 Case Study 2: Implementation of a Custom Coprocessor

A coprocessor is a specific processor aimed to accelerate an algorithm. The
hardware architecture is designed to efficiently speed-up a specific computation.
The rest of computations and tasks are carried on the general-purpose
15.3
Case Study 2: Implementation of a Custom Coprocessor
415
3
1
2
4
Fig. 15.28 Simulation of a read/write accesses to a peripherals register
Fig. 15.29 Simulation of the users action on a switch
416
15
Plain
block
AddExpKey
SubBytes
Encrypted
block
AddExpKey
InvShiftRows
Encrypted
block
AddExpKey
ShiftRows
Plain
block
AddExpKey
InvSubBytes
Expanded
Key
AddExpKey
MixColumns
key
Expanded
Key
key
AddExpKey
InvMixColumns
Fig. 15.30 Block encryption (left) and decryption (right) in the AES-128 cipher
microprocessor. The microprocessor writes commands and data to the coprocessors registers in order to perform a computation. When the coprocessor
completes, the microprocessor retrieves the computed data in order to continue
the algorithm.
This case study presents an embedded system which communicates with an
external PC in order to set the state of some led diodes and to read the state of some
switches. The system is connected through the serial port, but the communications
are ciphered using the Advanced Encryption Standard (AES-128). The system
decrypts the commands received from the PC and encrypts the answer messages.
The MicroBlaze cannot decrypt the messages at the desired speed, therefore, a
coprocessor is developed to accelerate the AES-128 computation.
15.3.1 A Brief Introduction to the AES-128 Cipher

The AES [1] standard comprises three symmetric-key block ciphers: AES-128,
AES-192 and AES-256, where the number denotes the key size. The cipher
encrypts a fixed-size block of plain text to get a block of ciphered text. It also
decrypts ciphered blocks in the reverse way using the same key. The block size of
the AES ciphers is 128 bits independently of the key size.
All the steps executed during AES encryption/decryption are done using a
variable named state. It is a 128-bit data arranged on an array of 4-rows x 4-columns
of 8-bit elements (bytes). During the AES-128 encryption the state is initially
written with a block of plain text. Then it is processed 9 rounds, and each round is
composed of 4 steps, as shown in Fig. 15.30. The initial and the final rounds are
slightly different since some steps are skipped. The AES state stores intermediary
computations and the block of ciphered text at the end of the final round. In a similar
way the decryption initializes the state with the ciphered block which is processed
with inverted computations and steps to get the block of plain text.
One of the AES steps applies the expanded key which must be computed from
the cipher key. The expansion of the key can be computed online or offline. The
15.3
417
Fig. 15.31 Enabling and configuring the profiling in the BSP
offline mode computes the expanded key before starting the rounds, therefore, it is
computed only when the cipher key changes.
The transmitter divides a message into blocks that are encrypted and transmitted. The receiver decrypts the received blocks to build the message. There are
several operation modes and padding schemes to permit block ciphers to work
with messages of any length [22]. This case study chooses the Electronic CodeBook (ECB) mode and null padding, due to its simplicity. The other modes require
the IV (Initialization Vector) generator and a different block partitioning, but this
fact does not affect the AES block cipher.
15.3.2 Software Implementation of the AES-128 Cipher

In order to study the applicability of a coprocessor on this application, the AES128 cipher is implemented as a C++ class named CAES128. The system receives/
transmits ciphered messages from/to the external PC and it uses the C++ class to
decrypt/encrypt blocks according to the cipher key. Microblaze reacts to the
messages to set the leds or to report the state of the switches.
The hardware specification is quite similar to the first case study, but it
pro-vides a single GPIO peripheral to read two switches and to set two leds.
Copy the folder which allocates the XPS project of the first case study and
rename it as co-processor. Edit the MHS file to add the 4-bit width GPIO and
connect its bidirectional port to an external FPGA port. The timer will not only
measure the time taken to encrypt/decrypt blocks, but it will also be used for the
ciphers profiling. The profiler requires the timer to be able to interrupt the
418
15
microprocessor, therefore, the MHS connects their interrupt ports. The system
removes the interrupt controller since there is a single interrupt source.
The UCF file is modified to attach the external port to the switches and leds
according the FPGA board. Implement the hardware and export it to SDK.
Open the SDK workspace c:\edk13.1\coprocessor\SDK\workspace. Modify the

BSP settings to enable the profiling and configure the profile timer (see Fig. 15.31).
Clean and build the new BSP which includes a new library which is required when
the application is profiled.
Delete the previous C++ projects and create a new project named server which
targets the new BSP. The server application is composed of four files.
The files caes128.h and caes128.cc are the source files of the C++ class
CAES128 which implements the cipher. It is a straightforward implementation in
order to facilitate the next steps. The class can be improved, in terms of speed and
security, but it requires a deeper knowledge of the AES. The class provides the
method SetKey to compute the expanded key at the beginning of the application.
Then, it provides methods to Encrypt or Decrypt a single 128-bit block.
15.3
419
3
2
5
4
Fig. 15.32 Settings to build the release target
The file app.cc implements the application. It waits for commands launched
from a PC in order to set the leds or to report the state of the switches. Any
command produces an answer message which is transmitted to the PC. Commands and answer messages are encrypted during their transmission on the serial
interface. The function send_rs232_cipher implements the ECB operation mode
and null padding. It divides an answer message into 128-bit blocks that are
individually encrypted and transmitted. In a similar way, the function
get_rs232_cipher builds a command message from the received and decrypted
blocks.
420
15
Fig. 15.33 Project cleaning

and building
1
2
The application stores into the variables EncryptMaxCycles/DecryptMaxCycles

the maximum number of clock cycles taken to encrypt/decrypt a block. It relies on
the file cchronometer.h which implements the C++ class to get the number of
clock cycles from the timer. The compiler can skip the code related to the
15.3
421
chronometer when the CIPHER_CHRONOMETER declaration is commented. The

application also declares the CIPHER_DISABLE which can be used to skip the
cipher to facilitate debugging.
Click the menu Project ? Properties to edit the settings of the build of the
target Release, as shown in Fig. 15.32. The compiler can optimize the code to
improve the execution speed, by changing the order of instructions or the unrolling
of loops. The ELF may not fit in the BRAM memory if the optimization level at
the maximum (-O3). In order to display the measured encryption/decryption time,
lower the optimization level (-O2) and set the debug to the maximum level (-g3).
Set the Release as the active target. Then, click the menu Project ? Clean to
clean the C++ project in order to build it from scratch (see Fig. 15.33).
In order to test or debug the application, the PC must also use the AES-128
cipher. There are many PC programs that permit the ciphering of messages
through a communication channel, but they usually employ sophisticated protocols (SSH, SSL, etc.) to authenticate users before transmitting the cipher keys.
The presented application is much simpler, since the PC and the embedded
system are programmed with the same cipher key. In order to test the application, the PC executes a Linux program to encrypt/decrypt messages from/to a
console which redirects the input/output to the serial port. The folder
linux_aes128 stores the executable files for Linux or Cygwin. Cygwin permits
the execution of applications written for Linux on Windows PC. Cygwin cannot
execute the Linux binaries, but it requires the building of the executable from the
Linux source files.
Install Cygwin or a Linux virtual machine and open a shell to launch a set of
commands. The first command changes to the directory which contains the path of
the executable (it may change). The next one configures the communication
parameters on the serial port (/dev/com1 may change). The third command maps a
file descriptor to the serial port. The last two commands decrypt or encrypt
messages between the console and the serial port. The cipher key is set to the
same value as the embedded system. Finally, program the FPGA and test
the application. The Linux console shows unreadable text if the cipher keys are
mismatched.
In order to display the time taken to encrypt/decrypt blocks, debug the ELF of
the target Release using SDK. Set a breakpoint at the last instruction of the C++
file to pause the execution when the user launches the quit command. Then resume
the application and launch several commands to the embedded system through the
Linux console. The variables EncryptMaxCycles and DecryptMaxCycles contain
422
15
the maximum number of clock cycles required to encrypt and decrypt a block. The
decryption of a block takes 1.76 ms (87,338 clock cycles, 50 MHz clock
frequency), but the serial communication requires 1.39 ms to transmit it (115,200
bps). The FIFO of the UART may overrun during the reception of large messages
since the MicroBlaze cannot decrypt blocks at the required speed. To solve the
problem, the transmission speed can be lowered or the system can implement a
flow control. A better solution is to accelerate the most time-spending
computations.
15.3.3 Profiling
The profiler [12] is an intrusive tool which is used to test the application performances. It reports the number of calls and the execution time of every function.
The profiling requires a dedicated timer able to interrupt the microprocessor in
order to sample the program counter. The source files of the application must be
compiled to add profiling ISR and code (compiler switchpg). The linker attaches
to the profiling library to build the executable.
The server application is not adequate to profile the since most of the execution
time is devoted to wait for user messages from the serial port. A better approach is
to execute a new application which continuously encrypts and decrypts messages.
Therefore, create a new C++ project named profiling. The application file, named
profiling_app.cc, uses the class CAES128 to encrypt/decrypt messages in a loop.
Change the compiler switches of the target Release to enable the profiler and the
same optimization level as the server application (switches -pg -O2). Next, clean
the project in order to build the application from scratch.
The profiler collects data and stores it in memory during the execution. Once
the application completes, the collected data is downloaded to the PC in order to
analyze it. The SDK must set the profiler memory which cannot be overlapped to
the application memory. Use the SDK to program the FPGA with the imported
bitstream and BMM, and set the ELF to bootloop. Open the XMD console to
launch some commands. The first command changes the working directory to the
applications folder. The second command establishes the connection to the
MicroBlazes MDM. The last command tries to download the ELF file into
memory. It fails since the profiler memory is not defined, but it reports the allocated memory of the ELF. The hardware implements the local memory from the
address 0 9 0000 to 0 9 3FFF, and the XMD reported there is available free
memory from address 0 9 2248.
15.3
423
1
2
Fig. 15.34 Run configuration to profile
The profiling data can be stored into any free space of memory. Click the menu
Run ? Run Configurations to open the dialog depicted in Fig. 15.34. Then add
the application to profile and set the profile memory from the 0 9 3000 address.
Finally, run the application and wait until the application ends. The PC downloads
424
15
Fig. 15.35 Collected data from the application profiling
the collected data which is stored in the file gmon.out. It is a binary file which is
interpreted by the GNU gprof tool.
Double click on the file gmon.out to display the results (see Fig. 15.35). The
SDK shows a graphical view of the collected data which can be arranged in several
ways, as the percentage of execution time devoted to each function. Two methods
of the CAES128 class take the 88% of the processing time: X and Multiply. They
are child functions called from the function InvMixColumn which is one of the
steps executed during the decryption. The child function X is also called from the
step MixColumn during the encryption.
The profiling information can be used to re-implement the most time-spending
functions to improve the execution time. However, a coprocessor can greatly
accelerate a specific computation.
15.3.4 Coprocessor Design

The coprocessor aims to accelerate the step InvMixColumn and its child functions.
The step MixColumn is quite similar, therefore, the hardware design of both steps
does not significantly increase the effort. The coprocessor will accelerate the
computations that take the 92% of the processing time obtained from the profiler.
15.3
425
Fig. 15.36 Attachment of the MicroBlaze and coprocessor (top), and FSL schematic (bottom)
Copy the previous XPS folder and rename it as coprocessor_ip in order to

develop the coprocessor, its driver, and to update the hardware and software of the
application.
15.3.4.1 Hardware Design

Coprocessors are attached to MicroBlaze through Fast Simplex Link (FSLs), as
depicted in Fig. 15.36 (top). A MicroBlazes master-FSL connects to the coprocessors slave-FSL in order to write the coprocessors registers. In the reverse way,
the MicroBlazes slave-FSL permits the reading of the coprocessors registers.
The FSL [21] is a point-to-point link which permits a low latency and fast
communication due to its simple handshake. The FSL does not provide an address
bus, as seen in Fig. 15.36 (bottom). Therefore, the coprocessors registers must be
sequentially accessed in a predetermined order. Each FSL provides, by default, a
16-depth FIFO which can temporally store data written from the master, when the
slave is not ready to read it. The FSL provides a single 32-bit width bus
(FLS_M_Data, FSL_S_Data) which can carry data or commands, depending on the
control signal (FSL_M_Control, FSL_S_Control). In order to read data, the FIFO
signals when data is available to read (FSL_S_Exists), and the slave acknowledges
426
15
when data is retrieved (FSL_S_Read). Similarly, in order to write data, the FIFO
signals when there is no free space (FSL_M_Full), and the master requests to write
data (FSL_M_Write).
XPS can create the template files for coprocessors through the same wizard
used for peripherals. Launch the wizard by clicking the menu Hardware ? Create
or Import Peripheral:
(1)
(2)
(3)
(4)
Choose create templates in the XPS directory

Set the name to fsl_mixcolumns and the version to v1.00.a
Select FSL attachment and go ahead with the default settings
Select to implement the driver template and finish the wizard
The coprocessor does not require additional parameters, ports or VHDL files,
therefore, the MPD and PAO files are not modified. The template file
fsl_mixcolumns.vhd is modified to design the coprocessor. It implements a 4x4
matrix of 8-bit registers (reg_state) to compute and store the AES state.
Additionally, a 1-bit register (reg_mode) configures the computation mode as
MixColumns or InvMixColumns. MicroBlaze will execute a control-type write
instruction to the coprocessors slave FSL to set the mode register. Next, it
follows four data-type write instructions to set the registers of the AES state.
Matrices are stored in rows in the microprocessors memory, and each 32-bit
data sets the 4 registers of a row in the coprocessor. The coprocessor starts the
computation when the registers reg_state and reg_mode have been written. The
coprocessor must acknowledge the read of data to the slave-FSL in order to
delete it from the FIFO.
The coprocessor computes and stores the result in the registers reg_state. The
coprocessor writes the resulting data to its master-FSL in a quite similar way. It
writes a new row on the FSL when the computation is completed and the FIFO is
not full. MicroBlaze executes four read instructions to the FSL in order to retrieve
the resulting AES state.
A key idea to accelerate a computation is in parallel processing. A full-parallel
implementation could compute the entire array of the AES state in a single clock
cycle, although it may occupy a large area. However, MicroBlaze would not
completely take profit of this architecture since it takes several more clock cycles
to execute the FSL instructions to set and retrieve the AES state. A semi-parallel
architecture offers a good trade-off between speed and area. The coprocessor
computes the left-side column of the reg_state in a clock cycle. The state register
shifts one column and the computation is repeated for the next 3 columns.
Therefore, it takes 4 clock cycles to complete.
15.3
427
428
15
15.3.4.2 Driver Design

The XPS wizard generates the H and C templates for the coprocessors driver. The
driver provides a single function which sends the AES state and gets the result
through FSLs. It executes a control-type write instruction followed by four datatype instructions to set the computation mode and the AES state. The state is
transmitted in rows to avoid rearrangement of data from memory to the FSL. The
last instructions read the resulting state.
The function uses macros to set/get data to/from FSLs. During the BSP
generation, the compiler replaces them by the MicroBlaze instructions that read
or write an FSL slot. MicroBlaze provides 16 master-FSLs and 16 slave-FSLs.
The input slot is the FSL index from which the coprocessor reads the input data
(coprocessors slave-FSL). The output slot is the FSL index which connects to
the coprocessors master-FSL. The slot names must match with the name of the
coprocessors instance (fsl_mixcolumns_0) which will be declared in the MHS
file, otherwise the BSP will fail.
15.3
429
15.3.5 Modification of the Embedded System

The hardware must be modified to attach MicroBlaze to the coprocessor through
FSLs. The class CAES128 must accelerate the computation of the steps MixColumns
and InvMixColumns using the coprocessors driver.
15.3.5.1 Hardware Modification

The XPS offers a wizard (see Fig. 15.37) to connect a coprocessor when clicking
the menu Hardware ? Configure Coprocessor. It modifies the MHS file to
attach the coprocessors instance (fsl_mixcolumns_0) to MicroBlaze through two
FSLs.
430
15
15.3.5.2 Software Modification

Export the new hardware to SDK and set the workspace to the path
c:\edk13.1\coprocessor_ip\SDK\workspace before continuing. By default, the
MSS associates a generic driver to the coprocessor. The BSP generation with
the coprocessors driver is similar to the case of the peripheral. First, edit the file
libgen.options to add the local repository which contains the drivers source files.
Then, edit the MSS file to change the driver associated to the coprocessor.
Clean and build the BSP project to generate the new BSP from scratch. The
software applications use the C++ class CAES128 which implements the block
15.3
431
cipher. The member methods MixColumns and InvMixColumns are modified in

order that the coprocessor computes these steps. The class also provides conditional compilation to permit the computation by software for testing purposes.
The rest of the source files are not modified, since the class CAES128 carries
out the encryption/decryption of blocks. Build the C++ projects to generate the
new executables.
15.3.6 Simulation
The coprocessor can be simulated in a very similar way as described for the
peripheral. The application profiling is used to build the simulation model since it
continuously encrypts/decrypts blocks using the coprocessor.
Search when the coprocessor asserts the signal start which launches the computation, as shown in Fig. 15.38. The waveform previously shows the execution of
five FSL write instructions by MicroBlaze. The coprocessor reads the data from its
slave-FSL. The first FSL instruction is a control-type access which writes the mode
register (reg_mode). The next four FSL instructions are data-type accesses to write
the rows of the state register (reg_state). The coprocessor starts the computation
after the completion of the five write accesses.
Figure 15.39 shows that the coprocessor takes 4 clock cycles to compute the
state register. Then, it writes the four rows of the resulting reg_state to its masterFSL, in order the MicroBlaze can read them. The coprocessor does not have to
wait for the MicroBlaze to read the resulting data, since the FIFO of the FSL is not
full.
432
15
2
3
Fig. 15.37 EDK wizard to connect a coprocessor
Fig. 15.38 Simulation of the coprocessors slave FSL
15.3
433
Fig. 15.39 Simulation of the coprocessor computation and the master FSL
Fig. 15.40 Profiling data with the coprocessor
434
15
15.3.7 Experimental Results

Follow the steps previously described to collect the new profiling data. Before
programming the FPGA, check the directories used to get the BIT, BMM and ELF
files are in the current SDK workspace. Figure 15.40 shows the functions
MixColumns and InvMixColumns currently represent only the 6.14% of the execution time. The time in these functions is mainly devoted to the transmitting and
receiving of the state variable through the FSL instructions.
The application server can be tested as described before. The measured number
of clock cycles to encrypt and decrypt a block are now 5,189 and 5,150, respectively.
Therefore the decryption time is greatly improved from 1.76 ms to 103 ls (50 MHz
clock frequency), which represents about 17 times faster.
References
1. Nist (2002) NIST Advanced Encryption Standard (AES) FIPS PUB 197
2. Xilinx (2005) Spartan-3 Starter Kit Board User Guide (UG130)
3. Xilinx (2010a) XST User Guide for Virttex-4, Virtex-5, Spartan-3, and Newer CPLD
Devices (UG627)
4. Xilinx (2010b) LogiCORE IP Processor Local Bus PLB v4.6 (DS531)
5. Xilinx (2010c) XPS General Purpose Input/Output (GPIO) v2.00.a (DS569)
6. Xilinx (2010d) LogiCORE IP XPS Timer/Counter v1.02.a (DS573)
7. Xilinx (2010e) LogiCORE UART Lite v1.01.a (DS571)
8. Xilinx (2010f), LogiCORE IP XPS Interrupt Controller v2.01.a (DS572)
9. Xilinx (2010g) Xilinx Processor IP Library. Software Drivers. uartlite v2.00.a
10. Xilinx (2010h) Xilinx Processor IP Library. Software Drivers. tmrctr v2.03.a
11. Xilinx (2010i) Xilinx Processor IP Library. Software Drivers. intc v2.02.a
12. Xilinx (2010j) EDK Profiling Use Guide (UG448)
13. Xilinx (2011a) Embedded System Tools Reference Manual (UG111)
14. Xilinx (2011b) Platform Specification Format Reference Manual (UG642)
15. Xilinx (2011c) Command Line Tools User Guide (UG628)
16. Xilinx (2011d) MicroBlaze Processor Reference Guide (UG081)
17. Xilinx (2011e) LogicCORE Local Memory Bus LMB v10 (DS445)
18. Xilinx (2011f) Standalone (v.3.0.1.a) (UG647)
19. Xilinx (2011g) ISim User Guide v13.1 (UG660)
20. Xilinx (2011h) Data2MEM User Guide (UG658)
21. Xilinx (2011i) LogiCORE IP Fast Simplex Link (FSL) V20 Bus v2.11d (DS449)
22. Menezes AJ, Oorschot PC, Vanstone SA (1996) Handbook of applied cryptography. CRC
Press, Boca Raton
Chapter 16
Partial Reconfiguration on Xilinx FPGAs
Partial Reconfiguration (PR) is the ability to change a portion (the reconfigurable

partition) of the device without disturbing the normal operation of the rest
(the static partition). A typical PR application is a reconfigurable coprocessor
which switches the configuration of the reconfigurable partition at run-time when
required by the application. The main advantage is the ability to map different
coprocessor configurations in the reconfigurable partition in a time-multiplexed
way, reducing the required area. The main drawbacks are the storage of partial
bitstreams, the reconfiguration time and the increased complexity of the design
flow. Reconfigurable systems must be implemented in a reconfigurable device,
such as FPGAs. Most FPGAs are not partially reconfigurable since the internal
architecture of the device must provide features to support it as well as the design
flow tools. Moreover, the steps required to implement a reconfigurable system on a
FPGA is highly dependent of the device architecture and cannot be easily adapted
to other FPGA vendors. Virtex-4/5/6 are the Xilinx families of devices that are
supported by the ISE 13.1 to perform partial reconfiguration [1].
16.1 Partial Reconfiguration on Xilinx FPGAs

Xilinx FPGAs that supports partial reconfiguration feature a set of common features at the hardware level:
ICAP (Internal Configuration Access Port). It is an internal version of the SelectMAP programming interface. It permits to write or read-back the FPGA
configuration from an embedded microprocessor or other internal circuit. The
ICAP is controlled through commands that are embedded in the bitstream file.
The configuration granularity is the minimum set of bit required to update FPGA
resources which is a frame in Virtex-4/5/6 devices.

435
436
16 Partial Reconfiguration on Xilinx FPGAs
Reconfigurable
routing line
Proxy LUT
Static
routing line
Partial
Reconfiguration
Reconfigurable
routing line
Proxy LUT
Reconfigurable
Partition
Static
Partition
Static
routingline
Fig. 16.1 FPGA layout displaying the proxy logic
Configuration frame. The configuration memory of the FPGA is arranged in

frames that are addressable memory spaces. Each frame is composed of a fixed
number of 32-bit configuration words. There are several types of frames, such as
the CLB or BRAM. A Virtex-5 frame [2] configures a single row of 20 CLBs,
4 BRAMs, etc.
Partial bitstream. Contains the ICAP commands and configuration frames
required to reconfigure a portion of the FPGA hardware resources. The configuration bits of the rest resources of the FPGA are not affected.
Glicthless reconfiguration. It guaranties that if a configuration bit has the same
value before and after the reconfiguration, the hardware resource controlled by
the bit does not undergo any discontinuities in its operation. Therefore a portion
of the static partition may be placed in the reconfigurable area and not be
affected by the reconfiguration.
At the design flow level, the PR terminology is:
HWICAP. It is an EDK peripheral which contains the ICAP plus FIFO memories and the PLB interface, in order to attach it to an embedded system.
Static partition or top-level logic. It is composed of an embedded processor
which drives the HWICAP with a partial bitstream. It cannot be affected during
the partial reconfiguration, in order to complete it.
Reconfigurable Partition (RP). It is composed of a subset of FPGA resources
that can be reconfigured at run-time from the static partition. A RP implements
several reconfigurable modules in a time-multiplexed way.
Reconfigurable Module (RM). It is an implemented module in the reconfigurable partition.
Partition pin. It connects a 1-bit net across the static and the dynamic partitions.
The proxy logic transfers the bit value of the partition pin.
Proxy logic. It is the interface logic between the static and the dynamic partitions. Each partition pin devotes a LUT placed in the RP which connects a
reconfigurable routing line to a static routing line (see Fig. 16.1). The LUT and
the reconfigurable routing line must be entirely contained in the RP. The
reconfiguration does not affect the static routing line, therefore, the communication is reestablished when the partial reconfiguration is completed.
16.1
Partial Reconfiguration on Xilinx FPGAs
437
Decoupling logic. Hardware resources that are changing their configuration bits
can drive unexpected transient values on crossing nets that may cause a malfunction in the static partition. The static partition can deassert the nets from the
proxy logic during the reconfiguration. Another approach is to reset the affected
circuits of the static partition when the reconfiguration completes.
16.2 Design Flow for Partial Reconfiguration

The Project Navigator tool, which is frequently used to implement designs on
Xilinx FPGA, does not support PR. The PR design flow is based on a technique
named modular design which can be run by using the command-line ISE tools as
well in the PlanAhead graphical tool. The modular design flow starts from a set of
module netlists to implement (map, place and route) the layout of modules separately, merging them to build the complete design layout. A previous phase to
modular design is to synthesize HDL sources to get the set of module netlists,
where the reconfigurable modules have to meet a set of requirements in order to
build a successful design.
16.3 Case Study

The case study exposes the design and PR flow of a self-reconfigurable embedded
system. It summarizes some steps related to the design flow of the embedded
system since they are detailed in the previous chapter. The case study is developed
for the AVNET Virtex-5 LX Evaluation Kit [3, 4], but it can be easily adapted to
other Virtex-4/5/6 boards.
Figure 16.2 depicts the embedded MicroBlaze which attaches to a reconfigurable coprocessor through FSLs. The coprocessor can map several configurations
sharing a devoted set of FPGA resources in a time-multiplexed way. The MicroBlaze switches the coprocessors configuration at run-time when required by
the application. It reads a partial bitstream to control the HWICAP accordingly to
perform the partial reconfiguration.
An important issue is the size of partial bitstreams which is usually larger than
the internal BRAM capacity. Therefore, the system architecture must retrieve
bitstreams from an external host or a non-volatile memory. An external host could
store and transmit bitstreams when requested. The approach is slow and it depends
on the communication channel. A better approach is to store the set of bitstreams
in a non-volatile memory since many prototyping boards provide some FLASH
which can be used for the required purpose. The case study provides autonomous
reconfiguration since it reads the partial bitstreams from a FLASH memory to
reconfigure the coprocessor.
438
External
oscillator
100 MHz
Reset
button
system
clock
reset
RP
DLMB
Controller
FSL mb_to_copro
DLMB
Reconfigurable
Coprocessor
MicroBlaze
BRAM
ILMB
Controller
FSL copro_to_mb
ILMB
RM
decoupling_rst
PLB
HWICAP
rs232
UART
MDM
flash_emc
MCH_EMC
Serial
communication
JTAG
programmer
FLASH
Memory
decoupling
GPIO
RM dummy
RM adder
RM xxx
Fig. 16.2 Overview of the system architecture
The reconfigurable coprocessor is defined as the reconfigurable partition (RP)

since it can allocate different reconfigurable modules (RM). The static partition is
composed of the rest of the system. Each RM performs a different computation on
3 9 3 matrices: addition, multiplication, scalar multiplication and the determinant.
There is also a dummy RM which does not perform any computation, but it is used
in the design flow, as explained later.
The followed steps are the hardware design on XPS and synthesis of the RMs,
the PR flow and the software development on SDK. The steps related to XPS and
SDK are similar to the cases of the previous chapter, but they meet a set of
requirements to permit the partial reconfiguration.
1. The hardware design starts with XPS to synthesize the embedded system in
order to get the set of netlist files. The next step synthesizes the different
configurations of the coprocessor to get the netlist files of the RMs.
2. PlanAhead runs the PR flow which starts declaring the reconfigurable partition
and modules from the set of netlist files. The PR flow computes the layout of
16.3
Case Study
439
the static partition and the RMs on the reconfigurable partition. The output is
the set of partial bitstreams required to perform the partial reconfiguration.
3. The software development on SDK. The partial bitstreams are programmed into
a FLASH memory. The application provides a C++ class which performs the
partial reconfiguration controlling the HWICAP according to the bitstreams
retrieved from the FLASH.
16.3.1 Hardware Design

The PR flow will start from a set of netlist files that are obtained from the synthesis of
the hardware design of the embedded system and the RMs. The XPS facilitates the
design and synthesis of the embedded system. The coprocessor will map different
circuit configurations, therefore, it is also necessary to synthesize the RMs.
16.3.1.1 Design of the Embedded System on XPS

The architecture of the embedded system includes a MicroBlaze microprocessor
attached to internal BRAMs through the data and instruction LMBs and memory
controllers. The PLB attaches MicroBlaze to the peripherals: RS232 serial interface, the MDM and the HWICAP. The system includes the FSL reconfigurable
coprocessor and the associated decoupling logic. The set of partial bitstreams will
be retrieved from an external FLASH memory, therefore, the system includes an
EMC (External Memory Controller) peripheral in order to permit the MicroBlaze
reading them to perform the reconfiguration.
In order to design the hardware of the embedded system, create a new XPS
project with the BSB wizard, as described in the previous chapter.
1. Choose the folder c:\edk13.1\reconfig\XPS to store the project file system.xmp.
2. Select the PLB, since the EDK does not support AXI for Virtex-5 devices.
3. Create a new design and select the xc5vlx50-ff676-1 device which is embedded
in the Virtex-5 prototyping board.
4. Select a single processor and continue to the end with the default options.
Edit the MHS file to change the polarity of reset button and the frequencies of
the reference clock and the systems clock to the appropriate values of the prototyping board. Set the frequency of the systems clock to 100 MHz as it will be
detailed later. The name of the systems clock is changed to sys_clk, therefore, it is
necessary that the clock inputs reference to the new name, including the LMBs and
the PLB. Finally, change the parameters C_HIGHADDR of the instruction and
data BRAM controllers to increase the processors local memory to 64 KB.
440
The system declares the instance rcopro which is the reconfigurable coprocessor. The source files of the coprocessor fsl_rcopro are stored in the default local
repository pcores created by the XPS wizard. The coprocessors instance attaches
to MicroBlaze through FSLs. There is a decoupling logic which can reset the FSLs
and the coprocessor. MicroBlaze will control a GPIO (instance decoupling_gpio)
in order to assert the 1-bit signal decoupling_rst when the reconfiguration is
completed. If not, the FIFOs of the FSLs would store unexpected data, and the
state machine of the coprocessor would start in an undesired state.
16.3
Case Study
441
The reconfigurable coprocessor can map several configurations. The parameter

C_CONFIG_IDX in the MHS file specifies the initial configuration which is
synthesized and implemented during the EDK flow. The value 0 sets the module
dummy as the initial configuration of the coprocessor.
MicroBlaze will control the reconfiguration through the HWICAP peripheral
[5]. The peripheral embeds the ICAP which is connected to read/write FIFOs and a
slave PLB interface. The maximum frequency of the ICAPs input clock is
100 MHz, therefore, the systems clock accomplishes this constraint. The system
442
will provide several clock domains if the ICAP sets a different clock frequency to
the systems clock; however, it will complicate the design.
MicroBlaze must read a partial bitstream from the external FLASH to feed the
ICAP, in order to reconfigure the coprocessor. EDK provides an external memory
controller (EMC) [6] which permits access to asynchronous memories such as
SRAM and FLASH. It can control several banks of different memories that share
the same input/output ports. Each bank is accessible from a configurable address
range and provides configuration parameters to set the width of the data bus and
the timing specifications. The EMC permits access to 8-bit/16-bit data width
memories, and it can match them to the 32-bit data bus of the PLB. The displayed
configuration is valid for the 28F128J3D [7] device, the 16 MB FLASH embedded
in the Virtex-5 development board [3]. The MicroBlaze can access to the FLASH
from the address 0xA0000000 to the 0xA0FFFFFF. The configuration can be
complicated but, the board designer usually provides a reference design which can
be used to get the parameters. The input/output ports of the EMC are externally
connected to the address, data and control busses of the FLASH. The width of
address bus of the FLASH is 24-bit, but the EMC generates 32-bit addresses.
Therefore, the MHS slices the 24 lower bits that are externally connected.
16.3
Case Study
443
The system must be synthesized in order to generate the netlist files required by the
PR flow. The system netlist contains the module dummy as the initial configuration of
the coprocessor. The module dummy just reads the slave FSL and writes the master
FSL with zeros, but it does not perform any computation. This way, MicroBlaze will
not hang up when it executes any read/write instructions to the coprocessors FSLs.
XPS can continue to implement the FPGA layout, generating the bitstream system.bit. Previously, the UCF file must be modified to map the external ports
depending on the development board. The implemented system is not valid to perform the PR but it can be exported to SDK in order to build the software. Moreover,
the system provides an initial configuration for the coprocessor which can be tested
and simulated, as described in the previous chapter. Finally, the system is able to
program the set of partial bitstreams into the FLASH, as is detailed later.
16.3.1.2 Synthesis of the RMs
The coprocessor will map several reconfigurable modules. All of them are connected to MicroBlaze through the same FSL ports. The master and the slave FSL
connections are quite similar to the case study exposed in the previous chapter.
The VHDL source files that contain the different configurations of the coprocessor
must be synthesized to obtain their netlist files. There are several ways, as with the
Xilinx Project Navigator or with a third-party synthesizer.
This case study executes the XST synthesizer [8] from a command shell. The
directory c:\edk13.1\reconfig\XST allocates the files required by XST. Open a
Xilinx console clicking the XPS menu Project?Launch Xilinx Shell, and type the
next commands:
444
The batch file fsl_rcopro.scr contains the XST commands to synthesize a set of
source VHDL files. The NGC file stores the netlist of a single module for the target
device.
The batch file relies on the file fsl_rcopro.prj which lists the files that are
synthesized, as in the PAO file. The VHDL file fsl_rcopro.vhd declares the toplevel entity of the coprocessor, and it contains the instance of the RM which is
synthesized. The rest of the VHDL files (fsl_rcopro_dummy.vhd, fsl_rcopro_adder.vhd, etc.) are the source files of the different RMs.
The XST must be executed for every RM in order to get the set of netlist files.
Change the parameter C_CONFIG_IDX and the name of the NGC output file to
synthesize a new RM. The XST folder will store the netlist files such as
fsl_rcopro_dummy.ngc, fsl_rcopro_adder.ngc, and so on.
16.3
Case Study
445
16.3.2 Partial Reconfiguration Flow on PlanAhead

PlanAhead is a complex tool to perform the design floorplanning on Xilinx
FPGAs, in order to improve clock frequency and delays, the placement of IOBs,
and more. PlanAhead can run in a PR flow mode in order to facilitate the
implementation of RMs on a reconfigurable partition. The PR flow requires an
additional license feature. The main steps to follow are:
1. Set the PR flow and the design files. The flow starts from the top-level netlist
file.
2. Set the reconfigurable partition and assign it the netlist of the RMs.
3. First layout implementation. It computes the FPGA layout of the static and the
reconfigurable partition with the initial RM. The static partition is promoted to
be imported during the next step.
4. Layout implementation for the rest of RMs. The flow continues importing the
promoted static partition, but implementing a new layout in the reconfigurable
partition for each new RM. The result is a set of FPGA layouts, where the
layout of the static partition remains invariable.
5. Bitstream generation. The last step is the bitstream generation phase to compute
the complete and partial bitstream of each implemented layout.
16.3.2.1 Set the PR Flow and the Design Files

Launch the PlanAhead and click the menu Create a New Project to open the
wizard (see Fig. 16.3).
1. Select the root folder c:\edk13.1\reconfig, and set the project name to
PlanAhead.
2. Choose the options Specify synthesized (EDIF or NGC) netlist and Set PR
Project. The PR license must be correctly installed in order to set the PR
project.
3. Select the top-level netlist file system.ngc which is stored in the folder
c:\edk13.1\reconfig\XPS\implementation. It is the synthesized embedded system designed in XPS. It relies on the rest of the netlist files, therefore, add the
same path to the Netlist directories. Do not select the copy option.
4. Continue with the default options to complete the wizard.
The BMM file system.bmm is required during the implementation of embedded
systems, in order to generate the initial bitstream which configures the FPGA and
the BRAM contents, as commented in the previous chapter. The BMM file is
stored in the same location as the set of netlist files. Therefore, copy the file
system.bmm to the folder c:\edk13.1\reconfig\PlanAhead which contains the
PlanAhead project.
446
1
3
2
5
6
6
7
Fig. 16.3 Wizard to create a new PlanAhead PR project
16.3.2.2 Set the Reconfigurable Partition and the Reconfigurable Modules

Click the menu Flow?Netlist Design to display the list of netlist files which
composes the systems hardware. The tab Netlist displays the instances of the
embedded system. In order to set the reconfigurable partition (see Fig. 16.4):
1. Select the instance rcopro and open the contextual menu (right button of the
mouse) and select Set Partition.
2. Select the option is a reconfigurable Partition, and change the name of the
reconfigurable module to module_dummy since it is the initial configuration
stored in the netlist files.
The tab Netlist now shows the module_dummy contained in the instance rcopro.
In order to add a new RM to the RP (see Fig. 16.5):
1. Open the contextual menu of the rcopro and click Add Reconfigurable Module
2. Change the RM name to module_adder.
3. Select the netlist file fsl_rcopro_adder.ngc which is stored in the XST directory. Continue with the default settings to end the wizard.
The previous steps are repeated for every RM. PlanAhead displays all the RM
modules in the tab Netlist. Finally, open the contextual menu of the configuration
module_dummy to select the option Set as Active Reconfigurable Module.
The next step defines the FPGA area devoted as the reconfigurable partition.
The RP must accomplish a set of design rules [1].
16.3
Case Study
447
Fig. 16.4 Set the RP and the initial RM to the coprocessors instance
2
4
1
Fig. 16.5 Adding a RM to the RP
The global clock logic must reside in the static partition

Some types of hardware resources (global buffers, boundary-scan, etc.) should
be placed in the static partition.
Bidirectional partition pins are not permitted.
In order to obtain a greater efficiency, the RP should align to the boundaries of
the configuration frames.
448
3
1
Fig. 16.6 Assigning the area of the RP
Select the instance rcopro and click the button Set Pblock Size in the device
view, as shown in Fig. 16.6. The number of hardware resources in the RP must be
higher than required by any of the synthesized RMs. This case study selects a
region which starts from the top-left edge of the FPGA. Set the partition height to
20 CLBs (a Virtex-5 configuration frame) and the width to 28 CLB columns.
The RP occupies the whole clock region of the FPGA which contains 560 SLICES,
4 RAMB36 and 8 DSP48 resources. Select all the resources in the next window to
permit them to be reconfigured, with the exception of the RAMB36 since no RM
requires them. The partial bitstream will reconfigure the hardware resources
contained in the RP, with the exception of the BRAM contents. This way, the size
of the partial bitstreams is reduced and the reconfiguration time improves.
Figure 16.7 displays the tab Statistics in the Properties of the Pblock
pblock_rcopro. It shows the size of a partial bitstream and the detailed number of
reconfigurable hardware resources. It also reports that the 100 % of the RP is contained in a single clock region, as desired. The tab Attributes shows the range of
resources as SLICE_X0Y100:SLICE_X27Y119 and DSP48_X0Y40:DSP48_X0Y47.
The numbers denote the xy coordinates starting from the bottom-left edge of the
FPGA. Therefore, the SLICES are arranged in a 28 9 20 matrix and the DSP48 are in
a single column. These parameters are annotated in a new UCF in order to implement
the layout of the reconfigurable system.
It is recommended to check the design rules for the PR flow before continuing.
Press Run DRC and unselect all rules except the related to the partial reconfiguration, as shown in Fig. 16.8. It should only report a warning for each RM
since it is not still implemented by a configuration.
16.3
Case Study
449
4
2
3
Fig. 16.7 Properties of the RP
Fig. 16.8 Design rule check for

the PR
16.3.2.3 First Layout Implementation

The tab Design Runs shows the settings to run the implementations (see Fig. 16.9).
The default run created by PlanAhead will implement the layouts of the static
partition and the dummy module into the reconfigurable partition. It creates a batch
file which calls the ISE tools starting from the annotated constraints and netlist
files.
1. Select the default run config_1 created by PlanAhead
2. Display the properties of the partitions and check the coprocessors RP will
implement the RM module_dummy.
450
12
5
4
3
6
10
8
9
11
Fig. 16.9 Configuring the implementation run for the first layout
3. Display the implementation options to add bm..\..\system.bmm to the NGDBuild tool. This option will load the BMM file which was copied from the XPS
project, in order to generate the annotated BMM file which will be required
later by the Data2MEM tool [9].
4. Display the general properties to change the run name to config_dummy.
5. Start the implementation run by pressing the button Launch Selected Runs.
The implementation of the layout takes some minutes to complete. PlanAhead
opens a dialog window when it completes. Promote the implemented partitions
since it is necessary in the next step.
16.3.2.4 Layout Implementation for the Next RMs
The next step implements the layouts for the rest of the reconfigurable modules. The
PR flow starts importing the layout of the static partition from the first implementation. Then, it computes the layout of a new RM implemented into the RP.
16.3
Case Study
451
1
2
4
5
8
Fig. 16.10 Configuring a new implementation run for the adder RM
1. Click the button Create New Runs to configure a new implementation

(see Fig. 16.10). Continue with the default netlist, constraints and device since
they are the same as the first implementation.
2. Change the default name to config_adder. Configure the RP rcopro to implement the RM module_adder.
3. Add the BMM file in the same way as exposed in the first implementation run.
4. Press the button More to add the new runs to implement the rest of the RMs and
repeat the previous two steps (but changing the name and the RM).
5. Finalize the wizard, but do not launch the run.
Select the new created runs to launch them. PlanAhead can take profit of a
multi-core PC incrementing the number of jobs to compute several layouts in
parallel. Each run imports the layout of the static partition and computes a new
layout in the reconfigurable partition. After finishing the implementation, do not
perform any action when asked by PlanAhead.
The ISE provides the FPGA Editor tool which displays an implemented layout.
Open the file config_dummy_routed.ncd which is stored in the folder c:\edk13.1\
reconfig\PlanAhead\PlanAhead.runs\config_dummy to display the FPGA first
layout. In a similar way, open another implemented layout and check that the
layouts are only different in the reconfigurable partition (see Fig. 16.11).
452

Reconfigurable
partition
config_dummy
config_adder
config_xxx
Fig. 16.11 The first implemented layout (left) and two other layouts for other RMs
16.3.2.5 Bitstreams Generation

It is the final step in the PR flow. Select all the implementation runs and open the
contextual menu to generate the bitstreams with the default options. It launches the
BitGen tool [10] from the implemented layouts. Each layout generates two bitstream
BIT files: the total and the partial. The total bitstream configures all the hardware
resources of the FPGA to program the reconfigurable embedded system. The partial
bitstream only configures the resources of the reconfigurable partition to program
the coprocessor. The BitGen also creates the file system_bd.bmm which annotates
the source BMM file with the placement of the BRAM blocks. The annotated BMM
file is required later to program the FPGA with an executable ELF file.
A change in the netlist files or in the PR settings requires updating the bitstream
files. PlanAhead permits to reset the implemented runs in order to update the
implemented layouts and bitstreams.
16.3.3 Software Development

Once the system hardware has been exported from XPS, the SDK can start, as
detailed in the previous chapter. Create a new SDK workspace in the folder
c:\edk13.1\reconfig\XPS\SDK\workspace. The imported bitstream file system.bit
16.3
Case Study
453
from XPS can program the FPGA with a static embedded system which cannot be
reconfigured since it does not provide a reconfigurable partition. However, SDK
can build the BSP and the ELF file for the imported system. Moreover, this static
system can be tested, debugged and simulated, with the initial configuration for the
coprocessor, in a similar way as in the previous chapter. The static system can be
used also to program the external FLASH with the partial bitstreams. The reconfigurable system will read the partial bitstreams to program the coprocessor.
16.3.3.1 Flashing the Partial Bitstreams

The folder bitstream contains a simple batch file bitstream.bat which copies the
required bitstream files from PlanAhead. Open a command shell by clicking the
SDK menu Xilinx Tools?Launch Shell, in order to execute the batch file.
The total bitstream of the reconfigurable system is copied and renamed to

system_dummy.bit since the module dummy is the initial coprocessor configuration.
The partial bitstreams are named partial_dummy.bit, partial_adder.bit and so on.
The file sizes of the partial bitstream are larger than those reported by PlanAhead
since they include header data and the ICAP commands [2] to perform the partial
reconfiguration. The batch also copies the BMM file system_bd.bmm which will be
necessary later.
The NOR FLASH 28F128J3D [7] provides 16 MB of non-volatile memory
arranged in 128 KB blocks. It provides the standard CFI (Common Flash Interface) which permits to erase and program the FLASH blocks. The application
devotes the last 1 MB (8 blocks) to storing the partial bitstreams and each single
block can allocate a partial bitstream file.
The easiest way to program the FLASH memory is from SDK. First, the FPGA
must map an embedded system which provides a memory controller able to
successfully access the external FLASH. Then, the SDK can load an executable
into the embedded system which downloads a data file and executes the CFI
queries to program the FLASH.
Therefore, program the FPGA with the bitstream system.bit imported from
XPS, and then click the menu Xilinx Tools?Program Flash to open the dialog
window shown in Fig. 16.12.
1. Set the file partial_dummy.bit which will be programmed into the FLASH.
2. Check that the EMC and bank are related to the external FLASH, and that the
base address is correct (0xA0000000).
3. Set the address offset to 15 MB (0xF00000). The file is programmed in the
resulting address of the addition the base address and the offset (0xA0F00000).
4. Program the FLASH. It will take some seconds to complete.
454
2
3
Fig. 16.12 Programming a partial bitstream into the FLASH
Repeat the process to program the rest of partial bitstreams into the next
FLASH blocks (0xA0F200000, 0xA0F400000, etc.). Therefore, the offset must be
incremented 128 KB (0x20000) to point to the next block.
16.3.3.2 BSP Generation
The BSP generation is similar to the study cases of the previous chapter. The SDK
provides a wizard to create the BSP project. The MSS file is modified to set the
specific driver to the reconfigurable coprocessor in order to build the BSP with it.
Moreover, the file libgen.options must add the path of the local repository
which contains the source files. They are stored in the default folder drivers
allocated in the XPS root directory.
16.3
Case Study
455
The coprocessor driver is composed of a different function for each module to

compute 3 9 3 matrices. They send data and retrieve results through the FSL slots,
in the same way as described in the previous chapter. The driver function for the
matrix addition and multiplication sends two matrices and it retrieves the resulting
matrix. The scalar multiplication function sends an integer and a matrix and
retrieves the computed matrix. The determinant function sends a matrix and it
returns the integer result. Finally, the driver for the dummy module sends any data
and it returns the data (should be zero) retrieved from the module.
16.3.3.3 Executable
Create a new C++ project named app1 to build the executable ELF file for the
generated BSP. The source file app1.cc contains the main application functions,
and it relies on the C++ class CBitStream. The class implements methods to read
the bitstream headers from FLASH and to perform the partial reconfiguration. The
application devotes an array of objects of the previous class which is initialized at
the beginning of the application. Three parameters define the settings used to
initialize the array reading the bitstream headers. Then, the application continues
communicating with the user through the serial port in order to load the desired
RM. It also performs a test of the active reconfigurable module which compares
the resulting data from the coprocessor and the computed data from MicroBlaze.
As commented previously, Microblaze must command the GPIO to reset the FSLs
and the reconfigurable coprocessor after the partial reconfiguration completes.
456
The class CBitStream contains the method ReadHeader to initialize its member
data from a bitstream file. The member data will take invalid values if the header
reading fails. The BIT files stored in the FLASH are composed of a header and raw
data. The header contains several fields [11] that provide information about the
16.3
Case Study
457
layout: target FPGA, date, time and name of the NCD file. The last field contains
the address and the size (in bytes) of the raw data.
The method Reconfig performs the partial reconfiguration after checking the
data collected from the header. The raw data of a bitstream embeds the set of ICAP
commands [2] that permit the partial reconfiguration of the FPGA. Therefore, the
task performed by the C++ class during the reconfiguration is reduced to write into
the HWICAP the 32-bit words retrieved from the raw data.
The HWICAP peripheral [5] embeds the ICAP and provides two FIFOs and a
set of registers to perform the partial reconfiguration or to read-back the configuration memory. The example uses three peripherals registers to perform the
partial reconfiguration: control, status, and write FIFO registers. MicroBlaze
accesses to the registers through C/C++ pointers as commented in the previous
chapter.
The C++ class implements the method Hwicap_WordWrite to write a configuration word to the HWICAP. First, the word is temporally stored in the write
458
FIFO. Then, it sets the control register to start the data transfer from the FIFO to
the ICAP. The status register indicates whether the write to the HWICAP was
completed, aborted or got an error. The method returns unsuccessfully in the last
two cases. The method waits until the write of the configuration word completes
before returning successfully.
The method Hwicap_InitWrite aborts and restarts the HWICAP before starting,
in order to assure that the peripheral is empty of data from a previous
reconfiguration.
The application would improve using an ISR associated to the HWICAP to

avoid waiting to the completion of the write. The ISR would read a new configuration word when the HWICAP is ready to get a new one. Moreover, the ISR
would write the FIFO with several configuration words before starting the data
transfer to the ICAP. They were not implemented since they would obfuscate the
basics about the partial reconfiguration.
16.3
Case Study
459
1
2
3
Fig. 16.13 Testing the reconfigurable system
16.3.3.4 Testing
In order to test the reconfigurable embedded system, the FPGA must be programmed with the bitstream generated by PlanAhead and the executable file.
Therefore, select the bitstream system_dummy.bit and the BMM system_bd.bmm
that were imported from PlanAhead (see Fig. 16.13). Then, select the executable
ELF file app1.elf. SDK calls the Data2MEM [9] tool to generate the bitstream file
download.bit which configures the FPGA and the BRAM contents.
The application can also be debugged in the reconfigurable system as described
in the previous chapter.
References
1.
2.
3.
4.
5.
6.
7.
8.
Xilinx (2011a) Partial reconfiguration users guide (UG702)

Xilinx (2010a) Virtex-5 FPGA configuration user guide (UG191)
Avnet (2009) Virtex-5 LX evaluation kit. User guide
Xilinx (2010b) Virtex-5 FPGA user guide (UG190)
Xilinx (2011b) LogiCORE IP XPS HWICAP (DS586)
Xilinx (2010c) LogiCORE IP XPS multi-channel external memory controller (DS575)
Numonyx (2007) NumonyxTM embedded flash memory J3 v.D datasheet
Xilinx (2010d) XST user guide for virtex-4, virtex-5, spartan-3 and newer CPLD devices
(UG627)
9. Xilinx (2011d) Data2MEM user guide (UG658)
10. Xilinx (2011c) Command line tools user guide (UG628)
11. PldTool (2010) Xilinx BIT bitstream files. http://www.pldtool.com/pdf/fmt_xilinxbit.pdf.
Accessed July 2011
About the Authors
Jean-Pierre Deschamps received his MS degree in electrical engineering from

the University of Louvain, Belgium, in 1967, his PhD in computer science from
the Autonomous University of Barcelona, Spain, in 1983, and his second PhD
degree in electrical engineering from the Polytechnic School of Lausanne,
Switzerland, in 1984. He has worked in several companies and universities. He is
currently a professor at the University Rovira i Virgili, Tarragona, Spain. His
research interests include ASIC and FPGA design, digital arithmetic and
cryptography. He is the author of nine books and of more than a hundred
international papers.
Gustavo Sutter received his MS degree in Computer Science from State
University UNCPBA of Tandil (Buenos Aires) Argentina, and his PhD degree
from the Autonomous University of Madrid, Spain. He has been a professor at the
UNCPBA, Argentina and is, currently, a professor at the Autonomous University
of Madrid, Spain. His research interests include ASIC and FPGA design, digital
arithmetic, and development of embedded systems. He is the author of three books
and about fifty international papers and communications.
Enrique Cant received his MS degree in electronic engineering (1995) and his
PhD in electronic engineering (2001), both from the Polytechnic University of
Barcelona, Spain. He has been a professor at the Polytechnic University of
Barcelona, Spain and is, currently, a professor at the University Rovira i Virgili,
Tarragona, Spain. His research interests include ASIC and FPGA design,
development of embedded systems, and dynamic reconfiguration of
programmable devices. He is the author of about fifty international papers and
communications.

Lecture Notes in Electrical Engineering 149, DOI: 10.1007/978-94-007-2987-2,
461
Index
100 s complement. ten0 s complement

90 s complement, 283
A
Activity interval, 41
Adder
base-B carry-chain, 278
decimal carry-chain, 279
decimal ripple-carry, 277
Adders, 153
addition array, 166
addition tree, 167, 180
binary, 155, 175
Brent-Kung, 160
B0 s complement, 174
carry select, 158
carry-lookahead, 160
carry-save, 165, 179
carry-select, 173
carry-skip, 158
combinational multioperand, 167
delay, 155, 165167, 172
FPGA implementation, 175
Ladner-Fisher, 160
Ling, 160
Logarithmic, 160
long operands, 162, 178
overflow, 174
radix-2k, 155, 158, 175
sequential multioperand, 163
stored-carry encoding, 165
Addersubtractors, 173
Addition
natural numbers, 153

Address decoders, 5
Address decoding
complete, 364
incomplete, 364
AES-128, 417
ECB, 419
padding, 419
state, 417
Alignment, 322
Alignment circuit, 324
ALU, 359, 365
API, 387
Arithmetic operations, 8
Array
carry-save, 167
multioperand addition, 166
arrival time, 139
ASIC, 96
asynchronous FIFO, 119
B
BCD (binary encoded digits). Decimal
BIT, 452
download.bit, 409, 459
partial_adder.bit, 453
partial_dummy.bit, 453
system.bit, 386, 453
system_dummy.bit, 459
Bit extension, 175
BitGen, 386, 452
BMM, 445
system.bmm, 445
system_bd.bmm, 387, 453, 459

Lecture Notes in Electrical Engineering 149, DOI: 10.1007/978-94-007-2987-2,
463
464
B (cont.)
Boolean equations, 1
BRAM, 366, 374, 437, 439
BSB, 376
BSP, 373, 384, 388, 454, 455
Bus, 368
master, 368
slave, 368
Bus-holder, 98
Bus-keeper, 98
C
Cache, 361
Carry computation, 155
Carry logic, 9
Carry-chain adder, 278, 279
Carry-chain adder-subtractor, 283
Carry-save adder, 34
Carry-save tree, 170
CISC, 359
Clock
distribution network, 108
gating, 113
jitter, 113
manager, 113
skew, 110
tree, 108
CMOS circuit, 95
Coloring
graph, 35
Command decoder, 87
Computation primitive, 37
Computation resource, 31
programmable, 27
Computation scheme, 37
Computation width, 41
Conditional assignments, 5, 7
Configuration frame, 436
Configuration granularity, 435
Connections
programmable, 31
Connectivity, 77
maximum, 77
minimum, 77, 79
Constrain
area, 113
timing, 113
Control register, 364
Control unit, 31
Control units
command encoding, 83
hierarchical decomposition, 83, 86
variable latency, 91
Index
Controllability, 138
Conversion
binary to radix-B, 251
radix-B to binary, 252
Converters
binary to decimal, 268
binary to radix-B, 252
decimal to binary, 253, 272
Coprocessor, 365, 425
driver, 429
hardware, 426
CORDIC, 267
Counter
3-to-2, 34
7-to-3, 34
Counters, 11, 167
3-to-2, 167
6-to-3, 171, 179
m-to-2, 170
parallel, 170
up/down, 12
Counters
Critical path, 139
Cryptoprocessor, 365
Cygwin, 421
D
Data bus, 7, 8
Data flow graph, 37
Data introduction interval, 58
Data path, 25, 29, 31
Data path connectivity, 77
Data register, 364
Data2MEM, 394, 450, 459
Data-flow model
DCM, 379
DDR, 367
Decimal
addition, 277
addition-subtration, 282
division, 286
multiplication, 286
non-restoring, 291
SRT-like algorithm, 295
Decoupling logic, 437, 440
Delay
clock-to-output delay, 104
derating factors, 101
fall time, 99
intrinsic delays, 99
propagation delay, 98
rise time, 99
slew rate, 99
Index
Delay Locked Loop (DLL), 113
Demultiplexersp, 5
Derating factors, 135
Digit extension, 174
Digit serial implementations
divider, 75
Digit serial processing, 72
Digital Clock Managers (DCM), 113
Dividers
binary SRT, 229, 232, 247
binary SRT with CSA, 232, 247
delay, 227, 234
digit selection, 236
digit-recurrence, 241, 247
non-restoring, 224, 247
on-the-fly conversion, 229, 236, 237
quotient selection, 227, 232, 235
radix-2k SRT, 234
radix-4 SRT, 238, 239
radix-B, 239
restoring, 227, 247
Division
binary digit-recurrence algorithm, 240
binary SRT algorithm, 229, 232, 247
convergence algorithms
digit-recurrence algorithm, 221, 223, 247
Goldschmidt algorithm, 239
P-D diagram, 223, 224, 236
quotient selection, 222
radix-2, 224
Robertson diagram, 222, 224
Division over GF(2m), 351
D-latch, 13
DMA, 363
Double-clocking, 111
Doubling circuit, 242
DRAM, 367, 368
Drive capabilities, 97
Driving
drive streng, 97
fan-in, 96
fan-out, 96
DSP, 215, 358, 360, 365
Dual port RAM, 119
E
Edge sensitive register, 104
EDIF, 130
EDK, 371
Electronic Automation Design, 1
ELF, 374, 453, 455, 459
465
Elliptic curve, 38
projective coordinates, 47
scalar product, 47
Embedded microprocessor, 359
Embedded system, 359
EMC, 439, 442
Encoding
commands, 83
redundant, 70
Endian, 359
big-endian, 360
little-endian, 359
Executable, 390, 455
Exponential function, 263
2x, 264, 273
accuracy, 264
ejz, 266
F
False paths, 113
Fan-out, 131
Fast or slow path, 113
FIFO, 363, 384
Finite fields, 337
Finite state machine
Mealy, 26, 30, 32
Moore, 32
registered, 33
Finite state machines, 13
Mealy, 13
Moore, 13
FLASH, 439, 443, 454
Flip-flops, 10, 104
temporal parameters, 104
Floating-point numbers
adder, 322
addersubtractor, 326, 330
addition, 309, 313
arithmetic operations, 309
division, 315
multiplication, 309
normalization, 309, 316
packing, 325
rounding, 309
truncation, 318
underflow, 318
unpacking, 322
Floating-point numeration system
unbiased, 319
FPU, 361, 365
FRAM
466
F (cont.)
Frequency synthesis, 113
FSL, 426, 437, 440
FIFO, 367, 426
slot, 428
FSM Encoding, 131
Index
IP cores, 372
ISE, 372
ISim, 372, 410
ISR, 374, 393
J
JTAG, 138
G
Generate function, 157, 278
Genetic algorithms, 144
Glicthless reconfiguration, 436
Glitch, 102
GNU compiler, 394
GNU gprof, 424
GNU tools chain, 372
Goldschmidt algorithm, 239, 245
GPIO, 363, 375, 392
Guard digits, 320
H
Halving circuit, 243
Handshaking protocol, 67, 69
Hard-core, 358, 360
Hardware Description Language, 1
Harvard architecture, 359
Hierarchical control units, 86, 88
Hierarchy preservation, 130
Hold time, 104
Hold time violation, 139
HWICAP, 436, 439, 441, 458
I
ICAP, 435, 442, 458
Implementations
dividers, 246
finite field arithmetic, 348
multipliers, 215
other operations, 268
Incompatibility graph, 48
Incompatibility relation, 48
Input buffer, 20
Input/output buffer, 130
Integers
Interrupt controller, 363, 383
Inversion
Newton-Raphson algorithm, 245
IO-port components, 19
IP, 357
L
Latch, 104
Latency, 56, 58
average, 67
LatticeMico32, 362
Layout implementation, 445, 449, 450
Level sensitive registers, 104
Libgen, 389
Lifetime, 45, 46
List scheduling, 44
LMB, 376, 439
Logarithm computation, 262
base-2, 252, 262, 273
Logic analyzer, 137
Long-path fault, 110
Look up tables, 2
Loop unrolling, 55, 72
LSB, 364
M
MAC, 358
MDD, 404
MDM, 375, 384, 439
Mean time between failures, 107
Memory, 366
Memory blocks, 17
Memory controller, 363
Mesosynchronous designs, 119
Metastability, 104, 116
recovery time, 107
setup-hold windows, 106
MHS, 373, 376
MicroBlaze, 362, 374, 437, 440
Minimum clock pulse, 104
MMU, 361
Modulo f(x) dividers, 350
binary algorithm, 347
Modulo f(x) division
binary algorithm, 351, 347
Modulo f(x) multiplication, 347
Modulo f(x) multipliers
Index
interleaved multiplier, 347, 348
multiply and reduce, 347, 348
Modulo f(x) operations, 347
addition and subtraction, 347
division, 351
squaring, 350
Modulo f(x) squaring, 348
Modulo m adder-subtractor, 339
Modulo m exponentiation
LSB-first, 346
MSB-first, 347
Modulo m multipliers
double, add and reduce, 341
interleaved multiplier, 340, 342
multiply and reduce
Modulo m operations, 337
addition, 338
Montgomery multiplication, 343
multiplication, 338
subtraction, 338
Modulo m reducer
Modulo p division, 347
binary algorithm, 347
Montgomery product, 343
carry-save addition, 344
MPD, 398
MSB, 364
MSS, 373
MTBF, 107
Multi-core, 360
Multi-cycle paths, 113
Multioperand adders
combinational, 166, 179
sequential, 163, 179
Multiplexers, 15
Multiplication
basic algorithm, 183
mod 2Bn+m, 199, 218
post correction, 206, 218
Multipliers
1-digit by 1-digit, 184
Booth, 204, 218
Booth-2, 210
carry-save, 187, 188, 215218
carry-save for integers, 201
constant, 211, 212, 215
Dadda, 189
delay, 185, 199, 201, 215
integers, 199, 205
mixed-radix, 191
multioperand adders, 189
n-digit by 1-digit, 185
467
parallel, 185, 214
post correction, 218
radix-2k, 217
radix-4, 210
right to left, 185
ripple-carry, 185
shift and add, 195, 201, 218
shift and add with CSA, 218
Wallace, 189
N
NCD, 385
Netlist, 130
Newton-Raphson algorithm, 245
Nios II, 362
Non-restoring algorithm, 291
Normalization
circuit, 324
O
Observability, 138
On-chip debugging, 137
Operations over Z2[x]/f(x), 347
addition and subtraction, 347
multiplication, 348
OS, 358, 361, 388, 408
Overflow, 174
P
Packing, 325
PAO, 400
Partial bitstream, 436, 453
Partial reconfiguration (PR), 435
Partition pin, 436
Peripheral, 362, 364
memory-mapped, 364
port-mapped, 364
Phase shifter, 113
Phase-locked loops (PLL), 113
Pipeline, 55, 67, 360
bandwidth, 58
period, 58
production, 58
rate, 58
speed, 58
Pipelined circuit, 60, 63
adder, 66
deskewing registers, 67
skewing registers, 67
468
P (cont.)
PlanAhead, 437, 445
Platgen, 385
PLB, 375, 398, 439
Power consumption, 119
dynamic, 112
sources, 120
static, 120
PowerPC, 358, 362, 365
Precedence graph, 37
scheduled, 42
Precedence relation, 34
Profiling, 423
Propagate function, 153, 278
Propagation delay, 99
Propagation delay of a register, 104
Proxy logic, 436
PSRAM, 367
Pull-down, 97, 135
Pull-up, 97, 135
R
Race-through, 110, 111
Random access memory, 18
Read only memory, 17
Reconfigurable coprocessor, 435, 438
Reconfigurable module (RM), 436, 438
Reconfigurable modules, 446
Reconfigurable partition, 446
Reconfigurable partition (RP), 436
Register balancing, 131
Register duplication, 131
Register transfer level (RTL), 103
Registers, 11
duplicate, 1, 131
parallel, 11
right shift, 11
Required time, 139
Resource assignment, 44
Response time, 58
Retiming, 131
Ripple-carry adder, 277
RISC, 359
Rounding schemes, 319, 320
RPM, 358
RTL, 358
RTL-level, 135
Runt pulse, 103
S
Scalar product, 38
Schedule
Index
admissible, 37, 38
ALAP, 37, 40
ASAP, 37, 40
Scheduling
operation scheduling, 34
SDK, 372, 388
SDRAM, 367
Segmentation, 58
admissible, 58, 60
Self-timed circuit, 67, 68, 69
adder, 70
Self-timing, 55
Setup time, 104
Setup time violation, 139
Sign bit, 174
Signed, 9
Simgen, 410
Simulated annealing, 44
Simulation, 135
in-circuit, 137
post place & route, 136
post-synthesis, 136
Single-core, 360
Skew, 135
Slack, 139
Slew rate, 135
SoC, 357
Soft-core, 360
Speedup factor, 58
Spikes, 103
Square root
floating-point, 316
Square rooters, 254
convergence methods, 260
fractional numbers, 259
Newton-Raphson method, 260, 273
non-restoring algorithm, 258, 272
restoring algorithm, 272
SRAM, 367
SRT-like algorithm, 295
Start signal, 67
Static partition, 436, 438
Static timing analysis, 139
Status register, 364
STD_LOGIC type, 20
Sticky bit, 322
Stuck-at fault model, 138
Subtractors, 173
Superscalar, 360
Synchronization
asynchronous FIFO, 117
failures, 110
Index
handshake signaling, 116
hold violation, 111
mesosynchronous, 119
registers, 108
setup violation, 110
synchronizer, 113, 116
true single phase clock (TSCP), 103
Synchronous design, 103
Synthesis, 130
constraints, 131
optimizations, 130
reports, 132
T
Tables, 3
TCL, 404
Ten0 s complement, 282
carry-chain adder-subtractor, 283
numeration system, 282
sign change, 283
Testbench, 410, 414
Throughput, 58
Timer, 393
Timing
analysis, 139
constrains, 113
false paths, 113
fast or slow paths, 113
multi-cycle paths, 113
TLB, 361
Transition time, 98
Tree
multioperand addition, 167
Trigonometric functions, 268
CORDIC, 267
469
pseudo-rotations, 267
Tri-state buffers, 5
Tri-states, 98
True single phase clock (TSCP), 103
U
UART, 363, 374
UCF, 385
Unpacking, 322
Unrolled loop implementations, 73
divider, 75
Unsigned, 9
V
Variable latency operations, 91
Video controller
Von Neumann architecture, 359
W
Wrapper, 414
X
XMD, 373, 396, 408
XMP, 373
XPS, 372, 376
Z
ZBT, 368
Zero-clocking, 110

Guide To FPGA

Uploaded by

Copyright:

Available Formats

Guide To FPGA

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Guide To FPGA

Uploaded by

Copyright:

Available Formats

Lecture Notes in Electrical Engineering

For further volumes:

Jean-Pierre Deschamps Gustavo D. Sutter

Springer Dordrecht Heidelberg New York London

It is a complement of the cited book; its aim is the generation of efficient

Basic Building Blocks . . . . . . . . . . . . .

Architecture of Digital Circuits . . . .

Special Topics of Data Path Synthesis . . . . . . . . . . . . . .

Control Unit Synthesis . . . . . . . .

Electronic Aspects of Digital Design . . . . . . . . . . . . . . . .

Implementation (Mapping, Placement

10.3 Square Rooters . . . . . . . . . . . . . . . . . . . . . . . . . .

12 Floating Point Arithmetic . . . . . . . . . . . . . . . . . .

15 Embedded Systems Development: Case Studies . . . . . . . . . . .

16 Partial Reconfiguration on Xilinx FPGAs . . . . . . . . . . .

About the Authors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Basic Building Blocks

Digital circuits are no longer defined by logical schemes but by Hardware

1.1 Combinational Components

J.-P. Deschamps et al., Guide to FPGA Implementation of Arithmetic Functions,

1 Basic Building Blocks

Fig. 1.1 Logic gates

More complex Boolean functions can also be defined. It is worthwhile to indicate

1.1 Combinational Components

Fig. 1.2 ANDORI gate

Fig. 1.3 One-bit full adder

The last choice of a WITH SELECT construction must be WHEN OTHERS in

1 Basic Building Blocks

Fig. 1.4 Four-digit to seven-segment decoder

Fig. 1.5 Four-input Look Up

Then the following component instantiation defines a 4-input XOR function.

1.1 Combinational Components

Fig. 1.6 Multiplexers

1.1.3 Controllable Connections

and the following conditional assignments a 1-to-4 demultiplexer (Fig. 1.7b)

1 Basic Building Blocks

Three-state buffers implement controllable switches.

1.1 Combinational Components

Fig. 1.8 Address decoder

Fig. 1.9 Three-state buffer

Fig. 1.10 Example of a data

1 Basic Building Blocks

Fig. 1.11 Data bus, second

000 001 01- 10- 11y 2y1y 0

Nevertheless, the same functionality can be implemented by an 8-bit 5-to-1

1.1.4 Arithmetic Circuits

1.1 Combinational Components

mod 2n+1 adder

Fig. 1.13 n-bit by m-bit

actual implementations, using special purpose FPGA resources (carry logic,

The following arithmetic equation defines an n-bit by m-bit multiplier where a is

1 Basic Building Blocks

1.2 Sequential Components

where en is the enable signal and d the data input.

1.2 Sequential Components

Fig. 1.15 Parallel register

1 Basic Building Blocks

Fig. 1.16 Right shift register

right shift register