Loops & Complexity
in
DIGITAL SYSTEMS
∗
Lecture Notes on Digital Design
in
Giga-Gate per Chip Era
(work in endless progress)
Gheorghe M. Ştefan
– 2014 version –
2
Introduction
... theories become clear and ‘reasonable’ only after incoherent parts of them have been used for a long time.
Paul Feyerabend1
The price for the clarity and simplicity of a ’reasonable’
approach is its incompleteness.
Few legitimate questions about how to teach digital systems in Giga-Gate Per Chip Era
(G2CE) are waiting for an answer.
1. What means a complex digital system? How complex systems are designed using small
and simple circuits?
2. How a digital system expands its size, increasing in the same time its speed? Are there
simple mechanisms to be emphasized?
3. Is there a special mechanism allowing a “hierarchical growing” in a digital system? Or,
how new features can be added in a digital system?
The first question occurs because already exist many different big systems which seem to have
different degree of complexity. For example: big memory circuits and big processors. Both are
implemented using a huge number of circuits, but the processors seem to be more “complicated”
than the memories. In almost all text books complexity is related only with the dimension of
the system. Complexity means currently only size, the concept being unable to make necessary
distinctions in G2CE. The last improvements of the microelectronic technologies allow us to put
on a Silicon die around a billion of gates, but the design tools are faced with more than the size
of the system to be realized in this way. The size and the complexity of a digital system must
be distinctly and carefully defined in order to have a more flexible conceptual environment for
designing, implementing and testing systems in G2CE.
The second question rises in the same context of the big and the complex systems. Growing
a digital system means both increasing its size and its complexity. How are correlated these two
growing processes? The dynamic of adding circuits and of adding adding features seems to be
very different and governed by distinct mechanisms.
The third question occurs in the hierarchical contexts in which the computation is defined.
For example, Kleene’s functional hierarchy or Chomsky’s grammatical hierarchy are defined to
1
Paul Feyerabend (b.1924, d.1994), having studied science at the University of Vienna, moved into philosophy
for his doctoral thesis. He became a critic of philosophy of science itself, particularly of “rationalist” attempts to
lay down or discover rules of scientific method. His first book, Against Method (1975), sets out “epistemological
anarchism”, whose main thesis was that there is no such thing as the scientific method.
3
4
explain how computation or formal languages used in computation evolve from simple to complex. Is this hierarchy reflected in a corresponding hierarchical organization of digital circuits?
It is obvious that a sort of similar hierarchy must be hidden in the multitude of features already
emphasized in the world of digital circuits. Let be the following list of usual terms: boolean
functions, storing elements, automata circuits, finite automata, memory functions, processing
functions, . . ., self-organizing processes, . . .. Is it possible to disclose in this list a hierarchy, and
more, is it possible to find similarities with previously exemplified hierarchies?
The first answer will be derived from the Kolmogorov-Chaitin algorithmic complexity: the
complexity of a circuit is related with the dimension of its shortest formal description. A big circuit (a circuit built using a big number o gates) can be simple or complex
depending on the possibility to emphasize repetitive patterns in its structure. A no pattern
circuit is a complex one because its description has the dimension proportional with its size.
Indeed, for a complex, no pattern circuit each gate must be explicitly specified.
The second answer associate the composition with sizing and the loop with featuring.
Composing circuits results biggest structures with the same kind of functionality, while closing
loops in a circuit new kind of behaviors are induced. Each new loop adds more autonomy to the
system, because increases the dependency of the output signals in the detriment of the input
signals. Shortly, appropriate loops means more autonomy that is equivalent sometimes with a
new level of functionality.
The third answer is given by proposing a taxonomy for digital systems based on the maximum number of included loops closed in a certain digital system. The old distinction between
combinational and sequential, applied only to circuits, is complemented with a classification
taking into the account the functional and structural diversity of the digital systems used in
the contemporary designs. More, the resulting classification provides classes of circuits having
direct correspondence with the levels belonging to Kleene’s and Chomsky’s hierarchies.
The first part of the book – Digital Systems: a Bird’s-Eye View – is a general introduction
in digital systems framing the digital domain in the larger context of the computational sciences,
introducing the main formal tool for describing, simulating and synthesizing digital systems, and
presenting the main mechanisms used to structure digital systems. The second part of the book
– Looping in Digital Systems – deals with the main effects of the loop: more autonomy and
segregation between the simple parts and the complex parts in digital systems. Both, autonomy
and segregation, are used to minimize size and complexity. The third part of the book – Loop
Based Morphisms – contains three attempts to make meaningful connections between the
domain of the digital systems, and the fields of recursive functions, of formal languages and of
information theories. The last chapter sums up the main ideas of the book making also some new
correlations permitted by its own final position. The book ends with a lot of annexes containing
short reviews of the prerequisite knowledge (binary arithmetic, Boolean functions, elementary
digital circuits, automata theory), compact presentations of the formal tools used (pseudo-code
language, Verilog HDL), examples, useful data about real products (standard cell libraries).
PART I: Digital Systems: a Bird’s-Eye View
The first chapter: What’s a Digital System? Few general questions are answered in this
chapter. One refers to the position of digital system domain in the larger class of the sciences of
computation. Another asks for presenting the ways we have to implement actual digital systems.
The importance is also to present the correlated techniques allowing to finalize a digital product.
The second chapter: Let’s Talk Digital Circuits in Verilog The first step in approaching
the digital domain is to become familiar with a Hardware Description Language (HDL) as the
5
main tool for mastering digital circuits and systems. The Verilog HDL is introduced and in the
same time used to present simple digital circuits. The distinction between behavioral descriptions
and structural descriptions is made when Verilog is used to describe and simulate combinational
and sequential circuits. The temporal behaviors are described, along with solutions to control
them.
The third chapter: Scaling & Speeding & Featuring The architecture and the organization of a digital system are complex objectives. We can not be successful in designing big
performance machine without strong tools helping us to design the architecture and the high
level organization of a desired complex system. These mechanisms are three. One helps us to
increase the brute force performance of the system. It is composition. The second is used to
compensate the slow-down of the system due to excessive serial composition. It is pipelining.
The last is used to add new features when they are asked by the application. It is about closing
loops inside the system in order to improve the autonomous behaviors.
The fourth chapter: The Taxonomy of Digital Systems A loop based taxonomy for
digital systems is proposed. It classifies digital systems in orders, as follows:
• 0-OS: zero-order systems - no-loop circuits - containing the combinational circuits;
• 1-OS: 1-order systems - one-loop circuits - the memory circuits, with the autonomy of the
internal state; they are used mainly for storing
• 2-OS: 2-order systems - two-loop circuits - the automata, with the behavioral autonomy
in their own state space, performing mainly the function of sequencing
• 3-OS: 3-order systems - three-loop circuits - the processors, with the autonomy in interpreting their own internal states; they perform the function of controlling
• 4-OS: 4-order systems - four-loop circuits - the computers, which interpret autonomously
the programs according to the internal data
• ...
• n-OS: n-order systems - n-loop circuits - systems in which the information is interpenetrated with the physical structures involved in processing it; the distinction between data
and programs is surpassed and the main novelty is the self-organizing behavior.
The fifth chapter: Our Final Target A small and simple programmable machine, called
toyMachine is defined using a behavioral description. In the last chapter of the second part a
structural design of this machine will be provided using the main digital structure introduced
meantime.
PART II: Looping in Digital Domain
The sixth chapter: Gates The combinational circuits (0-OS) are introduced using a functional approach. We start with the simplest functions and, using different compositions, the
basic simple functional modules are introduced. The distinction between simple and complex
combinational circuits is emphasized, presenting specific technics to deal with complexity.
6
The seventh chapter: Memories There are two ways to close a loop over the simplest
functional combinational circuit: the one-input decoder. One of them offers the stable structure
on which we ground the class of memory circuits (1-OS) containing: the elementary latches,
the master-slave structures (the serial composition), the random access memory (the parallel
composition) and the register (the serial-parallel composition). Few applications of storing
circuits (pipeline connection, register file, content addressable memory, associative memory) are
described.
The eight chapter: Automata Automata (2-OS) are presented in the fourth chapter. Due
to the second loop the circuit is able to evolve, more or less, autonomously in its own state
space. This chapter begins presenting the simplest automata: the T flip-flop and the JK flipflop. Continues with composed configurations of these simple structures: counters and related
structures. Further, our approach makes distinction between the big sized, but simple functional
automata (with the loop closed through a simple, recursive defined combinational circuit that
can have any size) and the random, complex finite automata (with the loop closed through a
random combinational circuit having the size in the same order with the size of its definition).
The autonomy offered by the second loop is mainly used to generate or to recognize specific
sequences of binary configurations.
The ninth chapter: Processors The circuits having three loops (3-OS) are introduced. The
third loop may be closed in three ways: through a 0-OS, through an 1-OS or through a 2-OS,
each of them being meaningful in digital design. The first, because of the segregation process
involved in designing automata using JK flip-flops or counters as state register. The size of
the random combinational circuits that compute the state transition function is reduced, in the
most of case, due to the increased autonomy of the device playing the role of the register. The
second type of loop, through a memory circuit, is also useful because it increases the autonomy
of the circuit so that the control exerted on it may be reduced (the circuit “knows more about
itself”). The third type of loop, that interconnects two automata (an functional automaton and
a control finite automaton), generates the most important digital circuits: the processor.
The tenth chapter: Computers The effects of the fourth loop are shortly enumerated in the
sixth chapter. The computer is the typical structure in 4-OS. It is also the support of the strongest
segregation between the simple physical structure of the machine and the complex structure of
the program (a symbolic structure). Starting from the fourth order the main functional up-dates
are made structuring the symbolic structures instead of restructuring circuits. Few new loops
are added in actual designs only for improving time or size performances, but not for adding
new basic functional capabilities. For this reason our systematic investigation concerning the
loop induced hierarchy stops with the fourth loop. The toyMachine behavioral description is
revisited and substituted with a pure structural description.
The eleventh chapter: Self-Organizing Structures ends the first part of the book with
some special circuits which belongs to n-OSs. The cellular automata, the connex memory and
the eco-chip are n-loop structures that destroy the usual architectural thinking based on the
distinction between the physical support for symbolic structures and the circuits used for processing them. Each bit/byte has its own processing element in a system which performs the
finest grained parallelism.
The twelfth chapter: Global-Loop Systems Why not a hierarchy of hierarchies of loops?
Having an n-order system how new features can be added? A possible answer: adding a
7
global loop. Thus, a new hierarchy of super-loops starts. It is not about science fiction.
ConnexArrayT M is an example. It is described, evaluated and some possible applications
are presented.
The main stream of this book deals with the simple and the complex in digital systems, emphasizing them in the segregation process that opposes simple structures of circuits to the complex
structures of symbols. The functional information offers the environment for segregating the
simple circuits from the complex binary configurations.
When the simple is mixed up with the complex, the apparent complexity of the system
increases over its actual complexity. We promote design methods which reduce the apparent
complexity by segregating the simple from the complex. The best way to substitute the apparent
complexity with the actual complexity is to drain out the chaos from order. One of the most
important conclusions of this book is that the main role of the loop in digital systems is to
segregate the simple from the complex, thus emphasizing and using the hidden resources of
autonomy.
In the digital systems domain prevails the art of disclosing the simplicity because there
exists the symbolic domain of functional information in which we may ostracize the complexity.
But, the complexity of the process of disclosing the simplicity exhausts huge resources of imagination. This book offers only the starting point for the architectural thinking: the art of finding
the right place of the interface between simple and complex in computing systems.
Acknowledgments
8
Contents
I
A BIRD’S-EYE VIEW ON DIGITAL SYSTEMS
1 WHAT’S A DIGITAL SYSTEM?
1.1 Framing the digital design domain . . . . . . . . .
1.1.1 Digital domain as part of electronics . . . .
1.1.2 Digital domain as part of computer science
1.2 Defining a digital system . . . . . . . . . . . . . . .
1.3 Our first target . . . . . . . . . . . . . . . . . . . .
1.4 Problems . . . . . . . . . . . . . . . . . . . . . . .
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
4
4
7
9
15
20
2 DIGITAL CIRCUITS
2.1 Combinational circuits . . . . . . . . . . . . . . . . . . . . .
2.1.1 Zero circuit . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Selection . . . . . . . . . . . . . . . . . . . . . . . .
2.1.3 Adder . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.4 Divider . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Sequential circuits . . . . . . . . . . . . . . . . . . . . . . .
2.2.1 Elementary Latches . . . . . . . . . . . . . . . . . .
2.2.2 Elementary Clocked Latches . . . . . . . . . . . . .
2.2.3 Data Latch . . . . . . . . . . . . . . . . . . . . . . .
2.2.4 Master-Slave Principle . . . . . . . . . . . . . . . . .
2.2.5 D Flip-Flop . . . . . . . . . . . . . . . . . . . . . . .
2.2.6 Register . . . . . . . . . . . . . . . . . . . . . . . . .
2.2.7 Shift register . . . . . . . . . . . . . . . . . . . . . .
2.2.8 Counter . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Putting all together . . . . . . . . . . . . . . . . . . . . . .
2.4 Concluding about this short introduction in digital circuits
2.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
22
23
24
26
28
28
28
32
33
35
37
38
40
41
42
43
43
3 GROWING & SPEEDING & FEATURING
3.1 Size vs. Complexity . . . . . . . . . . . . . .
3.2 Time restrictions in digital systems . . . . . .
3.2.1 Pipelined connections . . . . . . . . .
3.2.2 Fully buffered connections . . . . . . .
3.3 Growing the size by composition . . . . . . .
3.4 Speeding by pipelining . . . . . . . . . . . . .
3.4.1 Register transfer level . . . . . . . . .
3.4.2 Pipeline structures . . . . . . . . . . .
3.4.3 Data parallelism vs. time parallelism .
3.5 Featuring by closing new loops . . . . . . . .
3.5.1 ∗ Data dependency . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
47
49
51
54
57
58
63
63
64
65
67
70
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
10
CONTENTS
3.6
3.7
3.8
3.5.2 ∗ Speculating to avoid limitations imposed by
Concluding about composing & pipelining & looping
Problems . . . . . . . . . . . . . . . . . . . . . . . .
Projects . . . . . . . . . . . . . . . . . . . . . . . . .
4 THE TAXONOMY OF DIGITAL SYSTEMS
4.1 Loops & Autonomy . . . . . . . . . . . . . . . .
4.2 Classifying Digital Systems . . . . . . . . . . .
4.3 # Digital Super-Systems . . . . . . . . . . . . .
4.4 Preliminary Remarks On Digital Systems . . .
4.5 Problems . . . . . . . . . . . . . . . . . . . . .
4.6 Projects . . . . . . . . . . . . . . . . . . . . . .
data
. . .
. . .
. . .
dependency
. . . . . . .
. . . . . . .
. . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
72
73
74
76
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
77
78
81
84
84
85
86
5 OUR FINAL TARGET
5.1 toyMachine: a small & simple computing machine
5.2 How toyMachine works . . . . . . . . . . . . . . . .
5.3 Concluding about toyMachine . . . . . . . . . . . .
5.4 Problems . . . . . . . . . . . . . . . . . . . . . . .
5.5 Projects . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
87
88
96
99
99
99
II
.
.
.
.
.
.
LOOPING IN THE DIGITAL DOMAIN
6 GATES:
Zero order, no-loop digital systems
6.1 Simple, Recursive Defined Circuits . . . . . . . .
6.1.1 Decoders . . . . . . . . . . . . . . . . . .
Informal definition . . . . . . . . . . . . .
Formal definition . . . . . . . . . . . . . .
Recursive definition . . . . . . . . . . . .
Structural description . . . . . . . . . . .
Arithmetic interpretation . . . . . . . . .
Application . . . . . . . . . . . . . . . . .
6.1.2 Demultiplexors . . . . . . . . . . . . . . .
Informal definition . . . . . . . . . . . . .
Formal definition . . . . . . . . . . . . . .
Recursive definition . . . . . . . . . . . .
6.1.3 Multiplexors . . . . . . . . . . . . . . . .
Informal definition . . . . . . . . . . . . .
Formal definition . . . . . . . . . . . . . .
Recursive definition . . . . . . . . . . . .
Structural aspects . . . . . . . . . . . . .
Application . . . . . . . . . . . . . . . . .
6.1.4 ∗ Shifters . . . . . . . . . . . . . . . . . .
6.1.5 ∗ Priority encoder . . . . . . . . . . . . .
6.1.6 ∗ Prefix computation network . . . . . . .
6.1.7 Increment circuit . . . . . . . . . . . . . .
6.1.8 Adders . . . . . . . . . . . . . . . . . . . .
Carry-Look-Ahead Adder . . . . . . . . .
∗ Prefix-Based Carry-Look-Ahead Adder
∗ Carry-Save Adder . . . . . . . . . . . .
101
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
103
104
105
105
105
106
107
108
108
109
109
110
110
111
111
111
112
112
113
114
116
117
120
120
123
124
127
CONTENTS
11
6.1.9
6.1.10
6.1.11
6.1.12
6.2
6.3
6.4
6.5
∗ Combinational Multiplier . . . . . . . . . . . . . . .
Arithmetic and Logic Unit . . . . . . . . . . . . . . .
Comparator . . . . . . . . . . . . . . . . . . . . . . . .
∗ Sorting network . . . . . . . . . . . . . . . . . . . .
Bathcer’s sorter . . . . . . . . . . . . . . . . . . . . . .
6.1.13 ∗ First detection network . . . . . . . . . . . . . . . .
6.1.14 ∗ Spira’s theorem . . . . . . . . . . . . . . . . . . . . .
Complex, Randomly Defined Circuits . . . . . . . . . . . . . .
6.2.1 An Universal circuit . . . . . . . . . . . . . . . . . . .
6.2.2 Using the Universal circuit . . . . . . . . . . . . . . .
6.2.3 The many-output random circuit: Read Only Memory
Concluding about combinational circuits . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
130
131
134
135
136
141
142
142
142
144
147
152
153
158
7 MEMORIES:
First order, 1-loop digital systems
7.1 Stable/Unstable Loops . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2 Elementary Structures . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3 The Serial Composition: the Edge Triggered Flip-Flop . . . . . . . .
7.3.1 The Serial Register . . . . . . . . . . . . . . . . . . . . . . . .
7.4 The Parallel Composition: the Random Access Memory . . . . . . .
7.4.1 The n-Bit Latch . . . . . . . . . . . . . . . . . . . . . . . . .
7.4.2 Asynchronous Random Access Memory . . . . . . . . . . . .
Expanding the number of bits per word . . . . . . . . . . . .
Expanding the number of words by two dimension addressing
7.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.1 Synchronous RAM . . . . . . . . . . . . . . . . . . . . . . . .
7.5.2 Register File . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.5.3 Field Programmable Gate Array – FPGA . . . . . . . . . . .
The system level organization of an FPGA . . . . . . . . . .
The IO interface . . . . . . . . . . . . . . . . . . . . . . . . .
The switch node . . . . . . . . . . . . . . . . . . . . . . . . .
The basic building block . . . . . . . . . . . . . . . . . . . . .
The configurable logic block . . . . . . . . . . . . . . . . . . .
7.5.4 ∗ Content Addressable Memory . . . . . . . . . . . . . . . . .
7.5.5 ∗ An Associative Memory . . . . . . . . . . . . . . . . . . . .
7.5.6 ∗ Beneš-Waxman Permutation Network . . . . . . . . . . . .
7.5.7 ∗ First-Order Systolic Systems . . . . . . . . . . . . . . . . .
7.6 Concluding About Memory Circuits . . . . . . . . . . . . . . . . . .
7.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.8 Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
159
161
162
162
163
163
164
164
166
167
168
170
171
171
172
173
174
174
175
175
177
179
182
185
186
191
8 AUTOMATA:
Second order, 2-loop digital systems
8.1 Optimizing DFF with an asynchronous automaton
8.2 Two States Automata . . . . . . . . . . . . . . . .
8.2.1 The Smallest Automaton: the T Flip-Flop .
8.2.2 The JK Automaton: the Greatest Flip-Flop
8.2.3 ∗ Serial Arithmetic . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
195
197
199
199
200
200
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
12
CONTENTS
8.2.4 ∗ Universal 2-input and 2-state automaton . .
Functional Automata: the Simple Automata . . . . . .
8.3.1 Counters . . . . . . . . . . . . . . . . . . . . .
8.3.2 ∗ Accumulator Automaton . . . . . . . . . . .
8.3.3 ∗ Sequential multiplication . . . . . . . . . . .
∗ Radix-2 multiplication . . . . . . . . . . . . .
∗ Radix-4 multiplication . . . . . . . . . . . . .
8.3.4 ∗ “Bit-eater” automaton . . . . . . . . . . . . .
8.3.5 ∗ Sequential divisor . . . . . . . . . . . . . . .
8.4 ∗ Composing with simple automata . . . . . . . . . . .
8.4.1 ∗ LIFO memory . . . . . . . . . . . . . . . . .
8.4.2 ∗ FIFO memory . . . . . . . . . . . . . . . . .
8.4.3 ∗ The Multiply-Accumulate Circuit . . . . . .
8.5 Finite Automata: the Complex Automata . . . . . . .
8.5.1 Basic Configurations . . . . . . . . . . . . . . .
8.5.2 Designing Finite Automata . . . . . . . . . . .
8.5.3 ∗ Control Automata: the First “Turning Point”
Verilog descriptions for CROM . . . . . . . . .
Binary code generator . . . . . . . . . . . . . .
8.6 ∗ Automata vs. Combinational Circuits . . . . . . . .
8.7 ∗ The Circuit Complexity of a Binary String . . . . .
8.8 Concluding about automata . . . . . . . . . . . . . . .
8.9 Problems . . . . . . . . . . . . . . . . . . . . . . . . .
8.10 Projects . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9 PROCESSORS:
Third order, 3-loop digital systems
9.1 Implementing finite automata with ”intelligent registers” . . . . . .
9.1.1 Automata with JK “registers” . . . . . . . . . . . . . . . .
9.1.2 ∗ Automata using counters as registers . . . . . . . . . . . .
9.2 Loops closed through memories . . . . . . . . . . . . . . . . . . . .
Version 1: the controlled Arithmetic & Logic Automaton .
Version 2: the commanded Arithmetic & Logic Automaton
9.3 Loop coupled automata: the second ”turning point” . . . . . . . .
9.3.1 ∗ Push-down automata . . . . . . . . . . . . . . . . . . . .
9.3.2 The elementary processor . . . . . . . . . . . . . . . . . . .
9.3.3 Executing instructions vs. interpreting instructions . . . . .
Von Neumann architecture / Harvard architecture . . . . .
9.3.4 An executing processor . . . . . . . . . . . . . . . . . . . .
The organization . . . . . . . . . . . . . . . . . . . . . . . .
The instruction set architecture . . . . . . . . . . . . . . . .
Implementing toyRISC . . . . . . . . . . . . . . . . . . . . .
The time performance . . . . . . . . . . . . . . . . . . . . .
9.3.5 ∗ An interpreting processor . . . . . . . . . . . . . . . . . .
The organization . . . . . . . . . . . . . . . . . . . . . . . .
Microarchitecture . . . . . . . . . . . . . . . . . . . . . . . .
Instruction set architecture (ISA) . . . . . . . . . . . . . . .
Implementing ISA . . . . . . . . . . . . . . . . . . . . . . .
Time performance . . . . . . . . . . . . . . . . . . . . . . .
Concluding about our CISC processor . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
201
202
202
204
205
205
206
210
210
212
212
214
216
217
218
219
225
230
232
234
237
239
240
244
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
247
249
249
252
254
255
255
256
256
258
262
264
265
265
266
268
268
268
268
273
273
274
282
283
CONTENTS
9.4
9.5
9.6
9.7
∗ The assembly language: the lowest programming level
Concluding about the third loop . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . .
Projects . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
283
283
284
284
10 COMPUTING MACHINES:
≥4–loop digital systems
10.1 Types of fourth order systems . . . . . . . . . . . . . . . . . . . .
10.1.1 The computer – support for the strongest segregation . .
10.2 ∗ The stack processor – a processor as 4-OS . . . . . . . . . . . .
10.2.1 ∗ The organization . . . . . . . . . . . . . . . . . . . . . .
10.2.2 ∗ The micro-architecture . . . . . . . . . . . . . . . . . . .
10.2.3 ∗ The instruction set architecture . . . . . . . . . . . . . .
10.2.4 ∗ Implementation: from micro-architecture to architecture
10.2.5 ∗ Time performances . . . . . . . . . . . . . . . . . . . . .
10.2.6 ∗ Concluding about our Stack Processor . . . . . . . . . .
10.3 Embedded computation . . . . . . . . . . . . . . . . . . . . . . .
10.3.1 The structural description of toyMachine . . . . . . . . .
The top module . . . . . . . . . . . . . . . . . . . . . . .
The interrupt . . . . . . . . . . . . . . . . . . . . . . . . .
The control section . . . . . . . . . . . . . . . . . . . . . .
The data section . . . . . . . . . . . . . . . . . . . . . . .
Multiplexors . . . . . . . . . . . . . . . . . . . . . . . . .
Concluding about toyMachine . . . . . . . . . . . . . .
10.3.2 Interrupt automaton: the asynchronous version . . . . . .
10.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 Projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
285
286
288
288
289
291
294
295
299
300
300
300
300
305
305
306
308
308
308
317
317
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
319
320
321
321
321
326
327
328
331
331
331
331
331
331
331
333
334
336
337
338
339
11 # ∗ SELF-ORGANIZING STRUCTURES:
N-th order digital systems
11.1 Push-Down Stack as n-OS . . . . . . . . . . . . . . . . .
11.2 Cellular automata . . . . . . . . . . . . . . . . . . . . .
11.2.1 General definitions . . . . . . . . . . . . . . . . .
The linear cellular automaton . . . . . . . . . . .
The two-dimension cellular automaton . . . . . .
11.2.2 Applications . . . . . . . . . . . . . . . . . . . .
11.3 Systolic systems . . . . . . . . . . . . . . . . . . . . . . .
11.4 Interconnection issues . . . . . . . . . . . . . . . . . . .
11.4.1 Local vs. global connections . . . . . . . . . . . .
Memory wall . . . . . . . . . . . . . . . . . . . .
11.4.2 Localizing with cellular automata . . . . . . . . .
11.4.3 Many clock domains & asynchronous connections
11.5 Neural networks . . . . . . . . . . . . . . . . . . . . . .
11.5.1 The neuron . . . . . . . . . . . . . . . . . . . . .
11.5.2 The feedforward neural network . . . . . . . . .
11.5.3 The feedback neural network . . . . . . . . . . .
11.5.4 The learning process . . . . . . . . . . . . . . . .
Unsupervised learning: Hebbian rule . . . . . . .
Supervised learning: perceptron rule . . . . . . .
11.5.5 Neural processing . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
14
CONTENTS
11.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
12 # ∗
12.1
12.2
12.3
GLOBAL-LOOP SYSTEMS
Global loops in cellular automata
ConnexArrayT M based system .
Low level functional description .
12.3.1 System initialization . . .
12.3.2 Addressing functions . . .
12.3.3 Setting functions . . . . .
12.3.4 Transfer functions . . . .
Basic transfer functions .
12.3.5 Arithmetic functions . . .
12.3.6 Reduction functions . . .
12.3.7 Test functions . . . . . . .
12.3.8 Control functions . . . . .
12.3.9 Stream functions . . . . .
12.4 Index . . . . . . . . . . . . . . .
12.5 Problems . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13 DESIGNING A COMPLEX DIGITAL SYSTEM
13.1 The Architectural Description . . . . . . . . . . . .
13.2 Complex Interconnections . . . . . . . . . . . . . .
13.2.1 Pipelined complex systems . . . . . . . . .
13.2.2 Buffered complex systems . . . . . . . . . .
13.3 Complex Applications . . . . . . . . . . . . . . . .
13.4 Problems . . . . . . . . . . . . . . . . . . . . . . .
III
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
341
342
342
349
350
352
352
353
353
354
356
357
358
359
361
362
.
.
.
.
.
.
363
364
364
364
364
364
364
ANNEXES
365
A The Verilog HDL
367
B Binary Arithmetic
B.1 Binary numbers . . . . . . . . . . .
B.1.1 Positive integers . . . . . .
B.1.2 Signed integers . . . . . . .
B.1.3 Real numbers approximated
B.2 Addition & subtract . . . . . . . .
B.2.1 Positive integer operations .
B.2.2 Signed integer operations .
B.2.3 Floating point operations .
B.3 Multiply . . . . . . . . . . . . . . .
B.3.1 Positive integer operations .
B.3.2 Signed integer operations .
B.3.3 Floating point operations .
B.4 Division . . . . . . . . . . . . . . .
B.4.1 Positive integer operations .
B.4.2 Signed integer operations .
B.4.3 Floating point operations .
369
369
369
369
369
369
369
369
369
369
369
369
369
369
369
369
369
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
as floating point numbers
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
15
C Boolean functions
C.1 Short History . . . . . . . . . . . . . . . . . . .
C.2 Elementary circuits: gates . . . . . . . . . . . .
C.2.1 Zero-input logic circuits . . . . . . . . .
C.2.2 One input logic circuits . . . . . . . . .
C.2.3 Two inputs logic circuits . . . . . . . . .
C.2.4 Many input logic circuits . . . . . . . .
C.3 How to Deal with Logic Functions . . . . . . .
C.4 Minimizing Boolean functions . . . . . . . . . .
C.4.1 Canonical forms . . . . . . . . . . . . .
C.4.2 Algebraic minimization . . . . . . . . .
Minimal depth minimization . . . . . .
Multi-level minimization . . . . . . . . .
Many output circuit minimization . . .
C.4.3 Veitch-Karnaugh diagrams . . . . . . .
Minimizing with V-K diagrams . . . . .
Minimizing incomplete defined functions
V-K diagrams with included functions .
C.5 Problems . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
371
371
371
371
372
372
373
373
375
375
376
376
377
377
378
379
380
380
381
D Basic circuits
D.1 Actual digital signals . . . . . . . . . . . . . . .
D.2 CMOS switches . . . . . . . . . . . . . . . . . .
D.3 The Inverter . . . . . . . . . . . . . . . . . . . .
D.3.1 The static behavior . . . . . . . . . . . .
D.3.2 Dynamic behavior . . . . . . . . . . . .
D.3.3 Buffering . . . . . . . . . . . . . . . . .
D.3.4 Power dissipation . . . . . . . . . . . . .
Switching power . . . . . . . . . . . . .
Short-circuit power . . . . . . . . . . . .
Leakage power . . . . . . . . . . . . . .
D.4 Gates . . . . . . . . . . . . . . . . . . . . . . .
D.4.1 NAND & NOR gates . . . . . . . . . . .
The static behavior of gates . . . . . . .
Propagation time . . . . . . . . . . . . .
Power consumption & switching activity
Power consumption & glitching . . . . .
D.4.2 Many-Input Gates . . . . . . . . . . . .
D.4.3 AND-NOR gates . . . . . . . . . . . . .
D.5 The Tristate Buffers . . . . . . . . . . . . . . .
D.6 The Transmission Gate . . . . . . . . . . . . .
D.7 Memory Circuits . . . . . . . . . . . . . . . . .
D.7.1 Flip-flops . . . . . . . . . . . . . . . . .
Data latches and their transparency . .
Master-slave DF-F . . . . . . . . . . . .
Resetable DF-F . . . . . . . . . . . . . .
D.7.2 # Static memory cell . . . . . . . . . .
D.7.3 # Array of cells . . . . . . . . . . . . .
D.7.4 # Dynamic memory cell . . . . . . . . .
D.8 Problems . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
383
383
384
385
385
386
388
389
389
390
390
390
391
391
392
392
393
394
395
396
396
397
397
397
397
397
397
397
397
397
16
E Standard cell libraries
E.1 Main parameters . . . . . . . . . . . . . . . . . .
E.2 Basic cells . . . . . . . . . . . . . . . . . . . . . .
E.2.1 Gates . . . . . . . . . . . . . . . . . . . .
Two input AND gate: AND2 . . . . . . .
Two input OR gate: OR2 . . . . . . . . .
Three input AND gate: AND3 . . . . . .
Three input OR gate: OR3 . . . . . . . .
Four input AND gate: AND4 . . . . . . .
Four input OR gate: OR4 . . . . . . . . .
Four input AND-OR gate: ANDOR22 . .
Invertor: NOT . . . . . . . . . . . . . . .
Two input NAND gate: NAND2 . . . . .
Two input NOR gate: NOR2 . . . . . . .
Three input NAND gate: NAND3 . . . .
Three input NOR gate: NOR3 . . . . . .
Four input NAND gate: NAND4 . . . . .
Four input NOR gate: NOR4 . . . . . . .
Two input multiplexer: MUX2 . . . . . .
Two input inverting multiplexer: NMUX2
Two input XOR: XOR2 . . . . . . . . . .
E.2.2 # Flip-flops . . . . . . . . . . . . . . . . .
CONTENTS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
F Memory compilers
G Finite Automata
G.1 Basic definitions in automata theory . . . . . . . . . . . . .
G.2 How behaves a finite automata . . . . . . . . . . . . . . . .
G.3 Representing finite automata . . . . . . . . . . . . . . . . .
G.3.1 Flow-charts . . . . . . . . . . . . . . . . . . . . . . .
The flow-chart for a half-automaton . . . . . . . . .
The flow-chart for a Moore automaton . . . . . . . .
The flow-chart for a Mealy automaton . . . . . . . .
G.3.2 Transition diagrams . . . . . . . . . . . . . . . . . .
Transition diagrams for half-automata . . . . . . . .
Transition diagrams Moore automata . . . . . . . .
Transition diagrams Mealy automata . . . . . . . . .
G.3.3 Procedures . . . . . . . . . . . . . . . . . . . . . . .
HDL representations for Moore automata . . . . . .
HDL representations for Mealy automata . . . . . .
G.4 State codding . . . . . . . . . . . . . . . . . . . . . . . . . .
G.4.1 Minimal variation encoding . . . . . . . . . . . . . .
G.4.2 Reduced dependency encoding . . . . . . . . . . . .
G.4.3 Incremental codding . . . . . . . . . . . . . . . . . .
G.4.4 One-hot state encoding . . . . . . . . . . . . . . . .
G.5 Minimizing finite automata . . . . . . . . . . . . . . . . . .
G.5.1 Minimizing the size by an appropriate state codding
G.5.2 Minimizing the complexity by one-hot encoding . . .
G.6 Parasitic effects in automata . . . . . . . . . . . . . . . . . .
G.6.1 Asynchronous inputs . . . . . . . . . . . . . . . . . .
401
401
402
403
403
403
403
403
403
403
404
404
404
404
404
404
405
405
405
405
405
405
407
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
409
409
410
412
412
412
413
413
414
415
416
417
418
418
419
419
421
421
422
422
423
423
425
426
426
CONTENTS
17
G.6.2 The Hazard . . . . . . . . . . . . . . . . .
Hazard generated by asynchronous inputs
Propagation hazard . . . . . . . . . . . .
Dynamic hazard . . . . . . . . . . . . . .
G.7 Fundamental limits in implementing automata .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
427
427
428
430
430
H FPGA
433
I
435
435
435
435
How to make a project
I.1 How to organize a project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.2 An example: FIFO memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
I.3 About team work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
J Designing a simple CISC processor
J.1 The project . . . . . . . . . . . . . . . . .
J.2 RTL code . . . . . . . . . . . . . . . . . .
The module cisc processor.v . .
The module cisc alu.v . . . . . .
The module control automaton.v
The module register file.v . .
The module mux4.v . . . . . . . .
The module mux2.v . . . . . . . .
J.3 Testing cisc processor.v . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
437
437
437
437
438
438
445
446
446
446
K # Meta-stability
447
Index
456
18
CONTENTS
Part I
A BIRD’S-EYE VIEW ON
DIGITAL SYSTEMS
1
Chapter 1
WHAT’S A DIGITAL SYSTEM?
In the previous chapter
we can not find anything because it does not exist, but we suppose the reader is familiar
with:
• fundamentals about what means computation
• basics about Boolean algebra and basic digital circuits (see Annexes Boolean Functions and Basic circuits for a short refresh)
• the usual functions supposed to be implemented by digital sub-systems in the current
audio, video, communication, gaming, ... market products
In this chapter
general definitions related with the digital domain are used to reach the following targets:
• to frame the digital system domain in the larger area of the information technologies
• to present different ways the digital approach is involved in the design of the real
market products
• to enlist and shortly present the related domains, in order to integrate better the
knowledge and skills acquired by studying the digital system design domain
In the next chapter
is a friendly introduction in both, digital systems and a HDLs (Hardware Description
Languages) used to describe, simulate, and synthesized them. The HDL selected for this
book is called Verilog. The main topics are:
• the distinction between combinational and sequential circuits
• the two ways to describe a circuit: behavioral or structural
• how digital circuits behave in time.
3
4
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
Talking about Apple, Steve said, “The system
is there is no system.” Then he added, “that
does’t mean we don’t have a process.” Making the distinction between process and system
allows for a certain amount of fluidity, spontaneity, and risk, while in the same time it
acknowledges the importance of defined roles
and discipline.
J. Young & W. Simon1
A process is a strange mixture of rationally established rules, of imaginatively driven chaos,
and of integrative mystery.
A possible good start in teaching about a complex domain is an informal one. The main
problems are introduced friendly, using an easy approach. Then, little by little, a more rigorous
style will be able to consolidate the knowledge and to offer formally grounded techniques. The
digital domain will be disclosed here alternating informal “bird’s-eye views” with simple, formalized real stuff. Rather than imperatively presenting the digital domain we intend to disclose
it in small steps using a project oriented approach.
1.1
Framing the digital design domain
Digital domain can be defined starting from two different, but complementary view points: the
structural view point or the functional view point. The first version presents the digital domain
as part of electronics, while the second version sees the digital domain as part of computer
science.
1.1.1
Digital domain as part of electronics
Electronics started as a technical domain involved in processing continuously variable signals.
Now the domain of electronics is divided in two sub-domains: analogue electronics, dealing with
continuously variable signals and digital electronics based on elementary signals, called bits,
which take only two different levels 0 and 1, but can be used to compose any complex signals.
Indeed, a sequence of n bits is used to represent any number between 0 and 2n − 1, while a
sequence of numbers can be used to approximate a continuously variable signal.
Let be, for example, the analogue, continuously variable, signal in Figure 1.1. It can be
approximated by values sampled in discrete moments of time determined by the positive transitions of a square wave periodic signal called clock. It switches with a frequency of 1/T . The
value of the signal is measured in units u (for example, u = 100mV or u = 10µA). The operation
is called analog to digital conversion, and it is performed by an analog to digital converter
– ADC. Results the following sequence of numbers:
s(0 × T )
s(1 × T )
s(2 × T )
s(3 × T )
s(4 × T )
= 1 units ⇒ 001,
= 4 units ⇒ 100,
= 5 units ⇒ 101,
= 6 units ⇒ 110,
= 6 units ⇒ 110,
1
They co-authored iCon. Steve Jobs. The Greatest Second Act in the History of Business, an unauthorized
portrait of the co-founder of Apple.
1.1. FRAMING THE DIGITAL DESIGN DOMAIN
5
s(5 × T ) = 6 units ⇒ 110,
s(6 × T ) = 6 units ⇒ 110,
s(7 × T ) = 6 units ⇒ 110,
s(8 × T ) = 5 units ⇒ 101,
s(9 × T ) = 4 units ⇒ 100,
s(10 × T ) = 2 units ⇒ 010,
s(11 × T ) = 1 units ⇒ 001,
s(12 × T ) = 1 units ⇒ 001,
s(13 × T ) = 1 units ⇒ 001,
s(14 × T ) = 1 units ⇒ 001,
s(15 × T ) = 2 units ⇒ 010,
s(16 × T ) = 3 units ⇒ 011,
s(17 × T ) = 4 units ⇒ 100,
s(18 × T ) = 5 units ⇒ 101,
s(19 × T ) = 5 units ⇒ 101,
s(20 × T ) = 5 units ⇒ 101,
...
s(t)
✻
6×u
5×u
4×u
3×u
2×u
1×u
✲
t
clock
✻
1
✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻
✲
✛T✲
t
0
Figure 1.1: Analogue to digital. The analog signal is sampled at each T using the unit measure u.
s(t)
13 × u/2
12 × u/2
11 × u/2
10 × u/2
9 × u/2
8 × u/2
7 × u/2
6 × u/2
5 × u/2
4 × u/2
3 × u/2
2 × u/2
1 × u/2
✻
✲
clock
1
0
✻
✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻✻
✲
t
t
Figure 1.2: More accurate analogue to digital. The analogous signal is sampled at each T /2
using the unit measure u/2.
6
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
If a more accurate representation is requested, then both, the sampling period, T and
the measure units u must be reduced. For example, in Figure 1.2 both, T and u are halved.
A better approximation is obtained with the price of increasing the number of bits used for
representation. Each sample is represented on 4 bits instead of 3, and the number of samples is
doubled. This second, more accurate, conversion provides the following stream of binary data:
<0011, 0110, 1000, 1001, 1010, 1011, 1011, 1100, 1100, 1100, 1100, 1100, 1100,
1100, 1011, 1010, 1010, 1001, 1000, 0101, 0100, 0011, 0010, 0001, 0001, 0001,
0001, 0001, 0010, 0011, 0011, 0101, 0110, 0111, 1000, 1001, 1001, 1001, 1010,
1010, 1010, ...>
An ADC is characterized by two main parameters:
• the sampling rate: expressed in samples per second – SPS – or by the sampling frequency
– 1/T
• the resolution: the number of bits used to represent the value of a sample
Commercial ADC are provided with resolution in the range of 6 to 24 bits, and the sample rate
exceeding 3 GSPS (giga SPS). At the highest sample rate the resolution is limited to 12 bits.
analogInput 1
ADC1
✲
✻
✲ DAC1
analogOutput 1
✻
DIGITAL
SYSTEM
ADCM
analogInput M
✻
clock
✲
✲ DACN
✻
analogOutput N
✻
Figure 1.3: Generic digital electronic system.
The generic digital electronic system is represented in Figure 1.3, where:
• analogInput i, for i = 1, . . . M , provided by various sensors (microphones, ...), are sent to
the input of M ADCs
• ADCi converts analogInput i in a stream of binary coded numbers, using an appropriate
sampling interval and an appropriate number of bits for approximating the level of the
input signal
• DIGITAL SYSTEM processes the M input streams of data providing on its outputs N
streams of data applied on the input of N Digital-to-Analog Converters (DAC)
• DACj converts its input binary stream to analogOutput j
• analogOutput j, for j = 1, . . . N , are the outputs of the electronic system used to drive
various actuators (loudspeakers, ...)
1.1. FRAMING THE DIGITAL DESIGN DOMAIN
7
• clock is the synchronizing signal applied to all the components of the system; it is used to
trigger the moments when the signals are ready to be used and the subsystems are ready
to use the signals.
While loosing something at conversion, we are able to win at the level of processing. The
good news is that the loosing process is under control, because both, the accuracy of conversion
and of digital processing are highly controllable.
1.1.2
Digital domain as part of computer science
The domain of digital systems is considered, form the functional view point, as part of computing
science. This, possible view point presents the digital systems as systems which compute their
associated transfer functions. A digital system is seen as a sort of electronic system because
of the technology used now to implement it. But, from a functional view point it is simply a
computational system, because future technologies will impose maybe different physical ways to
implement it (using, for example, different kinds of nano-technologies, bio-technologies, photonbased devices, . . ..). Therefore, we decided to start our approach using a functionally oriented
introduction in digital systems, considered as a sub-domain of computing science. Technology
dependent knowledge is always presented only as a supporting background for various design
options.
Where can be framed the domain of digital systems in the larger context of computing
science? A simple, informal definition of computing science offers the appropriate context for
introducing digital systems.
ALGORITHMS
digital systems
❘
✠
HARDWARE
✠
TECHNOLOGY
abstract
❘
LANGUAGES
❘
APPLICATIONS
❄
actual
Figure 1.4: What is computer science? The domain of digital systems provides techniques for
designing the hardware involved in computation.
Definition 1.1 Computer science (see also Figure 1.4) means to study:
• algorithms,
• their hardware embodiment
• and their linguistic expression
with extensions toward
• hardware technologies
• and real applications. ⋄
8
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
The initial and the most abstract level of computation is represented by the algorithmic level.
Algorithms specify what are the steps to be executed in order to perform a computation. The
most actual level consists in two realms: (1) the huge and complex domain of the application
software and (2) the very tangible domain of the real machines implemented in a certain technology. Both contribute to implement real functions (asked, or aggressively imposed, my the
so called free market). An intermediate level provides the means to be used for allowing an
algorithm to be embodied in a physical structure of a machine or in an informational structure
of a program. It is about (1) the domain of the formal programming languages, and (2) the
domain of hardware architecture. Both of them are described using specific and rigorous formal
tools.
The hardware embodiment of computations is done in digital systems. What kind of formal tools are used to describe, in the most flexible and efficient way, a complex digital system?
Figure 1.5 presents the formal context in which the description tools are considered. Pseudocode language is an easy to understand and easy to use way to express algorithms. Anything
about computation can be expressed using this kind of languages. By the rule, in a pseudo-code
language we express, for our (human) mind, preliminary, not very well formally expressed, ideas
about an algorithm. The “main user” of this kind of language is only the human mind. But,
for building complex applications or for accessing advanced technologies involved in building big
digital systems, we need refined, rigorous formal languages and specific styles to express computation. More, for a rigorous formal language we must take into account that the “main user” is
a merciless machine, instead of a tolerant human mind. Elaborated programming languages
(such as C++, Java, Prolog, Lisp) are needed for developing complex contexts for computation
and to write using them real applications. Also, for complex hardware embodiments specific
hardware description languages, HDL, (such as Verilog, VHDL, SystemC) are proposed.
PSEUDO-CODE
LANGUAGE
✠
HARDWARE DESCRIPTION
LANGUAGES
❘
PROGRAMMING
LANGUAGES
Figure 1.5: The linguistic context in computer science. Human mind uses pseudo-code
languages to express informally a computation. To describe the circuit associated with the computation
a rigorous HDL (hardware description language) is needed, and to describe the program executing the
computation rigorous programming languages are used.
Both, general purpose programming languages and HDLs are designed to describe something
for another program, mainly for a compiler. Therefore, they are more complex and rigorous than
a simple pseudo-code language.
The starting point in designing a digital system is to describe it using what we call a specification, shortly, a spec. There are many ways to specify a digital system. In real life a
hierarchy of specs are used, starting from high-level informal specs, and going down until the
most detailed structural description is provided. In fact, de design process can be seen as a
1.2. DEFINING A DIGITAL SYSTEM
9
stream of descriptions which starts from an idea about how the new object to be designed behaves, and continues with more detailed descriptions, in each stage more behavioral descriptions
being converted in structural descriptions. At the end of the process a full structural description
is provided. The design process is the long way from a spec about what we intend to do to
another spec describing how our intention can be fulfilled.
At one end of this process there are innovative minds driven by the will to change the world.
In these imaginative minds there is no knowledge about “how”, there is only willingness about
“what”. At the other end of this process there are very skilled entities “knowing” how to do very
efficiently what the last description provides. They do not care to much about the functionality
they implement. Usually, they are machines driven by complex programs.
In between we need a mixture of skills provided by very well instructed and trained people.
The role of the imagination and of the very specific knowledge are equally important.
How can be organized optimally a designing system to manage the huge complexity of this
big chain, leading from an idea to a product? There is no system able to manage such a complex
process. No one can teach us about how to organize a company to be successful in introducing,
for example, a new processor on the real market. The real process of designing and imposing a
new product is trans-systemic. It is a rationally adjusted chaotic process for which no formal
rules can ever provided.
Designing a digital system means to be involved in the middle of this complex process,
usually far away from its ends. A digital system designer starts his involvement when the
specs start to be almost rigorously defined, and ends its contribution before the technological
borders are reached.
However, a digital designer is faced in his work with few level of descriptions during the
execution of a project. More, the number of descriptions increases with the complexity of the
project. For a very simple project, it is enough to start from a spec and the structural description
of the circuit can be immediately provided. But for a very complex project, the spec must be
split in specs for sub-systems, each sub-system must be described first by its behavior. The
process continue until enough simple sub-systems are defined. For them structural descriptions
can be provided. The entire system is simulated and tested. If it works synthesisable descriptions
are provided for each sub-system.
A good digital designer must be well trained in providing various description using an HDL.
She/he must have the ability to make, both behavioral and structural descriptions for circuits
having any level of complexity. Playing with inspired partitioning of the system, a skilled
designer is one who is able to use appropriate descriptions to manage the complexity of the
design.
1.2
Defining a digital system
Digital systems belong to the wider class of the discrete systems (systems having a countable
number of states). Therefore, a general definition for digital system can be done as a special
case of discrete system.
Definition 1.2 A digital system, DS, in its most general form is defined by specifying the five
components of the following quintuple:
DS = (X, Y, S, f, g)
where: X ⊆ {0, 1}n is the input set of n-bit binary configurations, Y ⊆ {0, 1}m is the output
set of m-bit binary configurations, S ⊆ {0, 1}q is the set of internal states of q-bit binary
configurations,
f : (X × S) → S
10
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
is the state transition function, and
g : (X × S) → Y
is the output transition function.
⋄
X
n
❄
❄
State
transition function
f : (X × S) → S
q
q
S+
❄
clock ✲
State memory
S
❄
❄
Output
transition function
g : (X × S) → Y
m
❄
Y
Figure 1.6: Digital system.
A digital system (see Figure 1.6) has two simultaneous evolutions:
• the evolution of its internal state which takes into account the current internal state and
the current input, generating the next state of the system
• the evolution of its output, which takes into account the current internal state and the
current input generating the current output.
The internal state of the system determines the partial autonomy of the system. The system
behaves on its outputs taking into account both, the current input and the current internal
state.
Because all the sets involved in the previous definition have the form {0, 1}b , each of the
b one-bit input, output, or state evolves in time switching between two values: 0 and 1. The
previous definition specifies a system having a n-bit input, an m-bit output and a q-bit internal
state. If xt ∈ X = {0, 1}n , yt ∈ Y = {0, 1}m , st ∈ S = {0, 1}q are values on input, output, and
of state at the discrete moment of time t, then the behavior of the system is described by:
st = f (xt−1 , st−1 )
yt = g(xt , st )
1.2. DEFINING A DIGITAL SYSTEM
11
While the current output is computed from the current input and the current state, the current state was computed using the previous input and the previous state. The two functions
describing a discrete system belong to two distinct class of functions:
sequential functions : used to generate a sequence of values each of them iterated from
its predecessor (an initial value is always provided, and the i-th value cannot be computed without computing all the previous i − 1 values); it is about functions such as
st = f (xt−1 , st−1 )
non-sequential functions : used to compute an output value starting only from the current
values applied on its inputs; it is about functions such as yt = g(xt , st ).
Depending on how the functions f and g are defined results a hierarchy of digital systems.
More on this in the next chapters.
The variable time is essential for the formal definition of the sequential functions, but for
the formal definition of the non-sequential ones it is meaningless. But, for the actual design of
both, sequential and non-sequential function the time is a very important parameter.
Results the following requests for the simplest embodiment of an actual digital systems:
• the elements of the sets X, Y and S are binary cods of n, m and q bits – 0s and 1s –
which are be codded by two electric levels; the current technologies work with 0 Volts for
the value 0, and with a tension level in the range of 1-2 Volts for the value 1; thus, the
system receives on its inputs:
Xn−1 , Xn−2 , . . . X0
stores the internal state of form:
Sq−1 , Sq−2 , . . . S0
and generate on its outputs:
Ym−1 , Ym−2 , . . . Y0
where: Xi , Sj , Yk ∈ {0, 1}.
• physical modules (see Figure 1.7), called combinational logic circuits – CLC –, to
compute functions like f (xt , st ) or g(xt , st ), which continuously follow, by the evolution of
their output values delayed with the propagation time tp , any change on the inputs xt and
st (the shaded time interval on the wave out represent the transient value of the output)
• a “master of the discrete time” must be provided, in order to make consistent suggestions
for the simple ideas as “previous”, “now”, “next”; it is about the special signal, already
introduced, having form of a square wave periodic signal, with the period T which swings
between the logic level 0 and the logic level 1; it is called clock, and is used to “tick” the
discrete time with its active edge (see Figure 1.8 where a clock signal, active on its positive
edge, is shown)
• a storing support to memorize the state between two successive discrete moments of time
is required; it is the register used to register, synchronized with the active edge of the
clock signal, the state computed at the moment t − 1 in order to be used at the next
moment, t, to compute a new state and a new output; the input must be stable a time
interval tsu (set-up time) before the active edge of clock, and must stay unchanged th (hold
time) after; the propagation time after the clock is tp .
12
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
in
✻
Xn−1 Xn−2 . . . X1 X0 Ym−1 Ym−2 . . .
0
0
0
0
0
0
...
...
...
0
0
1
0
1
0
...
1010
...
0100
...
0101
...
0
1
1
1
0
1
0101
0110
1010
...
1
1
1
1
1
1
...
...
...
in
in1
in2
✲
❄
time
out
CLC
✻
...
...
...
❄
F(in1)
out
a.
✛
F(in2)
✲
tp
✲
time
c.
b.
Figure 1.7: The module for non-sequential functions. a. The table used to define the function
as a correspondence between all input binary configurations in and binary configurations out. b. The
logic symbol for the combinatorial logic circuit – CLC – which computes out = F(in). c. The wave
forms describing the time behaviour of the circuit.
clock ✛
✻
Tclock
✲
✻
✻
ti−2
ti−1
✻
✻
✻
ti+1
ti+2
✲
time
ti
Figure 1.8: The clock. This clock signal is active on its positive edge (negative edge as active edge is
also possible). The time interval between two positive transitions is the period Tclock of the clock signal.
Each positive transition marks a discrete moment of time.
(More complex embodiment are introduced later in this text book. Then, the state will have a
structure and the functional modules will result as multiple applications of this simple definition.)
The most complex part of defining a digital system is the description of the two functions
f and g. The complexity of defining how the system behaves is managed by using various
Hardware Description Languages – HDLs. The formal tool used in this text book is the Verilog
HDL. The algebraic description of a digital system provided in Definition 1.2 will be expressed
as the Verilog definition.
Definition 1.3 A digital system is defined by the Verilog module digitalSystem, an object
which consists of:
external connections : lists the type, the size and the name of each connection
internal resources : of two types, as follows
storage resources : one or more registers used to store (to register) the internal state
of the system
functional resources : of two types, computing the transition functions for
state : generating the nextState value from the current state and the current input
output : generating the current output value from the current state and the current
input
1.2. DEFINING A DIGITAL SYSTEM
clock
✻
13
ti+1
✻
✲
dataIni = dataOuti+1
time
dataIn
✻
r
✲ tsu ✛
in1
✲
❄
✲
th ✛
time
clock
dataOut
in
register
out
r
❄
✻
dataOuti = dataIni−1
✛ tp ✲
in1
✲
time
a.
b.
Figure 1.9: The register. a. The wave forms describing timing details about how the register swithces
around the active edge of clock. b. The logic symbol used to define the static behaviour of the register
when both, inputs and outputs are stable between two active edges of the clock signal.
The simplest Verilog definition of a digital system follows (see Figure 1.10). It is simple
because the state is defined only by the content of a single q-bit register (the state has no structure)
and the functions are computed by combinational circuits..
There are few keywords which any text editor emphasize using bolded and colored letters:
• module and endmodule are used to delimit the definition of an entity called module
which is an object with inputs and outputs
• input denotes an input connection whose dimension, in number of bits, is specified in the
associated square brackets as follows: [n-1:0] which means the bits are indexed from n-1
to 0 from left to right
• output denotes an output connection whose dimension, in number of bits, is specified in
the associated square brackets as follows: [n-1:0] which means the bits are indexed from
n-1 to 0 from left to right
• reg [n-1:0] defines a storage element able to store n bits synchronized with the active edge
of the clock signal
• wire [n-1:0] defines a n-bit internal connection used to interconnect two subsystems in
the module
• always @(event) action specifies the action action triggered by the event event; in our
first example the event is the positive edge of clock (posedge clock) and the action is:
the state register is loaded with the new state stateRegister <= nextState
• ‘include is the command used to include the content of another file
• "fileName.v" specifies the name of a Verilog file
The following two dummy modules are used to synthesize the top level of the system; their
content is not specified, because we do not define a specific system; only the frame of a possible
definition is provided.
14
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
Figure 1.10: The top module for the general form of a digital system.
module stateTransition #(‘include "0_parameter.v")
(output [q-1:0] next ,
input [q-1:0] state,
input [n-1:0] in
);
// describe here the state transition function
endmodule
module outputTransition #(‘include "0_parameter.v")
(output [m-1:0] out ,
input [q-1:0] state,
input [n-1:0] in
);
// describe here the output transition function
endmodule
where the content of the file 0 parameter.v is:
parameter n = 8, // the input is coded on 8 bits
m = 8, // the output is coded on 8 bits
q = 4 // the state is coded on 4 bits
It must be actually defined for synthesis reasons. The synthesis tool must “know” the size of the
internal and external connections, even if the actual content of the internal modules is not yet
specified.
⋄
The synthesis of the generic structure, just defined, is represented in Figure 1.11, where there
are represented three (sub-)modules:
• the module fd which is a 4-bit state register whose output is called state[3:0]; it stores
the internal state of the system
1.3. OUR FIRST TARGET
15
Figure 1.11: The result of the synthesis for the module digitalSystem.
• the module stateTransition instantiated as stateTrans; it computes the value of the
state to be loaded in the state register in triggered by the next active (positive, in our
example) edge of clock; this module closes a loop over the state register
• the module outputTransition instantiated as outTrans; it computes the output value
from the current states and the current input (for some applications the current input
is not used directly to generate the output, its contribution to the output being delayed
through the state register).
The internal modules are interconnected using also the wire called next. The clock signal
is applied only to the register. The register module and the module stateTransition compute a sequential function, while the outputTransition module computes a non-sequential,
combinational function.
1.3
Our first target
We will pass twice through the matter of digital systems. Every time we have a specific target.
In this section the first target is presented. It consists of a simple system used to introduce the
basic knowledge about simple and small digital circuits. Our target has the form of a simple
specific circuit. It is about a digital pixel corrector.
A video sensor is a circuit built as a big array of cells which provides the stream of binary
numbers used to represent a picture or a frame in a movie. To manufacture such a big circuit
without any broken cell is very costly. Therefore, circuits with a small number of isolated wrong
cells, providing the erroneous signal zero, are accepted, because it is easy to make few corrections
on an image containing millions of pixels. The error manifests by providing a zero value for the
light intensity. The stream of numbers generated by Video Sensor is applied to the input of
Digital Pixel Corrector (see Figure 1.12) which performs the correction. It consists of detecting
the zeroes in the digital video stream (s(t) = 0) and of replacing them with the corrected value
s’(t) obtained by interpolation. The simplest interpolation uses s(t-1) and s(t+1) as follows:
s’(t) = if (s(t) = 0)
then (s(t-1) + s(t+1))/2
else s(t)
The circuits checks if the input is zero (if (s(t) = 0)). If not, s(t) goes through. If the input is
wrong the circuit provides the arithmetic mean computed in the smallest neighborhood, s(t-1)
and s(t+1).
16
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
Video Sensor
Sensor & ADC
✲ in[3:0]
Digital
out[3:0]
Pixel
Corrector
Digital System
✲
clock
Figure 1.12: Pixel correction system. The digital stream of pixels is corrected by substituting the
wrong values by interpolation.
The previous simple computation is made by the circuit Digital Pixel Corrector. In Figure
1.13 the wave form s(t) represents the light intensity, while s(t) represents the discrete stream
of samples to be converted in numbers. The stream s(t) follows the wave s(t), excepting in one
point where the value provided by the conversion circuit is, by error, zero.
In our simple example, the circuit Digital Pixel Corrector receives from the convertor a
stream of 4-bit numbers. The data input of the circuit is in[3:0]. It receives also the clock
signal. The active edge of clock is, in this example, the positive transition.
The stream of numbers received by the circuit is: 7, 9, 10, 11, 12, 11, 10, 8, 5, 3,
2, 2, 0, 5, 7, 9, 11, 12, 12, 13, .... On the 13-th position the wrong value, 0, is received. It will be substituted, in the output sequence, with the integer part of (2 + 5)/2.
In the first clock cycle, in[3:0] takes the value 0111, i.e., in[3] = 0, in[2] = 1, in[1]
= 1, in[0] = 1. In the second clock cycle in[3:0] = 1010, and so on. Thus, on each binary
input, in[3], ... in[0], a specific square wave form is applied. They consists of transitions
between 0 and 1. In real circuits, 0 is represented by the voltage 0, while 1 by a positive voltage
DDD = 1 . . . 2V .
A compact representation of the four wave form, in[3], in[2], in[1], in[0], is shown in
the synthetic wave form in[3:0]. It is a conventional representation. The two overlapped wave
suggest the transition of the input value in[3:0], while ”inside” each delimited time interval
the decimal value of the input is inserted. The simultaneous transitions, used to delimit a time
interval, signify the fact that some bits could switch form 0 to 1, while others from 1 to 0.
The output behaviour is represented in the same compact way. Our circuit transmits the
input stream to the output of the circuit with a delay of two clock cycles! Why? Because, to
compute the current output value out(t) the circuit needs the previous input value and the next
input value. Any value must “wait” the next one to be “loaded” in the correction circuit, while
the previous is already memorized. “Waiting” and “already memorized” means to be stored in
the internal state of the system. Thus, the internal state consists of three sub-states: ss1[3:0]
= in(t-1), ss2[3:0] = in(t-2) and ss3[3:0] = in(t-3), i.e., state is the concatenation of
the three sub-states :
state = {ss3[3:0], ss2[3:0], ss1[3:0]}
In each clock cycle the state is updated with the current input, as follows:
if state(t) = {ss3, ss2, ss1} then state(t+1) = {ss2, ss1, in}
Thus, the circuit takes into account simultaneously three successive values from the input stream,
all stored, concatenated, in the internal state register. The previously stated interpolation
relation is now reconsidered, for an actual implementation using a digital system, as follows:
out(t) = if (ss2 = 0)
1.3. OUR FIRST TARGET
17
s(t), s(t)
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
15 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
14 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
13 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
12 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
11 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
.. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
10 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . .error
9 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
8 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
7 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
6 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
5 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
4 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
3 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
2 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
1 . . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . . . .. . . . .
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
0
...
...
...
...
...
...
...
... time
...
...
...
...
...
...
...
...
...
...
...
...
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
clock
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
...
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
...
.
.
.
.
.
.
.. time
.
.
.
.
.
.
.
.
.
.
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
✻
❄
✲
✻
in[3]
in[2]
in[1]
in[0]
in[3:0]
..
..
..
.. 0
..
..
..
..
..
.. 1
..
..
..
..
..
.. 1
..
..
..
..
..
.. 1
..
..
..
..
..
..
..
7..
..
..
..
..
..
..
X
...
✻
✻
✻
✻
✻
out[3:0]
✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻ ✻✲
✻
..
..
..
..
..
..
..
...
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
.
.. 1 .. 1 .. 1 .. 1 .. 1 .. 1 .. 1 ... 0
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
..
.. 0 .. 0 .. 0 .. 1 .. 0 .. 0 .. 0 ... 1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
..
.. 0 .. 1 .. 1 .. 0 .. 1 .. 1 .. 0 ... 0
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
..
..
..
..
.. 1 .. 0 .. 1 .. 0 .. 1 .. 0 .. 0 ... 1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
9..
10
11
12
11
10
8..
5..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
7.
9.
10
11
12
11
10
X
..
..
..
..
..
..
..
..
.
.
.
.
..
.
.
.
.
.
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
3..
..
..
..
..
..
8..
..
.
0
0
1
1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
2..
..
..
..
..
..
5..
..
.
0
0
1
0
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
2..
..
..
..
..
..
3..
..
.
..
..
..
..
0 “0”
..
..
..
..
..
..
0 “0”
..
..
..
..
..
..
1 “0”
..
..
..
..
..
..
0 “0”
..
..
error
...
..
..
..
..
“0”
..
..
..
..
..
.
2..
..
.
❄
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
5..
..
..
..
..
..
2..
..
.
0
1
0
1
..
..
..
.. 0
..
..
..
..
..
.. 1
..
..
..
..
..
.. 1
..
..
..
..
..
.. 1
..
..
..
..
..
..
..
7..
..
..
..
..
..
..
3
...
✻
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
9..
..
..
..
..
..
5..
..
.
1
0
0
1
..
..
..
..
..
..
..
..
..
..
.
..
.. 1 .. 1 .. 1 ... 1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
.. 0 .. 1 .. 1 ... 1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
.. 1 .. 0 .. 0 ... 0
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
..
..
..
.. 1 .. 0 .. 0 ... 1
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
11
12
12
13
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
.7.
.9.
11
12
..
..
..
..
.
.
.
.
.
✲
time
✲
time
✲
time
✲
time
✲
time
✲
time
corrected
Figure 1.13: The interpolation process.
then (ss1 + ss3)/2
else ss2
If no wrong value on the stream, then the current output takes the value of the input received
two cycles before: one to load it as ss1 and another to move it in ss2(two clock cycles delay, or
two-clock latency). Else, the current output value is (partially) “restored” from the other two
sub-states of the system ss1 and ss3, first just received triggered by the last active edge of the
clock, and the second loaded two cycles before.
Now, let us take the general definition of a digital system and adapt it for designing the
Digital Pixel Corrector circuit. First, the file used to define the parameters – 0 parameter.v –
is modified according to the size of the external connections and of state. For our application,
the file takes the name 0 paramPixelCor.v, having the following content:
parameter n = 4, // the input is coded on 4 bits
m = 4, // the output is coded on 4 bits
q = 12 // the state is coded on 12 bits (to store three 4-bit values)
In the top module digitalSystem little changes are needed. The module’s name is changed to
pixelCorrector, the included parameter file is substituted, and the module outputTransition
18
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
is simplified, because the output value does not depend directly by the input value. The resulting
top module is:
module pixelCorrector #(‘include "0_paramPixelCor.v")
(output [m-1:0] out ,
input [n-1:0] in
,
input
clock);
reg [q-1:0] state;
wire [q-1:0] next ;
stateTransition stateTrans(next ,
state,
in
);
always @(posedge clock) state <= next;
outputTransition outTrans(out ,
state);
endmodule
Now, the two modules defining the transition functions must be defined according to the functionality desired for the system. State transition means to shift left the content of the state
register n positions and on the freed position to put the input value. The output transition
is a conditioned computation. Therefore, for our pixelCorrector module the combinational
modules have the following form:
module stateTransition #(‘include "0_paramPixelCor.v")
(output [q-1:0] next ,
input [q-1:0] state,
input [n-1:0] in
);
//state[2*n-1:0] is {in(t-2), in(t-1)}
assign next = {state[2*n-1:0], in};
endmodule
module outputTransition #(‘include "0_paramPixelCor.v")
(output reg [m-1:0] out ,
input
[q-1:0] state);
//state[n-1:0]
is in(t-1)
//state[2*n-1:n] is in(t-2)
//state[q-1:2*n] is in(t-3)
always @(state) // simpler: "always @(*)"
if (state[2*n-1:n] == 0) out = (state[n-1:0] + state[q-1:2*n])/2;
else
out = state[2*n-1:n]
;
endmodule
In order to verify the correctness of our design, a simulation module is designed. The clock signal
and the input stream are generated and applied to the input of the top module, instantiated
under the name dut (devices under test). A monitor is used to access the behavior of the circuit.
module testPixelCorrector #(‘include "0_paramPixelCor.v");
reg
clock;
reg [3:0] in
;
wire [3:0] out ;
initial begin
clock = 0
;
forever #1 clock = ~clock;
1.3. OUR FIRST TARGET
end
initial begin
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
in = 4’b0111;
in = 4’b1001;
in = 4’b1010;
in = 4’b1011;
in = 4’b1100;
in = 4’b1011;
in = 4’b1010;
in = 4’b1000;
in = 4’b0101;
in = 4’b0011;
in = 4’b0010;
in = 4’b0010;
in = 4’b0000; // the error
in = 4’b0101;
in = 4’b0111;
in = 4’b1001;
in = 4’b1011;
in = 4’b1100;
in = 4’b1100;
in = 4’b1101;
$stop;
end
pixelCorrector dut(out ,
in
,
clock);
initial $monitor ("time = %d state = %b_%b_%b out = %b",
$time, dut.state[q-1:2*n], dut.state[2*n-1:n],dut.state[n-1:0], out);
endmodule
The monitor provides the following stream of data:
posedge clock
ss3 ss2 ss1
out = if (ss2=0) (ss3+ss1)/2 else ss2
----------------------------------------------------------------------------# time = 0
state = xxxx_xxxx_xxxx
out = xxxx
# time = 1
state = xxxx_xxxx_0111
out = xxxx
# time = 3
state = xxxx_0111_1001
out = 0111
# time = 5
state = 0111_1001_1010
out = 1001
# time = 7
state = 1001_1010_1011
out = 1010
# time = 9
state = 1010_1011_1100
out = 1011
# time = 11
state = 1011_1100_1011
out = 1100
# time = 13
state = 1100_1011_1010
out = 1011
# time = 15
state = 1011_1010_1000
out = 1010
# time = 17
state = 1010_1000_0101
out = 1000
# time = 19
state = 1000_0101_0011
out = 0101
# time = 21
state = 0101_0011_0010
out = 0011
# time = 23
state = 0011_0010_0010
out = 0010
# time = 25
state = 0010_0010_0000
out = 0010
# time = 27
state = 0010_0000_0101
out = 0011 // = (2+5)/2 = 3; ERROR!
# time = 29
state = 0000_0101_0111
out = 0101
# time = 31
state = 0101_0111_1001
out = 0111
19
20
#
#
#
#
CHAPTER 1. WHAT’S A DIGITAL SYSTEM?
time
time
time
time
=
=
=
=
33
35
37
39
state
state
state
state
=
=
=
=
0111_1001_1011
1001_1011_1100
1011_1100_1100
1100_1100_1101
out
out
out
out
=
=
=
=
1001
1011
1100
1100
while the wave form of the same simulation are presented in Figure 1.14. The output values
Figure 1.14: The wave forms provided by simulation.
correspond to the input values with a two clock cycle latency. Each nonzero input goes through
the output in two cycles, while the wrong, zero inputs generate the intepolated values in the
same two cycles (in our example 0 generates 3, as the integer mean value of 2 and 5). The
functions involved in solving the pixel correction are:
• the predicate function, state[2*n-1:n] == 0, used to detect the wrong value zero
• the addition, used to compute the mean value of two numbers
• the division, used to compute the mean value of two numbers
• the selection function, which accorting to the a predicate sends to the output one value or
another value
• storage function, triggered by the active edge of the clock signal
Our first target is to provide the knowledge about the actual circuits used in the previous
simple application. The previous description is a behavioral description. We must acquire the
ability to provide the corresponding structural description. How to add, how to compare, how to
select, how to store, and few other similar function will be investigated before going to approach
the final, more complex target.
1.4
Problems
Problem 1.1 How behaves the pixelCorrector circuit if the very first value received is zero?
How can be improved the circuit to provide a better response?
Problem 1.2 How behaves the pixelCorrector circuit if if two successive wrong zeroes are
received to the input? Provide an improvement for this situation.
Problem 1.3 What is the effect of the correction circuit when the zero input comes form an
actual zero light intensity?
Problem 1.4 Synthesize the module pixelCorrector and identify in the RTL Schematic provided by the synthesis tool the functional components of the design. Explain the absences, if
any.
Problem 1.5 Design a more accurate version of the pixel correction circuit using a more complex interpolation rule, which takes into account an extended neighborhood. For example, apply
the the following interpolation:
s′ (t) = 0.2s(t − 2) + 0.3s(t − 1) + 0.3s(t + 1) + 0.2s(t + 2)
Chapter 2
DIGITAL CIRCUITS
In the previous chapter
the concept of digital system was introduced by:
• differentiating it from analog system
• but integrating it, in the same time, in a hybrid electronic system
• defining formally what means a digital system
• and by stating the first target of this text book: the introducing the basic small and
simple digital circuits
In this chapter
general definitions related with the digital domain are used to reach the following targets:
• to frame the digital system domain in the larger area of the information technologies
• to present different ways the digital approach is involved in the design of the real
market products
• to enlist and shortly present the related domains, in order to integrate better the
knowledge and skills acquired by studying the digital system design domain
In the next chapter
is a friendly introduction in both, digital systems and a HDLs (Hardware Description
Languages) used to describe, simulate, and synthesized them. The HDL selected for this
book is called Verilog. The main topics are:
• the distinction between combinational and sequential circuits
• the two ways to describe a circuit: behavioral or structural
• how digital circuits behave in time.
21
22
CHAPTER 2. DIGITAL CIRCUITS
In the previous chapter we learned, from an example, that a simple digital system, assimilated
with a digital circuit, is built using two kinds of circuits:
• non-sequential circuits, whose outputs follow continuously, with a specific delay, the evolution of input variable, providing a “combination” of input bits as the output value
• sequential circuits, whose output evolve triggered by the active edge of the special signal
called clock which is used to determine the “moment” when the a storage element changes
its content.
Consequently, in this chapter are introduced, by simple examples and simple constructs, the two
basic types of digital circuits:
• combinational circuits, used to compute fcomb : X → Y , defined in X = {0, 1}n with
values in Y = {0, 1}m , where fcomb (x(t)) = y(t), with x(t) ∈ X, y(t) ∈ Y representing two
values generated in the same discrete unit of time t (discrete time is “ticked” by the active
edge of clock)
• storage circuits, used to design sequential circuits, whose outputs follow the input values
with the delay of one clock cycle; fstore : X → X, defined in X = {0, 1}n with values in
X = {0, 1}n , where fstore (x(t)) = x(t − 1), with x(t), x(t − 1) ∈ X, representing the same
value considered in two successive units of time, t − 1 and t.
While a combinational circuit computes continuously its outputs according to each input change,
the output of the storage circuit changes only triggered by the active edge of clock.
In this chapter, the first section is for combinational circuits which are introduced by examples, while, in the second section, the storage circuit called register is generated step by step
starting from the simplest combinational circuits.
2.1
Combinational circuits
Revisiting the Digital Pixel Corrector circuit, lets take the functional description of the output
function:
if (state[2*n-1:n] == 0) out = (state[n-1:0] + state[q-1:2*n])/2;
else
out = state[2*n-1:n]
;
The previous form contains the following elementary functions:
• test function: state[2*n-1:n] == 0, defined in {0, 1}n with value in {0, 1}
• selection function:
if (test) out = action1;
else
out = action2;
defined in the Cartesian product ({0, 1} × {0, 1}n × {0, 1}n ) with values in {0, 1}n
2.1. COMBINATIONAL CIRCUITS
23
• Add function: state[n-1:0] + state[q-1:2*n], defined in ({0, 1}n ×{0, 1}n ) with value
in {0, 1}n
• Divide by 2 function: defined in {0, 1}n with value in {0, 1}n .
In Appendix D, section Elementary circuits: gates basic knowledge about Boolean logic and
the associated logic circuits are introduced. We use simple functions and circuits, like AND,
OR, NOT, XOR, ..., to design the previously emphasized combinational functions.
2.1.1
Zero circuit
The simplest test function tests if a n-bit binary configuration represents the number 0. The
function OR provides 1 if at least one of its inputs is 1, which means it provides 0 if all its inputs
are 0. Then, inverting – negating – the output of a n-input OR we obtain a circuit NOR – not
OR – whose output is 1 only when all its inputs are 0.
Definition 2.1 The n-input Zero circuit is a n-input NOR.
⋄
The Figure 2.1 represents few embodiment of the Zero circuit. The elementary, 2-input, Zero
circuit is represented in Figure 2.1a as a two-input NOR. For the n-input Zero circuit a n-input
NOR is requested (see Figure 2.1b) which can be implemented in two different ways (see Figure
2.1c and Figure 2.1d). One level NOR (see Figure 2.1b) with more than 4 inputs are impractical
(for reasons disclosed when we will enter in the physical details of the actual implementations).
in[1] in[0]
in[n-1:0]
in[7:0]
in[7:0]
in[n-1:0]
n
out=Zero(in[1:0])
out=Zero(in[n-1:0])
❄
Zero
❄
out = Zero(in[n-1:0])
out = Zero(in[7:0])
out = Zero(in[7:0])
a.
b.
c.
d.
e.
Figure 2.1: The Zero circuit. a. The 2-input Zero circuit is a 2-input NOR. b. The n-input Zero
circuit is a n-input NOR. c. The 8-input Zero circuit as a degenerated tree of 2-input ORs. d. The
8-input Zero circuit as a balanced tree of 2-input ORs. e. The logic symbol for the Zero circuit.
The two solution for the n-input NOR come from the two ways to expand an associative
logic function. It is about how the parenthesis are used. The first form (see Figure 2.1c) comes
from:
(a + b + c + d + e + f + g + h)′ = (((((((a + b) + c) + d) + e) + f ) + g) + h)′
generating a 7 level circuit (7 included parenthesis), while, the second form (see Figure 2.1d)
comes from:
(a+b+c+d+e+f +g+h)′ = ((a+b)+(c+d)+(e+f )+(g+h))′ = (((a+b)+(c+d))+((e+f )+(g+h)))′
24
CHAPTER 2. DIGITAL CIRCUITS
providing a 3 level circuit (3 included parenthesis). The number of gates used is the same for
the two solution. We expect that the second solution provide a faster circuit.
2.1.2
Selection
The selection circuit, called also multiplexer, is a three input circuit: a one-bit selection input
– sel –, and two selected inputs, one – in0 – selected to the output when sel=0 and another
– in1 – selected for sel=1. Let us take first the simplest case when both selected inputs are of
1 bit. This is the case for the elementary multiplexer, EMUX. If the input are: sel, in0, in1,
then the logic equation describing the logic circuit is:
out = sel’ · in0 + sel · in1
Then the circuit consists of one NOT, two ANDs and one OR as it is shown in Figure 2.2. The
AND gates are opened by selection signal, sel, allowing to send out the value applied on the
input in1, and by the negation of the selection signal signal, sel’, allowing to send out the value
applied on the input in0. The OR circuit “sum up” the outputs of the two ANDs, because only
one is “open” at a time.
in1
in0
sel
❄ ❄
✲
1
0
EMUX
out
sel
❄
out
a.
in1[n-1] in0[n-1]
❄
in1[n-2] in0[n-2]
❄
✲
EM U Xn−1
❄
b.
in1[0]
in0[0]
❄
❄
❄
✲
EM U Xn−2
✲
n
n
❄
❄
in1
EM U X0
✲ sel
in0
M U Xn
out
n
sel
❄
out[n-1]
❄
out[n-2]
c.
❄
❄
out[0]
d.
Figure 2.2: The selection circuit. a. The logic schematic for the elementary selector, EMUX
(elementary multiplexer). b. The logic symbol for EMUX. c. The selector (multiplexor) for n-bit words,
M U Xn . d. The logic symbol for M U Xn .
The selection circuit for two n-bit inputs is functionally (behaviorally) described by the
following Verilog module:
module ifThenElse #(parameter n = 4)
(output [n-1:0] out,
input
sel,
input [n-1:0] in1, in0);
assign out = sel ? in1 : in0;
endmodule
2.1. COMBINATIONAL CIRCUITS
25
In the previous code we decided to design a circuit for 4-bit data. Therefore, the parameter
n is set to 4 only in the header of the module.
The structural description is much more complex because it specifies all the details until
the level of elementary gates. The description has two modules: the top module – ifThenElse
– and the module describing the simplest select circuit – eMux.
module eMux(output out,
input sel, in1, in0);
wire
invSel;
not inverter(invSel, sel);
and and1(out1, sel, in1),
and0(out0, invSel, in0);
or outGate(out, out1, out0);
endmodule
module ifThenElse #(parameter n = 4)
(output [n-1:0] out,
input
sel,
input [n-1:0] in1, in0);
genvar i ;
generate for (i=0; i<n; i=i+1)
begin: eMUX
eMux selector(.out(out[i]),
.sel(sel
),
.in1(in1[i]),
.in0(in0[i]));
end
endgenerate
endmodule
The repetitive structure of the circuit is described using the generate form.
To verify the design a test module is designed. This module generate stimuli for the input
of the device under test (dut), and monitors the inputs and the outputs of the circuit.
module testIfThenElse #(parameter n = 4);
reg [n-1:0] in1, in0;
reg
sel;
wire [n-1:0] out;
initial begin
in1 = 4’b0101;
in0 = 4’b1011;
sel = 1’b0;
#1 sel = 1’b1;
#1 in1 = 4’b1100;
#1 $stop;
end
ifThenElse dut(out,
sel,
in1, in0);
initial $monitor
("time = %d sel = %b_in1 = %b in0 = %b out = %b",
$time, sel, in1, in0, out);
endmodule
26
CHAPTER 2. DIGITAL CIRCUITS
The result of simulation is:
# time = 0 sel = 0_in1 = 0101 in0 = 1011 out = 1011
# time = 1 sel = 1_in1 = 0101 in0 = 1011 out = 0101
# time = 2 sel = 1_in1 = 1100 in0 = 1011 out = 1100
The result of synthesis is represented in Figure 2.3.
Figure 2.3: The result of the synthesis for the module ifThenElse.
2.1.3
Adder
A n-bit adder is defined as follows, using a Verilog behavioral description:
module adder #(parameter n = 4)//
(output [n-1:0] sum,
//
output
carry, //
input
c,
//
input [n-1:0] a, b); //
defines a n-bit adder
the n-bit result
carry output
carry input
the two n-bit numbers
assign {carry, sum} = a + b + c;
endmodule
Fortunately, the previous module is synthesisable by the currently used synthesis tools. But,
in this stage, it is important for us to define the actuala internal structure of an adder. We start
from a 1-bit adder, whose output are described by the following Boolean equations:
sum = a ⊕ b ⊕ c
carry = a · b + a · c + b · c
where, a, b and c are one bit Boolean variables. Indeed, the sum output results as the sum of
three bits: the two numbers, a and b, and the carry bit, c, coming from the previous binary
range. As we know, the modulo 2 sum is performed by a XOR circuit. Then, a ⊕ b is the sum of
the two one-bit numbers. The result must be added with c – (a ⊕ b) ⊕ c – using another XOR
circuit. The carry signal is used by the next binary stage. The expression for carry is written
taking into account that the carry signal is one if at least two of the input bits are one: carry
2.1. COMBINATIONAL CIRCUITS
27
is 1 if a and b or a and c or b and c (the function is the majority function). Its expression is
embodied also in logic circuits, but not before optimizing its form as follows:
carry = a · b + a · c + b · c = a · b + c · (a + b) = a · b + c · (a ⊕ b)
Because (a ⊕ b) is already computed for sum, the circuit for carry requests only two ANDs and
an OR. In Figure 2.4a the external connections of the 1-bit adder are represented. The input c
receives the carry signal from the previous binary range. The output carry generate the carry
signal for the next binary or range.
a
carry
✛
a
b
❄
❄
OneBitAdd
b
a[n-1] b[n-1]
c
✛
❄
c
OneBitAdd
❄
❄
carryOut
a[1]
❄
✛
❄
a[0]
b[0]
❄
carryIn
❄
✛ OneBitAdd ✛ OneBitAdd ✛
❄
sum[n-1]
sum
b[1]
❄
❄
❄
sum[1]
sum[0]
carry
sum
a.
c.
b.
Figure 2.4: The adder circuit. a. The logic symbol for one-bit adder. b. The logic schematic for
the one-bit adder. c. The block schematic for the n-bit adder.
The functions for the one bit adder are obtained formally, without any trick, starting from
the truth table defining the operation (see Figure 2.5).
a
0
0
0
0
1
1
1
1
b
0
0
1
1
0
0
1
1
c
0
1
0
1
0
1
0
1
sum
0
1
1
0
1
0
0
1
carry
0
0
0
1
0
1
1
1
Figure 2.5: The truth table for the adder circuit. The first three columns contains all the
three-bit binary configuration the circuit could receive. The two last columns describe the behavior of
the sum and carry output.
The two expressions are extracted from the truth table as “sum” of “products”. Only the
“products” generating 1 to output are “summed”. Results:
sum = a′ b′ c + a′ bc′ + ab′ c′ + abc
carry = a′ bc + ab′ c + abc′ + abc
and, using the Boolean algebra rules the form are reduced to the previously written expressions.
28
CHAPTER 2. DIGITAL CIRCUITS
For a circuit with more than one output, minimizing it means to minimize the overall design,
not only each expressions associated to its outputs. For our design – the one bit adder – the
expression used to implement the sum output is not minimal. It is more complex that the
minimal form (instead of an OR gate we used the more complicated gate XOR), but it contains
a sub-circuit shared with the circuit associated to carry output. It is about the first XOR
circuit (see Figure 2.4b).
2.1.4
Divider
The divide operation – a/b – is, in the general case, a complex operation. But, in our application – Digital Pixel Correction – it is about dividing by 2 a binary represented number. It is
performed, without any circuit, simply by shifting the bits of the binary number one position to
right. The number number[n-1:0] divided by two become {1’b0, number[n-1:1]}.
2.2
Sequential circuits
In this section we intend to introduce the basic circuits used to build the sequential parts of a
digital system. It is about the sequential digital circuits. These circuits are mainly used to build
the storing sub-systems in a digital system. To store in a digital circuit means to maintain the
value of a signal applied on the input of the circuit. Simply speaking, the effect of the signal to
be stored must be “re-applied” on another input of the circuit, so as the effect of the input signal
to be memorized is substituted. Namely, the circuit must have a loop closed form one of its
output to one of its input. The resulting circuit, instead of providing the computation it performs
without loop, it will provide a new kind of functionality: the function of memorizing. Besides the
function of memorizing, sequential circuits are used to design simple or complex automata (in
this section we provide only examples of simple automata). The register, the typical sequential
circuit, is used also in designing complex systems allowing efficient interconnections between
various sub-systems.
2.2.1
Elementary Latches
This subsection is devoted to introduce the elementary structures whose internal loop allow the
simplest storing function: latching an event.
The reset-only latch is the AND loop circuit represented in Figure 2.6a. The passive input
value for AND loop is 1 ((Reset)’ = 1), while the active input value is 0 ((Reset)’ = 0).
If the passive input value is applied, then the output of the circuits is not affected (the output
depends only by the other input of the AND circuit). It can be 0 or 1, depending by the previous
values applied on the input. When the active value is temporary applied, then the state of the
circuit (the value of its output) switches in 0, with a delay of tpHL (propagation time from high
to low) and remains forever in this state, independent on the following the input value. We
conclude that the circuit is sensitive to the signal 0 temporarily applied on its input, i.e., it is
able to memorize forever the event 0. The circuit “catches” and “latches” the input value only
if the input in maintained on 0 until the second input of the AND circuit receives the value 0,
with a delay time tpHL . If the temporary input transition in 0 is too short the loop is unable to
latch the event.
The set-only latch is the OR loop circuit represented in Figure 2.6b. The passive value for
OR loop is 0 (Set = 0) while the active input value is 1 (Set = 1). If the passive input value
is applied, then the output of the circuits is not affected (the output depends only by the other
2.2. SEQUENTIAL CIRCUITS
29
input of the OR circuit). It can be 0 or 1, depending by the previous values applied on the
input. When the active value is temporary applied, then the state of the circuit (the value of
its output) switches in 1 and remains forever in this state, independent on the input value. We
conclude that the circuit is sensitive to the signal 1 temporarily applied on its input, i.e., it is
able to memorize forever the event 1. The only condition, similar to that applied for AND loop,
is to have an enough long duration of temporary input transition in 1.
(Reset)’
✻
(Reset)’
✲
t
Set
✻
andLoopOut
a.
✲
Set
t
andLoopOut
t
pHL
✻✲ ✛
✲
orLoopOut
t
b.
orLoopOut
✲ ✛
✻
R’
tpLH
✲
S
t
setRestLatchOut
✻✲
2t
✛pHL
t
✲ ✛pLH
✲
setRestLatchOut
c.
t
d.
Figure 2.6: The elementary latches. Using the loop, closed from the output to one input, elementary storage elements are built. a. AND loop provides a reset-only latch. b. OR loop provides the
set-only version of a storage element. c. The heterogeneous elementary set-reset latch results combining
the reset-only latch with the set-only latch. d. The wave forms describing the behavior of the previous
three latch circuits.
The heterogenous set-reset latch results by combining the previous two latches (see Figure
2.6c). The circuit has two inputs: one active-low (active on 0) input, R’, to reset the circuit (out
= 0), and another active-high (active on 1) input, S, to set the circuit (out = 0). The value 0
must remain to the input R’ at least 2tpHL for a stable switching of the circuit into the state 0,
because the loop depth in the state 1 is given by the propagation time through both gates that
switch from high to low. For a similar reason, the value 1 must remain to the input S at least
2tpLH when the circuit must switch in 1. However, the output of the circuit reacts faster to the
set signal, because from the input set to the output of the circuit there is only one gate, while
from the other input to output the depth of the circuit is doubled.
The symmetric set-reset latches are obtained by applying De Morgan’s law to the heterogenous elementary latch. In the first version, the OR circuit is transformed by De Morgan’s
law (the form a + b = (a’ b’)’ is used) resulting the circuit from Figure 2.7a. The second
version (see Figure 2.7b) is obtained applying the other form of the same law to the AND circuit
30
CHAPTER 2. DIGITAL CIRCUITS
S’
R’
Q
S
R
Q’
Q
Q’
a.
b.
Figure 2.7: Symmetric elementary latches. a. Symmetric elementary NAND latch with lowactive commands S’ and R’. b. Symmetric elementary NOR latch with high-active commands S and
R.
(ab = (a’ + b’)’). The passive input value for the NAND elementary latch is 1, while for
the NOR elementary latch it is 0. The active input value for the NAND elementary latch is 0,
while for the NOR elementary latch it is 1. The symmetric structure of these latches have two
outputs, Q and Q’.
Although, the structural description in an actual design does not go until such detail, it is
useful to use a simulation for understand how this small, simple, but fundamental circuit works.
For the sake of simulation only, the description of the circuit contains time assignment. If the
module is supposed to by eventually synthesised, then the time assignment must be removed.
VeriSim 2.1 The Verilog sttructural description of NAND latch is:
module elementary_latch(output out, not_out,
input not_set, not_reset);
nand
#2 nand0(out, not_out, not_set);
nand
#2 nand1(not_out, out, not_reset);
endmodule
The two NAND gates considered in this simulation have the propagation time equal with 2
unit times – #2.
For testing the behavior of the NAND latch just described, the following module is used:
module test_shortest_input;
reg
not_set, not_reset;
initial begin
not_set
not_reset
#10 not_reset
#10 not_reset
#10 not_set
#10 not_set
//#1 not_set
//#2 not_set
//#3 not_set
#10 not_set
#10 not_set
#10 not_reset
#10 not_reset
#10 $stop;
=
=
=
=
=
=
=
=
=
=
=
=
=
1;
1;
0;
1;
0;
1;
1;
1;
1;
0;
1;
0;
1;
// reset
//
//
//
//
//
//
set
1-st experiment
2-nd experiment
3-rd experiment
4-th experiment
another set
// reset
end
elementary_latch
endmodule
dut(out, not_out, not_set, not_reset);
2.2. SEQUENTIAL CIRCUITS
31
In the first experiment the set signal is activated on 0 during 10ut (ut stands for unit time).
In the second experiment (comment the line 9 and de-comment the line 10 of the test module),
a set signal of 1ut is unable to switch the circuit. The third experiment, with 2ut set signal,
generate an unstable simulated, but non-actual, behavior (to be explained by the reader). The
fourth experiment, with 3ut set signal, determines the shortest set signal able to switch the latch
(to be explained by the reader).
⋄
In order to use these latches in more complex applications we must solve two problems.
The first latch problem : the inputs for indicating how the latch switches are the same as
the inputs for indicating when the latch switches; we must find a solution for declutching the
two actions building a version with distinct inputs for specifying “how” and “when”.
The second latch problem : if we apply synchronously S’=0 and R’=0 on the inputs of
NAND latch (or S=1 and R=1 on the inputs of OR latch), i.e., the latch is commanded “to
switch in both states simultaneously”, then we can not predict what is the state of the latch
after the ending of these two active signals.
The first latch problem will be partially solved in the next subsection, introducing the clocked
latch, but the problem will be completely solved only by introducing the master-slave structure.
The second latch problem will be solved, only in one of the chapter that follow, with the JK
flip-flop, because the circuit needs more autonomy to “solve” the contradictory command that
“says him” to switch in both states simultaneously.
Application: de-bouncing circuit Interfacing digital systems with the real world involves
sometimes the use of mechanical switching contacts. The bad news is that this kind of contact
does not provide an accurate transition. Usually when it closes, a lot of parasitic bounces come
with the main transition (see wave forms S’ and R’ in Figure 2.8).
VDD
S’
✻
✲
S’
■
...
...
...
.
✠ .....
.......
.
Q
time
Q
✻
✲
time
R’
R’
✻
✲
VDD
time
Figure 2.8: The de-bouncing circuit.
The debouncing circuit provide clean transitions when digital signals must generated by
electro-mechanical switches. In Figure 2.8 an RS latch is used to clear up the bounces generated
by a two-position electro-mechanical switch. The elementary latch latches the first transition
32
CHAPTER 2. DIGITAL CIRCUITS
from VDD to 0. The bounces that follow have no effect on the output Q because the latch is
already switched, by the first transition, in the state they intend to lead the circuit.
2.2.2
Elementary Clocked Latches
In order to start solving the first latch problem the elementary latch is supplemented with two
gates used to validate the data inputs only during the active level of clock. Thus the clocked
elementary latch is provided.
active level
S
③
CK
R
❄ ❄ ❄
S
CK
R
R’
S’
RSL
Q
❄
a.
Q
Q’
Q’
❄
b.
Figure 2.9: Elementary clocked latch. The transparent RS clocked latch is sensitive (transparent)
to the input signals during the active level of the clock (the high level in this example). a. The internal
structure. b. The logic symbol.
The NAND latch is used to exemplify (see Figure 2.9a) the partial separation between how
and when. The signals R’ and S’ for the NAND latch are generated using two 2-input NAND
gates. If the latch must be set, then on the input S we apply 1, R is maintained in 0 and, only
after that, the clock is applied, i.e., the clock input CK switches temporary in 1. In this case
the active level of the clock is the high level. For reset, the procedure is similar: the input R is
activated, the input S is inactivated, and then the clock is applied.
We said that this approach allows only a partial declutching of how by when because on the
active level of CK the latch is transparent, i.e., any change on the inputs S and R can modify
the state of the circuit. Indeed, if CK = 1 and S or R is activated the latch is set or reset, and
in this case how and when are given only by the transition of these two signals, S for set or R
for reset. The transparency will be avoided only when, in the next subsection, the transition of
the output will be triggered by the active edge of clock.
The clocked latch does not solve the second latch problem, because for R = S = 1 the end of
the active level of CK switches the latch in an unpredictable state.
VeriSim 2.2 The following Verilog code can be used to understand how the elementary clocked
latch works.
module clocked_nand_latch(output
out, not_out,
input
set, reset, clock);
elementary_latch the_latch(out, not_out, not_set, not_reset);
nand
#2 nand2(not_set, set, clock);
nand
#2 nand3(not_reset, reset, clock);
endmodule
⋄
2.2. SEQUENTIAL CIRCUITS
2.2.3
33
Data Latch
The second latch problem can be only avoided, not removed in this stage of our approach,
by introducin a restriction on the inputs of the clocked latch. Indeed, introducing an inverter
circuit between the inputs of the RS clocked latch, as is shown in Figure 2.10a, the ambiguous
command (simultaneous set and reset) can not be applied. Now, the situation R = S = 1
becomes impossible. The output is synchronized with the clock only if on the active level of CK
the input D is stable.
We call the resulting one input with D (from Data). The circuit is called Data Latch, or
simple D-latch.
D
D
❄
S
❄
R
D
Q’
Q
RSL
Q
a.
❄
❄
CK
CK
DL
❄
b.
❄
Q’
❄
c.
Q’
Q
Figure 2.10: The data latch. Imposing the restriction R = S ′ to an RS latch results the D latch
without non-predictable transitions (R = S = 1 is not anymore possible). a. The structure. b. The
logic symbol. c. An improved version for the data latch internal structure.
The output of this new circuit follows continuously the input D during the active level of
clock. Therefore, the autonomy of this circuit is questionable because act only in the time when
the clock is inactive (on the inactive level of the clock). We say D latch is transparent on the
active level of the clock signal, i.e, the output is sensitive, to any input change, during the active
level of clock.
VeriSim 2.3 The following Verilog code can be used to describe the behavior of a D latch.
module data_latch( output reg out,
output
not_out,
input
data, clock);
always @(*) if (clock) out = data;
assign not_out = ~out;
endmodule
⋄
The main problem when data input D is separated by the timing input CK is the correlation
between them. When this two inputs change in the same time, or, more precisely, during the
same small time interval, some behavioral problems occur. In order to obtain a predictable
behavior we must obey two important time restrictions: the set-up time and the hold time.
In Figure 2.10c an improved version of the circuit is presented. The number of components
are minimized, the maximum depth of the circuit is maintained and the input load due to the
input D is reduced from 2 to 1, i.e., the circuit generating the signal D is loaded with one input
instead of 2, in the original circuit.
VeriSim 2.4 The following Verilog code can be used to understand how a D latch works.
module test_data_latch;
reg
data, clock;
34
CHAPTER 2. DIGITAL CIRCUITS
initial begin
clock = 0;
forever #10 clock = ~clock;
end
initial begin
data = 0;
#25 data = 1;
#10 data = 0;
#20 $stop;
end
data_latch dut(out, not_out, data, clock);
endmodule
module data_latch(output
input
not #2
out, not_out,
data, clock);
data_inverter(not_data, data);
clocked_nand_latch
endmodule
rs_latch(out, not_out, data, not_data, clock);
The second initial construct from test data latch module can be used to apply data in
different relation with the clock.
⋄
D
CK
CK
D
D
in1
in0
CK
sel
C
.. . . . .
...
.
..
...
.
..
. . . . . . ..
Q
EMUX
out
Q
Q’
Q’
a.
b.
Q
c.
Figure 2.11: The optimized data latch. An optimized version is implemented closing the loop
over an elementary multiplexer, EMUX. a. The resulting minimized structure for the circuit represented
in Figure 2.10a. b. Implementing the minimized form using only inverting circuits.
The internal structure of the data latch (4 2-input NANDs and an inverter in Figure 2.10a)
can be minimized opening the loop by disconnecting the output Q from the input of the gate
generating Q′ , and renaming it C. The resulting circuit is described by the following equation:
Q = ((D · CK)′ · (C(D′ · CK)′ )′ )′
which can be successively transformed as follows:
Q = ((D · CK) + (C(D′ · CK)′ )
Q = ((D · CK) + (C(D + CK ′ ))
2.2. SEQUENTIAL CIRCUITS
35
Q = D · CK + C · D + C · CK ′ (anti − hasard redundancy 1 )
Q = D · CK + C · CK ′
The resulting circuit is an elementary multiplexor (the selection input is CK and the selected
inputs are D, by CK = 1, and C, by CK = 0. Closing back the loop, by connecting Q to C,
results the circuit represented in Figure 2.11a. The actual circuit has also the inverted output
Q′ and is implemented using only inverted gates as in Figure 2.11b. The circuit from Figure
2.10a (using the RSL circuit from Figure 2.9a) is implemented with 18 transistors, instead of 12
transistors supposed by the minimized form Figure 2.11b.
VeriSim 2.5 The following Verilog code can be used as one of the shortest description for a D
latch represented in Figure 2.11a.
module mux_latch(
assign
endmodule
output q
input
d, ck
q = ck ? d : q;
,
);
In the previous module the assign statement, describing an elementary multiplexer, contains
the loop. The variable q depends by itself. The code is synthesisable.
⋄
We ended using the elementary multiplexer to describe the most complex latch. This latch
is used in structuring almost any storage sub-system in a digital system. Thus, one of the basic
combinational function, associated to the main control function if-then-else, is proved to be the
basic circuit in designing storage elements.
2.2.4
Master-Slave Principle
In order to remove the transparency of the clocked latches, disconnecting completely the how
from the when, the master-slave principle was introduced. This principle allows us to build a
two state circuit named flip-flop that switches synchronized with the rising or falling edge of the
clock signal.
The principle consists in serially connecting two clocked latches and in applying the clock
signal in opposite on the two latches (see Figure 2.12a). In the exemplified embodiment the
first latch is transparent on the high level of clock and the second latch is transparent on the
low level of clock. (The symmetric situation is also possible: the first latch is transparent of
the low level value of clock and the second no the high value of clock.) Therefore, there is no
time interval in which the entire structure is transparent. In the first phase, CK = 1, the first
latch is transparent - we call it the master latch - and it switches according to the inputs S
and R. In the second phase CK = 0 the second latch - the slave latch - is transparent and
it switches copying the state of the master latch. Thus the output of the entire structure is
modified only synchronized with the negative transition of CK, i.e., only at the transition from
1 to 0 of the clock, because the state of the master latch freezes until the clock switches back
1
Anti-hasard redundancy equivalence: f(a,b,c) = ab + ac + bc’ = ac + bc’
Proof:
f(a,b,c) = ab + ac + bc’ + cc’, cc’ is ORed because xx’ = 0 and x = x + 0
f(a,b,c) = a(b + c) + c’(b + c) = (b + c)(a + c’)
f(a,b,c) = ((b + c)’ + (a + c’)’)’, applying De Morgan law
f(a,b,c) = (b’c’ + a’c)’, applying again De Morgan law
f(a,b,c) = (ab’c’ + a’b’c’ + a’bc + a’b’c)’ = (m4 + m0 + m3 + m1)’, expanding to the disjunctive normal form
f(a,b,c) = m2 + m5 + m6 + m7 = a’bc’ + ab’c + abc’ + abc, using the “complementary” minterms
f(a,b,c) = bc’(a + a’) + ac(b + b’) = ac + bc’, q.e.d.
36
CHAPTER 2. DIGITAL CIRCUITS
S
❄
✾
S CK R
RSL
Q
Q’
❄
Q
❄
❄
S
Q’
❄
✾
✻
❄
❄
R
S
Q’
Q
RSF-F
❄
S CK R
RSL
Q
Q’
a.
active edge
active edge
R
❄❄❄
CK
Q
❄
b.
❄
R
RSF-F
❄
c.
❄
Q’
❄
Figure 2.12: The master-slave principle. Serially connecting two RS latches, activated with
different levels of the clock signal, results a non-transparent storage element. a. The structure of a RS
master-slave flip-flop, active on the falling edge of the clock signal. b. The logic symbol of the RS flip-flop
triggered by the negative edge of clock. c. The logic symbol of the RS flip-flop triggered by the positive
edge of clock.
to 1. We say the RS master-slave flip-flop switches always at (always @ expressed in Verilog)
the falling (negative) edge of the clock. (The version triggered by the positive edge of clock is
also possible.)
The switching moment of a master-slave structure is determined exclusively by the active
edge of clock signal. Unlike the RS latch or data latch, which can sometimes be triggered (in the
transparency time interval) by the transitions of the input data (R, S or D), the master-slave
flip-flop flips only at the positive edge of clock (always @(posedge clock)) or at the negative
edge of clock (always @(negedge clock)) edge of clock, according with the values applied on
the inputs R and S. The how is now completely separated from the when. The first latch problem
is finally solved.
VeriSim 2.6 The following Verilog code can be used to understand how a master-slave flip-flop
works.
module master_slave(output out, not_out, input set, reset, clock);
wire
master_out, not_master_out;
clocked_nand_latch
master_latch(
slave_latch(
.out
(master_out
.not_out(not_master_out
.set
(set
.reset (reset
.clock (clock
.out
(out
.not_out(not_out
.set
(master_out
.reset (not_master_out
.clock (~clock
),
),
),
),
)),
),
),
),
),
));
endmodule
⋄
There are some other embodiments of the master-slave principle, but all suppose to connect
latches serially.
Three very important time intervals (see Figure 2.13) must catch our attention in designing
digital systems with edge triggered flip-flops:
2.2. SEQUENTIAL CIRCUITS
37
set-up time – (tSU ) – the time interval before the active edge of clock in which the inputs R
and S must stay unmodified allowing the correct switch of the flip-flop
edge transition time – (t+ or t− ) – the positive or negative time transition of the clock signal
hold time – (tH ) – the time interval after the active edge of CK in which the inputs R and S
must be stable (even if this time is zero or negative).
100%
90%
clock
✲
10%
time
in
✛ ✲✛✲✛
✲
tsu
t+
✲
time
th
Figure 2.13: Magnifying the transition of the active edge of the clock signal. The input
data must be stable around the active transition of the clock tsu (set-up time) before the beginning of
the clock transition, during the transition of the clock, t+ (active transition time), and th (hold time)
after the end of the active edge.
In the switching “moment”, that is approximated by the time interval tSU + t+ + tH or
tSU + t− + tH “centered” on the active edge (+ or −), the data inputs must evidently be stable,
because otherwise the flip-flop “does not know” what is the state in which it must switch.
Now, the problem of decoupling the how by the when is better solved. Although, this solution
is not perfect, because the ”moment” of the switch is approximated by the short time interval
tSU + t+/− + tH . But the ”moment” does not exist for a digital designer. Always it must be a
time interval, enough over-estimated for an accurate work of the designed machine.
2.2.5
D Flip-Flop
Another tentative to remove the second latch problem leads to a solution that again avoids only
the problem. Now the RS master-slave flip-flop is restricted to R = S ′ (see Figure 2.14a).
The new input is named also D, but now D means delay. Indeed, the flip-flop resulting by
this restriction, besides avoiding the unforeseeable transition of the flip-flop, gains a very useful
function: the output of the D flip-flop follows the D input with a delay of one clock cycle.
Figure 2.14c illustrates the delay effect of this kind of flip-flop.
Warrning! D latch is a transparent circuit during the active level of the clock, unlike the
D flip-flop which is no time transparent and switches only on the active edge of the clock.
VeriSim 2.7 The structural Verilog description of a D flip-flop, provided only for simulation
purpose, follows.
module dff(output
input
wire
not_d;
out, not_out,
d, clock
);
38
CHAPTER 2. DIGITAL CIRCUITS
❄
D
CK
R
RSF-F
Q
Q’
✻
S
a.
Q
❄
Q’
❄
✲
✻
t
Q
✻
Q’
❄
❄
✲
DF-F
❄
❄
❄
D
b.
❄
t
D
❄
❄
Q
❄
c.
✲
t
Figure 2.14: The delay (D) flip-flop. Restricting the two inputs of an RS flip-flop to D = S = R′ ,
results an FF with predictable transitions. a. The structure. b. The logic symbol. c. The wave forms
proving the delay effect of the D flip-flop.
not #2
data_inverter(not_d, d);
master_slave
endmodule
rs_ff(out, not_out, d, not_d, clock);
The functional description currently used for a D flip-flop active on the negative edge of clock
is:
module dff(output
input
reg out
,
d, clock);
always @(negedge clock) out <= d;
endmodule
⋄
The main difference between latches and flip-flops is that over the D flip-flop we can close
a new loop in a very controllable fashion, unlike the D latch which allows a new loop, but
the resulting behavior is not so controllable because of its transparency. Closing loops over D
flip-flops result in synchronous systems. Closing loops over D latches result in asynchronous
systems. Both are useful, but in the first kind of systems the complexity is easiest manageable.
2.2.6
Register
One of the most representative and useful storage circuit is the register. The main application
of register is to support the synchronous sequential processes in a digital system. There are two
typical use of the register:
• provides a delayed connection between sub-systems
• stores the internal state of a system (see section 1.2); the register is used to close of the
internal loop in a digital system.
The register circuit store synchronously the value applied on its inputs. Register is used
mainly to support the design of control structures in a digital system.
The skeleton of any contemporary digital design is based on registers, used to store, synchronously with the system clock, the overall state of the system. The Verilog (or VHDL)
description of a structured digital design starts by defining the registers, and provides, usually,
2.2. SEQUENTIAL CIRCUITS
39
an Register Transfer Logic (RTL) description. An RTL code describe a set of registers interconnected through more or less complex combinational blocks. For a register is a non-transparent
structure any loop configurations are supported. Therefore, the design is freed by the care of
the uncontrollable loops.
In−1
❄
D
DF-F
Q
a.
❄
On−1
In−2
❄
I0
...
D
DF-F
...
Q
❄
On−2
❄
CK
I
❄
❄
D
DF-F
Rn
CK
❄O
Q
...
❄
O0
b.
Figure 2.15: The n-bit register. a. The structure: a bunch of DF-F connected in parallel. b. The
logic symbol.
Definition 2.2 An n-bit register, Rn , is made by parallel connecting a Rn−1 with a D (masterslave) flip-flop (see Figure 2.15). R1 is a D flip-flop.
⋄
VeriSim 2.8 An 8-bit enabled and resetable register with 2 unit time delay is described by the
following Verilog module:
module register #(parameter n = 8)(output
input
input
always
reg [n-1:0] out
[n-1:0] in
reset, enable, clock)
@(posedge clock) #2 if (reset)
else if (enable)
else
out <= 0
out <= in
out <= out
,
,
;
;
;
;
endmodule
The time behavior specified by #2 is added only for simulation purpose. The synthesizable
version must avoid this non-sinthesizable representation.
⋄
The main feature of the register assures its non-transparency, excepting an ”undecided transparency” during a short time interval, tSU + tedge + tH , centered on the active edge of the clock
signal. Thus, a new loop can be closed carelessly over a structure containing a register. Due
to its non-transparency the register will be properly loaded with any value, even with a value
depending on its own current content. This last feature is the main condition to close the loop
of a synchronous automata - the structure presented in the next chapter.
The register is used at least for the following purposes: to store, to buffer, to synchronize,
to delay, to loop, . . ..
Storing The enable input allows us to determine when (i.e., in what clock cycle) the input is
loaded into a register. If enable = 0, the registers stores the data loaded in the last clock cycle
when the condition enable = 1 was fulfilled. This means we can keep the content once stored
into the register as much time as it is needed.
40
CHAPTER 2. DIGITAL CIRCUITS
clock
my in
❄
✰
in
enable
register
reset
out
clk
✻
my reg
✛
✛
❄
✻
✻
✻
✲
time
rst
in
my out
✻
“1”
✻
a
c
b
e
d
✲
time
register my reg( .out
.in
.enable
.reset
.clock
(my out ),
(my in ),
(1’b1 ),
(rst
),
(clk
));
out
✻
a
b
c
e
✲
time
Figure 2.16: Register at work. At each active edge of clock (in this example it is the positive edge)
the register’s output takes the value applied on its inputs if reset = 0 and enable = 1.
Buffering The registers can be used to buffer (to isolate, to separate) two distinct blocks so
as some behaviors are not transmitted through the register. For example, in Figure 2.16 the
transitions from c to d and from d to e at the input of the register are not transmitted to the
output.
Synchronizing For various reasons the digital signals are generated “unaligned in time” to
the inputs of a system, but they are needed to be received very well controlled in time. We say
usually, the signals are applied asynchronously but they must be received synchronously. For
example, in Figure 2.16 the input of the register changes somehow chaotically related to the
active edge of the clock, but the output of the register switches with a constant delay after the
positive edge of clock. We say the inputs are synchronized to the output of the register. Their
behavior is “time tempered”.
Delaying The input value applied in the clock cycle n to the input of a register is generated
to the output of the register in the clock cycle n+1. In other words, the input of a register is
delayed one clock cycle to its output. See in Figure 2.16 how the occurrence of a value in one
clock cycle to the register’s input is followed in the next clock cycle by the occurrence of the
same value to the register’s output.
Looping Structuring a digital system means to make different kind of connections. One of
the most special, as we see in what follows, is a connection from some outputs to certain inputs
in a digital subsystem. This kind of connections are called loops. The register is an important
structural element in closing controllable loops inside a complex system.
2.2.7
Shift register
One of the simplest application of register is to perform shift operations. The numerical interpretation of a shift is the multiplication by the power of 2, for left shift, or division with the of
2, for right shift. A register used as shifter must be featured with four operation modes:
00: nop – no operation mode; it is mandatory for any set of function associated with a circuit
(the circuit must be “able to stay doing nothing”)
2.2. SEQUENTIAL CIRCUITS
41
01: load – the register’s state is initialized to the value applied on its inputs
10: leftShift – shift left with one binary position; if the register’s state is interpreted as a
binary number, then the operation performed is a multiplication by 2
11: rightShift – shift right with one binary position; if the register’s state is interpreted as a
binary number, then the operation performed is a division by 2
A synthesisable Verilog description of the circuit is:
module shiftRegister(output reg [15:0] out ,
input
[15:0] in
,
input
[1:0] mode ,
input
clock);
always @(posedge clock)
case(mode)
2’b00: out <= out;
2’b01: out <= in ;
2’b10: out <= out << 1;
2’b11: out <= out >> 1; // for positive integers
//2’b11: out <= {out[15], out[15:1]}; // for signed integers
endcase
endmodule
The case construct describes a 4-input multiplexor, M U X4 . Two versions are provided in
the previous code, one for positive integer numbers and another for signed integers. The second
is “commented”.
2.2.8
Counter
Let be the following circuit: its output is identical with its internal state, its state can take the
value received on its data input, its internal state can be modified incrementing the number
which represents its state or can stay unchanged. Let us call this circuit: presetable counter.
Its Verilog behavioral description, for an 8-bit state, is:
module counter(
output reg [7:0]
out
,
input
[7:0]
in
,
input
init
, // initialize with in
input
count
, // increment state
input
clock
);
always @(posedge clock)
// always at the positive edge of clock
if (init)
out <= in;
else if (count) out <= out + 1;
endmodule
The init input has priority to the input count, if it is active (init = 1) the value of count
is ignored and the value of state is initialized to in. If init in not activated, then if count = 1
then the value of counter is incremented modulo 256.
The actual structure of the circuit results (easy) from the previous Verilog description. Indeed, the structure
if (init) ...
else
...
suggests a selector (a multiplexor), while
out <= out + 1;
42
CHAPTER 2. DIGITAL CIRCUITS
in
❄
in
INC8
out
inc en
init
✲
❄
in1
sel
MUX8
out
❄
our inc
in0
❦
✯
reg in
clock
clock
sel
❄
in
R8
out
count
❨
✶
inc out
✛
reset
enable
✐
our reg
”1”
COUNTER8
❄
out
Figure 2.17: The internal structure of a counter. If init = 1, then the value of the register is
initialized to in, else if count = 1 each active edge of clock loads in register its incremented value.
imposes an increment circuit. Thus, the schematic represented in Figure 2.17 pops up in our
mind.
The circuit INC8 in Figure 2.17 represents an increment circuit which outputs the input in
incremented when the input inc en (increment enable) is activated.
2.3
Putting all together
Now, going back to our first target enounced in section 1.3, let us put together what we learned
about digital circuits in this section. The RTL code for the Digital Pixel Corrector circuit can
be written now “more directly” as follows:
module pixelCorrector #(‘include "0_paramPixelCor.v")
(output [m-1:0] out ,
input [n-1:0] in
,
input
clock);
reg
[q-1:0] state; // the state register
always @(posedge clock) state <= {state[7:0], in}; // state transition
assign out = (state[7:4] == 0) ? ({1’b0, state[3:0]} + state[11:8]) >> 1 :
state[7:4];
endmodule
The schematic we have in mind while writing the previous code is represented in Figure 2.18,
where:
• the state register, state, has three sections of 4 bits each; in each cycle the positive edge
of clock shifts left the content of state 4 binary positions, and in the freed locations loads
the input value in
• the middle section is continuously tested, by the module Zero, if its value is zero
2.4. CONCLUDING ABOUT THIS SHORT INTRODUCTION IN DIGITAL CIRCUITS 43
• the first and the last sections of the state are continuously added and the result is divided
by 2 (shifted one position right) and applied to the input 1 of the selector circuit
• the selector circuit, ifThenElse (the multiplexer), selects to the output, according to the
test performed by the module Zero, the middle section of state or the shifted output of
the adder.
in
❄
state[11:8]
❄
state[7:4]
state
❄
✮
state[3:0]
clock
divide circuit
❄
❄
Zero
carry
✲ sel
❄
adder
❄❄❄❄
❄
❄
1 ifThenElse 0
❄
sum[0]
✢
M U X4
✙
❄
out
Figure 2.18: The structure of the pixelCorrector circuit.
Each received value is loaded first as state[3:0], then it moves in the next section. Thus,
an input value cames in the position to be sent out only with a delay of two clock cycles. This
two-cycle latency is imposed by the interpolation algorithm which must wait for the next input
value to be loaded as a stable value.
2.4
Concluding about this short introduction in digital circuits
A digital circuit is build of combinational circuits and storage registers
Combinational logic can do both, control and arithmetic
Logic circuits, with appropriate loops, can memorize
HDL, as Verilog or VHDL, must be used to describe digital circuits
Growing, speeding and featuring digital circuits digital systems are obtained
2.5
Problems
Combinational circuits
Problem 2.1 Design the n-input Equal circuit which provides 1 on its output only when its
two inputs, of n-bits each, are equal.
Problem 2.2 Design a 4-input selector. The selection code is sel[1:0] and the inputs are of
1 bit each: in3, in2, in1, in0.
44
CHAPTER 2. DIGITAL CIRCUITS
Problem 2.3 Provide the proof for the following Boolean equivalence:
a · b + c · (a + b) = a · b + c · (a ⊕ b)
Problem 2.4 Provide the Verilog structural description of the adder module for n = 8. Synthesise the design and simulate it.
Problem 2.5 Draw the logic schematic for the XOR circuit.
Problem 2.6 Provide the proof that: a ⊕ b = (a′ ⊕ b)′ = (a ⊕ b′ )′ .
Problem 2.7 Define the truth table for the one-bit subtractor and extract the two expressions
describing the associated circuit.
Problem 2.8 Design the structural description of a n-bit subtractor. Synthesise and simulate
it.
Problem 2.9 Provide the structural Verilog description of the adder/subtractor circuit behaviorally defined as follows:
module addSub #(parameter n = 4)// defines a n-bit adder
(output [n-1:0] sum,
// the n-bit result
output
carry, // carry output (borrow output for subtract)
input
sub,
// sub=1 ? sub : add
input
c,
// carry input (borrow input for subtract)
input [n-1:0] a, b); // the two n-bit numbers
assign {carry, sum} = sub ? a - b - c : a + b + c;
endmodule
Simulate and synthesise the resulting design.
Flip-flops
Problem 2.10 Why, in Figure 2.6, we did not use a XOR gate to close a latching loop?
Problem 2.11 Design the structural description, in Verilog, for a NOR elementary latch. Simulate the circuit in order to determine the shortest signal which is able to provide a stable transition of the circuit.
Try to use NOR gates with different propagation time.
Problem 2.12 When it is necessary to use a NOR elementary latch for a de-bouncing circuit?
Problem 2.13 Try to use the structural simulation of an elementary latch to see how behaves
the circuit when the two inputs of the circuit are activated simultaneously (the second latch
problem). If you are sure the simulation is correct, but it goes spooky, then go to office ours to
discuss with your teacher.
Problem 2.14 What if the NAND gates from the circuit represented in Figure 2.9 are substituted with NOR gates?
Problem 2.15 Design and simulate structurally an elementary clocked latch, using only 4 gates,
which is transparent on the level 0 of the clock signal.
2.5. PROBLEMS
45
Problem 2.16 Provide the test module for the module data latch (see subsection 2.2.3) in
order to verify the design.
Problem 2.17 Draw, at the gate level, the internal structure of a master-slave RS flip-flop
using
• NAND gates and an inverter
• NOR gates and an inverter
• NAND and NOR gates, for two versions:
– triggered by the positive edge of clock
– triggered by the negative edge of clock.
Applications
Problem 2.18 Draw the block schematic for the circuit performing pixel correction according
to the following interpolation rule:
s′ (t) = (2 × s(t − 2) + 6 × s(t − 1) + 6 × s(t + 1) + 6 × s(t + 2))/16
Using the schematic, write the Verilog code describing the circuit. Simulate and synthesise it.
Problem 2.19 Design a circuit which receives a stream of 8-bit numbers and sends, with a minimal latency, instead of each received number the mean value of the last three received numbers.
Problem 2.20 Design a circuit which receives a stream of 8-bit signed numbers and sends, with
one clock cycle latency, instead of each received number its absolute value.
Problem 2.21 Draw the block schematic for the module shiftRegister, described in subsection 2.2.7, using a register an two input multiplexers, M U X2 . Provide a Verilog structural
description for the resulting circuit. Simulate and synthesise it.
Problem 2.22 Define a two-input DF-F using a DF-F and an EMUX. Use the new structure
to describe structurally a presetable shift right register. Add the possibility to perform logic or
arithmetic shift.
Problem 2.23 Write the structural description for the increment circuit INC8 introduces in
subsection 2.2.8.
Problem 2.24 Write the structural description for the module counter defined in subsection
2.2.8. Simulate and synthesize it.
Problem 2.25 Define and design a reversible counter able to count-up and count-down. Simulate and synthesize it.
Problem 2.26 Design the accumulator circuit able to add sequentially a clock synchronized
sequence of up to 256 16-bit signed integers. The connections of the circuit are
module accumulator
(output reg [?:0]
input
[15:0]
input
[1:0]
input
...
endmodule
acc
, // the output register used to accumulate
number, // the input receiving the stream of numbers
com
, // the command: 00=nop, 01=init, 10=accumulate
clock );
46
CHAPTER 2. DIGITAL CIRCUITS
The init command initializes the state, by clearing the register, in order to start a new
accumulation process.
Problem 2.27 Design the two n-bit inputs combinational circuit which computes the absolute
difference of two numbers.
Problem 2.28 Define and design a circuit which receives a one-bit wave form and shows on
its three one-bit outputs, by one clock cycle long positive impulses, the following events:
• any positive transition of the input signal
• any negative transition of the input signal
• any transition of the input signal.
Problem 2.29 Design the combinational circuit which compute the absolute value of a signed
number.
Chapter 3
GROWING & SPEEDING &
FEATURING
In the previous chapter
starting from simple algorithms small combinational and sequential circuits were designed,
using the Verilog HDL as tool to describe and simulate. From the first chapter is ggod to
remember:
• Verilog can be used for both behavioral (what does the circuit?) and structural (how
looks the circuit?) descriptions
• the outputs of a combinational circuits follow continuously (with as small as possible
delay) any change of its inputs, while a sequential one takes into account a shorter
or a longer history of the input behavior
• the external time dependencies must be minimized if not avoided; each circuit must
have its own and independent time behavior in order to allow global optimizations
In this chapter
the three main mechanisms used to generate a digital system are introduced:
• composition: the mechanism allowing a digital circuit to increase its size and its
computational power
• pipeline: is the way of interconnecting circuits to avoid the increase of the delays
generated by too many serial compositions
• loop: is a kind of connection responsible for adding new type of behaviors, mainly by
increasing the autonomy of the system
In the next chapter
a taxonomy based on the number of loops closed inside a digital system is proposed. Each
digital order, starting from 0, is characterized by the degree of the autonomy its behavior
develops. While digital circuits are combinational or sequential, digital systems will be:
• 0 order, no-loop circuits (the combinational circuits)
• first order, 1-loop circuits (simple flip-flops, ...)
• second order, 2-loop circuits (finite automata, ...)
• third order, 3-loop circuits (processors, ...)
• ...
47
48
CHAPTER 3. GROWING & SPEEDING & FEATURING
... there is no scientific theory about what can
and can’t be built.
David Deutsch1
Engineering only uses theories, but it is art.
In this section we talk about simple things which have multiple, sometime spectacular,
followings. What can be more obvious than that a system is composed by many subsystems and
some special behaviors are reached only using appropriate connections.
Starting from the ways of composing big and complex digital systems by appropriately interconnecting simple and small digital circuits, this book introduces a more detailed classification
of digital systems. The new taxonomy classifies digital systems in orders, based on the maximum number of included loops closed inside each digital system. We start from the basic idea
that a new loop closed in a digital system adds new functional features in it. By composition,
the system grows only by forwarded connections, but by appropriately closed backward connections it gains new functional capabilities. Therefore, we will be able to define many
functional levels, starting with time independent combinational functions and continuing with
memory functions, sequencing functions, control functions and interpreting functions. Basically,
each new loop manifests itself by increasing the degree of autonomy of the system.
Therefore, the main goal of this section is to emphasize the fundamental developing mechanisms in digital systems which consist in compositions & loops by which digital systems gain
in size and in functional complexity.
In order to better understand the correlation between functional aspects and structural aspect
in digital systems we need a suggestive image about how these systems grow in size and how
they gain new functional capabilities. The oldest distinction between combinational circuits and
sequential circuits is now obsolete because of the diversity of circuits and the diversity of their
applications. In this section we present a new idea about a mechanism which emphasizes a
hierarchy in the world of digital system. This world will be hierarchically organized in orders
counted from 0 to n. At each new level a functional gain is obtained as a consequence of the
increased autonomy of the system.
Two are the mechanisms involved in the process of building digital systems. The first allows
of system to grow in size. It is the composition, which help us to put together, using only
forward connections, many subsystems in order to have a bigger system. The second mechanism
is a special connection that provides new functional features. It is the loop connection, simply
the loop. Where a new loop is closed, a new kind of behavior is expected. To behave means,
mainly, to have autonomy. If a system use a part of own outputs to drive some of its inputs,
then “he drives himself” and an outsider receives this fact as an autonomous process.
Let us present in a systematic way, in the following subsections, the two mechanisms. Both
are very simple, but our goal is to emphasize, in the same time, some specific side effects as
consequences of composing & looping, like the pipeline connection – used to accelerate the speed
of the too deep circuits – or the speculative mechanisms – used to allow loops to be closed in
pipelined structures.
Building a real circuit means mainly to interconnect simple and small components in order
to grow an enough fast system appropriately featured. But, growing is a concept with no precise
meaning. Many people do not make distinction between “growing the size” and “growing the
complexity” of a system, for example. We will start making the necessary distinctions between
“size” and “complexity” in the process of growing a digital system.
1
David Deutch’s work on quantum computation laid the foundation for that field, grounding new approaches
in both physics and the theory of computation. He is the author of The Fabric of Reality.
3.1. SIZE VS. COMPLEXITY
3.1
49
Size vs. Complexity
The huge size of the actual circuits implemented on a single chip imposes a more precise distinction between simple circuits and complex circuits. When we can integrated on a single chip
more than 109 components, the size of the circuits becomes less important than their complexity.
Unfortunately we don’t make a clear distinction between size and complexity. We say usually:
“the complexity of a computation is given by the size of memory and by the CPU time”. But, if
we have to design a circuit of 100 million transistors it is very important to distinguish between
a circuit having an uniform structure and a randomly structured ones. In the first case the
circuit can be easy specified, easy described in an HDL, easy tested and so on. Otherwise, if the
structure is completely random, without any repetitive substructure inside, it can be described
using only a description having a similar dimension with the circuit size. When the circuit is
small, it is not a problem, but for million of components the problem has no solution. Therefore,
if the circuit is very big, it is not enough to deal only with its size, the most important becomes
also the degree of uniformity of the circuit. This degree of uniformity, the degree of order inside
the circuit can be specified by its complexity.
As a consequence we must distinguish more carefully the concept of size by the concept of
complexity. Follow the definitions of these terms with the meanings we will use in this book.
Definition 3.1 The size of a digital circuit, Sdigital circuit , is given by the dimension of the
physical resources used to implement it. ⋄
In order to provide a numerical expression for size we need a more detailed definition which
takes into account technological aspects. In the ’40s we counted electronic bulbs, in the ’50s we
counted transistors, in the ’60s we counted SSI2 and MSI3 packages. In the ’70s we started to
use two measures: sometimes the number of transistors or the number of 2-input gates on the
Silicon die and other times the Silicon die area. Thus, we propose two numerical measures for
the size.
Definition 3.2 The gate size of a digital circuit, GSdigital circuit , is given by the total number
of CMOS pairs of transistors used for building the gates (see the appendix Basic circuits) used
to implement it4 . ⋄
This definition of size offers an almost accurate image about the Silicon area used to implement the circuit, but the effects of lay-out, of fan-out and of speed are not catched by this
definition.
Definition 3.3 The area size of a digital circuit, ASdigital circuit , is given by the dimension of
the area on Silicon used to implement it. ⋄
The area size is useful to compute the price of the implementation because when a circuit
is produced we pay for the number of wafers. If the circuit has a big area, the number of the
circuits per wafer is small and the yield is low5 .
Definition 3.4 The algorithmic complexity of a digital circuit, simply the complexity,
Cdigital circuit , has the magnitude order given by the minimal number of symbols needed to express
its definition. ⋄
2
Small Size Integrated circuits
Medium Size Integrated circuits
4
Sometimes gate size is expressed in the total number of 2-input gates necessary to implement the circuit.
We prefer to count CMOS pairs of transistors (almost identical with the number of inputs) instead of equivalent
2-input gates because is simplest. Anyway, both ways are partially inaccurate because, for various reasons, the
transistors used in implementing a gate have different areas.
5
The same number of errors make useless a bigger area of the wafer containing large circuits.
3
50
CHAPTER 3. GROWING & SPEEDING & FEATURING
Definition 2.2 is inspired by Gregory Chaitin’s definition for the algorithmic complexity of a
string of symbols [Chaitin ’77]. The algorithmic complexity of a string is related to the dimension
of the smallest program that generates it. The program is interpreted by a machine (more in
Chapter 12). Our Cdigital circuit can be associated to the shortest unambiguous circuit description
in a certain HDL (in the most of cases it is about a structural description).
Definition 3.5 A simple circuit is a circuit having the complexity much smaller than its size:
Csimple circuit << Ssimple circuit .
Usually the complexity of a simple circuit is constant: Csimple circuit ∈ O(1). ⋄
Definition 3.6 A complex circuit is a circuit having the complexity in the same magnitude
order with its size:
Ccomplex circuit ∼ Scomplex circuit .⋄
Example 3.1 The following Verilog program describes a complex circuit, because the size of its
definition (the program) is
Sdef. of random
module random_circ(output
wire
w1, w2;
and and1(w1, a, b),
and2(w2, c, d);
or
or1(f, w1, c),
or2(g, w1, w2);
endmodule
circ
= k1 + k2 × Srandom
f, g, input
circ
∈ O(Srandom
circ ).
a, b, c, d, e);
⋄
Example 3.2 The following Verilog program describes a simple circuit, because the program
that define completely the circuit is the same for any value of n.
module or_prefixes #(parameter n = 256)(output reg [0:n-1] out, input [0:n-1] in);
integer k;
always @(in) for (k=1; k<n; k=k+1) begin
out[0] = in[0];
out[k] = in[k] | out[k-1];
end
endmodule
The prefixes of OR circuit consists in n OR2 gates connected in a very regular form. The
definition is the same for any value of n6 . ⋄
Composing circuits generate not only biggest structures, but also deepest ones. The depth
of the circuit is related with the associated propagation time.
Definition 3.7 The depth of a combinational circuit is equal with the total number of serially
connected constant input gates (usually 2-input gates) on the longest path from inputs to the
outputs of the circuit. ⋄
6
A short discussion occurs when the dimension of the input is specified. To be extremely rigorous, the parameter
n is expressed using a string o symbols in O(log n). But usually this aspect can be ignored.
3.2. TIME RESTRICTIONS IN DIGITAL SYSTEMS
51
The previous definition offers also only an approximate image about the propagation time
through a combinational circuit. Inspecting the parameters of the gates listed in Appendix
Standard cell libraries you will see more complex dependence contributing to the delay introduced
by a certain circuit. Also, the contribution of the interconnecting wires must be considered when
the actual propagation time in a combinational circuit is evaluated.
Some digital functions can be described starting from the elementary circuit which performs
them, adding a recursive rule for building a circuit that executes the same function for any size
of the input. For the rest of the circuits, which don’t have such type of definitions, we must
use a definition that describes in detail the entire circuit. This description will be non-recursive
and thus complex, because its dimension is proportional with the size of circuit (each part of
the circuit must be explicitly specified in this kind of definition). We shall call random circuit a
complex circuit, because there is no (simple) rule for describing it.
The first type of circuits, having recursive definitions, are simple circuits. Indeed, the elementary circuit has a constant (usually small) size and the recursive rule can be expressed using
a constant number of signs (symbolic expressions or drawings). Therefore, the dimension of
the definition remains constant, independent by n, for this kind of circuits. In this book, this
distinction, between simple and complex, will be exemplified and will be used to promote useful
distinctions between different solutions.
At the current technological level the size becomes less important than the complexity,
because we can produce circuits having an increasing number of components, but we can describe
only circuits having the range of complexity limited by our mental capacity to deal efficiently
with complex representations. The first step to have a circuit is to express what it must do in
a behavioral description written in a certain HDL. If this ”definition” is too large, having the
magnitude order of a huge multi-billion-transistor circuit, we don’t have the possibility to write
the program expressing our desire.
In the domain of circuit design we passed long ago beyond the stage of minimizing the number
of gates in a few gates circuit. Now, the most important thing, in the multi-billion-transistor
circuit era, is the ability to describe, by recursive definitions, simple (because we can’t write
huge programs), big (because we can produce more circuits on the same area) sized circuits. We
must take into consideration that the Moore’s Law applies to size not to complexity.
3.2
Time restrictions in digital systems
The most general form of a digital circuit (see Figure 3.1) includes both combinational and
sequential behaviors. It includes two combinational circuits – (comb circ 1 and comb circ 2)
– and register. There are four critical propagation paths in this digital circuit:
1. form input to register through comb circ 1, which determines minimum input arrival
time before clock: tin reg
2. from register to register through comb circ 1, which determines minimum period of
clock: treg reg = Tmin , or maximum frequency of clock: fmax = 1/T
3. from input to output through comb circ 2, which determines maximum combinational
path delay: tin out
4. from register to output through comb circ 2, which determines maximum output required time after clock: treg out .
If the active transition of clock takes place at t0 and the input signal changes after t0 −tin reg ,
then the effect of the input change will be not registered correctly at t0 in register. The input
52
CHAPTER 3. GROWING & SPEEDING & FEATURING
input
✲
✛
comb circ 1
tin
reg
❘
✠
treg
reg
clock
❄
register
❄
treg
out
❄
✲
comb circ 2
tin
out
❘
❄output
Figure 3.1: The four critical propagation paths in digital circuits. Input-to-register time
(tin reg ) is recommended to be as small as possible in order to reduce the time dependency from the
previous sub-system. Register-to-register time (Tmin ) must be minimal to allow a high frequency for the
clock signal. Input-to-output time (tin out ) is good to be undefined to avoid hard to manage sub-systems
interdependencies. Register-to-output time (treg out ) must be minimal to reduce the time dependency for
the next sub-system
must be stable in the time interval from t0 − tin reg to t0 in order to have a predictable behavior
of the circuit.
The loop is properly closed only if Tmin > treg + tcc2 + tsu and th < treg + tcc2 , where: treg
is the propagation time through register from active edge of clock to output, and tcc2 is the
propagation time through comb circ 1 on the path 2.
If the system works with the same clock, then tin out < Tmin , preferably tin out << Tmin .
Similar conditions are imposed for tin reg and treg out , because we suppose there are additional
combinational delays in the circuits connected to the inputs and to the outputs of this circuit,
or at least a propagation time through a register or set-up time to the input of a register.
Example 3.3 Let us compute the propagation times for the four critical propagation paths of
the counter circuit represented in Figure 2.17. If we consider #1 = 100ps results:
• tin reg = tp (mux2 8) = 0.1ns
(the set-up time for the register is considered too small to be considered)
• fmax = 1/T = 1/(tp (reg) + tp (inc) + tp (mux2 8)) = 1/(0.2 + 0.1 + 0.1)ns = 2.5 GHz
• tin
out
is not defined
3.2. TIME RESTRICTIONS IN DIGITAL SYSTEMS
• treg
out
53
= tp (reg) = 0.2ns ⋄
Example 3.4 Let be the circuit from Figure 3.2, where:
• register is characterized by: tp (register) = 150ps, tsu (register) = 35ps, th = 27ps
• adder with tp (adder) = 550ps
• selector with tp (selector) = 85ps
• comparator with tp (comparator) = 300ps
The circuit is used to accumulate a stream of numbers applied on the input data, and to compare
it against a threshold applied on the input thr. The accumulation process is initialized by the
signal reset, and is controlled by the signal acc.
thr
data
❄
a
❄
b
adder
550ps a+b
❄
❄
1
0
selector
85ps out
35ps
sel
✛
rst
✛
❄
in
register
150ps out
❄
acc
reset
clock
❄
a
b
comparator
300ps a<b
❄
over
❄sum
Figure 3.2: Accumulate & compare circuit.
In the left-down corner of each rectangle
is written the propagation time of each module. If acc = 1 the circuit accumulates, else the
content of register does not change.
The propagation time for the four critical propagation path of this circuit are:
• tin
reg
= tp (adder) + tp (selector) + tsu (register) = (550 + 85 + 35)ps = 670ps
• fmax = 1/T = 1/(tp (register) + tp (adder) + tp (selector) + tsu (register)) =
1/(150 + 550 + 85 + 35)ps = 1.21GHz
• tin
• treg
out
= tp (comparator) = 300ps
out
= tp (register) + tp (comparator) = 450ps
⋄
While at the level of small and simple circuits no additional restriction are imposed, for
complex digital systems there are mandatory rules to be followed for an accurate design. Two
main restrictions occur:
54
CHAPTER 3. GROWING & SPEEDING & FEATURING
1. the combinational path through the entire system must be completely avoided,
2. the combinational, usually called asynchronous, input and output path must be avoided
as much as possible if not completely omitted.
Combinational paths belonging to distinct modules are thus avoided. The main advantage
is given by the fact that design restrictions imposed in one module do not affect time restriction
imposed in another module. There are two ways to consider these restrictions, a weak one and a
strong one. The first refers to the pipeline connections, while the second to the fully buffered
connections.
3.2.1
Pipelined connections
For the pipelined connection between two complex modules the timing restrictions are the following:
1. from input to output through: it is not defined
2. from register to output through: treg out = treg – it does not depend by the internal
combinational structure of the module, i.e., the outputs are synchronous, because they
are generated directly from registers.
✲
✲
in2
out2
comb1
in1
✲
✲
pr1
✲
out1
sr1
in2
out2
comb2
in1
✛
sys1
pr2
✲
out1
sr2
nextState
✲
✛
nextState
sys2
clock
Figure 3.3: Pipelined connections.
Only two combinational paths are accepted: (1) from register to register, and (2) form input
to register. In Figure 3.3 a generic configuration is presented. It is about two systems, sys1 and
sys2, pipeline connected using the output pipeline registers (pr1 between sys1 and sys2, and
pr2 between sys2 and an external system). For the internal state are used the state registers
sr1 and sr2. The timing restrictions for the two combinational circuits comb1 and comb2 are not
correlated. The maximum clock speed for each system does not depend by the design restrictions
imposed for the other system.
The pipeline connection works well only if the two systems are interconnected
with short wires, i.e., the two systems are implemented on adjacent areas on the silicon die.
No additional time must be considered on connections because they a very short.
The system from Figure 3.3 is descried by the following code.
module pipelineConnection( output
input
[15:0]
[15:0]
out
in
,
,
3.2. TIME RESTRICTIONS IN DIGITAL SYSTEMS
input
wire
[15:0] pipeConnect ;
sys sys1(
.pr
(pipeConnect),
.in
(in
),
.clock (clock
)),
sys2(
.pr
(out
),
.in
(pipeConnect),
.clock (clock
));
endmodule
module sys(output reg [15:0] pr
input
[15:0] in
input
clock
reg
[7:0]
sr
;
wire
[7:0]
nextState
;
wire
[15:0] out
;
comb myComb(.out1
(nextState ),
.out2
(out
),
.in1
(sr
),
.in2
(in
));
always @ (posedge clock)
begin
clock
55
);
,
,
);
pr <= out
;
sr <= nextState ;
end
endmodule
module comb(
output
output
input
input
[7:0]
[15:0]
[7:0]
[15:0]
out1,
out2,
in1 ,
in2 );
// ...
endmodule
Something very important is introduced by the last Verilog code: the distinction between
blocking and non-blocking assignment:
• the blocking assignment, = : the whole statement is done before control passes to the
next
• the non-blocking assignment, <= : evaluate all the right-hand sides in the project for
the current time unit and assign the left-hand sides only at the end of the time unit.
Let us use the following simulation to explain the very important difference between the two
kinds of assignment.
VeriSim 3.1 The following simulation used 6 clocked registers. All of them switch on the
positive edge. But, the code is written for three of them using the blocking assignment, while
for the other three using the non-blocking assignment. The resulting behavior show us the
difference between the two clock triggered assignment. The blocking assignment seems to be
useless, because propagates the input through all the three registers in one clock cycle. The nonblocking assignment shifts the input along the three serially connected registers clock by clock.
This second behavior can be used in real application to obtain clock controlled delays.
module blockingNonBlocking(output
output
input
input
reg [1:0]
reg [1:0]
[1:0]
blockingOut
nonBlockingOut
in
clock
,
,
,
);
56
CHAPTER 3. GROWING & SPEEDING & FEATURING
reg [1:0]
reg1, reg2, reg3, reg4;
always @(posedge clock) begin
always @(posedge clock) begin
reg1
reg2
blockingOut
= in
= reg1
= reg2
;
;
;
end
reg3
reg4
nonBlockingOut
<= in
;
<= reg3 ;
<= reg4 ;
end
endmodule
module blockingNonBlockingSimulation;
reg
clock
;
reg
[1:0]
in
;
wire
[1:0]
blockingOut, nonBlockingOut ;
initial begin
clock = 0
;
forever #1 clock = ~clock ;
end
initial begin
in = 2’b01 ;
#2 in = 2’b10 ;
#2 in = 2’b11 ;
#2 in = 2’b00 ;
#7 $stop
;
end
blockingNonBlocking dut(blockingOut
,
nonBlockingOut ,
in
,
clock
);
initial
$monitor
("clock=%b in=%b reg1=%b reg2=%b blockingOut=%b reg3=%b reg4=%b nonBlockingOut=%b",
clock, in, dut.reg1, dut.reg2, blockingOut, dut.reg3, dut.reg4, nonBlockingOut);
endmodule
/*
clock=0 in=01 reg1=xx reg2=xx blockingOut=xx
reg3=xx reg4=xx nonBlockingOut=xx
clock=1 in=01 reg1=01 reg2=01 blockingOut=01
reg3=01 reg4=xx nonBlockingOut=xx
clock=0 in=10 reg1=01 reg2=01 blockingOut=01
reg3=01 reg4=xx nonBlockingOut=xx
clock=1 in=10 reg1=10 reg2=10 blockingOut=10
reg3=10 reg4=01 nonBlockingOut=xx
clock=0 in=11 reg1=10 reg2=10 blockingOut=10
reg3=10 reg4=01 nonBlockingOut=xx
clock=1 in=11 reg1=11 reg2=11 blockingOut=11
reg3=11 reg4=10 nonBlockingOut=01
clock=0 in=00 reg1=11 reg2=11 blockingOut=11
reg3=11 reg4=10 nonBlockingOut=01
clock=1 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=11 nonBlockingOut=10
clock=0 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=11 nonBlockingOut=10
clock=1 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=00 nonBlockingOut=11
clock=0 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=00 nonBlockingOut=11
clock=1 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=00 nonBlockingOut=00
clock=0 in=00 reg1=00 reg2=00 blockingOut=00
reg3=00 reg4=00 nonBlockingOut=00
*/
It is obvious that the registers reg1 and reg2 are useless because they are somehow “transparent”. ⋄
The non-blocking version of assigning the content of a register will provide a clock controlled
delay. Anytime in a design there are more than one registers the non-blocking assignment must
be used.
VerilogSummary 1 :
3.2. TIME RESTRICTIONS IN DIGITAL SYSTEMS
57
= : blocking assignment the whole statement is done before control passes to the next
<= : non-blocking assignment evaluate all the right-hand sides in the project for the current time unit and assign the left-hand sides only at the end of the time unit.
3.2.2
Fully buffered connections
The most safe approach, the synchronous one, supposes fully registered inputs and outputs
(see Figure 3.4 where the functionality is implemented using combinatorial circuits and registers
and the interface with the rest of the system is implemented using only input register and
output register).
The modular synchronous design of a big and complex system is the best approach for a
robust design, and the maximum modularity is achieved removing all possible time dependency
between the modules. Then, take care about the module partitioning in a complex system
design!
Two fully buffered modules can be placed on the silicon die with less restrictions, because even
if the resulting wires are long the signals have time to propagate because no gates are connected
between the output register of the sender system and the input register of the receiver system..
clock
input
❄
input register
❄
comb circuits & registers
❄
output register
❄output
Figure 3.4: The general structure of a module in a complex digital system. If any big
module in a complex design is buffered with input and output registers, then we are in the ideal situation
when: tin reg and treg out are minimized and tin out is not defined.
For the synchronously interfaced module represented in Figure 3.4 the timing restrictions
are the following:
1. form input to register: tin
module
reg
= tsu – it does not depend by the internal structure of the
2. from register to register: Tmin , and fmax = 1/T – it is a system parameter
58
CHAPTER 3. GROWING & SPEEDING & FEATURING
3. from input to output through: it is not defined
4. from register to output through: treg
structure of the module.
out
= treg – it does not depend by the internal
Results a very well encapsuled module easy to be integrate in a complex system. The price
of this approach consists in an increasing number of circuits (the interface registers) and some
restrictions in timing imposed by the additional pipeline stages introduced by the interface
registers. These costs can be reduced by a good system level module partitioning.
3.3
Growing the size by composition
The mechanism of composition is well known to everyone who worked at least a little in mathematics. We use forms like:
f (x) = g(h1 (x), . . . , hm (x))
to express the fact that computing the function f requests to compute first all the functions hi (x)
and after that the m-variable function g. We say: the function g is composed with the functions
hi in order to have computed the function f . In the domain digital systems a similar formalism
is used to “compose” big circuits from many smaller ones. We will define the composition
mechanism in digital domain using a Verilog-like formalism.
Definition 3.8 The composition (see Figure 3.5) is a two level construct, which performs the
function f using on the second level the m-ary function g and on the first level the functions
h 1, h 2, ... h m, described by the following, incompletely defined, but synthesisable, Verilog
modules.
in
❄
❄
h 1
out 1
❄
h 2
h m
out m
out 2
❄ ❄
❄
g
❄ out
= f(in)
Figure 3.5: The circuit performing composition. The function g is composed with the functions
h 1, ... h m using a two level circuit. The first level contains m circuits computing in parallel the
functions h i, and on the second level there is the circuit computing the reduction-like function g.
module f #(‘include "parameters.v")(
wire
[size_1-1:0] out_1;
wire
[size_2-1:0] out_2;
// ...
wire
[size_m-1:0] out_
g
second_level(
.out
(out
.in_1
(out_1
output
input
),
),
[sizeOut-1:0]
[sizeIn-1:0]
out ,
in );
3.3. GROWING THE SIZE BY COMPOSITION
59
.in_2
(out_2 ),
// ...
.in_m
(out_m ));
h_1 first_level_1(.out(out_1), .in(in));
h_2 first_level_2(.out(out_2), .in(in));
// ...
h_m first_level_m(.out(out_m), .in(in));
endmodule
module g #(‘include "parameters.v")(
output
input
input
// ...
input
[sizeOut-1:0]
[size_1-1:0]
[size_2-1:0]
out ,
in_1,
in_2,
[size_m-1:0]
in_m);
// ...
endmodule
module h_1 #(‘include "parameters.v")( output [size_1-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
module h_2 #(‘include "parameters.v")( output [size_2-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
// ...
module h_m #(‘include "parameters.v")( output [size_m-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
The content of the file parameters.v is:
parameter
sizeOut
sizeIn
size_1
size_2
// ...
size_m
=
=
=
=
32,
8 ,
12,
16,
= 8
⋄
The general form of the composition, previously defined, can be called the serial-parallel
composition, because the modules h 1, ... h m compute in parallel m functions, and all are
serial connected with the module g (we can call it reduction type function, because it reduces the
vector generated by the previous level to a value). There are two limit cases. One is the serial
composition and another is the parallel composition. Both are structurally trivial, but represent
essential limit aspects regarding the resources of parallelism in a digital system.
Definition 3.9 The serial composition (see Figure 3.6a) is the composition with m = 1.
Results the Verilog description:
module f #(‘include "parameters.v")(
output
input
[sizeOut-1:0]
[sizeIn-1:0]
out ,
in );
60
CHAPTER 3. GROWING & SPEEDING & FEATURING
wire
[size_1-1:0] out_1;
g
second_level(
.out
(out
),
.in_1
(out_1 ));
h_1 first_level_1(.out(out_1), .in(in));
endmodule
module g #(‘include "parameters.v")(
output
input
[sizeOut-1:0]
[size_1-1:0]
out ,
in_1);
// ...
endmodule
module h_1 #(‘include "parameters.v")( output [size_1-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
⋄
in
in
❄
h1
❄
h1
❄
g
h2
❄
b.
❄
❄
out 1
❄
hm
❄
out 2
❄
out m
out
a.
Figure 3.6: The two limit forms of composition. a. The serial composition, for m = 1, imposing an inherent sequential computation. b. The parallel composition, with no reduction-like function,
performing data parallel computation.
Definition 3.10 The parallel composition (see Figure 3.6b) is the composition in the particular case when g is the identity function. Results the following Verilog description:
module f #(‘include "parameters.v")(
wire
wire
// ...
wire
[size_1-1:0]
[size_2-1:0]
out_1;
out_2;
[size_m-1:0]
out_m;
assign out = {
output
input
[sizeOut-1:0]
[sizeIn-1:0]
out_m,
// ...
out_2,
out_1}; // g is identity function
h_1 first_level_1(.out(out_1), .in(in));
h_2 first_level_2(.out(out_2), .in(in));
// ...
h_m first_level_m(.out(out_m), .in(in));
endmodule
out ,
in );
3.3. GROWING THE SIZE BY COMPOSITION
61
module h_1 #(‘include "parameters.v")( output [size_1-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
module h_2 #(‘include "parameters.v")( output [size_2-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
// ...
module h_m #(‘include "parameters.v")( output [size_m-1:0] out ,
input [sizeIn-1:0] in );
// ...
endmodule
The content of the file parameters.v is now:
parameter
sizeIn
size_1
size_2
// ...
size_m
sizeOut
=
=
=
8
12
16
,
,
,
=
=
8
,
size_1 +
size_2 +
// ...
size_m
⋄
Example 3.5 Using the mechanism described in Definition 1.3 the circuit computing the scalar
product between two 4-component vectors will be defined, now in true Verilog. The test module
for n = 8 is also defined allowing to test the design.
module inner_prod #(‘include "parameter.v")
(output [((2*n+2)-1):0] out, input [n-1:0] a3, a2, a1, a0, b3, b2, b1, b0);
wire[2*n-1:0] p3, p2, p1, p0;
mult m3(p3, a3, b3),
m2(p2, a2, b2),
m1(p1, a1, b1),
m0(p0, a0, b0);
add4 add(out, p3, p2, p1, p0);
endmodule
module mult #(‘include "parameter.v")
(output [(2*n-1):0] out, input [n-1:0]
assign out = m1 * m0;
endmodule
m1, m0);
module add4 #(‘include "parameter.v")
(output [((2*n+2)-1):0] out, input [(2*n-1):0] t3, t2, t1, t0);
assign out = t3 + t2 + t1 + t0;
endmodule
The content of the file parameter.v is:
62
CHAPTER 3. GROWING & SPEEDING & FEATURING
a3
b3
a2
❄ ❄
a1
❄ ❄
m3
p3
b2
b1
❄ ❄
m2
m1
p2
a0
b0
❄ ❄
m0
p1
p0
❄❄❄❄
add
❄
out = a3*b3 + a2*b2 + a1*b1 + a0*b0
Figure 3.7: An example of composition. The circuit performs the scalar vector product for
4-element vectors. The first level compute in parallel 4 multiplications generating the vectorial product,
and the second level reduces the resulting vector of products to a scalar.
parameter n = 8
The simulation is done by running the module:
module test_inner_prod;
reg[7:0] a3, a2, a1, a0, b3, b2, b1, b0;
wire[17:0] out;
initial
begin
{a3, a2, a1, a0} = {8’d1, 8’d2, 8’d3, 8’d4};
{b3, b2, b1, b0} = {8’d5, 8’d6, 8’d7, 8’d8};
end
inner_prod dut(out, a3, a2, a1, a0, b3, b2, b1, b0);
initial $monitor("out=%0d", out);
endmodule
The test outputs: out = 70
The description is structural at the top level and behavioral for the internal sub-modules
(corresponding to our level of understanding digital systems). The resulting circuit is represented
in Figure 3.7. ⋄
VerilogSummary 2 :
• The directive ‘include is used to add in any place inside a module the content of the file
xxx.v writing: ‘include "xxx.v"
• We just learned how to concatenate many variables to obtain a bigger one (in the definition
of the parallel composition the output of the system results as a concatenation of the
outputs of the sub-systems it contains)
• Is good to know there is also a risky way to specify the connections when a module is
instantiated into another: to put the name of connections in the appropriate positions in
the connection list (in the last example)
By composition we add new modules in the system, but we don’t change the class to which
the system belongs. The system gains the behaviors of the added modules but nothing more.
By composition we sum behaviors only, but we can not introduce in this way a new kind of
behavior in the world of digital machines. What we can’t do using new modules we can do with
an appropriate connection: the loop.
3.4. SPEEDING BY PIPELINING
3.4
63
Speeding by pipelining
One of the main limitation in applying the composition is due to the increased propagation time
associated to the serially connected circuits. Indeed, the time for computing the function f is:
tf = max(th 1 , . . . , th
m)
+ tg
In the last example, the inner product is computed in:
tinner
product
= tmultiplication + t4
number add
If the 4-number add is also composed using 2-number add (as in usual systems) results:
tinner
product
= tmultiplication + 2 × taddition
For the general case of n-components vectors the inner product will be computed, using a similar
approach, in:
tinner
product (n)
= tmultiplication + taddition × log2 n ∈ O(log n)
For this simple example, of computing the inner product of two vectors, results for n ≥ n0 a
computational time bigger than can be accepted in some applications. Having enough multipliers, the multiplication will not limit the speed of computation, but even if we have infinite
2-input adders the computing time will remain dependent by n.
The typical case is given by the serial composition (see Figure 3.6a), where the function
out = f (in) = g(h 1(in)) must be computed using 2 serial connected circuits, h 1(in) and
g(int out), in time:
tf = th 1 + tg .
A solution must be find to deal with the too deep circuits resulting from composing to many
or to “lazy” circuits.
First of all we must state that fast circuits are needed only when a lot of data is waiting to
be computed. If the function f (in) is rarely computed, then we do not care to much about the
speed of the associated circuit. But, if there is an application supposing a huge stream of data
to be successively submitted to the input of the circuit f , then it is very important to design a
fast version of it.
Golden rule: only what is frequently computed must be accelerated!
3.4.1
Register transfer level
The good practice in a digital system is: any stream of data is received synchronously and it is
sent out synchronously. Any digital system can be reduced to a synchronous machine receiving
a stream of input data and generating another stream of output results. As we already stated,
a “robust” digital design is a fully buffered one because it provides a system interfaced to the
external “world” with registers.
The general structure of a system performing the function f (x) is shown in Figure 3.8a,
where it is presented in the fully buffered version. This kind of approach is called register
transfer level (RTL) because data is transferred, modified by the function f , from a register,
input reg, to another register, output reg. If f = g(h 1(x)), then the clock frequency is limited
to:
1
1
=
treg + tf + tsu
treg + th 1 + tg + tsu
The serial connection of the module computing h 1 and g is a fundamental limit. If f
computation is not critical for the system including the module f , then this solution is very
good, else you must read and assimilate the next, very important, paragraph.
fclock
max
=
64
3.4.2
CHAPTER 3. GROWING & SPEEDING & FEATURING
Pipeline structures
To increase the processing speed of a long stream of data the clock frequency must be increased.
If the stream has the length n, then the processing time is:
Tstream (n) =
clock
1
× (n + 2) = (treg + th
fclock
in
1
+ tg + tsu ) × (n + 2)
in
clock
❄
❄
input reg
input reg
sync in
sync in
❄
❄
h 1(x)
f(x) = g(h 1(x))
int out
❄
pipeline reg
out
❄
sync int out
❄
output reg
❄
g(z)
sync out
out
❄
output reg
❄
sync out
a.
b.
Figure 3.8: Pipelined computation. a. A typical Register Transfer Logic (RTL) configuration.
Usually it is supposed a “deep” combinational circuit computes f (x). b. The pipeline structure splits the
combinational circuit associated with function f (x) in two less “deep” circuits and inserts the pipeline
register in between.
The only way to increase the clock rate is to divide the circuit designed for f in two serially
connected circuits, one for h 1 and another for g, and to introduce between them a new register.
Results the system represented in Figure 3.8b. Its clock frequency is:
fclock
max
=
1
max(th 1 , tg ) + treg + tsu
and the processing time for the same string is:
Tstream (n) = (max(th 1 , tg ) + treg + tsu ) × (n + 3)
The two systems represented in Figure 3.8 are equivalent. The only difference between them
is that the second performs the processing in n + 3 clock cycles instead of n + 2 clock cycles
for the first version. For big n, the current case, this difference is a negligible quantity. We call
latency the number of the additional clock cycle. In this first example latency is: λ = 1.
This procedure can be applied many times, resulting a processing “pipe” with a latency
equal with the number of the inserted register added to the initial system. The resulting system
is called a pipelined system. The additional registers are called pipeline registers.
The maximum efficiency of a pipeline system is obtained in the ideal case when, for an
(m + 1)-stage pipeline, realized inserting m pipeline registers:
max(tstage 0 , tstage 1 , . . . , tstage
m)
=
tstage
0
+ tstage 1 + . . . + tstage
m+1
m
3.4. SPEEDING BY PIPELINING
65
max(tstage 0 , tstage 1 , . . . , tstage
m)
>> treg + tsu
λ = m << n
In this ideal case the speed is increased almost m times. Obviously, no one of these condition
can be fully accomplished, but there are a lot of real applications in which adding an appropriate
number of pipeline stages allows to reach the desired speed performance.
Example 3.6 The pseudo-Verilog code for the 2-stage pipeline system represented in Figure 3.8
is:
module pipelined_f(
output reg [size_in-1:0]
input
[size_out-1:0]
reg
[size_in-1:0]
sync_in;
wire
[size_int-1:0] int_out,
reg
[size_int-1:0] sync_int_out;
wire
[size_out-1:0] out;
h_1 this_h_1(
.out
(int_out),
.in
(sync_in));
g
this_g( .out
(out),
.in
(sync_int_out));
always @(posedge clock) begin
sync_in
sync_int_out
sync_out
end
endmodule
module h_1(
output [size_int-1:0]
input
[size_in-1:0]
assign #15 out = ...;
endmodule
module g(
output [size_out-1:0]
input
[size_int-1:0]
assign #18 out = ...;
endmodule
sync_out,
in);
<= #2 in;
<= #2 int_out;
<= #2 out;
out,
in);
out,
in);
Suppose, the unit time is 1ns. The maximum clock frequency for the pipeline version is:
fclock =
1
GHz = 50M Hz
max(15, 18) + 2
This value must be compared with the frequency of the non-pipelined version, which is:
fclock =
1
GHz = 28.57M Hz
15 + 18 + 2
Adding only a simple register and accepting a minimal latency (λ = 1), the speed of the system
increased with 75%. ⋄
3.4.3
Data parallelism vs. time parallelism
The two limit cases of composition correspond to the two extreme cases of parallelism in digital
systems:
• the serial composition will allow the pipeline mechanism which is a sort of parallelism
which could be called diachronic parallelism or time parallelism
66
CHAPTER 3. GROWING & SPEEDING & FEATURING
• the parallel composition is an obvious form of parallelism, which could be called synchronic
parallelism or data parallelism.
The data parallelism is more obvious: m functions, h 1, . . . , h m, are performed in parallel
by m circuits (see Figure 3.6b). But, time parallelism is not so obvious. It acts only in a
pipelined serial composition, where the first stage is involved in computing the most recently
received data, the second stage is involved in computing the previously received data, and so
on. In an (m + 1)-stage pipeline structure m + 1 elements of the input stream are in different
stages of computation, and at each clock cycle one result is provided. We can claim that in such
a pipeline structure m + 1 computations are done in parallel with the price of a latency λ = m.
✛
a3
a2
b3
❄
❄
register
a1
b2
❄ ❄
∗
✲
✲
DATA PARALLELISM
∗
✲
❄
❄
✲
register
register
❄
b0
❄
∗
❄
∗
✲
✻
register
❄ ❄
+
✲
a0
b1
❄
+
✲
register
❄
TIME PARALLELISM
register
❄
+
✲
clock
register
❄
❄
Figure 3.9: The pipelined inner product circuit for 4-component vectors. Each multiplier
and each adder send its result in a pipeline register. For this application results a three level pipeline
structure with different degree of parallelism. The two kind of parallelism are exemplified. Data parallel
has the maximum degree on the first level. The degree of time parallelism is three: in each clock cycle
three pairs of 4-element vectors are processed. One pair in the first stage of multiplications, another pair
is the second stage of performing two additions, and one in the final stage of making the last addition.
The previous example of a 2-stage pipeline accelerated the computation because of the time
parallelism which allows to work simultaneously on two input data, on one applying the function
h 1 and in another applying the function g. Both being simpler than the global function f , the
increase of clock frequency is possible allowing the system to deliver results at a higher rate.
Computer scientists stress on both type of parallelism, each having its own fundamental
limitations. More, each form of parallelism bounds the possibility of the other, so as the parallel
processes are strictly limited in now a day computation. But, for us it is very important to
emphasize in this stage of the approach that:
circuits are essentially parallel structures with both the possibilities and
the limits given by the mechanism of composition.
The parallel resources of circuits will be limited also, as we will see, in the process of closing
loops one after another with the hope to deal better with complexity.
Example 3.7 Let us revisit the problem of computing the scalar product. We redesign the circuit
in a pipelined version using only binary functions.
3.5. FEATURING BY CLOSING NEW LOOPS
module pipelined_inner_prod(
output
input
input
[17:0]
[7:0]
67
out,
a3, a2, a1, a0, b3, b2, b1, b0,
clock);
wire[15:0] p3, p2, p1, p0;
wire[17:0] s1, s0;
mult
mult3(p3, a3, b3, clock),
mult2(p2, a2, b2, clock),
mult1(p1, a1, b1, clock),
mult0(p0, a0, b0, clock);
add
add11(s1, {1’b0, p3}, {1’b0, p2}, clock),
add10(s0, {1’b0, p1}, {1’b0, p0}, clock),
add0(out, s1[16:0], s0[16:0], clock);
endmodule
module mult(
output reg [15:0]
input
[7:0]
input
always @(posedge clock)
out
endmodule
module add(output reg [17:0]
input
[16:0]
input
always @(posedge clock)
endmodule
out,
m1, m0,
clock);
<= m1 * m0;
out,
t1, t0,
clock);
out <= t1 + t0;
⋄
The structure of the pipelined inner product (dot product) circuit is represented in Figure
3.9. It shows us the two dimensions of the parallel computation. The horizontal dimension
is associated with data parallelism, the vertical dimension is associated with time parallelism.
The first stage allows 4 parallel computation, the second allows 2 parallel computation, and the
last consists only in a single addition. The mean value of the degree of data parallelism is
2.33. The system has latency 2, allowing 7 computations in parallel. The peak performance
of this system is the whole degree of parallelism which is 7. The peak performance is the
performance obtained if the input stream of data is uninterrupted. If it is interrupted because
of the lack of data, or for another reason, the latency will act reducing the peak performance,
because some or all pipeline stages will be inactive.
3.5
Featuring by closing new loops
A loop connection is a very simple thing, but the effects introduced in the system in which it is
closed are sometimes surprising. All the time are beyond the evolutionary facts. The reason for
these facts is the spectacular effect of the autonomy whenever it manifests. The output of the
system starts to behave less conditioned by the evolution of inputs. The external behavior of the
system starts to depend more and more by something like an “internal state” continuing with
a dependency by an “internal behavior”. In the system starts to manifest internal processes
seem to be only partially under the external control. Because the loop allows of system to act
on itself, the autonomy is the first and the main effect of the mechanism of closing loops. But,
the autonomy is only a first and most obvious effect. There are others, more subtle and hidden
consequences of this apparent simple and silent mechanism. This book is devoted to emphasize
deep but not so obvious correlations between loops and complexity. Let’s start with the definition
and a simple example.
68
CHAPTER 3. GROWING & SPEEDING & FEATURING
in
❄
❄
in1
in0
out1
out0
❨
the loop
loop
❄out
Figure 3.10: The loop closed into a digital system. The initial system has two inputs, in1 and
in0, and two outputs, out1 and out0. Connecting out0 to in0 results a new system with in and out
only.
Definition 3.11 The loop consists in connecting some outputs of a system to some of its inputs
(see Figure 3.10), as in the pseudo-Verilog description that follows:
module loop_system #(‘include "parameters.v")( output
input
wire
[loop_dim-1:0]
the_loop;
no_loop_system our_module( .out1
(out)
,
.out0
(the_loop) ,
.in1
(in)
,
.in0
(the_loop) );
endmodule
module no_loop_system #(‘include "parameters.v")(
[out_dim-1:0]
[in_dim-1:0]
output
output
input
input
out ,
in );
[out_dim-1:0]
[loop_dim-1:0]
[in_dim-1:0]
[loop_dim-1:0]
out1
out0
in1
in0
,
,
,
);
/* The description of ’no_loop_system’ module */
endmodule
⋄
The most interesting thing in the previous definition is a “hidden variable” occurred in
module loop system(). The wire called the loop carries the non-apparent values of a variable
evolving inside the system. This is the variable which evolves only internally, generating the
autonomous behavior of the system. The explicit aspect of this behavior is hidden, justifying
the generic name of the “internal state evolution”.
The previous definition don’t introduce any restriction about how the loop must be closed. In
order to obtain desired effects the loop will be closed keeping into account restrictions depending
by each actual situation. There also are many technological restrictions that impose specific
modalities to close loops at different level in a complex digital system. Most of them will be
discussed later in the next chapters.
Example 3.8 Let be a synchronous adder. It has the outputs synchronized with an positive
edge clocked register (see Figure 3.11a). If the output is connected back to one of its input, then
results the structure of an accumulator (see Figure 3.11b). The Verilog description follows.
module acc(output[19:0] out, input[15:0] in, input clock, reset);
sync_add our_add(out, in, out, clock, reset);
endmodule
module sync_add(
output reg
[19:0]
out
,
3.5. FEATURING BY CLOSING NEW LOOPS
input
[15:0]
input
[19:0]
input
always @(posedge clock) if (reset)
else
endmodule
in1
69
in1
,
in2
,
clock
, reset );
out = 0;
out = in1 + in2;
in
in2
❄
❄
❄
adder
reset
❄
clock
❄
register
❄
adder
reset
❄ ❄
clock
register
❄out
a.
❄ out
b.
Figure 3.11: Example of loop closed over an adder with synchronized output. If the
output becomes one of the inputs, results a circuit that accumulates at each clock cycle. a. The initial
circuit: the synchronized adder. b. The resulting circuit: the accumulator.
In order to make a simulation the next test acc module is written:
module test_acc;
reg clock, reset;
reg[15:0] in;
wire[19:0] out;
initial begin
clock = 0;
forever #1 clock = ~clock;
end // the clock
initial begin
reset = 1;
#2 reset = 0;
#10 $stop;
end
always @(posedge clock) if (reset)
in = 0;
else
in = in + 1;
acc dut(out, in, clock, reset);
initial $monitor("time=%0d clock=%b in=%d out=%d",
$time, clock, in, dut.out);
endmodule
By simulation results the following behavior:
#
#
#
#
#
#
#
#
#
#
time=0
time=1
time=2
time=3
time=4
time=5
time=6
time=7
time=8
time=9
clock=0
clock=1
clock=0
clock=1
clock=0
clock=1
clock=0
clock=1
clock=0
clock=1
in=
in=
in=
in=
in=
in=
in=
in=
in=
in=
x
0
0
1
1
2
2
3
3
4
out=
out=
out=
out=
out=
out=
out=
out=
out=
out=
x
0
0
1
1
3
3
6
6
10
70
CHAPTER 3. GROWING & SPEEDING & FEATURING
# time=10 clock=0 in=
# time=11 clock=1 in=
4 out=
5 out=
10
15
⋄
The adder becomes an accumulator. What is spectacular in this fact? The step made by
closing the loop is important because an “obedient” circuit, whose outputs followed strictly
the evolution of its inputs, becomes a circuit with the output depending only partially by the
evolution of its inputs. Indeed, the the output of the circuit depends by the current input but,
in the same time, depends by the content of the register, i.e., by the “history accumulated” in
it. The output of adder can be predicted starting from the current inputs, but the output of
the accumulator supplementary depends by the state of circuit (the content of the register). It
was only a simple example, but I hope, useful to pay more attention to loop.
3.5.1
∗ Data dependency
The good news about loop is its “ability” to add new features. But any good news is accompanied by its
own bad news. In this case is about the limiting of the degree of parallelism allowed in a system with a
just added loop. It is mainly about the necessity to stop sometimes the input stream of data in order to
decide, inspecting an output, how to continue the computation. The input data waits for data arriving
from an output a number of clock cycles related with the system latency. To do something special the
system must be allowed to accomplish certain internal processes.
Both, data parallelism and time parallelism are possible because when the data arrive the system
“knows” what to do with them. But sometimes the function to be applied on certain input data is decided
by processing previously received data. If the decision process is to complex, then new data can not be
processed even if the circuits to do it are there.
Example 3.9 Let be the system performing the following function:
procedure cond_acc(a,b, cond);
out = 0;
end = 0;
loop
if (cond = 1)
out = out + (a + b);
else
out = out + (a - b);
until (end = 1) // the loop is unending
end
For each pair of input data the function is decided according to a condition input.
The Verilog code describing an associated circuit is:
module cond_acc0(
output reg [15:0] out,
input
[15:0] a, b,
input
cond, reset, clock);
always @(posedge clock)
if (reset)
out <= 0;
else if (cond)
out <= out + (a + b);
else
out <= out + (a - b);
endmodule
In order to increase the speed of the circuit a pipeline register is added with the penalty of λ = 1.
Results:
module cond_acc1(
output
input
input
reg[15:0] pipe;
always @(posedge clock)
reg [15:0]
[15:0]
out,
a, b,
cond, reset, clock);
3.5. FEATURING BY CLOSING NEW LOOPS
71
a
cond
a
cond
b
✲ ❄ ❄
+/-
b
✲ ❄ ❄
❄
+/-
pipe reg
❄ ❄
❄ ❄
+
+
❄
out reg
clock
❄
out reg
clock
out
❄
out
❄
a.
b.
a
b
✲ ❄ ❄
a
+/-
b
✲ ❄ ❄
❄
+/-
pipe reg
❄ ❄
❄ ❄
+
clock
+
❄
❄
out reg
out reg
clock
❄
out[15]
out
out[15]
c.
out
❄
d.
Figure 3.12: Data dependency when a loop is closed in a pipelined structure. a. The
non-pipelined version. b. The pipelined version. c. Adding a loop to the non-pipelined version. d.
To the pipelined version the loop can not be added without supplementary precautions because data
dependency change the overall behavior. The selection between add and sub, performed by the looped
signal comes too late.
if (reset)
else begin
begin
out <= 0;
pipe <= 0;
end
if (cond)
pipe <= a + b;
else
pipe <= a - b;
out <= out + pipe;
end
endmodule
Now let us close a loop
input takes the value of the
on each pair of input data
the computation performed
addapt acc.
in the first version of the system (without pipeline register). The condition
sign of the output. The loop is: cond = out[15]. The function performed
in each clock cycle is determined by the sign of the output resulted from
with the previously received pairs of data. The resulting system is called
module addapt_acc0(output [15:0] out,
input
[15:0] a, b,
input
reset, clock);
cond_acc0 cont_acc0( .out
(out),
72
CHAPTER 3. GROWING & SPEEDING & FEATURING
.a
.b
.cond
.reset
.clock
(a),
(b),
(out[15]),
(reset),
(clock));
// the loop
endmodule
Figure 3.12a represents the first implementation of the cond acc circuit, characterized by a low clock
frequency because both the adder and the adder/subtracter contribute to limiting the clock frequency:
fclock =
1
t+/− + t+ + treg
Figure 3.12b represents the pipelined version of the same circuit working faster because only one from
adder and the adder/subtracter contributes to limiting the clock frequency:
fclock =
1
max(t+/− , t+ ) + treg
A small price is paid by λ = 1.
The 1-bit loop closed from the output out[15] to cond input (see Figure 3.12c) allows the circuit to
decide itself if the sum or the difference is accumulated. Its speed is identical with the initial, no-loop,
circuit.
Figure 3.12d warns us against the expected damages of closing a loop in a pipelined system. Because
of the latency the “decision comes” to late and the functionality is altered. ⋄
In the system from Figure 3.12a the degree of parallelism is 1, and in Figure 3.12b the system has
the degree of parallelism 2, because of the pipeline execution. When we closed the loop we where obliged
to renounce to the bigger degree of parallelism because of the latency associated with the pipe. We have
a new functionality – the circuit decides itself regarding the function executed in each clock cycle – but
we must pay the price of reducing the speed of the system.
According to the algorithm the function performed by the block +/- depends on data received in the
previous clock cycles. Indeed, the sign of the number stored in the output register depends on the data
stream applied on the inputs of the system. We call this effect data dependency. It is responsible for
limiting the degree of parallelism in digital circuits.
The circuit from Figure 3.12d is not a solution for our problem because the condition cond comes to
late. It corresponds to the operation executed on the input stream excepting the most recently received
pair of data. The condition comes too late, with a delay equal with the latency introduced by the pipeline
execution.
3.5.2
∗ Speculating to avoid limitations imposed by data dependency
How can we avoid the speed limitation imposed by a new loop introduced in a pipelined execution? It is
possible, but we must pay a price enlarging the structure of the circuit.
If the circuit does not know what to do, addition or subtract in our previous example, then in it will be
compute both in the first stage of pipeline and will delay the decision for the next stage so compensating
the latency. We use the same example to be more clear.
Example 3.10 The pipelined version of the circuit addapt acc is provided by the following Verilog code:
module addapt_acc1(output reg [15:0]
input
[15:0]
input
reg
[15:0] pipe1, pipe0;
always @(posedge clock)
if (reset) begin
out <= 0;
pipe1 <= 0;
pipe0 <= 0;
out,
a, b,
reset, clock);
3.6. CONCLUDING ABOUT COMPOSING & PIPELINING & LOOPING
b
a
❄ ❄
73
❄ ❄
-
+
❄❄
❄❄
pipe1
reset
pipe0
clock
✲
❄ ❄
1
0
mux2 16
❄ ❄
+
❄❄
out
out[15]
❄
out
Figure 3.13: The speculating solution to avoid data dependency. In order to delay the
moment of decision both addition and subtract are computed on the first stage of pipeline. Speculating
means instead to decide what to do, addition or subtract, we decide what to consider after doing both.
else
end
begin
pipe1 <= a + b;
pipe0 <= a - b;
if (out[15])
out <= out + pipe1;
else
out <= out + pipe0;
end
endmodule
The execution time for this circuit is limited by the following clock frequency:
fclock =
1
1
≃
max(t+ , t− , (t+ + tmux )) + treg
t− + treg
The resulting frequency is very near to the frequency for the pipeline version of the circuit designed in the
previous example.
Roughly speaking, the price for the speed is: an adder & two registers & a multiplexer (see for
comparing Figure 3.12c and Figure 3.13). Sometimes it deserves! ⋄
The procedure applied to design addapr acc1 involves the multiplication of the physical resources.
We speculated, computing on the first level of pipe both the sum and the difference of the input values.
On the second state of pipe the multiplexer is used to select the appropriate value to be added to out.
We call this kind of computation speculative evaluation or simply speculation. It is used to accelerate
complex (i.e., “under the sign” of a loop) computation. The price to be paid is an increased dimension
of the circuit.
3.6
Concluding about composing & pipelining & looping
The basic ideas exposed in this section are:
• a digital system develops applying two mechanisms: composing functional modules and
closing new loops
74
CHAPTER 3. GROWING & SPEEDING & FEATURING
• by composing the system grows in size improving its functionality with the composed
functions
• by closing loops the system gains new features which are different from the previous
functionality
• the composition generates the conditions for two kinds of parallel computation:
– data parallel computation
– time parallel computation (in pipeline structures)
• the loop limits the possibility to use the time parallel resources of a system because of
data dependency
• a speculative approach can be used to accelerate data dependent computation in pipeline
systems; it means the execution of operation whose result may not actually be needed; it
is an useful optimization when early execution accelerates computation justifying for the
wasted effort of computing a value which is never used
• circuits are mainly parallel systems because of composition (some restriction may apply
because of loops).
Related with the computing machines Flynn [Flynn ’72] introduced three kind of parallel
machines:
• MIMD (multiple-instructions-multiple data), which means mainly having different programs working on different data
• SIMD (single-instructions-multiple-data), which means having one program working on
different data,
• MISD (multiple-instructions-single-data), which means having different programs working
on the same data.
Related with the computing a certain function also three kind of almost the same parallelism
can be emphasized:
• time parallelism, which is somehow related with MIMD execution, because in each temporal
stage a different operation (instruction) can be performed
• data parallelism, which is identic with SIMD execution
• speculative parallelism, which is a sort of MISD execution.
Thus, the germs of parallel processes, developed at the computing machine level, occur, at an
early stage, at the circuit level.
3.7
Problems
Speculative circuits
Problem 3.1 Let be the circuit described by the following Verilog module:
3.7. PROBLEMS
module xxx( output
input
input
input
input
input
75
reg [31:0]
[31:0]
[31:0]
[31:0]
a
x1
x2
x3
clock
reset
,
,
,
,
,
);
always @(posedge clock) if (reset) a <= 0;
else if (a > x3)
else
endmodule
a <= a + (x1 + x2);
a <= a + (x1 - x2);
The maximum frequency of clock is limited by the propagation time through the internal loop
(tregr eg ) or by tin reg . To maximize the frequency a speculative solution is asked.
Problem 3.2 Provide the speculative solution for the next circuit.
module yyy( output
output
input
input
input
input
reg [31:0]
reg [31:0]
[31:0]
[31:0]
a
b
x1
x2
clock
reset
,
,
,
,
,
);
always @(posedge clock)
if (reset) begin
a <= 0;
b <= 0;
end
else
case({a + b > 8’b101, a - b < 8’b111})
2’b00: {a,b} <= {a + (x1-x2),
2’b01: {a,b} <= {a + (8’b10 * x1+x2),
2’b10: {a,b} <= {a + (8’b100 * x1-x2),
2’b11: {a,b} <= {a + (x2-x1),
endcase
endmodule
b
b
b
b
+
+
+
+
(x1+x2)
};
(x1+8’b10 * x2) };
(8’b100 * x2)
};
(8’b100 * x1+x2)};
Problem 3.3 The following circuit has two included loops. The speculation will increase the
dimension of the circuit accordingly. Provide the speculative version of the circuit.
module zzz( output
input
input
input
input
input
reg
[15:0]
reg [31:0]
[15:0]
[15:0]
[15:0]
out,
x1,
x2,
x3,
clock,
reset);
acc;
always @(posedge clock)
if (reset)
begin
out <= 0;
acc <= 0;
end
else
begin
out <= acc * x3;
if (out[15])
acc <= acc + x1 + x2 + out[31:0];
else
acc <= acc + x1 - x2 + out[15:0];
end
endmodule
76
3.8
CHAPTER 3. GROWING & SPEEDING & FEATURING
Projects
Use Appendix How to make a project to learn how to proceed in implementing a project.
Project 3.1
Chapter 4
THE TAXONOMY OF DIGITAL
SYSTEMS
In the previous chapter
the basic mechanisms involved in defining the architecture of a digital system were
introduced:
• the parallel composing and the serial composing are the mechanism allowing two kind
of parallelism in digital systems – data parallelism & time parallelism – both involved
in increasing the “brute force” of a computing machine
• the pipeline connection supports the time parallelism, accelerating the inherent serial
computation
• closing loops new kinds of functionality are allowed (storing, behaving, interpreting,
... self-organizing)
• speculating is the third type of parallelism introduced to compensate the limitations
generated by loops closed in pipelined systems
In this chapter
loops are used to classify digital systems in orders, takeing into account the increased
degree of autonomy generated by each new added loop. The main topics are:
• the autonomy of a digital system depends on the number of embedded loops closed
inside
• the loop based taxonomy of digital systems developed to match the huge diversity of
the systems currently developed
• some preliminary remarks before starting to describe in detail digital circuits and
how they can be used to design digital systems
In the next chapter
the final target of our lessons on digital design is defined as the structure of the simplest
machine able to process a stream of input data providing another stream of output data.
The functional description of the machine is provided emphasizing:
• the external connections and how they are managed
• the internal control functions of the machine
• the internal operations performed on the received and internally stored data.
77
78
CHAPTER 4. THE TAXONOMY OF DIGITAL SYSTEMS
A theory is a compression of data; comprehension is compression.
Gregory Chaitin1
Any taxonomy is a compressed theory, i.e., a
compression of a compression. It contains,
thus, illuminating beauties and dangerous insights for our way to comprehend a technical
domain. How can we escape from this attractive trap? Trying to comprehend beyond what
the compressed data offers.
4.1
Loops & Autonomy
The main and the obvious effect of the loop is the autonomy it can generate in a digital system.
Indeed, the first things we observe in a circuit in which a new loop is introduced are new and
independent behaviors. Starting with a simple example the things will become more clear in an
easy way. We use an example with a system initially defined by a transition table. Each output
corresponds to an input with a certain delay (one time unit, #1, in our example). After closing
the loop, starts a sequential process, each sequence taking time corresponding with the delay
introduced by the initial system.
Example 4.1 Let be the digital system initSyst from Figure 4.1a, with two inputs, in, lp,
and one output, out. What hapend when is closed the loop from the output out to the input lp?
Let’s make it. The following Verilog modules describe the behavior of the resulting circuit.
module loopSyst(
output [1:0]
out,
input
in);
initSyst noLoopSyst(.out(out), .in(in), .loop(out));
endmodule
module initSyst(
output
input
input
reg [1:0]
[1:0]
out ,
in ,
loop);
initial out = 2’b11; // only for simulation purpose
always @(in or loop) #1 case ({in, loop})
3’b000: out
3’b001: out
3’b010: out
3’b011: out
3’b100: out
3’b101: out
3’b110: out
3’b111: out
endcase
endmodule
1
From [Chaitin ’06]
=
=
=
=
=
=
=
=
2’b01;
2’b00;
2’b00;
2’b10;
2’b01;
2’b10;
2’b11;
2’b01;
4.1. LOOPS & AUTONOMY
79
In order to see how behave loopSyst we will use the following test module which initialize
(for this example in a non-orthodox fashion because we don’t know nothing about the internal
structure of initSyst) the output of initSyst in 11 and put on the input in for 10 unit time
the value 0 and for the next 10 unit time the value 1.
module test;
reg in;
wire[1:0] out;
initial begin
in = 0;
#10
in = 1;
#10
$stop;
end
loopSyst dut(out, in);
initial $monitor(
"time=%0d in=%b out=%b",
$time, in, dut.out);
endmodule
{in, lp} out
in
❄
lp
❄
❨
initSyst
loopSyst
❄out
a.
b.
❄
00
01
00
✻♦
01
00
00
10
01
10
11
01
00
01
10
11
00
01
10
11
✲
❄
01
❄
10
10
✻
❄
11
c.
0
0
0
0
1
1
1
1
11
d.
Figure 4.1: Example illustrating the autonomy. a. A system obtained from an initial system
in which a loop is closed from output to one of its input. b. The transition table of the initial system
where each output strict corresponds to the input value. c. The output evolution for constant input: in
= 0. d. The output evolution for a different constant input: in = 1.
The simulation offers us the following behavior:
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
time=0
time=1
time=2
time=3
time=4
time=5
time=6
time=7
time=8
time=9
time=10
time=11
time=12
time=13
time=14
time=15
in=0
in=0
in=0
in=0
in=0
in=0
in=0
in=0
in=0
in=0
in=1
in=1
in=1
in=1
in=1
in=1
out=11
out=10
out=00
out=01
out=00
out=01
out=00
out=01
out=00
out=01
out=00
out=01
out=10
out=11
out=01
out=10
80
CHAPTER 4. THE TAXONOMY OF DIGITAL SYSTEMS
#
#
#
#
time=16
time=17
time=18
time=19
in=1
in=1
in=1
in=1
out=11
out=01
out=10
out=11
The main effect we want to emphasize is the evolution of the output under no variation of
the input in. The initial system, defined in the previous case, has an output that switches only
responding to the input changing (see also the table from Figure 4.1b). The system which results
closing the loop has its own behavior. This behavior depends by the input value, but is triggered
by the events coming through the loop. Figure 4.1c shows the output evolution for in = 0 and
Figure 4.1d represents the evolution for in = 1. ⋄
VerilogSummary 3 :
• the register reg[1:0] out defined in the module initSyst is nor a register, it is a Verilog
variable, whose value is computed by a case procedure anytime at least one of the two
inputs change (always @(in or lp))
• a register which changes its state “ignoring” a clock edge is not a register, it is a variable
evolving like the output of a combinational circuit
• what is the difference between an assign and an always (a or b or ...)? The body
of assign is continuously evaluated, rather than the body of always which is evaluated
only if at least an element of the list of sensitivity ((a or b or ...)) changes
• in running a simulation an assign is more computationally costly in time than an always
which is more costly in memory resources.
Until now we used in a non-rigorous manner the concept of autonomy. It is necessary for
our next step to define more clearly this concept in the digital system domain.
Definition 4.1 In a digital system a behavior is called autonomous iff for the same input
dynamic there are defined more than one distinct output transitions, which manifest in distinct
moments. ⋄
If we take again the previous example we can see in the result of the simulation that in the
moment time = 2 the input switches from 0 to 0 and the output from 10 to 00. In the next
moment input switches the same, but output switches from 00 to 10. The input of the system
remains the same, but the output behaves distinctly. The explanations is for us obvious because
we have access to the definition of the initial system and in the transition table we look for the
first transition in the line 010 and we find the output 00 and for the second in the line 000
finding there 00. The input of the initial system is changed because of the loop that generates
a distinct response.
In our example the input dynamic is null for a certain output dynamic. There are example
when the output dynamic is null for some input transitions (will be found such examples when
we talk about memories).
Theorem 4.1 In the respect of the previous definition for autonomy, closing an internal loop
generates autonomous behaviors. ⋄
Proof
Let be The description of ’no loop system’ module from Definition 3.4 described, in the general form, by the following pseudo-Verilog construct:
4.2. CLASSIFYING DIGITAL SYSTEMS
always @(in1 or in0) #1
case (in1)
... :
case (in0)
... : {out1,
...
... : {out1,
endcase
...
... :
case (in0)
... : {out1,
...
... : {out1,
endcase
endcase
81
out0} = f_00(in1, in0);
out0} = f_0p(in1, in0);
out0} = f_q0(in1, in0);
out0} = f_qp(in1, in0);
The various occurrences of {out1, out2} are given by the functions f ij(in1, in2) defined
in Verilog.
When the loop is closed, in0 = out0 = state, the in1 remains the single input of the
resulting system, but the internal structure of the system continue to receive both variable, in1
and in0. Thus, for a certain value of in1 there are more Verilog functions describing the next
value of {out1, out0}. If in1 = const, then the previous description is reduced to:
always @(state) #1 case (state)
... : {out1, state} = f_i0(const, state);
...
... : {out1, state} = f_ip(const, state};
endcase
The output of the system, out1, will be computed for each change of the variable state,
using the function f ji selected by the new value of state, which function depends by state.
For each constant value of in1 another set of functions is selected. In the two-level case, which
describe no loop system, this second level is responsible for the autonomous behavior.
⋄
4.2
Classifying Digital Systems
The two mechanisms, of composing and of ”looping”, give us a very good instrument for a
new classification of digital systems. If the system grows by different compositions, then it
allows various kinds of connections. In this context the loops are difficult to be avoided. They
occur sometimes in large systems without the explicit knowledge of the designer, disturbing
the design process. But, usually we design being aware of the effect introduced by this special
connection – the loop. This mechanism leads us to design a complex network of loops which
include each other. Thus, in order to avoid ambiguities in using the loops we must define what
means ”included loop”. We shall use frequently in the next pages this expression for describing
how digital systems are built.
Definition 4.2 A loop includes another loop only when it is closed over a serial or a serialparallel composition which have at least one subsystem containing an internal loop, called an
included loop. ⋄
Attention! In a parallel composition a loop going through one of the parallel connected
subsystem does not include a loop closed over another parallel connected subsystem. A new
loop of the kind “grows” only a certain previously closed loop, but does not add a new one.
82
CHAPTER 4. THE TAXONOMY OF DIGITAL SYSTEMS
Example 4.2 In Figure 4.2 the loop (1) is included by the loop (2). In a serial composition
built with S1 and S2 interconnected by (3), we use the connection (2) to add a new loop. ⋄
X
❄
✯
the loop which includes
(2)
the serial connection
❄
S1
✶ (3)
❄
❄ ✐
S2
(1)
the included loop
Y
❄
Figure 4.2: Included loops. The loop (2) includes loop (1), closed over the subsystem S2 , because
S2 is serially connected with the subsystem S1 and loop (2) includes both S1 and S2 .
Now we can use the next recursive definition for a new classification of digital systems. The
classification contains orders, from 0 to n.
Definition 4.3 Let be a n-order system, n-OS. A (n+1)-OS can be built only adding a new loop
which includes the first n loops. The 0-OS contains only combinational circuits (the loop-less
circuits). ⋄
This classification in orders is very consistent with the nowadays technological reality for
n < 5. Over this order the functions of digital systems are imposed mainly by information, this
strange ingredient who blinks in 2-OS, is born in 3-OS and grows in 4-OS monopolizing the
functional control in digital systems (see Chapter 16 in this book). But obviously, a function
of a circuit belonging of certain order can be performed also by circuits from any higher ones.
For this reason we use currently circuits with more than 4 loops only for they allow us to apply
different kind of optimizations. Even if a new loop is not imposed by the desired functionality,
we will use it sometimes because of its effect on the system complexity. As will be exemplified,
a good fitted loop allows the segregation of the simple part from an apparent complex system,
having as main effect a reduced complexity.
Our intention in the second part of this book is to propose and to show how works the
following classification:
0-OS - combinational circuits, with no autonomy
1-OS - memories, having the autonomy of internal state
2-OS - automata, with the autonomy to sequence
3-OS - processors, with the autonomy to control
4-OS - computers, with the autonomy to interpret
...
n-OS - systems with the highest autonomy: to self-organize.
This new classification can be exemplified2 (see also Figure 4.3) as follows:
2
For almost all the readers the following enumeration is now meaningless. They are kindly invited to revisit
this end of chapter after assimilating the first 7 chapter of this book.
4.2. CLASSIFYING DIGITAL SYSTEMS
83
❄ ❄
CLC
❄
0-OS: Combinational Circuit
1-OS: Memory Circuit
commands
clock
❄
❄
CLC
Register
❄
state
✻ ❄
2-OS: Automaton
❄
❄
Effecrory
Automaton
❄
❄
Controll
Automaton
✻ ❄
flags
3-OS: Processor
✻
✻
✻
❄
❄
❄
✛✲
✛✲
✛✲
✛✲
Memory
Processor
✻✻
address
✻✻✻
data
controll
✻
✻
✻
❄
❄
❄
✛✲
✛✲
✛✲
✛✲
✻
✻
✻
❄
❄
❄
✛✲
✛✲
✛✲
✛✲
✻
❄
4-OS: Computer
✻
❄
✻
❄
n-OS: Cellular Automaton
Figure 4.3: Examples of circuits belonging to different orders. A combinational circuit
is in 0-OS class because has no loops. A memory circuit contains one-loop circuits and therefore it is
in 1-OS class. Because the register belongs to 1-OS class, closing a loop containing a register and a
combinational circuit (which is in 0-OS class) results an automaton: a circuit in 2-OS class. Two loop
connected automata – a circuit in 3-OS class – can work as a processor. An example of 4-OS is a simple
computer obtained loop connecting a processor with a memory. Cellular automata contains a number of
loops related with the number of automata it contains.
• 0-OS: gate, elementary decoder (as the simplest parallel composition), buffered elementary
decoder (the simplest serial-parallel composition), multiplexer, adder, priority encoder, ...
• 1-OS: elementary latch, master-slave flip-flop (serial composition), random access memory
(parallel composition), register (serial-parallel composition), ...
• 2-OS: T flip-flop (the simplest two states automaton), J-K flip-flop (the simplest two input
automaton), counters, automata, finite automata, ...
• 3-OS: automaton using loop closed through K-J flip-flops or counters, stack-automata,
elementary processors, ...
• 4-OS: micro-controller, computer (as Processor & RAM loop connected), stack processor,
co-processor
• ...
• n-OS: cellular automaton.
84
CHAPTER 4. THE TAXONOMY OF DIGITAL SYSTEMS
The second part of this book is devoted to sketch a digital system theory based on these
two-mechanism principle of evolving in digital circuits: composing & looping. Starting with
combinational, loop-less circuits with no autonomy, the theory can be developed following the
idea of the increasing system autonomy with each additional loop. Our approach will be a
functional one. We will start with simple functions and we will end with complex structures
with emphasis on the relation between loops and complexity.
4.3
# Digital Super-Systems
When a global loop is introduced in an n-order system results a digital super-system (DSS).
4.4
Preliminary Remarks On Digital Systems
The purpose of this first part of the book is to run over the general characteristics of digital
systems using an informal high level approach. If the reader become accustomed with the basic
mechanisms already described, then in the second part of this book he will find the necessary
details to make useful the just acquired knowledge. In the following paragraphs the governing
ideas about digital systems are summed up.
Combinational circuits vs. sequential circuits Digital systems receive symbols or stream
of symbols on their inputs and generate other symbols or stream of symbols on their outputs
by computation. For combinational systems each generated symbol depends only by the last
recently received symbol. For sequential systems at least certain output symbols are generated
taking into account, instead of only one input symbol, a stream of more than one input symbols.
Thus, a sequential system is history sensitive, memorizing the meaningful events for its own
evolution in special circuits – called registers – using a special synchronization signal – the clock.
Composing circuits & closing loops A big circuit results composing many small ones. A
new kind of feature can be added only closing a new loop. The structural composing corresponds
the the mathematical concept of composition. The loop corresponds somehow to the formal
mechanism of recursion. Composing is an “additive” process which means to put together
different simple function to obtain a bigger or a more complex one. Closing a loop new behaviors
occur. Indeed, when a snake eats a mouse nothing special happens, but if the Orouboros3 serpent
bits its own tail something very special must be expected.
Composition allows data parallelism and time parallelism Digital systems perform
in a “natural” way parallel computation. The composition mechanism generate the context
for the most frequent forms of parallelism: data parallelism (in parallel composition) and time
parallelism (in serial composition). Time parallel computation is performed in pipeline systems,
where the only limitation is the latency, which means we must avoid to stop the flow of data
through the “pipe”. The simplest data parallel systems can be implemented as combinational
circuits. The simplest time parallel systems must be implemented as sequential circuits.
3
This symbol appears usually among the Gnostics and is depicted as a dragon, snake or serpent biting its own
tail. In the broadest sense, it is symbolic of time and the continuity of life. The Orouboros biting its own tail
is symbolic of self-fecundation, or the ”primitive” idea of a self-sufficient Nature - a Nature, that is continually
returning, within a cyclic pattern, to its own beginning.
4.5. PROBLEMS
85
Closing loops disturbs time parallelism The price we pay for the additional features we
get when a new loop is closed is, sometimes, the necessity to stop the data flow through the
pipelined circuits. The stop is imposed by the latency and the effect can be loosing, totaly or
partially, the benefit of the existing time parallelism. Pipelines & loops is a bad mixture, because
the pipe delays the data coming back from the output of the system to its own input.
Speculation can restore time parallelism If the data used to decide comes back to late,
the only solution is to delay also the decision. Follows, instead of selecting what to do, the
need to perform all the computations envisaged by the decision and to select later only the
desired result according to the decision. To do all the computations means to perform speculative
parallel computation. The structure imposed for this mechanism is a MISD (multiple instruction
single data) parallel computation on certain pipeline stage(s). Concluding, three kind of parallel
processes can be stated in a digital system: data parallelism, time parallelism and speculative
parallelism.
Closed loops increase system autonomy The features added by a loop closed in a digital
system refer mainly to different kinds of autonomy. The loop uses the just computed data to
determine how the computation must be continued. It is like an internal decision is partially
driven by the system behavior. Not all sort of autonomy is useful. Some times the increased
autonomy makes the system too “stubborn”, unable to react to external control signals. For this
reason, only an appropriately closed loop generates an useful autonomy, that autonomy which
can be used to minimize the externally exercised control. More about how to close proper loops
in the next chapters.
Closing loops induces a functional hierarchy in digital systems The degree of autonomy is a good criteria to classify digital systems. The proposed taxonomy establishes the degree
of autonomy counting the number of the included loops closed inside a system. Digital system
are classified in orders: the 0-order systems contain no loop circuits, and n-order systems contain
at least one circuit with n included loops. This taxonomy corresponds with the structural and
functional diversity of the circuits used in the actual digital systems.
The top view of the digital circuits domain is almost completely characterized by the previous
features. Almost all of them are not technology dependent. In the following, the physical
embodiment of these concepts will be done using CMOS technology. The main assumptions
grounding this approach may change in time, but now they are enough robust and are simply
stated as follows: computation is an effective formally defined process, specified using finite
descriptions, i.e., the length of the description is not related with the dimension of the processed
data, with the amount of time and of physical resources involved.
Important question: What are the rules for using composition and looping? No rules restrict
us to compose or to loop. The only restrictions come from our limited imagination.
4.5
Problems
Autonomous circuits
Problem 4.1 Prove the reciprocal of Theorem 1.1.
86
CHAPTER 4. THE TAXONOMY OF DIGITAL SYSTEMS
Problem 4.2 Let be the circuit from Problem 1.25. Use the Verilog simulator to prove its
autonomous behavior. After a starting sequence applied on its inputs, keep a constant set of
values on the input and see if the output is evolving.
Can be defined an input sequence which brings the circuit in a state from which the autonomous behavior is the longest (maybe unending)? Find it if it exists.
Problem 4.3 Design a circuit which after the reset generates in each clock cycle the next Fibbonaci number starting from zero, until the biggest Fibbonaci number smaller than 232 . When
the biggest number is generated the machine will start in the next clock cycle from the beginning
with 0. It is supposed the biggest Fibbonaci number smaller than 232 in unknown at the design
time.
Problem 4.4 To the previously designed machine add a new feature: an additional output
generating the index of the current Fibbonaci number.
4.6
Projects
Use Appendix How to make a project to learn how to proceed in implementing a project.
Project 4.1
Chapter 5
OUR FINAL TARGET
In the previous chapter
a new, loop based taxonomy was introduced. Because each newly added loop increases the autonomy of
the system, results a functional circuit hierarchy:
• history free, no-loop, combinational circuits performing logic and arithmetic functions (decoders,
multiplexors, adders, comparators, ...)
• one-loop circuits used mainly as storage support (registers, random access memories, register files,
shift registers, ...)
• two-loop, automata circuits used for recognition, generation, control, in simple (counters, ...) or
complex (finite automata) embodiments
• three-loop, processors systems: the simplest information & circuit entanglement used to perform
complex functions
• four-loop, computing machines: the simplest digital systems able to perform complex programmable
functions, because of the segregation between the simple structure of the circuit and the complex
content of the program memory
• ...
In this chapter
a very simple programmable circuit, called toyMachine, is described using the shortest
Verilog description which can be synthesized using the current tools. It is used to delimit
the list of circuits that must be taught for undergraduates students. This version of a
programmable circuit is selected because:
• its physical implementation contains only the basic structures involved in defining a
digital system
• it is a very small & simple entangled structure of circuits & information used for
defining, designing and building a digital system with a given transfer function
• it has a well weighted complexity so as, after describing all the basic circuits, an
enough meaningful structure can be synthesized.
In the next chapter
starts the second part of this book which describes digital circuits closing a new loop after each chapter.
It starts with the chapter about no-loop digital circuits, discussing about:
• simple (and large sized) uniform combinational circuits, easy to be described using a recursive pattern
• complex and size limited random combinational circuits, whose description’s size is in the same range
with their size
87
88
CHAPTER 5. OUR FINAL TARGET
We must do away with all explanation, and description alone must take its place. . . . The problems are
solved, not by giving new information, but by arranging what we have always known.
Ludwig Wittgenstein
1
Before proceeding to accomplish our targeted project
we must describe it using what we have always
known.
Our final target, for these lessons on Digital Design, is described in this chapter as an
architecture. The term is borrowed from builders. They use it to define the external view and
the functionality of a building. Similarly, in computer science the term of architecture denotes
the external connections and the functions performed by a computing machine. The architecture
does not tell anything about how the defined functionality is actually implemented inside the
system. Usually there are multiple possible solutions for a given architecture.
The way from “what” to “how” is the content of the next part of this book. The architecture
we will describe here states what we intend to do, while for learning how to do, we must know a
lot about simple circuits and the way they can be put together in order to obtain more complex
functions.
5.1
toyMachine: a small & simple computing machine
The architecture of one of the simplest meaningful machine will be defined by (1) its external
connections, (2) its internal state and (3) its transition functions. The transition functions
refer to how both, the internal state (the function f from the general definition) and the outputs
(the function g from the general definition) switch.
Let us call the proposed system toyMachine. It is almost the simplest circuit whose functionality can be defined by a program. Thus, our target is to provide the knowledge for building
a simple programmable circuit in which both, the physical structure of the circuit and the informational structure of the program contribute to the definition of a certain function.
inStream
❄
dataMemory
✛
✲
Programmable Logic Controller
toyMachine
✛
✲ programMemory
❄
outStream
Figure 5.1: Programmable Logic Controller designed with toyMachine.
1
From Witgenstein’s Philosophical Investigation (#109). His own very original approach looked for an alternative way to the two main streams of the 20th Century philosophy: one originated in Frege’s formal positivism,
and another in Husserl’s phenomenology. Wittgenstein can be considered as a forerunner of the architectural
approach, his vision being far beyond his contemporary fellows were able to understand.
5.1. TOYMACHINE: A SMALL & SIMPLE COMPUTING MACHINE
89
The use of such a programmable circuit is presented in Figure 5.1, where inputStream[15:0]
represents the stream of data which is received by the toyMachine processor, it is processed
according to the program stored in programMemory, while, f needed, dataMemory stores intermediate data or support data. The result is issued as the data stream outputStream[15:0].
The use of such a programmable circuit is presented in Figure 5.1, where inputStream[15:0]
represents the stream of data which is received by the toyMachine processor, it is processed
according to the program stored in programMemory, while, if needed, dataMemory stores intermediate data or support data. The result is issued as the data stream outputStream[15:0].
For the purpose of this chapter, the internal structure of toyMachine is presented in Figure
5.2. The internal state of toyMachine is stored in:
programCounter : is a 32-bit register which stores the current address in the program memory;
it points in the program memory to the currently executed instruction; the reset signal
sets its value to zero; during the execution of each instruction its content is modified in
order to read the next instruction
intEnable : is a 1-bit state register which enable the action of the input int; the reset signal
sets it to 0, thus disabling the interrupt signal
regFile : the register file is a collection of 32 32-bit registers organized as a three port small
memory (array of storage elements):
• one port for write to the address destAddr
• one port for read the left operand from the address leftAddr
• one ort for read the right operand from the address rightAddr
used to store the most frequently used variables involved in each stage of the computation
carry : is a 1-bit register to store the value of the carry signal when an arithmetic operation is
performed; the value can be used for one of the next arithmetic operation
The external connections are:
inStream : the input stream of data
readyIn : the input data on inStream is valid
readIn : acknowledge for the sender on input data that the current value is received
outStream : the output stream of data
readyOut : the receiver of the data of the outStream is ready to receive
writeOut
: send the date to the receiver of outStream
int : interrupt signal is an “intrusive” signal used to trigger a “special event”; the signal int
acts, only if the interrupt is enabled (intEnable = 1), as follows:
begin
regFile[30]
programCounter
<= programCounter
<= regFile[31]
; // one-level stack
;
end
The location regFile[30] is loaded with the current value of programCounter when the
interrupt is acknowledged, and the content of regFile[31] is loaded as the next program
counter, i.e., the register 31 contains the address of the routine started by the occurrence
of the interrupt when it is acknowledged. The content of regFile[30] will be used to
restore the state of the machine when the program started by the acknowledged signal int
ends.
90
CHAPTER 5. OUR FINAL TARGET
clock
❄
reset
readyIn
inStream
readIn
int
✻
❄
❄
inRegister
dataIn
regFile
✛
❘
✲
❄
❘
programCounter
✠
✛
✻
✛
Combinational logic
✲
instruction
✛
❄
❄
carry
dataAddr
dataOut
write
intEnable
✒
✛
✛
✛
progAddr
✲
❄
outRegister
toyMachine
❄
readyOut
outStream
❄
writeOut
❄
inta
Figure 5.2: The internal state of toyMachine.
inta : interrupt acknowledge
progAddr : the address for the programm memory
instruction : the instruction received from the program memory
dataAddr : the addres for data memory
dataOut : the data sent to the data memory
dataIn : the data received from the data memroy
write : the write singnal for the data memory
reset : the synchronous reset signal is activated to initialize the system
clock : the clock signal
The stream of data is buffered in two registers:
inRegister : is a16-bit input register used as buffer
outRegister : is a 16-bit output register used as buffer
For the interconnections between the buffer registers, internal registers and the external
signals area is responsible the unspecified bloc Combinatorial logic.
The transition function is given by the program stored in an external memory called Program Memory (reg[31:0] programMemory[0:1023], for example). The program “decides”
(1) when a new value of the inStream is received, (2) when and how a new state of the machine
is computed and (3) when the output outStram is actualized. Thus, the output outStream
evolves according to the inputs of the machine and according to the history stored in its internal
state.
5.1. TOYMACHINE: A SMALL & SIMPLE COMPUTING MACHINE
91
The internal state of the above described engine is processed using combinational circuits,
whose functionality will be specified in this section using a Verilog behavioral description. At
the end of the next part of this book we will be able to synthesize the overall system using a
Verilog structural description.
The toyMachine’s instruction set architecture (ISA) is a very small subset of any 32-bit
processor (for example, the MicroBlaze processor [MicroBlaze]).
Each location in the program memory contains one 32-bit instruction organized in two formats, as follows:
instruction =
where:
{opCode[5:0], destAddr[4:0], leftAddr[4:0], rightAddr[4:0], 11’b0}
{opCode[5:0], destAddr[4:0], leftAddr[4:0], immValue[15:0]
opCode[5:0]
:
destAddr[4:0] :
leftAddr[4:0] :
rightAddr[4:0]:
immValue[15:0]:
|
;
operation code
selects the destination in the register file
selects the left operand from the register file
selects the right operand from the register file
immediate value
The actual content of the first field – opCode[5:0] – determines how the rest of the instruction is interpreted, i.e., what kind of instruction format has the current instruction. The first
format applies the operation coded by opCode to the values selected by leftAddr and rightAddr
from the register file; the result is stored in register file to the location selected by destAddr.
The second format uses immValue extended with sign as a 32-bit value to be stored in register
file at destAddr or as a relative address for jump instructions.
parameter
nop
= 6’b000000, //
//
jmp
= 6’b000001, //
zjmp
= 6’b000010, //
nzjmp
= 6’b000011, //
ei
= 6’b000110, //
di
= 6’b000111, //
halt
= 6’b001000, //
//
//
neg
= 6’b010000, //
bwand
= 6’b010001, //
bwor
= 6’b010010, //
bwxor
= 6’b010011, //
add
= 6’b010100, //
sub
= 6’b110101, //
addc
= 6’b110100, //
subc
= 6’b010101, //
ashr
= 6’b010110, //
val
= 6’b010111, //
hval
= 6’b011111, //
//
receive = 6’b100000, //
issue
= 6’b100001, //
get
= 6’b111000, //
send
= 6’b111001, //
datard = 6’b111010, //
datawr = 6’b111011; //
no operation: increment programCounter only
CONTROL INSTRUCTIONS
programCounter take the value form a register
jump if the selected regiter is 0
jump if the selected regiter is not 0
enable interrupt
disable interrupt
programCounter does not change
DATA INSTRUCTIONS: pc = pc + 1
Arithmetic & logic instructions
bitwise not
bitwise and
bitwise or
bitwise exclusive or
add
subtract
add with carry
subtract with carry
arithmetic shift right with one position
load immediate value with sign extension
append immediate value on high positions
Input output instructions
load inRegister if readyIn = 1
send the content of outRegister if readyOut = 1
load in file register the inRegister
load in outRegister the content of register
read from data memory
write to data memory
Figure 5.3: toyMachine’s ISA defined by the file 0 toyMachineArchitecture.v.
92
CHAPTER 5. OUR FINAL TARGET
The Instruction Set Architecture (ISA) of toyMachine is described in Figure 5.3, where
are listed two subset of instructions:
• control instructions: used to control the program flow by different kinds of jumps performed
conditioned, unconditioned or triggered by the acknowledged interrupt interrupt
• data instructions: used to modify the content of the file register, or to exchange data
with the external systems (each execution is accompanied with programCounter <=
programCounter + 1).
The file is used to specify opCode, the binary codes associated to each instruction.
The detailed description of each instruction is given by the Verilog behavioral descriptions
included in the module toyMachine (see Figure 5.4).
The first ‘include includes the binary codes defined in 0 toyMachineArchitecture.v (see
Figure 5.3) for each instruction executed by our simple machine.
module toyMachine(
input [15:0] inStream
, //
input
readyIn
, //
output
readIn
, //
output [15:0] outStream , //
input
readyOut
, //
output
writeOut
, //
input
int
, //
output
inta
, //
output [31:0] dataAddr
, //
output [31:0] dataOut
, //
output
write
, //
input [31:0] dataIn
, //
output [31:0] progAddr
, //
input [31:0] instruction, //
input
reset
, //
input
clock
);//
// INTERNAL STATE
reg [15:0] inRegister ;
reg [15:0] outRegister ;
reg [31:0] programCounter ;
reg [31:0] regFile[0:31] ;
reg
carry ;
reg
intEnable ;
‘include
‘include
‘include
‘include
endmodule
input stream of data
input stream is ready
read one element from the input stream
output stream of data
ready to receive an element of the output stream
write the element form outRegister
interrupt input
interrupt acknowledge
address for data memory
data of be stored in data memory
write in data memory
data from data memory
address for program memory
instruction form the program memory
reset input
clock input // 2429 LUTs
"0_toyMachineArchitecture.v"
"instructionStructure.v"
"controlFunction.v"
"dataFunction.v"
Figure 5.4: The file toyMachine.v containing the toyMachine’s behavioral description.
The second ‘include includes the file used to describe how the instruction fields are structured.
The last two ‘include lines include the behavioral description for the two subset of instructions performed by our simple machine. These last two files reflect the ignorance of the reader
in the domain of digital circuits. They are designed to express only what the designer intent to
build, but she/he doesn’t know yet how to do what must be done. The good news: the resulting description can be synthesized. The bad news: the resulting structure is very big (far from
5.1. TOYMACHINE: A SMALL & SIMPLE COMPUTING MACHINE
wire
wire
wire
wire
wire
assign
assign
assign
assign
assign
[5:0]
[4:0]
[4:0]
[4:0]
[31:0]
opCode
destAddr
leftAddr
rightAddr
immValue
opCode
destAddr
leftAddr
rightAddr
immValue
=
=
=
=
=
93
;
;
;
;
;
instruction[31:26]
;
instruction[25:21]
;
instruction[20:16]
;
instruction[15:11]
;
{{16{instruction[15]}}, instruction[15:0]};
Figure 5.5: The file instructionStructure.v.
optimal) and has a very complex form, i.e., no pattern can be emphasized. In order to provide
a small & simple circuit, in the next part of this book we will learn how to segregate the simple
part from the complex part of the circuits used to provide an optimal actual structure. Then,
we will learn how to optimize both, the simple, pattern-dominated circuits and the complex,
pattern-less ones.
The file instructionStructure.v (see Figure 5.5) defines the fields of the instruction. For
the two forms of the instruction appropriate fields are provided, i.e., the instruction content is
divided in many forms, thus allowing different interpretation of it. The bits instruction[15:0]
are used in two ways according to the opCode. If the instruction uses two operands, and both
are supplied by the content of the register file, then instruction[15:0] = rightAddr, else the
same bits are the most significant 5 bits of the 16-bit immediate value provided to be used as
signed operand or as a relative jump address.
The file controlFunction.v (see Figure 5.6) describes the behavior of the control instructions. The control of toyMachine refers to both, interrupt mechanism and the program flow
mechanism.
The interrupt signal int is acknowledged, activating the signal inta only if intEnable = 1
(see the assign on the first line in Figure 5.6). Initially, the interrupt is not allowed to act: reset
signal forces intEnable = 0. The program decides when the system is “prepared” to accept
interrupts. Then, the execution of the instruction ei (enable interrupt) determines intEnable
= 1. When an interrupt is acknowledged, the interrupt is disabled, letting the program decide
when another interrupt is welcomed. The interrupt is disabled by executing the instruction di
– disable interrupt.
The program flow is controlled by unconditioned and conditioned jump instructions. But,
the inta signal once activated, has priority, allowing the load of the program counter with the
value stored in regFile[31] which was loaded, by the initialization program of the system, with
the address of the subroutine associated to the interrupt signal.
The value of the program counter, programCounter, is by default incremented with 1, but
when a control instruction is executed its value can be incremented with the signed integer
instruction[15:0] or set to the value of a register contained in the register file. The program
control instructions are:
jmp : absolute jump with the value selected from the register file by the field leftAddr; the
register programCounter takes the value contained in the selected register
zjmp : relative jump with the signed value immValue if the content of the register selected by
leftAddr from the register file is 0, else programCounter = programCounter + 1
nzjmp : relative jump with the signed value immValue if the content of the register selected by
leftAddr from the register file is not 0, else programCounter = programCounter + 1
94
CHAPTER 5. OUR FINAL TARGET
assign inta = intEnable & int;
always @(posedge clock) if (reset)
intEnable <= 0
else if (inta)
intEnable <= 0
else if (opCode == ei)
intEnable <= 1
else if (opCode == di) intEnable <= 0
always @(posedge clock)
if (reset)
programCounter <= 0
else if (inta)
programCounter <= regFile[31]
else case(opCode)
jmp
:
programCounter <= regFile[leftAddr]
zjmp
: if (regFile[leftAddr] == 0)
programCounter <= programCounter + immValue
else programCounter <= programCounter + 1
nzjmp
: if (regFile[leftAddr] !== 0)
programCounter <= programCounter + immValue
else programCounter <= programCounter + 1
receive : if (readyIn)
programCounter <= programCounter + 1
else programCounter <= programCounter
issue
: if (readyOut)
programCounter <= programCounter + 1
else programCounter <= programCounter
halt :
programCounter <= programCounter
default
programCounter <= programCounter + 1
endcase
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
assign progAddr = programCounter;
Figure 5.6: The file controlFunction.v.
receive : relative jump with the signed value immValue if readyIn is 1, else programCounter
= programCounter + 1
issue : relative jump with the signed value immValue if readyOut is 1, else programCounter =
programCounter + 1
halt : the program execution halts, programCounter = programCounter (it is a sort of nop
instruction without incrementing the register programCounter = programCounter).
Warning! If intEnable = 0 when the instruction halt is executed, then the overall system
is blocked. The only way to turn it back to life is to activate the reset signal.
The file dataFunction.v (see Figure 5.7) describes the behavior of the data instructions.
The signal inta has the highest priority. It forces the register 30 of the register file to store
the current state of the register programCounter. It will be used to continue the program,
interrupted by the acknowledged interrupt signal int, by executing a jmp instruction with the
content of regFile[30].
The following data instructions are described in this file:
add : the content of the registers selected by leftAddr and rightAddr are added and the
result is stored in the register selected by destAddr; the value of the resulted carry is
stored in the carry one-bit register
sub : the content of the register selected by rightAddr is subtracted form the content of the
register selected by leftAddr, the result is stored in the register selected by destAddr;
the value of the resulted borrow is stored in the carry one-bit register
5.1. TOYMACHINE: A SMALL & SIMPLE COMPUTING MACHINE
95
addc : add with carry - the content of the registers selected by leftAddr and rightAddr and
the content of the register carry are added and the result is stored in the register selected
by destAddr; the value of the resulted carry is stored in the carry one-bit register
subc : subtract with carry - the content of the register selected by rightAddr and the content
of carry are subtracted form the content of the register selected by leftAddr, the result
is stored in the register selected by destAddr; the value of the resulted borrow is stored
in the carry one-bit register
always @(posedge clock)
if (inta)
regFile[30]
<= programCounter
;
else
case(opCode)
add
: {carry, regFile[destAddr]}
<= regFile[leftAddr] + regFile[rightAddr]
;
sub
: {carry, regFile[destAddr]}
<= regFile[leftAddr] - regFile[rightAddr]
;
addc
: {carry, regFile[destAddr]}
<= regFile[leftAddr] + regFile[rightAddr] + carry
;
subc
: {carry, regFile[destAddr]}
<= regFile[leftAddr] - regFile[rightAddr] - carry
;
ashr
: regFile[destAddr]
<= {regFile[leftAddr][31], regFile[leftAddr][31:1]}
;
neg
: regFile[destAddr] <= ~regFile[leftAddr]
;
bwand
: regFile[destAddr] <= regFile[leftAddr] & regFile[rightAddr]
;
bwor
: regFile[destAddr] <= regFile[leftAddr] | regFile[rightAddr]
;
bwxor
: regFile[destAddr] <= regFile[leftAddr] ^ regFile[rightAddr]
;
val
: regFile[destAddr] <= immValue
;
hval
: regFile[destAddr] <= {immValue[15:0], regFile[leftAddr][15:0]};
get
: regFile[destAddr] <= {{16{inRegister[15]}}, inRegister}
;
send
: outRegister
<= regFile[leftAddr][15:0]
;
receive : if (readyIn)
inRegister
<= inStream
;
datard : regFile[destAddr] <= dataIn
;
default
regFile[0]
<= regFile[0]
;
endcase
assign
assign
assign
assign
assign
assign
readIn
writeOut
write
dataAddr
dataOut
outStream
=
=
=
=
=
=
(opCode == receive) & readyIn
(opCode == issue) & readyOut
(opCode == datawr)
regFile[leftAddr]
regFile[rightAddr]
outRegister
;
;
;
;
;
;
Figure 5.7: The file dataFunction.v.
ashr : the content of the register selected by leftAddr is arithmetically shifted right one
position and stored in the register selected by destAddr
neg : every bit contained in the register selected by leftAddr are inverted and the result is
stored in the register selected by destAddr
bwand : the content of the register selected by leftAddr is AND-ed bit-by-bit with the
content of the register selected by rightAddr and the result is stored in the register
selected by destAddr
96
CHAPTER 5. OUR FINAL TARGET
bwor : the content of the register selected by leftAddr is OR-ed bit-by-bit with the content
of the register selected by rightAddr and the result is stored in the register selected by
destAddr
bwxor : the content of the register selected by leftAddr is XOR-ed bit-by-bit with the content
of the register selected by rightAddr and the result is stored in the register selected by
destAddr
val : the register selected by destAddr is loaded with the signed integer immValue
hval : is used to construct a 32-bit value placing instruction[15:0] on the 16 highest binary
position in the content of the register selected by leftAddr; the result is stored at destAddr
in the register file
get : the register selected by destAddr are loaded with the content of inRegister
send : the outRegister register is loaded with the least 15 significant bits of the register
selected by leftAddr
receive : if readyIn = 1, then the inRegister is loaded with the current varue applied on
the input inStream and the readIn signal is activated for the sender to “know” that the
current value was received
datard : the data accessed at the address dataAddr = leftOp = regFile[leftAddr] is
loaded in register file at th elocatioin destAddr
issue : generate, only when readyOut = 1, the signal writeOut used by the receiver to take
the value from the outRegister register
datawr : generate the signal write used by the data memory to write at the address
regFile[leftAddr] the data stored in regFile[rightAddr]
5.2
How toyMachine works
The simplest, but not the easiest way to use toyMachine is to program it in machine language2 , i.e., to write programs as sequence of binary coded instructions stored in programMemory
starting from the address 0.
The general way to solve digital problems using toyMachine, or a similar device, is (1)
to define the input stream, (2) the output stream, and (3) the program which transforms the
input stream into the corresponding output stream. Usually, we suppose an input signal which
is sampled at a program controlled rate, and the results is an output stream of samples which
is interpreted as the output signal. The transfer function of the system is programmed in the
binary sequence of instructions stored in the program memory.
The above described method to implement a digital system is called programmed logic,
because a general purpose programmable machine is used to implement a certain function which
generate an output stream of data starting from an input stream of data. The main advantage
of this method is its flexibility, while the main disadvantages are the reduced speed and the
increased size of the circuit. If the complexity, price and time to market issues are important,
then it can be the best solution.
2
The next levels are to use an assembly language or a high level language (for example: C), but these
approaches are beyond our goal.
5.2. HOW TOYMACHINE WORKS
97
Example 5.1 Let us revisit the pixel correction problem whose solution as circuit was presented
in Chapter 1. Now we consider a more elaborated environment (see Figure 5.8). The main
difference is that the transfers of the streams of data are now conditioned by specific dialog
signals. The subSystem generating pixels is interrogated, by readyIn, before its output is
loaded in inRegister; the takeing of its output is notified back by the signal readIn. Similarly,
there is a dialog with the subSystem using pixels. It is interrogated by readyOut and notified
by writeOut.
programMemory
instruction
reset
subSystem
generating
pixels
✲
✲
✛
✻
❄❄
readyIn
inStream
readIn
int
❄
progAddr
toyMachine
✲
✲
writeOut
outStream
readyOut
✛
subSystem
using
pixels
clock
Figure 5.8: Programmed logic implementation for the interpol circuit.
In this application the data memory is not needed and the int signal is not used. The
corresponding signals are omitted or connected to fix values in Figure 5.8.
The program (see Figure 5.9) is structured to use three sequences of instructions called
pseudo-macros3 . The first, called input, reads the input dealing with the dialog signals, readyIn
and readIn. The second, called output, controls the output stream dealing with the signals
readyOut and writeOut. The third pseudo-macro tests if the correction is needed, and apply it
if necessary.
The pseudo-macro input: The registers 0, 1, and 2 from regFile are used to store three
successive values of pixels from the input stream. Then, before receiving a new value, the content
of register 1 is moved in the register 2 and the content of register 0 is moved in register 1 (see
the first two line in code). The move instruction is emulated OR-ing the content of the source
register in the destination register (bwor 2, 1, 1). Now the register 0 form the register file is
ready to receive a new value. The next instruction loads in inRegister the value of the input
stream when it is valid. The instruction receive loops with the same value in programCounter
until the readyIn signal becomes 1. The input value, once buffered in the inRegister, could be
loaded in regFile[0].
The sequence of instructions just described performs the shift operation defined in the module
stateTransition which is instantiated in the module pixelCorrector used as example in our
introductory first chapter.
The pseudo-macro output: The value to be sent out as the next pixel is stored in the register
1. The outRegister register must be loaded with regFile[1] (send 1) and the signal writeOut
must be activated when readyOut is valid (issue. The instruction issue loops with the same
value in programCounter until the readyOut signal becomes 1.
3
The true macros are used in assembly languages. For this level of the machine language a more rudimentary
form of macro is used.
98
CHAPTER 5. OUR FINAL TARGET
‘insert input
‘insert input
val 3, 0
;
loop
‘insert input
‘insert compute
‘insert output
jmp loop
;
// PSEUDO-MACRO input
‘define
input
bwor 2, 1, 1;
bwor 1, 0, 0;
receive
;
get 0
;
// PSEUDO-MACRO output
‘define
output
send 1
;
issue
;
// PSEUDO-MACRO compute
‘define
compute
nzjmp 1, vld;
add 1, 0, 2 ;
ashr 1, 1
;
vld
nop
;
//
//
//
//
//
//
//
receive a dummy pixel
receive the first pixel
for unconditioned relative jump
receive the current pixel
make the correction if necessary
emit the current pixel
unconditioned jump to ’loop’
//
//
//
//
move
move
load
load
regFile[1]
regFile[0]
inRegister
regFile[0]
in regFile[2]
in regFile[1]
when readyIn = 1
with inRegister
// load outRegister with regFile[1]
// wait for readyOut and writeOut
//
//
//
//
jump to the line labeled "vld" if regFile[1] !== 0
and the next and previous pixel
divide by two the summ
for the purpose of modularity
Figure 5.9: Machine language program for interpol.
The pseudo-macro compute: This pseudo-macro first perform the test on the content of the
register 1 (nzjmp 1, vld), then makes the correction if necessary adding and dividing by two
(by an arithmetic shift right). The last nop instruction is imposed by the modularity of our pour
macro-like mechanism.
00
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
loop
vld
010010_00010_00001_00001_00000000000
010000_00000_00000_00000_00000000000
010010_00010_00001_00001_00000000000
010010_00001_00000_00000_00000000000
010000_00000_00000_00000_00000000000
010111_00011_00000_00000_00000000000
010010_00010_00001_00001_00000000000
010010_00001_00000_00000_00000000000
010000_00000_00000_00000_00000000000
000011_00000_00001_0000000000000011
010100_00001_00000_00010_00000000000
010110_00001_00001_00000_00000000000
000000_00000_00000_00000_00000000000
111001_00000_00001_00000_00000000000
100001_00000_00000_00000_00000000000
000001_00000_00011_1111111111110111
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
bwor 2, 1, 1
receive
bwor 2, 1, 1
bwor 1, 0, 0
receive
val 3, 0
bwor 2, 1, 1; loop
bwor 1, 0, 0
receive
nzjmp 1, vld
add 1, 0, 2
ashr 1, 1
nop
send 1
issue
zjmp 3, loop
Figure 5.10: The binary form of the program for interpol.
The actual program stored in the internal program memory (see // THE PROGRAM in Figure
5.9) has a starting part receiving two input values, followed by an unending (because gpi[11] =
0) loop which receives a new value, compute the value to be sent out and sends it.
⋄
5.3. CONCLUDING ABOUT TOYMACHINE
5.3
99
Concluding about toyMachine
Our final target is to be able to describe the actual structure of toyMachine using as much
as possible simple circuits. Maybe we just catched a glimpse about the circuits we must learn how
to design. It is almost obvious that the following circuits are useful for building toyMachine:
adders, subtractors, increment circuits, selection circuits, various logic circuits, registers, fileregisters, memories, read-only memories (for fix program memory). The next chapters present
detailed descriptions of all above circuits, and a little more.
The behavioral description of toyMachine is synthesisable, but the resulting structure
is too big and has a completely unstructured shape. The size increases the price, and the lack
of structure make impossible any optimization of area, of speed or of the energy consumption.
The main advantages of the just presented behavioral description is its simplicity, and the
possibility to use it as a more credible “witness” when the structural description will be verified.
Pros & cons for programmed logic
:
• it is a very flexible tool, but can not provide hi-performance solutions
• very good time to market, but not for mass production
• good for simple one chip solution, but not to be integrated as an IP on a complex SoC
solution
5.4
Problems
Problem 5.1 Write for toyMachine the program which follows-up as fast as possible by the
value on outStream the number of 1s on the inputs inStream.
Problem 5.2 Redesign the interpol program for a more accurate interpolation rule:
pi = 0.2 × pi−2 + 0.3 × pi−1 + 0.3 × pi+1 + 0.2 × pi+2
Problem 5.3
5.5
Projects
Project 5.1 Design the test environment for toyMachine, and use it to test the example from
this chapter.
100
CHAPTER 5. OUR FINAL TARGET
Part II
LOOPING IN THE DIGITAL
DOMAIN
101
Chapter 6
GATES:
Zero order, no-loop digital systems
In the previous chapter
ended the first part of this book, where we learned to “talk” in the Verilog HDL about
how to build big systems composing circuits and smaller systems, how to accelerate the
computation in a lazy system, and how to increase the autonomy of a system closing
appropriate loops. Where introduced the following basic concepts:
• serial, parallel, and serial-parallel compositions used to increase the size of a digital
system, maintaining the functional capabilities at the same level
• data (synchronic) parallelism and time (diachronic) parallelism (the pipeline connection) as the basic mechanism to improve the speed of processing in digital systems
• included loops, whose effect of limiting the time parallelism is avoided by speculating
– the third form of parallelism, usually ignored in the development of the parallel
architectures
• classifying digital circuits in orders, the n-th order containing circuits with n levels
of embedded loops
The last chapter of the first part defines the architecture of the machine whose components
will be described in the second part of this book.
In this chapter
the zero order, no-loop circuits are presented with emphasis on:
• how to expand the size of a basic combinational circuit
• the distinction between simple and complex combinatorial circuits
• how to deal with the complexity of combinatorial circuits using “programmable”
devices
In the next chapter
the first order, memory circuits are introduced presenting
• how a simple loop allows the occurrence of the memory function
• the basic memory circuits: elementary lathes, clocked latches, master-slave flip-flops
• memories and registers as basic systems composed using the basic memory circuits
103
104
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Belief #5: That qualitative as well as quantitative aspects of information systems will be
accelerated by Moore’s Law. . . . In the minds
of some of my colleagues, all you have to do
is identify one layer in a cybernetic system
that’s capable of fast change and then wait for
Moore’s Law to work its magic.
Jaron Lanier1
The Moore’s Law applies to size not to complexity.
In this chapter we will forget for the moment about loops. Composition is the only mechanism
involved in building a combinational digital system. No-loop circuits generate the class of history
free digital systems whose outputs depend only by the current input variables, and are reassigned
“continuously” at each change of inputs. Anytime the output results as a specific “combination”
of inputs. No autonomy in combinational circuits, whose outputs obey “not to say a word” to
inputs.
The combinational functions with n 1-bit inputs and m 1-bit outputs are called Boolean
function and they have the following form:
f : {0, 1}n → {0, 1}m .
For n = 1 only the NOT function is meaningful in the set of the 4 one-input Boolean functions.
For n = 2 from the set of 16 different functions only few functions are currently used: AND, OR,
XOR, NAND, NOR, NXOR. Starting with n = 3 the functions are defined only by composing
2-input functions. (For a short refresh see Appendix Boolean functions.)
Composing small gates results big systems. The growing process was governed in the last
40 years by Moore’s Law2 . For a few more decades maybe the same growing law will act.
But, starting from millions of gates per chip, it is very important what kind of circuits grow
exponentially!
Composing gates results two kinds of big circuits. Some of them are structured following
some repetitive patterns, thus providing simple circuits. Others grow patternless, providing
complex circuits.
6.1
Simple, Recursive Defined Circuits
The first circuits used by designers were small and simple. When they were grew a little they
were called big or complex. But, now when they are huge we must talk, more carefully, about
big sized simple circuits or about big sized complex circuits. In this section we will talk about
simple circuits which can be actualized at any size, i.e., their definitions don’t depend by the
number, n, of their inputs.
n
In the class of n-inputs circuits there are 22 distinct circuits. From this tremendous huge
number of logical function we use currently an insignificant small number of simple functions.
What is strange is that these functions are sufficient for almost all the problem which we are
confronted (or we are limited to be confronted).
1
2
Jaron Lanier coined the term virtual reality. He is a computer scientist and a musician.
The Moore’s Law says the physical performances in microelectronics improve exponentially in time.
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
105
One fact is clear: we can not design very big complex circuits because we can not specify
them. The complexity must get away in another place (we will see that this place is the world
of symbols). If we need big circuit they must remain simple.
In this section we deal with simple, if needed big, circuits and in the next with the complex
circuits, but only with ones having small size.
From the class of the simple circuits we will present only some very usual such as decoders,
demultiplexors, multiplexors, adders and arithmetic-logic units. There are many other interesting
and useful functions. Many of them are proposed as problems at the end of this chapter.
6.1.1
Decoders
The simplest problem to be solved with a combinational logic circuit (CLC) is to answer the
question: “what is the value applied to the input of this one-input circuit?”. The circuit which
solves this problem is an elementary decoder (EDCD). It is a decoder because decodes its onebit input value by activating distinct outputs for the two possible input values. It is elementary
because does this for the smallest input word: the one-bit word. By decoding, the value applied
to the input of the circuit is emphasized activating distinct signals (like lighting only one of n
bulbs). This is one of the main functions in a digital system. Before generating an answer to
the applied signal, the circuit must “know” what signal arrived on its inputs.
Informal definition
The n-input decoder circuit – DCDn – (see Figure 6.1) performs one of the basic function in
digital systems: with one of its m one-bit outputs specifies the binary configuration applied
on its inputs. The binary number applied on the inputs of DCDn takes values in the set
X = {0, 1, ...2n − 1}. For each of these values there is one output – y0 , y1 , ...ym−1 – which is
activated on 1 if its index corresponds with the current input value. If, for example, the input
of a DCD4 takes value 1010, then y10 = 1 and the rest 15 one-bit outputs take the value 0.
✲ x0
✲ x1
...
✲ xn−1
DCDn
y0 y1
❄ ❄
...
ym−1
❄
Figure 6.1: The n-input decoder (DCDn ).
Formal definition
In order to rigorously describe and to synthesize a decoder circuit a formal definition is requested.
Using Verilog HDL, such a definition is very compact certifying the non-complexity of this circuit.
Definition 6.1 DCDn is a combinational circuit with the n-bit input X, xn−1 , . . . , x0 , and the
m-bit output Y , ym−1 , . . . , y0 , where: m = 2n , with the behavioral Verilog description:
module dec #(parameter inDim = n)(input [inDim - 1:0]
sel,
output [(1 << inDim) - 1:0] out);
assign out = 1 << sel;
endmodule
⋄
106
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
The previous Verilog description is synthesisable by the current software tools which provide
an efficient solution. It happens because this function is simple and it is frequently used in
designing digital systems.
Recursive definition
The decoder circuit DCDn for any n can be defined recursively in two steps:
• defining the elementary decoder circuit (EDCD = DCD1 ) as the smallest circuit performing the decode function
• applying the divide & impera rule in order to provide the DCDn circuit using DCDn/2
circuits.
For the first step EDCD is defined as one of the simplest and smallest logical circuits.
Two one-input logical function are used to perform the decoding. Indeed, parallel composing
(see Figure 6.2a) the circuits performing the simplest functions: f21 (x0 ) = y1 = x0 (identity
function) and f11 (x0 ) = y0 = x′0 (NOT function), we obtain an (EDCD). If the output y0 is
active, it means the input is zero. If the output y1 is active, then the input has the value 1.
x0
x0
y0
y1
y1
y0
EDCD
a.
b.
Figure 6.2: The elementary decoder (EDCD). a. The basic circuit. b. Buffered EDCD, a
serial-parallel composition.
In order to isolate the output from the input the buffered EDCD version is considered serial
composing an additional inverter with the previous circuit (see Figure 6.2b). Hence, the fan-out
of EDCD does not depend on the fan-out of the circuit that drives the input.
y0
y0
y1
yp−1
y1
DCDn/2
yp
yp−1
✻
ym−1
n/2
n
n/2
✲
y0
yp−1
y1
DCDn/2
xn−1 . . . x0
Figure 6.3: The recursive definition of n-inputs decoder (DCDn ). Two DCDn/2 are used
to drive a two dimension array of AN D2 gates. The same rule is applied for the two DCDn/2 , and so on
until DCD1 = EDCD is needed.
The second step is to answer the question about how can be build a (DCDn ) for decoding
an n-bit input word.
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
107
Definition 6.2 The structure of DCDn is recursive defined by the rule represented in Figure
6.3. The DCD1 is an EDCD (see Figure 6.2b). ⋄
The previous definition is a constructive one, because provide an algorithm to construct a
decoder for any n. It falls into the class of the “divide & impera” algorithms which reduce the
solution of the problem for n to the solution of the same problem for n/2.
The quantitative evaluation of DCDn offers the following results:
Size: GSDCD (n) = 2n GSAN D (2) + 2GSDCD (n/2) = 2(2n + GSDCD (n/2))
GSDCD (1) = GSEDCD = 2
GSDCD (n) ∈ O(2n )
Depth: DDCD (n) = DAN D (2) + DDCD (n/2) = 1 + DDCD (n/2) ∈ O(log n)
DDCD (1) = DEDCD = 2
Complexity: CDCD ∈ O(1) because the definition occupies a constant drown area (Figure 6.3)
or a constant number of symbols in the Verilog description for any n.
The size, the complexity and the depth of this version of decoder is out of discussion because
the order of the size can not be reduced under the number of outputs (m = 2n ), for complexity
O(1) is the minimal order of magnitude, and for depth O(log n) is optimal takeing into account
we applied the “divide & impera” rule to build the structure of the decoder.
Structural description
The current HDLs, such as VHDL and Verilog, do not support recursive descriptions. Therefore,
for structural descriptions, when the behavioral ones do not provide an appropriate solution,
the recursive definition must be translated into an iterative one.
x0
x1
xn−1
...
y0
...
y1
...
ym−1
Figure 6.4: “Constant depth” DCD Applying the associative rule into the hierarchical network of
AN D2 gates results the one level AN Dn gates circuit driven by n EDCDs.
An iterative structural version of the previous recursive constructive definition is possible,
because the outputs of the two DCDn/2 from Figure 6.3 are also 2-input AND circuits, the same
as the circuits on the output level. In this case we can apply the associative rule, implementing
the last two levels by only one level of 4-input ANDs. And so on, until the output level of the
2n n-input ANDs is driven by n EDCDs. Now we have the decoder represented in Figure 6.4).
Apparently it is a constant depth circuit, but if we take into account that the number of inputs
in the AND gates is not constant, then the depth is given by the depth of an n-input gate which
is in O(log n). Indeed, an n-input AND has an efficient implementation as as a binary tree of
2-input ANDs.
This “constant depth” DCD version – CDDCD – is faster than the previous for small values
of n (usually for n < 6; for more details see Appendix Basic circuits), but the size becomes
108
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
SCDDCD (n) = n × 2n + 2n ∈ O(n2n ). The price is over-dimensioned related to the gain, but for
small circuits sometimes it can be accepted.
The pure structural description for DCD3 is:
module dec3(output [7:0] out,
input [2:0] in );
// internal connections
wire in0, nin0, in1, nin1, in2, nin2;
// EDCD for in[0]
not not00(nin0, in[0]);
not not01(in0, nin0) ;
// EDCD for in[1]
not not10(nin1, in[1]);
not not11(in1, nin1) ;
// EDCD for in[2]
not not20(nin2, in[2]);
not not21(in2, nin2) ;
// the second level
and and0(out[0], nin2, nin1, nin0); //
and and1(out[1], nin2, nin1, in0 ); //
and and2(out[2], nin2, in1, nin0); //
and and3(out[3], nin2, in1, in0 ); //
and and4(out[4], in2, nin1, nin0); //
and and5(out[5], in2, nin1, in0 ); //
and and6(out[6], in2, in1, nin0); //
and and7(out[7], in2, in1, in0 ); //
endmodule
output
output
output
output
output
output
output
output
0
1
2
3
4
5
6
7
For n = 3 the size of this iterative version is identical with the size which results from the recursive
definition. There are meaningful differences only for big n. In real designs we do not need this kind of
pure structural descriptions because the current synthesis tools manage very well even pure behavioral
descriptions such that from the formal definition of the decoder.
Arithmetic interpretation
The decoder circuit is also an arithmetic circuit. It computes the numerical function of exponentiation:
Y = 2X . Indeed, for n = i only the output yi takes the value 1 and the rest of the outputs take the value
0. Then, the number represented by the binary configuration Y is 2i .
Application
Because the expressions describing the m outputs of DCDn are:
y0 = x′n−1 · x′n−2 · . . . x′1 · x′0
y1 = x′n−1 · x′n−2 · . . . x′1 · x0
y2 = x′n−1 · x′n−2 · . . . x1 · x′0
...
ym−2 = xn−1 · xn−2 · . . . x1 · x′0
ym−1 = xn−1 · xn−2 · . . . x1 · x0
the logic interpretation of these outputs is that they represent all the min-terms for an n-input function.
Therefore, any n-input logic function can be implemented using a DCDn and an OR with maximum
m − 1 inputs.
Example 6.1 Let be the 3-input 2-output function defined in the table from Figure 6.5. A DCD3 is used
to compute all the min-terms of the 3 variables a, b, and c. A 3-input OR is used to “add” the min-terms
for the function X, and a 4-input OR is used to “add” the min-terms for the function Y.
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
a
0
0
0
0
1
1
1
1
b
0
0
1
1
0
0
1
1
c
0
1
0
1
0
1
0
1
X
0
1
0
0
0
1
1
0
✲ x0
✲ x1
✲ x2
c
b
a
Y
1
1
0
1
0
0
1
0
109
DCD3
y0 y1 y2 y3 y4 y5 y6 y7
X
Y
Figure 6.5:
Each min-term is computed only once, but it can be used as many times as the implemented functions
suppose.
⋄
6.1.2
Demultiplexors
The structure of the decoder is included in the structure of the other usual circuits. Two of them are the
demultiplexor circuit and the multiplexer circuit. These complementary functions are very important in
digital systems because of their ability to perform “communication” functions. Indeed, demultiplexing
means to spread a signal from a source to many destinations, selected by a binary code and multiplexing
means the reverse operation to catch signals from distinct sources also selected using a selection code.
Inside of both circuits there is a decoder used to identify the source of the signal or the destination of
the signal by decoding the selection code.
Informal definition
The first informally described solution for implementing the function of an n-input demultiplexor is to
use a decoder with the same number of inputs and m 2-input AND connected as in Figure 6.6. The value
of the input enable is generated to the output of the gate opened by the activated output of the decoder
DCDn . It is obvious that a DCDn is a DM U Xn with enable = 1. Therefore, the size, depth of DMUXs
are the same as for DCDs, because the depth is incremented by 1 and to the size is added a value which
is in O(2n ).
✲
✲
x0
x1
DCDn
✲ xn−1
y0
y1
ym−1
enable
y0
y1
ym−1
Figure 6.6: Demultiplexor. The n-input demultiplexor (DM U Xn ) includes a DCDn and 2n AN D2
gates used to distribute the input enable in 2n different places according to the n-bit selection code.
For example, if on the selection input X = s, then the outputs yi take the value 0 for i ̸= s and
ys = enable. The inactive value on the outputs of this DMEX is 0.
110
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Formal definition
Definition 6.3 The n-input demultiplexor – DM U Xn – is a combinational circuit which transfers the
1-bit signal from the input enable to the one of the outputs ym−1 , . . . , y0 selected by the n-bit selection
code X = xn−1 , . . . , x0 , where m = 2n . It has the following behavioral Verilog description:
module dmux #(parameter inDim = n)(input [inDim - 1:0]
sel
,
input
enable,
output [(1 << inDim) - 1:0] out
);
assign out = enable << sel;
endmodule
⋄
Recursive definition
The DMUX circuit has also a recursive definition. The smallest DMUX, the elementary DMUX
– EDMUX –, is a 2-output one, with a one-bit selection input. EDMUX is represented in Figure
6.7. It consists of an EDCD used to select, with its two outputs, the way for the signal enable.
Thus, the EDMUX is a circuit that offers the possibility to transfer the same signal (enable) in
two places (y0 and y1 ), according with the selection input (x0 ) (see Figure 6.7.
❄
EDCD
x0
enable
✲
enable
x0 EDMUX
y0
❄
y1
❄
y1
y0
a.
b.
Figure 6.7: The elementary demultiplexor. a. The internal structure of an elementary demultiplexor (EDMUX) consists in an elementary decoder, 2 AN D2 gates, and an inverter circuit as input
buffer. b. The logic symbol.
xn−1
✲
x0
enable
EDMUX
y0
y1
✛enable
xn−2 , . . . , x0
n−1
✲
enable
DM U Xn−1
y m −1
y0
❄
y0
✛
2
...
❄
y m −1
2
✲
enable
DM U Xn−1
y m −1
y0
❄
ym
2
✛
2
...
❄
ym−1
Figure 6.8: The recursive definition of DM U Xn . Applying the same rule for the two DM U Xn−1
a new level of 2 EDMUXs is added, and the output level is implemented using 4 DM U Xn−2 .
And so on until the output level is implemented using 2n−1 EDMUXs. The resulting circuit
contains 2n − 1 EDMUXs.
The same rule – divide & impera – is used to define an n-input demultiplexor, as follows:
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
111
Definition 6.4 DM U Xn is defined as the structure represented in Figure 6.8, where the two
DM U Xn−1 are used to select the outputs of an EDMUX. ⋄
If the recursive rule is applied until the end the resulting circuit is a binary tree of EDMUXs.
It has SDM U X (N ) ∈ O(2n ) and DDM U X (n) ∈ O(n). If this depth is considered too big for the
current application, the recursive process can be stopped at a convenient level and that level is
implemented with a “constant depth” DMUXs made using “constant depth” DCDs. The mixed
procedures are always the best. The previous definition is a suggestion for how to use small
DMUXs to build bigger ones.
6.1.3
Multiplexors
Now about the inverse function of demultiplexing: the multiplexing, i.e., to take a bit of information from a selected place and to send in one place. Instead of spreading by demultiplexing,
now the multiplexing function gathers from many places in one place. Therefore, this function is
also a communication function, allowing the interconnecting between distinct places in a digital
system. In the same time, this circuit is very useful for implementing random, i.e. complex,
logical functions, as we will see at the end of this chapter. More, in the next chapter we will see
that the smallest multiplexor is used to build the basic memory circuits. Looks like this circuit
is one of the most important basic circuit, and we must pay a lot of attention to it.
Informal definition
The direct intuitive implementation of a multiplexor with n selection bits – M U Xn – starts also
from a DCDn which is now serially connected with an AND-OR structure (see Figure 6.9). The
outputs of the decoder open, for a given input code, only one AND gate that transfers to the
output the corresponding selected input which, by turn, is OR-ed to the output y.
ym−1
DCDn
y1
y0
i0
im−1
i1
...
✻
n
...
xn−1 , . . . , x0
...
y
Figure 6.9: Multiplexer. The n selection inputs multiplexer M U Xn is made serial connecting a
DCDn with an AND-OR structure.
Applying in this structure the associativity rule, for the AND gates to the output of the
decoder and the supplementary added ANDs, results the actual structure of MUX. The structure
AND-OR maintains the size and the depth of MUX in the same orders as for DCD.
Formal definition
As for the previous two circuits – DCD and DMUX –, we can define the multiplexer using a
behavioral (functional) description.
112
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Definition 6.5 A multiplexer M U Xn is a combinational circuit having n selection inputs
xn−1 , . . . , x0 that selects to the output y one input from the m = 2n selectable inputs, im−1 , . . . , i0 .
The Verilog description is:
module mux #(parameter inDim = n)(input [inDim-1:0]
sel, // selection inputs
input [(1<<inDim)-1:0] in , // selected inputs
output
out);
assign out = in[sel];
endmodule
⋄
The MUX is obviously a simple function. Its formal description, for any number of inputs has
a constant size. The previous behavioral description is synthesisable efficiently by the current
software tools.
Recursive definition
There is also a rule for composing large MUSs from the smaller ones. As usual, we start from
an elementary structure. The elementary MUX – EMUX – is a selector that connects the signal
i1 or i0 in y according to the value of the selection signal x0 . The circuit is presented in Figure
6.10a, where an EDCD with the input x0 opens only one of the two ANDs ”added” by the OR
circuit in y. Another version for EMUX uses tristate inverting drivers (see Figure 6.10c).
i0
c
i1
i0
x0
c’
✲ x0
c’
❄
❄
i0
i1
i1
EMUX
y
a.
❄
y
b.
y
c
x0
c’
c
c.
Figure 6.10: The elementary multiplexer (EMUX). a. The structure of EMUX containing an
EDCD and the smallest AND-OR structure. b. The logic symbol of EMUX. c. A version of EMUX
using transmission gates (see section Basic circuits).
The definition of M U Xn starts from EMUX, in a recursive manner. This definition will
show us that MUX is also a simple circuit (CM U X (n) ∈ O(1)). In the same time this recursive
definition will be a suggestion for the rule that composes big MUXs from the smaller ones.
Definition 6.6 M U Xn can be made by serial connecting two parallel connected M U Xn/2 with
an EMUX (see Figure 6.11 that is part of the definition), and M U X1 = EM U X. ⋄
Structural aspects
This definition leads us to a circuit having the size in O(2n ) (very good, because we have m = 2n
inputs to be selected in y) and the depth in O(n). In order to reduce the depth we can apply
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
✲
2
2
❄
❄
...
i0
im
i m −1
...
i0
113
...
im−1
...
i m −1
❄
i m −1
2
✲
M U Xn−1
❄
i0
2
M U Xn−1
y
y
xn−2 , . . . , x0
✲
xn−1
❄ ❄
i
x0 0
i1
EMUX
y
❄y
Figure 6.11: The recursive definition of M U Xn . Each M U Xn−1 has a similar definition (two
M U Xn−2 and one EMUX), until the entire structure contains EMUXs. The resulting circuit is a binary
tree of 2n − 1 EMUXs.
step by step the next procedure: for the first two levels in the tree of EMUXs we can write the
equation
y = x1 (x0 i3 + x′0 i2 ) + x′1 (x0 i1 + x′0 i0 )
that becomes
y = x1 x0 i3 + x1 x′0 i2 + x′1 x0 i1 + x′1 x′0 i0 .
Using this procedure two or more levels (but not too many) of gates can be reduced to one.
Carefully applied this procedure accelerate the speed of the circuit.
Application
Because the logic expression of a n selection inputs multiplexor is:
y = xn−1 . . . x1 x0 im−1 + . . . + x′n−1 . . . x′1 x0 i1 + x′n−1 . . . x′1 x′0 i0
any n-input logic function is specified by the binary vector {im−1 , . . . i1 , i0 }. Thus any n input
logic function can be implemented with a M U Xn having on its selected inputs the binary vector
defining it.
Example 6.2 Let be function X defined in Figure 6.12 by its truth table. The implementation
with a M U X3 means to use the right side of the table as the defining binary vector.
a
0
0
0
0
1
1
1
1
b
0
0
1
1
0
0
1
1
c
0
1
0
1
0
1
0
1
X
0
1
0
0
0
1
1
0
0 1 0 0 0 1 1 0
❄❄❄❄❄❄❄❄
c
b
a
✲ x0
✲ x1
✲ x2
i0 i1 i 2 i 3 i 4 i5 i6 i7
y
❄
X
Figure 6.12:
⋄
M U X3
114
CHAPTER 6. GATES:
6.1.4
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
∗ Shifters
One of the simplest arithmetic circuit is a circuit able to multiply or to divided with a number equal
with a power of 2. The circuit is called also shifter because these operations do not change the relations
between the bits of the number, they change only the position of the bits. The bits are shifted a number
of positions to the left, for multiplication, or to the right for division.
The circuit used for implementing a shifter for n-bit numbers with m − 1 positions is an m-input
multiplexor having the selected inputs defined on n bits.
In the previous subsection were defined multiplexors having 1-bit selected inputs. How can be expanded the number of bits of the selected inputs? An elementary multiplexor for p-bit words, pEMUX,
is made using p EMUXs connected in parallel. If the two words to be multiplexed are ap−1 , . . . a0 and
bp−1 , . . . b0 , then each EMUX is used to multiplex a pair of bits (ai , bi ). The one-bit selection signal is
shared by the p EMUXs. nEMUX is a parallel extension of EMUX.
ap−1
bp−1
ap−2
bp−2
a0
b0
x
yp−1
yp−2
y0
Figure 6.13: The structure of pEMUX. Because the selection bit is the same for all EMUXs one
EDCD is shared by all of them.
Using pEMUXs an pM U Xn can be designed using the same recursive procedure as for designing
M U Xn starting from EMUXs.
An 2n − 1 positions left shifter for p-bit numbers, pLSHIF Tn , is designed connecting the selected
inputs of an pEMUX, i0 , . . . im−1 where m = 2n , to the number to be shifted N = {an−1 , an−2 , . . . a0 }
(ai ∈ {0.1} for i = 0, 1, . . . (m − 1)) in 2n − 1 ways, according to the following rule:
ij = {{an−j−1 , an−j−2 , . . . a0 }, {j{0}}}
for: j = 0, 1, . . . (m − 1).
Example 6.3 For 4LSHIF T2 an 4M U X2 is used. The binary code specifying the shift dimension is
shiftDim[1:0], the number to be shifted is in[3:0], and the ways the selected inputs, in0, in1, in2,
in3, are connected to the number to be shifted are:
in0
in1
in2
in3
=
=
=
=
{in[3],
{in[2],
{in[1],
{in[0],
in[2],
in[1],
in[0],
0
,
in[1],
in[0],
0
,
0
,
in[0]}
0
}
0
}
0
}
Figure 6.14 represents the circuit.
The Verilog description of the shifter is done by instantiating a 4-way 4-bit multiplexor with its inputs
connected according to the previously described rule.
module leftShifter(output [3:0] out
,
input [3:0] in
,
input [1:0] shiftDim);
mux4_4 shiftMux(.out(out
),
.in0(in
),
.in1({in[2:0], 1’b0}),
.in2({in[1:0], 2’b0}),
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
115
.in3({in[0] , 3’b0}),
.sel(shiftDim
));
endmodule
in[0]
in[1]
in[2]
in[3]
❄ ❄ ❄ ❄
✲
shiftDim
in0
sel
in1 in2
mux4
in3
mux3
❄ ❄ ❄ ❄
✲
in0
sel
in1 in2
mux4
in3
mux2
❄ ❄ ❄ ❄
✲
in0
sel
in1 in2
mux4
in3
mux1
❄ ❄ ❄ ❄
✲
in0
sel
in1 in2
mux4
in3
mux0
❄
❄
❄
❄
out[3]
out[2]
out[1]
out[0]
Figure 6.14: The structure of 4LSHIF T2 , a maximum 3-position, 4-bit number left
shifter.
The multiplexor used in the previous module is built using 3 instantiations of an elementary 4-bit
multiplexors. Results the two level tree of elementary multiplexors interconnected as the following Verilog
code describes.
module mux4_4(output [3:0] out,
input [3:0] in0, in1, in2, in3,
input [1:0] sel); // 4-way 4-bit multiplexor (4MUX_2)
wire[3:0]
out1, out0;
// internal connections between the two levels
mux2_4 mux(out, out0, out1, sel[1]), // output multiplexor
mux1(out1, in2, in3, sel[0]), // input multiplexor for in3 and in2
mux0(out0, in0, in1, sel[0]); // input multiplexor for in1 and in0
endmodule
Any n-bit elementary multiplexer is described by the following parameterized module:
module mux2_4 #(parameter n = 4)(output [n-1:0] out,
input [n-1:0] in0, in1,
input
sel); // 2-way 4-bit mux: 4EMUX
assign out = sel ? in1 : in0; // if (sel) then in1, else in0
endmodule
⋄
The same idea helps us to design a special kind of shifter, called barrel shifter which performs a
rotate operation described by the following rule: if the input number is N = {an−1 , an−2 , . . . a0 }
(ai ∈ {0.1} for i = 0, 1, . . . (m − 1)), then rotating it with i positions will provide:
ii = {an−i−1 , an−i−2 , . . . a0 , an−1 , an−2 , . . . an−i }
for: i = 0, 1, . . . (m − 1). This first solution for the rotate circuit is very similar with the shift circuit.
The only difference is: all the inputs of the multiplexor are connected to an input value. No 0s on any
inputs of the multiplexor.
A second solution uses only elementary multiplexors. A version for 8-bit numbers is presented in the
following Verilog code.
module leftRotate(output [7:0] out
,
input [7:0] in
,
input [2:0] rotateDim);
116
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
wire [7:0] out0, out1;
mux2_8
level0(.out(out0), .in0(in ), .in1({in[6:0], in[7]} ),
.sel(rotateDim[0])),
level1(.out(out1), .in0(out0), .in1({out0[5:0], out0[7:6]}), .sel(rotateDim[1])),
level2(.out(out ), .in0(out1), .in1({out1[3:0], out1[7:4]}), .sel(rotateDim[2]));
endmodule
module mux2_8(output [7:0] out,
input [7:0] in0, in1,
input
sel);
assign out = sel ? in1 : in0;
endmodule
While the first solution uses for n bit numbers n M U Xlog2 (rotateDim) , the second solution uses
log2 (rotateDim) nEM U Xs. Results:
Sf irstSolutionOf Lef tRotate = (n × (rotateDim − 1)) × SEM U X
SsecondSolutionOf Lef tRotate = (n × log2 (rotateDim)) × SEM U X
6.1.5
∗ Priority encoder
An encoder is a circuit which connected to the outputs of a decoder provides the value applied on the
input of the decoder. As we know only one output of a decoder is active at a time. Therefore, the encoder
compute the index of the activated output. But, a real application of an encoder is to encode binary
configurations provided by any kind of circuits. In this case, more than one input can be active and the
encoder must have a well defined behavior. One of this behavior is to encode the most significant bit and
to ignore the rest of bits. For this reason the encoder is a priority encoder.
The n-bit input, enabled priority encoder circuit, P E(n), receives xn−1 , xn−2 , . . . x0 and, if the enable
input is activated, en = 1, it generates the number Y = ym−1 , ym−2 , . . . y0 , with n = 2m , where Y is
the biggest index associated with xi = 1 if any, else zero output is activated. (For example: if en = 1,
for n = 8, and x7 , x6 , . . . x0 = 00110001, then y2 , y1 , y0 = 101 and zero = 0) The following Verilog code
describe the behavior of P E(n).
module priority_encoder #(parameter m = 3)(input
[(1’b1<<m)-1:0] in
input
enable
output reg [m-1:0]
out
output reg
zero
integer i;
always @(*) if (enable) begin
out = 0;
for(i=(1’b1 << m)-1; i>=0; i=i-1)
if ((out == 0) && in[i]) out = i;
if (in == 0)
zero = 1;
else
zero = 0;
end
else
begin
out = 0;
zero = 1;
end
endmodule
For testing the previous description the following test module is used:
module test_priority_encoder #(parameter
reg
[(1’b1<<m)-1:0]
in
;
reg
enable ;
wire
[m-1:0]
out
;
wire
zero
;
m = 3);
,
,
,
);
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
initial
begin
#1
#1
#1
#1
#1
#1
#1
#1
#1
#1
#1
end
priority_encoder
initial $monitor
117
enable = 0;
in = 8’b11111111;
enable = 1;
in = 8’b00000001;
in = 8’b0000001x;
in = 8’b000001xx;
in = 8’b00001xxx;
in = 8’b0001xxxx;
in = 8’b001xxxxx;
in = 8’b01xxxxxx;
in = 8’b1xxxxxxx;
in = 8’b110;
$stop;
dut(in
,
enable ,
out
,
zero
);
($time, "enable=%b in=%b out=%b zero=%b",
enable, in, out, zero);
endmodule
Running the previous code the simulation provides the following result:
time
time
time
time
time
time
time
time
time
time
time
= 0
= 1
= 2
= 3
= 4
= 5
= 6
= 7
= 8
= 9
=10
enable
enable
enable
enable
enable
enable
enable
enable
enable
enable
enable
=
=
=
=
=
=
=
=
=
=
=
0
1
1
1
1
1
1
1
1
1
1
in
in
in
in
in
in
in
in
in
in
in
=
=
=
=
=
=
=
=
=
=
=
11111111
11111111
00000001
0000001x
000001xx
00001xxx
0001xxxx
001xxxxx
01xxxxxx
1xxxxxxx
00000110
out
out
out
out
out
out
out
out
out
out
out
=
=
=
=
=
=
=
=
=
=
=
000
111
000
001
010
011
100
101
110
111
010
zero
zero
zero
zero
zero
zero
zero
zero
zero
zero
zero
=
=
=
=
=
=
=
=
=
=
=
1
0
0
0
0
0
0
0
0
0
0
It is obvious that this circuit computes the integer part of the base 2 logarithm. The output zero is
used to notify that the input value is unappropriate for computing the logarithm, and “prevent” us from
takeing into account the output value.
6.1.6
∗ Prefix computation network
There is a class of circuits, called prefix computation networks, P CNf unc (n), defined for n inputs
and having the characteristic function f unc. If f unc is expressed using the operation ◦, then the
function of P CN◦ (n) is performed by a circuit having the inputs x0 , . . . xn−1 and the outputs
y0 , . . . yn−1 related as follows:
y0 = x 0
y 1 = x0 ◦ x1
y2 = x 0 ◦ x 1 ◦ x 2
...
yn−1 = x0 ◦ x1 ◦ . . . ◦ xn−1
where the operation “◦” is an associative and commutative operation. For example, ◦ can be
the arithmetic operation add, or the logic operation AND. In the first case xi is an m-bit binary
number, and in the second case it is a 1-bit Boolean variable.
118
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Example 6.4 If ◦ is the Boolean function AND, then P CNAN D (n) is described by the following
behavioral description:
module and_prefixes #(parameter n = 64)(input
[0:n-1] in ,
output reg [0:n-1] out);
integer k;
always @(in) begin out[0] = in[0];
for (k=1; k<n; k=k+1) out[k] = in[k] & out[k-1];
end
endmodule
⋄
There are many solutions for implementing P CNAN D (n). If we use AND gates with up to
n inputs, then there is a first direct solution for P CNAN D starting from the defining equations
(it consists in one 2-input gate, plus one 3-input gate, . . . plus one (n-1)-input gate). A very
large high-speed circuit is obtained. Indeed, this direct solution offers a circuit with the size
S(n) ∈ O(n2 ) and the depth D(n) ∈ O(1). We are very happy about the speed (depth), but the
price paid for this is too high: the squared size. In the same time our design experience tells us
that this speed is not useful in current applications because of the time correlations with other
subsystems. (There is also a discussion about gate having n inputs. These kind of gates are not
realistic.)
x0
x1
x2
x3
x4
xn−2 xn−1
x5
P CNAN D (n/2)
y0
y1
y2
y3
yn−3 yn−2 yn−1
y4
Figure 6.15: The internal structure of P CNAN D (n). It is recursively defined: if P CNAN D (n/2)
is a prefix computation network, then the entire structure is P CNn .
A second solution offers a very good size but a too slow circuit. If we use only 2-input AND
gates, then the definition becomes:
y0 = x0
y1 = y0 & x 1
y2 = y1 & x 2
...
yn−1 = yn−2 & xn−1
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
119
A direct solution starting from this new form of the equations (as a degenerated binary tree of
ANDs) has S(n) ∈ O(n) and D(n) ∈ O(n). This second solution is also very inefficient, now
because of the speed which is too low.
x0
x1
x2
x3
x4
x5
x6
x7
x8
x9 x10 x11 x12 x13 x14 x15
▼
EP Nand
P Nand (4)
P Nand (8)
y0
y1
y2
y3
y4
y5
y6
y7
y8
y9 y10 y11 y12 y13 y14 y15
P Nand (16)
Figure 6.16: P CNAN D (16)
The third implementation is a optimal one. For P CNAN D (n) is used the recursive defined
network represented in Figure 6.15 [Ladner ’80], where in each node, for our application, there is a
2-inputs AN D gate. If P CNAN D (n/2) is a well-functioning prefix network, then all the structure
works as a prefix network. Indeed, P CNAN D (n/2) computes all even outputs because of the n/2
input circuits that perform the y2 function between successive pairs of inputs. On the output
level the odd outputs are computed using even outputs and odd inputs. The P CNAN D (n/2)
structure is built upon the same rule and so on until P CNAN D (1) that is a system without any
circuit. The previous recursive definition is applied for n = 16 in Figure 6.16.
The size of P CNAN D (n), S(n) (with n a power of 2), is evaluated starting from: S(1) = 0,
S(n) = S(n/2) + (n − 1)Sy1 where Sy1 is the size of the elementary circuit that defines the
network (in our case is a 2-inputs AND gate). The next steps leads to:
S(n) = S(n/2i ) + (n/2i−1 + . . . + n/20 − i)Sy1
and ends, for i = log2 n (the rule is recursively applied log n times), with:
S(n) = (2n − 2 − log2 n)Sy1 ∈ O(n).
(The network consists in two binary trees of elementary circuits. The first with the bottom
root having n/2 leaves on the first level. The second with the top root having n/2 − 1 leaves
120
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
on the first level. Therefore, the first tree has n − 1 elementary circuits and the second tree
has n − 1 − log n elementary circuits.) The depth is D(n) = 2Dy2 log2 n ∈ O(log n) because
D(n) = D(n/2) + 2 (at each step two levels are added to the system).
6.1.7
Increment circuit
The simplest arithmetic operation is the increment. The combinational circuit performing this
function receives an n-bit number, xn−1 , . . . x0 , and a one-bit command, inc, enabling the operation. The outputs, yn−1 , . . . y0 , and crn−1 behaves according to the value of the command:
If inc = 1, then
{crn−1 , yn−1 , . . . y0 } = {xn−1 , . . . x0 } + 1
else
{crn−1 , yn−1 , . . . y0 } = {0, xn−1 , . . . x0 }.
in
xn−1
cr
xn−2
❄
inc
crn−1 ✛
in
cr
EINC inc
❄
✛ crn−2
out
EINC
❄
yn−1
out
a.
x0
❄
✛
INCn−1
inc
INCn
❄
yn−2
❄
y0
b.
Figure 6.17: Increment circuit. a. The elementary increment circuit (called also half adder). b.
The recursive definition for an n-bit increment circuit.
The increment circuit is built using as “brick” the elementary increment circuit, EINC,
represented in Figure 6.17a, where the XOR circuit generate the increment of the input if inc = 1
(the current bit is complemented), and the circuit AND generate the carry for the the next binary
order (if the current bit is incremented and it has the value 1). An n-bit increment circuit, IN Cn
is recursively defined in Figure 6.17b: IN Cn is composed using an IN Cn−1 serially connected
with an EINC, where IN C0 = EIN C.
6.1.8
Adders
Another usual digital functions is the sum. The circuit associated to this function can be also
made starting from a small elementary circuits, which adds two one-bit numbers, and looking
for a simple recursive definitions for n-bit numbers.
The elementary structure is the well known full adder which consists in two half adders and
an OR2 . An n-bit adder could be done in a recursive manner as the following definition says.
Definition 6.7 The full adder, FA, is a circuit which adds three 1-bit numbers generating a
2-bit result:
F A(in1, in2, in3) = {out1, out0}
FA is used to build n-bit adders. For this purpose its connections are interpreted as follows:
• in1, in2 represent the i-th bits if two numbers
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
121
• in3 represents the carry signal generated by the i − 1 stage of the addition process
• out0 represents the i-th bit of the result
• out1 represents the carry generated for the i + 1-th stage of the addition process
Follows the Verilog description:
module full_adder(output sum, carry_out, input in1, in2, carry_in);
half_adder ha1(sum1, carry1, in1, in2),
ha2(sum, carry2, sum1, carry_in);
assign carry_out = carry1 | carry2;
endmodule
module half_adder(output sum, carry, input
assign sum = in1 ^ in2,
carry = in1 & in2;
endmodule
in1, in2);
⋄
Note: The half adder circuit is also an elementary increment circuit (see Figure 6.17a).
Definition 6.8 The n-bits ripple carry adder, (ADDn ), is made by serial connecting on the
carry chain an ADDn−1 with a FA (see Figure 6.18). ADD1 is a full adder.
module adder #(parameter n = 4)(output [n-1:0] out ,
output
cry ,
input
[n-1:0] in1 ,
input
[n-1:0] in2 ,
input
cin );
wire
[n:1]
carry
;
assign cry = carry[n]
;
generate
if (n == 1) fullAdder firstAder(.out(out[0]
.cry(carry[1]
.in1(in1[0]
.in2(in2[0]
.cin(cin
else
begin
adder #(.n(n-1)) partAdder(
),
),
),
),
));
.out(out[n-2:0]
.cry(carry[n-1]
.in1(in1[n-2:0]
.in2(in2[n-2:0]
.cin(cin
fullAdder lastAdder(.out(out[n-1]
),
.cry(carry[n]
),
.in1(in1[n-1]
),
.in2(in2[n-1]
),
.cin(carry[n-1] ));
end
endgenerate
endmodule
module fullAdder(
output
output
input
out ,
cry ,
in1 ,
),
),
),
),
));
122
CHAPTER 6. GATES:
assign
assign
endmodule
input
input
cry = in1 & in2
out = in1 ^ in2
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
in2 ,
cin);
| (in1 ^ in2) & cin ;
^ cin
;
⋄
An−1 Bn−1
❄
❄
A
✛
Cn
An−2
❄
...
A0 Bn−2
B0
...
❄❄
❄
B
FA
C+
S
C
✛
ADDn−1
C+
❄
C
❄
Sn−2
Sn−1
...
✛
C0
❄
S0
Figure 6.18: The recursive defined n-bit ripple-carry adder (ADDn ). ADDn is simply
designed adding to an ADDn−1 a full adder (FA), so as the carry signal ripples from one FA to the next.
The previous definition used the conditioned generation block.3 The Verilog code from the
previous recursive definition can be used to simulate and to synthesize the adder circuit. For
this simple circuit this definition is too sophisticated. It is presented here only to provide a
simple example of how a recursive definition is generated.
A simpler way to define an adder is provided in the next example where a generate block is
used.
Example 6.5 Generated n-bit adder:
module add #(parameter n=8)(
input
input
output
output
[n-1:0] in1, in2,
cIn
,
[n-1:0] out
,
cOut
);
wire
[n:0]
cr ;
assign cr[0] = cIn ;
assign cOut = cr[n] ;
genvar i
;
generate for (i=0; i<n;
fa adder(
.in1
.in2
.cIn
.out
.cOut
end
endgenerate
endmodule
module fa( input
output
wire
xr ;
i=i+1) begin: S
(in1[i] ),
(in2[i] ),
(cr[i] ),
(out[i] ),
(cr[i+1]));
in1, in2, cIn
out, cOut
,
);
3
The use of the conditioned generation block for recursive definition was suggested to me by my colleague
Radu Hobincu.
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
assign xr
= in1 ^ in2
assign out = xr ^ cIn
assign cOut = in1 & in2
endmodule
| cIn
123
;
;
& xr ;
⋄
Because the add function is very frequently used, the synthesis and simulation tools are able
to ”understand” the simplest one-line behavioral description used in the following module:
module add #(parameter n=8)(
input
[n-1:0]
input
output [n-1:0]
output
assign {cOut, out} = in1 + in2 + cIn
;
endmodule
in1, in2,
cIn
,
out
,
cOut
);
Carry-Look-Ahead Adder
The size of ADDn is in O(n) and the depth is unfortunately in the same order of magnitude.
For improving the speed of this very important circuit there was found a way for accelerating
the computation of the carry: the carry-look-ahead adder (CLAn ). The fast carry-look-ahead
adder can be made using a carry-look-ahead (CL) circuit for fast computing all the carry signals
Ci and for each bit an half adder and a XOR (the modulo two adder)(see Figure 6.19). The half
adder has two roles in the structure:
Ai
Bi
❄
❄
A
...
B
S
Pi
...
HA
CR
✛
Gn−1
❄
G0
...
Pn−1
❄
❄
P0
...
❄
Gi
Ci
❄
✛ C0
Carry-Lookahead Circuit
❄❄
Si
CnCn−1
...
❄
C1
Figure 6.19: The fast n-bit adder. The n-bit Carry-Lookahead Adder (CLAn ) consists in n HAs,
n 2-input XORs and the Carry-Lookahead Circuit used to compute faster the n Ci , for i = 1, 2, . . . n.
• sums the bits Ai and Bi on the output S
• computes the signals Gi (that generates carry as a local effect) and Pi (that allows the
propagation of the carry signal through the binary level i) on the outputs CR and P .
The XOR gate adds modulo 2 the value of the carry signal Ci to the sum S.
In order to compute the carry input for each binary order an additional fast circuit must
be build: the carry-look-ahead circuit. The equations describing it start from the next rule: the
carry toward the level (i + 1) is generated if both Ai and Bi inputs are 1 or is propagated
from the previous level if only one of Ai or Bi are 1. Results:
Ci+1 = Ai Bi + (Ai + Bi )Ci = Ai Bi + (Ai ⊕ Bi )Ci = Gi + Pi Ci .
124
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Applying the previous rule we obtain the general form of Ci+1 :
Ci+1 = Gi + Pi Gi−1 + Pi Pi−1 Gi−2 + Pi Pi−1 Pi−2 Gi−3 + . . . + Pi Pi−1 . . . P1 P0 C0
for i = 0, . . . , n.
Computing the size of the carry-look-ahead circuit results SCL (n) ∈ O(n3 ), and the theoretical depth is only 2. But, for real circuits an n-input gates can not be considered as a one-level
circuit. In Basic circuits appendix (see section Many-Input Gates) is shown that an optimal
implementation of an n-input simple gate is realized as a binary tree of 2-input gates having the
depth in O(log n). Therefore, in a real implementation the depth of a carry-look ahead circuit
has DCLA ∈ O(log n).
For small n the solution with carry-look-ahead circuit works very good. But for larger n
the two solutions, without carry-look-ahead circuit and with carry-look-ahead circuit, must be
combined in many fashions in order to obtain a good price/performance ratio. For example, the
ripple carry version of ADDn is divided in two equal sections and two carry look-ahead circuits
are built for each, resulting two serial connected CLAn/2 . The state of the art in this domain is
presented in [Omondi ’94].
It is obvious that the adder is a simple circuit. There exist constant sized definition for all
the variants of adders.
∗ Prefix-Based Carry-Look-Ahead Adder
Let us consider the expressions for C1 , C2 , . . . Ci , . . .. It looks like each product from Ci−1 is a prefix of
a product from Ci . This suggests a way to reduce the too big size of the carry-look-ahead circuit from
the previous paragraph. Following [Cormen ’90] (see section 29.2.2), an optimal carry-look-ahead circuit
(with asymptotically linear size and log depth) is described in this paragraph.
Let be xi , the carry state, the information used in the stage i to determine the value of the carry.
The carry state takes three values, according to the table represented in Figure 6.20, where Ai and Bi
are the i-th bits of the numbers to be added. If Ai = Bi = 0, then the carry bit is 0 (it is killed). If
Ai = Bi = 1, then the carry bit is 1 (it is generated). If Ai = 0, and Bi = 1 or Ai = 1 and Bi = 0,
then the carry bit is equal with the carry bit generated in the previous stage, Ci−1 (the carry from the
previous range propagates). Therefore, in each binary stage the carry state has three values: k, p, g.
Ai−1
0
0
1
1
Bi−1
0
1
0
1
Ci
0
Ci−1
Ci−1
1
xi
kill
propagate
propagate
generate
Figure 6.20: Kill-Propagate-Generate table.
We define the function ⊗ which composes the carry states of the two binary ranges. In Figure 6.21
the states xi and xi−1 are composed generating the carry state of two successive bits, i − 1 an i. If xi = k,
then the resulting composed state is independent by xi−1 and it takes the value k. If xi = g, then the
resulting composed state is independent by xi−1 and it takes the value g. If xi = p, then the resulting
composed state propagates and it is xi−1 .
The carry state for the binary range i is determined using the expression:
yi = yi−1 ⊗ xi = x0 ⊗ x1 ⊗ . . . xi
starting with:
y0 = x0 = {k, g}
which is associated with the carry state used to specify the carry input C0 (for k is no carry, for g is
carry, while p has no meaning for the input carry state).
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
⊗
k
p
g
xi−1
k
k
k
k
xi
p
k
p
g
125
g
g
g
g
Figure 6.21: The elementary Carry-Look-Ahead (eCLA) function [Cormen ’90].
Thus, carry states are computed as follows:
y 0 = x0
y 1 = x0 ⊗ x 1
y 2 = x0 ⊗ x 1 ⊗ x2
...
y i = x0 ⊗ x 1 ⊗ x2 . . . x i
It is obvious that we have a prefix computation. Let us call the function ⊗ eCLA (elementary CarryLook-Ahead). Then, the circuit used to compute the carry states y0 , y1 , y2 , . . . , yi is prefixCLA.
Theorem 6.1 If x0 = {k, g}, then y1 = {k, g} for i = 1, 2, . . . i and the value of Ci is set to 0 by k and
to 1 by g.
Proof: in the table from Figure 6.21 if the line p is not selected, then the value p does not occur in
any carry state. Because, x0 = {k, g} the previous condition is fulfilled.
⋄
An appropriate codding of the three values of x will provide a simple implementation. We propose
the codding represented in the table from Figure 6.22. The two bits used to code the state x are P and
G (with the meaning used in the previous paragraph). Then, yi [0] = Ci .
xi
k
p
g
P
0
1
0
G
0
0
1
Figure 6.22: Coding the carry state.
The first, direct form of the adder circuit, for n = 7, is represented in Figure 6.23, where there are
three levels:
• the input level of half-adders which compute xi = {Pi , Gi } for 1 = 1, 2, . . . 6, where:
Pi = Ai−1 ⊕ Bi−1
Gi = Ai−1 · Bi−1
• the intermediate level of prefixCLA which computes the carry states y1 , y2 , . . . y7 with a chain of
eCLA circuits
• the output level of XORs used to compute the sum Si starting from Pi+1 and Ci .
The size of the resulting circuit is in O(n), while the depth is in the same order of magnitude. The next
step is to reduce the depth of the circuit to O(log n).
The eCLA circuit is designed using the eCLA function defined inFigure 6.21 and the codding of the
carry state presented in Figure 6.22. The resulting logic table is represented in Figure 6.24. It is build
takeing into consideration that the code 11 is not used to define the carry state x. Thus in the lines
126
CHAPTER 6. GATES:
A6
B6
A5
❄ ❄
A4
❄ ❄
HA
x7
✛y7
B5
❄
y6 [0]
eCLA
x6 [1]
A3
y5 [0]
eCLA
x5 [1]
A2
y4 [0]
eCLA
x4 [1]
A1
y3 [0]
eCLA
x3 [1]
A0
y2 [0]
HA
x1
❄
✛y2
eCLA
x2 [1]
B0
❄ ❄
HA
x2
❄
✛y3
B1
❄ ❄
HA
x3
❄
✛y4
B2
❄ ❄
HA
x4
❄
✛y5
B3
❄ ❄
HA
x5
❄
✛y6
B4
❄ ❄
HA
x6
eCLA
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
❄
✛y1
eCLA
y1 [0]
x1 [1]
x0 = {0, C0 }
✛
♦
prefixCLA
❄
y7 [0] = C7 = cOut
S5
S4
S3
S2
S1
S0
Figure 6.23: The 7-bit adder with ripple eCLA.
containing xi = {Pi , Gi } = 11 and xi−1 = {Pi−1 , Gi−1 } = 11 the function is not defined (instead of 0 or
1 the value of the function is “don’t care”). The expressions for the two outputs of the eCLA circuit are:
Pout = Pi · Pi−1
Gout = G − i + P − i · Gi−1
Pi
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
Gi
0
0
0
0
1
1
1
1
0
0
0
0
1
1
1
1
Pi−1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
Gi−1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
Pout
0
0
0
0
0
0
0
0
1
-
Gout
0
0
0
1
1
1
0
1
0
-
Figure 6.24: The logic table for eCLA.
For the pure serial implementation of the prefixCLA circuit from Figure 6.23, because x0 = {0, C0 },
in each eCLA circuit Pout = 0. The full circuit will be used in the log-depth implementation of the
prefixCLA circuit. Following the principle exposed in the paragraph Prefix Computation network the
prefixCLA circuit is redesigned optimally for the adder represented in Figure 6.25. If we take from Figure
6.16 the frame labeled P Nand (8) and the 2-input AND circuits are substituted with eCLA circuits, then
results the prefixCLA circuit from Figure 6.25.
Important notice: the eCLA circuit is not a symmetric one. The most significant inputs, Pi and
Gi correspond to the i-th binary range, while the other two correspond to the previous binary range.
This order should be taken into consideration when the log-depth prefixCLA (see Figure 6.25) circuit
is organized.
While in the prefixCLA circuit from Figure 6.23 all eCLA circuits have the output Pout = 0, in the
log-depth prefixCLA from Figure 6.25 some of them are fully implemented. Because x0 = {0, C0 },
eCLA indexed with 0 and all the eCLAs enchained after it, indexed with 4,6, 7, 8, 9, and 10, are of the
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
127
simple form with Pout = 0. Only the eCLA indexed with 1, 2, 3, and 5 are fully implemented, because
their inputs are not restricted.
A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 A0 B0
❄❄ ❄❄ ❄❄ ❄❄ ❄❄ ❄❄ ❄❄
HA
x7
HA
x6
HA
x5
HA
x4
HA
x3
HA
x2
HA
x1
x0 =
{0, C0 }
log-depth prefixCLA
❄❄
❄❄
eCLA
3
❄❄
eCLA
5
❄❄
eCLA
1
❄❄
eCLA
2
x7 [1]
x6 [1]
x5 [1]
x4 [1]
x3 [1]
x2 [1]
x1 [1]
eCLA
0
❄❄
eCLA
4
❄❄
eCLA
6
❄❄
eCLA
7
❄❄
eCLA
10
y6 [0]
❄❄
eCLA
9
y5 [0]
❄❄
eCLA
8
y4 [0]
y3 [0]
y2 [0]
y1 [0]
y0 [0]
S6
y7 [0] = C7
S5
S4
S3
S2
S1
S0
Figure 6.25: The adder with the prefix based carry-look-ahead circuit.
This version of n-bit adder is asymptotically optimal, because the size of the circuit is in O(n) and
the depth is in O(log n).
∗ Carry-Save Adder
For adding m n-bit numbers there is a faster solution than the one which supposes to use the direct
circuit build as a tree of m − 1 2-number adders. The depth and the size of the circuit is reduced using
a carry-save adder circuit.
The carry save adder receives three n-bit numbers:
x = {xn−1 , xn−2 , . . . x1 , x0 }
y = {yn−1 , yn−2 , . . . y1 , y0 }
z = {zn−1 , zn−2 , . . . z1 , z0 }
and generate two (n + 1)-bit numbers:
c = {cn−1 , cn−2 , . . . c1 , c0 , 0}
s = {0, sn−1 , sn−2 , . . . s1 , s0 }
where:
c + s = x + y + z.
128
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
The function of the circuit is described by the following transfer function applied for i = 0, . . . n − 1:
xi + yi + zi = {ci , si }.
The internal structure of the carry-save adders contains n circuits performing the previous function which
is the function of a full adder. Indeed, the binary variables xi and yi are applied on the two inputs of the
FA, zi is applied on the carry input of the same FA, and the two inputs ci , si are the carry-out and the
sum outputs. Therefore, an elementary carry-save adder, ECSA, has the structure of a FA (see Figure
6.26a).
z = {zn−1 , . . . z0 }
y = {yn−1 , . . . y0 }
x[i]
y[i]
❄
❄❄❄ ❄❄❄
❄
A
C+
z[i]
x = {xn−1 , . . . x0 }
B
FA
C
✛
c
y
x
s
c
c = {cn−1 , . . . c0 , 0}
a.
c
s
❄
❄
❄
❄
3CSAn
❄
s = {0, sn−1 , . . . s0 }
b.
n0 n1 n2
z
ECSA0
❄
s[i]
y
x
s
❄
❄
❄❄❄
z
ECSAn−2
❄
ECSA
c[i]
z
ECSAn−1
S
❄
y
x
n3
❄❄❄
y
x
z
3CSAn
c
s
❄❄❄
y
x
❄❄❄❄
{0, n3}
n0 n1 n2 n3
z
4REDAn
3CSAn+1
c
s
s
❄❄
❄
4CSAn
n0 + n1 + n2 + n3
d.
ADDn+1
❄n0 + n1 + n2 + n3
c.
Figure 6.26: Carry-Save Adder. a. The elementary carry-save adder. b. CSA for 3 n-bit numbers:
3CSAn . c. How to use carry-save adders to add 4 n-bit numbers. d. The logic symbol for a 4 n-bit
inputs reduction adder (4REDAn ) implemented with two 3CSAn and a (n + 1)-bit adder.
To prove for the functionality of CSA we write for each ECSA:
{cn−1 , sn−1 } = xn−1 + yn−1 + zn−1
{cn−2 , sn−2 } = xn−2 + yn−2 + zn−2
...
{c1 , s1 } = x1 + y1 + z1
{c0 , s0 } = x0 + y0 + z0
from which results that:
x + y + z = {cn−1 , sn−1 } × 2n−1 + {cn−2 , sn−2 } × 2n−2 + . . . {c1 , s1 } × 21 + {c0 , s0 } × 20 =
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
129
{cn−1 , cn−2 , . . . c1 , c0 } × 2 + {sn−1 , sn−2 , . . . s1 , s0 } = {cn−1 , cn−2 , . . . c1 , c0 , 0} + {0, sn−1 , sn−2 , . . . s1 , s0 }.
Figure 6.26c shows a 4-number adder for n-bit numbers. Instead of 3 adders, 2 3CSAn circuits and
a (n + 1)-bit adder are used. The logic symbol for the resulting 4-input reduction adder – 4REDAn –
is represented in Figure 6.26d. The depth of the circuit is of 2 FAs and an (n + 1)-bit adder, instead of
the depth associated with one n-bit adder and an (n + 1)-bit adder. The size is also minimized if in the
standard solution carry-look-ahead adders are used.
Example 6.6 The module 3CSAm is generated using the following template:
module csa3_m #(parameter n=m)(input
output
wire
[n-1:0] out ;
wire
[n-1:0] cr ;
[n-1:0] in0, in1, in2
[n:0]
sOut, cOut
,
);
genvar i
;
generate for (i=0; i<n; i=i+1) begin: S
fa adder(in0[i], in1[i], out[i], cr[i]);
end
endgenerate
assign sOut = {1’b0, out}
assign cOut = {cr, 1’b0}
endmodule
;
;
where: (1) an actual value for m must be provided and (2) the module fa is defined in the previous
example. ⋄
Example 6.7 For the design of a 8 1-bit input adder (8REDA1 ) the following modules are used:
module reda8_1(
output [3:0]
out ,
input
[7:0]
in );
wire
[2:0]
sout, cout ;
csa8_1 csa( in[0],in[1],in[2],in[3],in[4],in[5],in[6],in[7],cout,sout);
assign out = cout + sout
;
endmodule
module csa8_1( input
in0, in1, in2, in3, in4, in5, in6, in7,
output [2:0]
cout, sout
);
wire
[1:0]
sout0, cout0, sout1, cout1;
csa4_1 csa0(in0, in1, in2, in3, sout0, cout0),
csa1(in4, in5, in6, in7, sout1, cout1);
csa4_2 csa2(sout0, cout0, sout1, cout1, cout, sout);
endmodule
where csa4 1 and csa4 2 are instances of the following generic module:
module csa4_p( input
[1:0]
in0, in1, in2, in3
output [2:0]
sout, cout
wire
[2:0]
s1, c1 ;
wire
[3:0]
s2, c2 ;
csa3_p csa0(in0, in1, in2, s1, c1);
csa3_q csa1(s1, c1, {1’b0, in3}, s2, c2);
assign sout = s2[2:0], cout = c2[2:0];
endmodule
,
);
for p = 1 and p = 2, while q = p + 1. The module csa3 m is defined in the previous example. ⋄
The efficiency of this method to add many numbers increases, compared to the standard solution,
with the number of operands.
130
CHAPTER 6. GATES:
6.1.9
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
∗ Combinational Multiplier
One of the most important application of the CSAs circuits is the efficient implementation of a combinational multiplier. Because multiplying n-bit numbers means to add n 2n − 1-bit partial products, a
nCSA2n−1 circuit provides the best solution. Adding n n-bit numbers using standard carry-adders is
done, in the best case, in time belonging to O(n × log n) on a huge area we can not afford. The proposed
solution provides a linear time.
In Figure 6.27 is an example for how for the binary multiplication for n = 4 works. Thus, the
combinational multiplication is done in the following three stages:
a3 a2 a1 a0 ×
b3 b2 b1 b0
p23
p33 p32
p03 p02 p01 p00
p13 p12 p11 p10
p22 p21 p20
p31 p30
p7 p6 p5 p4 p3 p2 p1 p0
1 0 1 1 ×
1 1 0 1
multiplicand: a
multiplier: b
partial
partial
partial
partial
product:
product:
product:
product:
pp0
pp1
pp2
pp3
final product: p
1 0 1 1
0 0 0 0
1 0 1 1
1 0 1 1
1 0 0 0 1 1 1 1
Figure 6.27: Multiplying 4-bit numbers.
1. compute n partial products – pp0 , . . . ppn−1 – using a two-dimension array of n × n AN D2 circuits
(see in Figure 6.28 the “partial products computation” dashed box); for n = 4 the following
relations describe this stage:
pp0 = {a3 · b0 , a2 · b0 , a1 · b0 , a0 · b0 }
pp1 = {a3 · b1 , a2 · b1 , a1 · b1 , a0 · b0 }
pp2 = {a3 · b2 , a2 · b2 , a1 · b2 , a0 · b0 }
pp3 = {a3 · b3 , a2 · b3 , a1 · b3 , a0 · b0 }
2. shift each partial product ppi i binary positions left; results n (2n − 1)-bit numbers; for n = 4:
n0 = pp0 << 0 = {0, 0, 0, pp0 }
n1 = pp1 << 1 = {0, 0, pp1 , 0}
n2 = pp2 << 2 = {0, pp2 , 0, 0}
n3 = pp3 << 3 = {pp3 , 0, 0, 0}
easy to be done, with no circuits, by an appropriate connection of the AND array outputs to the
next circuit level (see in Figure 6.28 the “hardware-less shifter” dashed box)
3. add the resulting n numbers using a nCSA2n−1 ; for n = 4:
p = {0, 0, 0, pp0 }plus{0, 0, pp1 , 0}plus{0, pp2 , 0, 0}plus{pp3 , 0, 0, 0}
The combinational multiplier circuit is presented in Figure 6.28 for the small case of n = 4. The first
stage – partial products computation – generate the partial products ppi using 2-input ANDs as one
bit multipliers. The second stage of the computation request a hardware-less shifter circuit, because the
multiplying n-bit numbers with a power of 2 no bigger than n − 1 is done by an appropriate connection of
each ppi to the (2n − 1)-bit inputs of the next stage, filling up the unused positions with zeroes. The third
stage consists of a reduction carry-save adder – nREDA2n−1 – which receives, as 2n − 1-bit numbers,
the partial products ppi , each multiplied with 2i .
The circuit represented in Figure 6.28 for n = 4, has in the general case the size in O(n2 ) and the
depth in O(n). But, the actual size and depth of the circuit is established by the 0s applied to the input
of the nREDA2n−1 circuit, because some full-adders are removed from design and others are reduced to
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
a3
a2
a1
131
a0
b0
pp0 = {a3 · b0 , a2 · b0 , a1 · b0 , a0 · b0 }
b1
4
b2
4
b3
4
pp1
pp2
pp3
4
✣
first stage: partial products computation
✯
3
n3
2
2
n2
n1
3
n0
4REDA7
second stage: hardware-less shifter
s = n0 + n1 + n2 + n3
8
❄p = a × b
Figure 6.28: 4-bit combinational multiplier. .
half-adders. The actual size of nREDA2n−1 results to be very well approximated with the actual size of
a nREDAn . The actual depth of the combinational multiplier is well approximated with the depth of a
2.5n-bit adder.
The decision to use a combinational multiplier must be done takeing into account (1) its area, (2) the
acceleration provided for the function it performs and (3) the frequency of the multiplication operation
in the application.
6.1.10
Arithmetic and Logic Unit
All the before presented circuits have had associated only one logic or one arithmetic function.
Now is the time to design the internal structure of a previously defined circuit having many
functions, which can be selected using a selection code: the arithmetic and logic unit – ALU –
(for the functional description see Example 2.8). ALU is the main circuit in any computational
device, such as processors, controllers or embedded computation structures.
The ALU circuit can be implemented in many forms. One of them is the speculative version
(see Figure 6.29) well described by the Verilog module from Example 2.8, where the case structure describes, in fact, an 8-input multiplexor for 33-bit words. We call this version speculative
because all the possible functions are computed in order to be all available to be select when the
function code arrives to the func input of ALU. This approach is efficient when the operands are
available quickly and the function to be performed “arrives” lately (because it is usually decoded
from the instruction fetched from a program memory). The circuit “speculates” computing all
the defined functions offering 8 results from which the func code selects one. (This approach
will be useful for the ALU designed for the stack processor described in Chapter 10.)
The speculative version provides a fast version in some specific designs. The price is the
big size of the resulting circuit (mainly because the arithmetic section contains and adder and
an subtractor, instead a smaller circuit performing add or subtract according to a bit used to
complement the right operand and the carryIn signal).
An area optimized solution is provided in the next example.
Example 6.8 Let be the 32-bit ALU with 8 functions described in Example 2.8. The imple-
132
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
right[n-1:0]
{2’b0, left[n-1:1]}
left[n-1:0]
carryIn
❄
❄❄
a+b
❄
❄❄
a−b
❄
❄
n × AN D2
❄
❄
n × OR2
❄
❄
❄
n×
n × XOR2
func
N OT
✲ sel
❄❄❄❄❄❄❄❄
01234567
(n + 1) × M U X8
❄
{carryOut, out[n-1:0]}
Figure 6.29: The internal structure of the speculative version of an arithmetic and
logic unit. Each function is performed by a specific circuit and the output multiplexer selects the
desired result.
mentation will be done using an adder-subtractor circuit and a 1-bit slice for the logic functions.
Results the following Verilog description:
module structuralAlu(output [31:0] out
,
output
carryOut,
input
carryIn ,
input
[31:0] left
, right
input
[2:0]
func
);
wire
[31:0] shift, add_sub, arith, logic;
addSub addSub(.out
(add_sub ),
.cout
(carryOut),
.left
(left
),
.right (right
),
.cin
(carryIn ),
.sub
(func[0] ));
logic log( .out
(logic
),
.left
(left
),
.right (right
),
.op
(func[1:0]));
mux2
shiftMux(.out(shift
),
.in0(left
),
.in1({1’b0, left[31:1]}),
.sel(func[0]
)),
arithMux(.out(arith ),
.in0(shift ),
.in1(add_sub),
.sel(func[1])),
outMux(.out(out
),
.in0(arith ),
.in1(logic ),
.sel(func[2]));
endmodule
module mux2(input
input
[31:0]
sel
,
in0, in1,
,
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
assign
endmodule
output [31:0] out
out = sel ? in1 : in0;
module addSub(output [31:0]
output
input [31:0]
input
assign {cout, out} = left
endmodule
133
);
out
,
cout
,
left, right ,
cin, sub
);
+ (right ^ {32{sub}}) + (cin ^ sub);
module logic(output reg [31:0] out
,
input
[31:0] left, right ,
input
[1:0] op
);
integer i;
wire [3:0] f;
assign f = {op[0], ~(~op[1] & op[0]), op[1], ~|op};
always @(left or right or f)
for(i=0; i<32; i=i+1) logicSlice(out[i], left[i], right[i], f);
task
logicSlice;
output
o;
input
l, r;
input
[3:0]
f;
o = f[{l,r}];
endtask
endmodule
The resulting circuit is represented in Figure 6.30. This version can be synthesized on a
smaller area, because the number of EM U Xs is smaller, instead of an adder and a subtractor
an adder/subtractor is used. The price for this improvement is a smaller speed. Indeed, the
add sub module “starts” to compute the addition or the subtract only when the signal sub =
func[0] is received. Usually, the code func results from the decoding of the current operation
to be performed, and, consequently, comes later. ⋄
right
left
✲
❄
❄
✲
n × M U XE
❄
❄
sub
add sub
✛
✛
carryOut
✲
✲
❄
❄
logic
carryIn
func[0]
✲
❄
❄
n × M U XE
func[1]
func[2]
✲
❄
❄
n × M U XE
❄out
Figure 6.30: The internal structure of an area optimized version of an ALU. The add sub
module is smaller than an adder and a subtractor, but the operation “starts” only when func[0] is valid.
134
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
We just learned a new feature of the Verilog language: how to use a task to describe a
circuit used many times in implementing a simple, repetitive structure.
The internal structure of ALU consists mainly in n slices, one for each input pair left[i],
rught[i] and a carry-look-ahead circuit(s) used for the arithmetic section. It is obvious that
ALU is also a simple circuit. The magnitude order of the size of ALU is given by the size of the
carry-look-ahead circuit because each slice has only a constant dimension and a constant depth.
Therefore, the fastest version implies a size in O(n3 ) because of the carry-look-ahead circuit.
But, let’s remind: the price for the fastest solution is always too big! For optimal solutions see
[Omondi ’94].
6.1.11
Comparator
Comparing functions are used in decisions. Numbers are compared to decide if they are equal
or to indicate the biggest one. The n-bit comparator, COM Pn , is represented in Figure 6.31a.
The numbers to be compared are the n-bit positive integers a and b. Three are the outputs of
the circuit: lt out, indicating by 1 that a < b, eq out, indicating by 1 that a = b, and gt out,
indicating by 1 that a > b. Three additional inputs are used as expanding connections. On
these inputs is provided information about the comparison done on the higher range, if needed.
If no higher ranges of the number under comparison, then these thre inputs must be connected
as follows: lt in = 0, eq in = 1, gt in = 0.
a[n-1:0]
lt in
eq in
gt in
b[n-1:0]
❄
✲
✲
✲
a[i]
❄
✲
✲
✲
COMP
lt out
lt in
eq out
eq in
gt out
gt in
a.
✲
✲
✲
b[i]
❄
❄
ECOMP
✲
✲
✲
lt out
eq out
gt out
b.
a[n-1] b[n-1]
✲
✲
✲
lt in
eq in
gt in
❄
❄
ECOMP
a[n-2:0]
✲
✲
✲
b[n-2:0]
❄
COM Pn−1
❄
✲
✲
✲
lt out
eq out
gt out
c.
Figure 6.31: The n-bit comparator, COM Pn . a. The n-bit comparator. b. The elementary
comparator. c. A recursive rule to built an COM Pn , serially connecting an ECOM P with a COM Pn−1
The comparison is a numerical operation which starts inspecting the most significant bits
of the numbers to be compared. If a[n − 1] = b[n − 2], then the result of the comparison is
given by comparing a[n − 2 : 0] with b[n − 1 : 0], else, the decision can be done comparing
only a[n − 1] with b[n − 1] (using an elementary comparator, ECOM P = COM P1 (see Figure
6.31b)), ignoring a[n − 2 : 0] and b[n − 2 : 0]. Results a recursive definition for the comparator
circuit.
Definition 6.9 An n-bit comparator, COM Pn , is obtained serially connecting an COM P1 with
a COM Pn−1 . The Verilog code describing COM P1 (ECOMP) follows:
module e_comp( input
a
b
,
,
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
assign
lt_in
,
eq_in
,
gt_in
,
output lt_out ,
eq_out ,
gt_out );
lt_out = lt_in | eq_in
eq_out = eq_in & ~(a ^
gt_out = gt_in | eq_in
135
// the previous e_comp decided lt
// the previous e_comp decided eq
// the previous e_comp decided gt
// a < b
// a = b
// a > b);
& ~a & b,
b),
& a & ~b;
endmodule
⋄
The size and the depth of the circuit resulting from the previous definition are in O(n). The
size is very good, but the depth is too big for a high speed application.
An optimal comparator is defined using another recursive definition based on the divide et
impera principle.
Definition 6.10 An n-bit comparator, COM Pn , is obtained using two COM Pn/2 , to compare the higher and the lower half of the numbers (resulting {lt out high, eq out high,
gt out high} and {lt out low, eq out low, gt out low}), and a COM P1 to compare
gt out low with lt out low in the context of {lt out high, eq out high, gt out high}. The
resulting circuit is represented in Figure 6.32. ⋄
a[n-1:n/2]
“1”
✲
✲
✲
b[n-1:n/2]
❄
a[n/2-1:0]
❄
“1”
COM Pn/2
✲
✲
✲
❄
✲
✲
✲
❄
ECOMP
b[n/2-1:0]
❄
COM Pn/2
✲
✲
✲
❄
✲
lt out
eq out
gt out
Figure 6.32: The optimal n-bit comparator. Applying the divide et impera principle a COM Pn
is built using two COM Pn/2 and an ECOM P . Results a log-depth circuit with the size in O(n).
The resulting circuit is a log-level binary tree of ECOMPs. The size remains in the same
order4 , but now the depth is in O(log n).
The bad news is: the HDL languages we have are unable to handle safely recursive definitions.
The good news is: the synthesis tools provide good solutions for the comparison functions
starting from a very simple behavioral description.
6.1.12
∗ Sorting network
In the most of the cases numbers are compared in order to be sorted. There are a lot of algorithms for
sorting numbers. They are currently used to write programs for computers. But, in G2CE the sorting
function will migrate into circuits, providing specific accelerators for general purpose computing machines.
4
The actual size of the circuit can be minimized takeing into account that: (1) the compared input of ECOMP
cannot be both 1, (2) the output eq out of one COM Pn/2 is unused, and (3) the expansion inputs of both
COM Pn/2 are all connected to fix values.
136
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
To solve in circuits the problem of sorting numbers we start again from an elementary module: the
elementary sorter (ESORT).
Definition 6.11 An elementary sorter (ESORT) is a combinational circuit which receives two n-bit
integers, a and b and generate outputs them as min(a,b) and max(a,b). The logic symbol of ESORT is
represented in Figure 6.33b. ⋄
The internal organization of an ESORT is represented in Figure 6.33a. If COM Pn is implemented
in an optimal version, then this circuit is optimal because its size is linear and its depth is logarithmic.
a
b
✲
“1”
✲
✲
✲
❄
❄
❄
n × EM U X
✲
❄
❄
n × EM U X
❄
COM Pn
a
b
✲
✲
min(a,b) max(a,b)
❄
❄
min(a,b)
max(a,b)
a.
b.
Figure 6.33: The elementary sorter. a. The internal structure of an elementary sorter. The
output lt out of the comparator is used to select the input values to output in the received order (if
lt out = 1) or in the crossed order (if lt out = 0). b. The logic symbol of an elementary sorter.
stage 1
d
stage 2
stage 3
max(max(a,b), max(c,d))
c
max(min(max(a,b)max(cd),max(min(a,b), min(c,d)))
b
min(min(max(a,b)max(cd),max(min(a,b), min(c,d)))
a
min(min(a,b), min(c,d))
Figure 6.34: The 4-input sorter. The 4-input sorter is a three-stage combinational circuit built by
5 elementary sorters.
The circuit for sorting a vector of n numbers is build by ESORTs organized on many stages. The resulting combinational circuit receives the input vector (x1 , x2 , x3 , . . . xn ) and generates the sorted version
of it. In Figure 6.34 is represented a small network of ESORTs able to sort the vector of integers (a, b, c, d).
The sorted is organized on three stages. On the first stage two ESORTs are used sort separately the
sub-vectors (a, b) and (c, d). On the second stage, the minimal values and the maximal values obtained
from the previous stage are sorted, resulting the the smallest value (the minimal of the minimals), the
biggest value (the maximal of the maximal) and the two intermediate values. For the last two the third
level contains the last ESORT which sorts the middle values.
The resulting 4-input sorting circuit has the depth Dsort (4) = 3 × Desort (n) and the size Ssort (4) =
5 × Sesort (n), where n is the number of bits used to represent the sorted integers.
Bathcer’s sorter
What is the rule to design a sorter for a n-number sequence? This topic will pe presented using [Batcher
’68], [Knuth ’73] or [Parberry ’87].
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
137
The n-number sequence sorter circuit, Sn , is presented in Figure 6.35. It is a double-recursive
construct containing two Sn/2 modules and the merge module Mn , which has also a recursive definition,
because it contains two Mn/2 modules and n/2 − 1 elementary sorters, S2 . The Mn module is defined
as a sorter which sorts the input sequence of n numbers only if it is formed by two n/2-number sorted
sequences.
In order to prove that Batcher’s sorter represented in Figure 6.35 sorts the input sequence we need
the following theorem.
Theorem 6.2 An n-input comparator network is a sorting network iff it works as sorter for all sequences
of n symbols of zeroes and ones. ⋄
The previous theorem is known as Zero-One Principle.
We must prove that, if Mn/2 are merge circuits, then Mn is a merge circuit. The circuit M2 is an
elementary sorter, S2 .
If {x0 , . . . , xn−1 } is a sequence of 0 and 1, and {a0 , . . . , an−1 } is a sorted sequence with g zeroes,
while {b0 , . . . , bn−1 } is a sorted sequence with h zeroes, then the left Mn/2 circuit receives ⌈g/2⌉ + ⌈h/2⌉
zeroes5 , and the right Mn/2 circuit receives ⌊g/2⌋ + ⌊h/2⌋ zeroes6 . The value:
Z = ⌈g/2⌉ + ⌈h/2⌉ − (⌊g/2⌋ + ⌊h/2⌋)
takes only three values with the following output behavior for Mn :
Z=0 : at most one S2 receives on its inputs the unsorted sequence {1,0} and does it work, while all the
above receive {0,0} and the bellow receive {1,1}
Z=1 : y0 = 0, follows a number of elementary sorters receiving {0,0}, while the rest receive {1,1} and
the last output is yn−1 = 1
Z=2 : y0 = 0, follows a number of elementary sorters receiving {0,0}, then one sorter with {0,1} on
its inputs, while the rest receive {1,1} and the last output is yn−1 = 1
Thus, no more than one elementary sorter is used to reverse the order in the received sequence.
The size and depth of the circuit is computed in two stages, corresponding to the two recursive levels
of the definition. The size of the n-input merge circuit, SM (n), is iteratively computed starting with:
SM (n) = 2SM (n/2) + (n/2 − 1)SS (2)
SM (2) = SS (2)
Once the size of the merge circuit is obtained, the size of the n-input sorter, SS (n), is computed using:
SS (n) = 2SS (n/2) + SM (n)
Results:
SS (n) = (n(log2 n)(−1 + log2 n)/4 + n − 1)SS (2)
A similar approach is used for the computation of the depth. The depth of the n-input merge circuit,
DM (n), is iteratively computed using:
DM (n) = DM (n/2) + DS (2)
DM (2) = DS (2)
while the depth of the n-input sorter, DS (n), is computed with:
DS (n) = DS (n/2) + DM (n)
Results:
DS (n) = (log2 n)(log2 n + 1)/2
We conclude that: SS (n) ∈ O(n × log 2 n) and DS (n) ∈ O(log 2 n).
5
6
⌈a⌉ means rounded up integer part of the number a.
⌊a⌋ means rounded down integer part of the number a.
138
CHAPTER 6. GATES:
x0
x1
x2
x3
✲
✲
✲
✲
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
a0
a1
a2
a3
Sn/2
xn/2−2
xn/2−1
xn/2
xn/2+1
xn/2+2
xn/2+3
✲
✲
an/2−2
an/2−1
✲
✲
✲
✲
b0
b1
b2
b3
Sn/2
xn−2
xn−1
✲
✲
bn/2−2
bn/2−1
Mn/2
m0 m1 m2
Mn/2
Mn
mn/2−1
m0 m1
y0
y1
y2
Sn
y3
y4
yn−3
yn−2
yn−1
Figure 6.35: Batcher’s sorter. The n-input sorter, Sn , is defined by a double-recursive construct:
“Sn = 2 × Sn/2 + Mn ”, where the merger Mn consists of “Mn = 2 × Mn/2 + (n/2 − 1)S2 ”.
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
4-input sorter
8-input merger
139
8-input sorter
✌
✢
level 1
level 2
✢
level 3
level 4
level 5
level 6
level 7
level 8
level 9
level 10
Figure 6.36: 16-input Batcher’s sorter.
level 1
level 2
level 3
level 4
level 5
Figure 6.37: 32-input Batcher’s merege circuit.
140
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
In Figure 6.36 a 16-input sorter designed according to the recursive rule is shown, while in Figure
6.37 a 32-input merge circuit is detailed. With 2 16-input sorters and a 32-input merger a 32-input sorted
can be build.
In [Ajtai ’83] is a presented theoretically improved algorithm, with SS (n) ∈ O(n×log n) and DS (n) ∈
O(log n). But, the constants associated to the magnitude orders are too big to provide an optimal solution
for currently realistic circuits characterized by n < 109 .
The recursive Verilog description is very useful for this circuit because of the difficulty to
describe in a HDL a double recursive circuit.
The top module describe the first level of the recursive definition: a n-input sorter is built using two
n/2-input sorters and a n-input merger, and the 2-input sorter is the elementary sorter. Results the
following description in Verilog:
module sorter #(‘include "0_parameters.v")
(
output [m*n-1:0]
out ,
input
[m*n-1:0]
in );
wire
wire
[m*n/2-1:0] out0;
[m*n/2-1:0] out1;
generate
if (n == 2) eSorter eSorter(.out0
(out[m-1:0]
),
.out1
(out[2*m-1:m]
),
.in0
(in[m-1:0]
),
.in1
(in[2*m-1:m]
));
else
begin
sorter #(.n(n/2))
sorter0(.out(out0
),
.in (in[m*n/2-1:0] )),
sorter1(.out(out1
),
.in (in[m*n-1:m*n/2]));
merger #(.n(n))
merger( .out(out
),
.in ({out1, out0}
));
end
endgenerate
endmodule
For the elementary sorter, eSorter, we have the following description:
module eSorter #(‘include "0_parameters.v")
(
output [m-1:0] out0,
output [m-1:0] out1,
input
[m-1:0] in0 ,
input
[m-1:0] in1 );
assign out0 = (in0 > in1) ? in1 : in0
assign out1 = (in0 > in1) ? in0 : in1
endmodule
;
;
The n-input merge circuit is described recursively using two n/2-input merge circuits and the elementary sorters as follows:
module merger #(‘include "0_parameters.v")
(
output [m*n-1:0]
out ,
input
[m*n-1:0]
in );
wire
wire
wire
wire
[m*n/4-1:0]
[m*n/4-1:0]
[m*n/4-1:0]
[m*n/4-1:0]
even0
odd0
even1
odd1
;
;
;
;
6.1. SIMPLE, RECURSIVE DEFINED CIRCUITS
wire
wire
[m*n/2-1:0] out0
[m*n/2-1:0] out1
141
;
;
genvar i;
generate
if (n == 2) eSorter eSorter(.out0
(out[m-1:0]
),
.out1
(out[2*m-1:m]
),
.in0
(in[m-1:0]
),
.in1
(in[2*m-1:m]
));
else
begin
for (i=0; i<n/4; i=i+1) begin : oddEven
assign even0[(i+1)*m-1:i*m] = in[2*i*m+m-1:2*i*m]
assign even1[(i+1)*m-1:i*m] = in[m*n/2+2*i*m+m-1:m*n/2+2*i*m]
assign odd0[(i+1)*m-1:i*m] = in[2*i*m+2*m-1:2*i*m+m]
assign odd1[(i+1)*m-1:i*m] = in[m*n/2+2*i*m+2*m-1:m*n/2+2*i*m+m]
end
merger #(.n(n/2))
merger0(.out(out0
),
.in ({even1, even0} )),
merger1(.out(out1
),
.in ({odd1, odd0}
));
for (i=1; i<n/2; i=i+1) begin : elSort
eSorter eSorter(.out0
(out[(2*i-1)*m+m-1:(2*i-1)*m]
),
.out1
(out[2*i*m+m-1:2*i*m]
),
.in0
(out0[i*m+m-1:i*m]
),
.in1
(out1[i*m-1:(i-1)*m]
));
end
assign out[m-1:0]
= out0[m-1:0]
;
assign out[m*n-1:m*(n-1)]
= out1[m*n/2-1:m*(n/2-1)]
;
end
endgenerate
endmodule
;
;
;
;
The parameters of the sorter are defined in the file 0 parameters.v:
parameter
6.1.13
n = 16, // number of inputs
m = 8
// number of bits per input
∗ First detection network
Let be the vector of Boolean variable: inVector = {x0, x1, ...
outputs three vectors of the same size:
x(n-1)}. The function firstDetect
first
= {0, 0, ... 0, 1, 0, 0, ... 0}
beforeFirst = {1, 1, ... 1, 0, 0, 0, ... 0}
afterFirst = {0, 0, ... 0, 0, 1, 1, ... 1}
indicating, by turn, the position of the first 1 in inVector, all the positions before the first 1, and all the
positions after the first 1.
Example 6.9 Let be a 16-bit input circuit performing the function firstDetect.
inVector
= {0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1}
first
= {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
beforeFirst = {1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
afterFirst = {0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}
⋄
142
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
The circuit performing the function firstDetect is described by the following Verilog program:
module firstDetect #(parameter n =4)( input
output
output
output
wire [0:n-1] orPrefixes;
[0:n-1]
[0:n-1]
[0:n-1]
[0:n-1]
in
,
first
,
beforeFirst,
afterFirst );
or_prefixes prefixNetwork(.in (in
),
.out(orPrefixes));
assign first
= orPrefixes & ~(orPrefixes >> 1),
beforeFirst = ~orPrefixes
,
afterFirst = (orPrefixes >> 1)
;
endmodule
module or_prefixes #(parameter n =4)(input
[0:n-1] in ,
output reg [0:n-1] out);
integer k;
always @(in) begin out[0] = in[0];
for (k=1; k<n; k=k+1) out[k] = in[k] | out[k-1];
end
endmodule
The function firstDetect classifies each component of a Boolean vector related to the first occurrence
of the value 17 .
6.1.14
6.2
6.2.1
∗ Spira’s theorem
Complex, Randomly Defined Circuits
An Universal circuit
Besides the simple, recursively defined circuits there is the huge class of the complex or random
circuits. Is there a general solution for these category of circuits? A general solution asks a
general circuit and this circuit surprisingly exists. Now rises the problem of how to catch the
huge diversity of random in this approach. The following theorem will be the first step in solving
the problem.
Theorem 6.3 For any n, all the functions of n binary-variables have a solution with a combinational logic circuit. ⋄
Proof Any Boolean function of n variables can be written as:
f (xn−1 , . . . , x0 ) = x′n−1 g(xn−2 , . . . , x0 ) + xn−1 h(xn−2 , . . . x0 ).
where:
g(xn−2 , . . . , x0 ) = f (0, xn−2 , . . . , x0 )
h(xn−2 , . . . , x0 ) = f (1, xn−2 , . . . , x0 )
Therefore, the computation of any n-variable function can be reduced to the computation
of two other (n − 1)-variables functions and an EMUX. The circuit, and in the same time the
algorithm, is represented in Figure 6.40. For the functions g and h the same rule may applies.
7
The function firstDetect becomes very meaningful related to the minimalization rule in Kleene’s computational model [Kleene ’36].
6.2. COMPLEX, RANDOMLY DEFINED CIRCUITS
x0
x1
x0
❄❄
f1
y1
y2
f1
x2
y1
y2
❄❄
fn/2+1
x2
❄❄
“0”
xn/2+1
❄❄
xn/2+2
❄❄
fn/2+1
xn/2+2
❄❄
f2
fn/2+2
fn/2+2
❄
❄
❄
x3
❄❄
xn/2+1
“1”
❄❄
❄❄
f2
x1
143
f3
y3
xn/2
❄
xn
❄❄
fn/2
xn
❄❄
❄❄
fn
fn
fn/2
❄❄
❄❄
✲
fn
yn
xn
❄
s
1 0
emux
yn
a.
❄
b.
Figure 6.38:
And so on until the two zero-variable functions: the value 0 and the value 1. The solution is an
n-level binary tree of EMUXs having applied to the last level zero-variable functions. Therefore,
solution is a M U Xn and a binary string applied on the 2n selected inputs. The binary sting has
n
the length 2n . Thus, for each of the 22 functions there is a distinct string defining it. ⋄
The universal circuit is indeed the best example of a big simple circuit, because it is described
by the following code:
module nU_circuit #(‘include "parameter.v")(
assign
endmodule
output
input
input
[(1’b1 << n)-1:0]
[n-1:0]
out ,
func,
in );
out = func[in];
The file parameter.v contains the value for n. But, attention! The size of the circuit is:
SnUc ircuit (n) ∈ O(2n ).
Thus, circuits are more powerful than Turing Machine (TM), because TM solve only problem
having the solution algorithmically expressed with a sequence of symbols that does not depend
by n. Beyond the Turing-computable function there are many functions for which the solution
is a family of circuits.
The solution imposed by the previous theorem is an universal circuit for computing the n
binary variable functions. Let us call it nU-circuit (see Figure 6.41). The size of this circuit is
Suniversal (n) ∈ O(2n ) and its complexity is Cuniversal (n) ∈ O(1). The functions is specified by
the “program” P = mp−1 , mp−2 , . . . m0 which is applied on the selected inputs of the n-input
multiplexer M U Xn . It is about a huge simple circuit. The functional complexity is associated
with the “program” P , which is a binary string.
This universal solution represents the strongest segregation between a simple physical structure - the n-input MUX - and a complex symbolic structure - the string of 2n bits applied on
the selected inputs which works like a “program”. Therefore, this is THE SOLUTION, MUX is
144
CHAPTER 6. GATES:
x0 x1
x0 x1
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
“1” x2
“0” x2
x3
x2
x0 x1
x2
x2
x3
x3
x3
x3
s
1 0
emux
s
a.
x0 x1
1 0
emux
c.
b.
x2 x3
x2 x3
x0 x1
x 2 x3
x 0 x1
x2 x3
XOR
✠
s
1 0
emux
s
1 0
emux
e.
d.
f.
Figure 6.39:
THE CIRCUIT and we can stop here our discussion about digital circuits!? ... But, no! There
are obvious reasons to continue our walk in the world of digital circuits:
• first: the exponential size of the resulting physical structure
• second: the huge size of the “programs” which are in a tremendous majority represented
by random, uncompressible, strings (hard or impossible to be specified).
The strongest segregation between simple and complex is not productive
in no-loop circuits. Both resulting structure, the simple circuit and the
complex binary string, grow exponentially.
6.2.2
Using the Universal circuit
We have a chance to use M U Xn to implement f (xn−1 , . . . , x0 ) only if one of the following
conditions is fulfilled:
1. n is small enough resulting realizable and efficient circuits
2. the “program” is a string with useful regularities (patterns) allowing strong minimization
of the resulting circuit
6.2. COMPLEX, RANDOMLY DEFINED CIRCUITS
145
xn−2 , . . . , x0
n−1
✲
✲
CLCg
❄
g(xn−2 , . . . , x0 )
CLCh
❄
h(xn−2 , . . . , x0 )
i0
i1
x0 emux
xn−1
y
✲
❄f (xn−1 , . . . , x0 )
Figure 6.40: The universal circuit. For any CLCf (n), where f (xn−1 , . . . , x0 ) this recursively
defined structure is a solution. EMUX behaves as an elementary universal circuit.
mp−1 mp−2 mp−3 mp−4 mp−5 mp−6 mp−7 mp−8
x0
✲
❄❄
1 0
emux
✲
✲
1 0
emux
x1
❄❄
✲
1 0
emux
❄❄
1 0
emux
✲
✲
1 0
emux
❄❄
x2
✲
m3 m2
❄❄
✲
1 0
emux
❄❄
m1 m0
❄❄
1 0
emux
✲
✲
1 0
emux
❄❄
1 0
emux
❄❄
❄❄
1 0
emux
✲
xn−1
❄❄
1 0
emux
nU-circuit
❄
f (xn−1 , xn−2 , . . . x0 )
Figure 6.41: The Universal Circuit as a tree of EMUXs. The depth of the circuit is equal
with the number, n, of binary input variables. The size of the circuit increases exponentially with n.
Follows few well selected examples. First is about an application with n enough small to provide
an useful circuit (it is used in Connection Machine as an “universal” circuit performing anyone
of the 3-input logic function [Hillis ’85]).
Example 6.10 Let be the following Verilog code:
module three_input_functions(
assign
endmodule
output
input
[7:0]
input
out = func[{in0, in1, in2}];
out
func
in0, in1, in2
,
,
);
The circuit three input functions can be programmed, using the 8-bit string func as “program”, to perform anyone 3-input Boolean function. It is obviously a M U X3 performing
out = f (in0, in1, in2)
where the function f is specified (“programmed”) by an 8-bit word (“program”). ⋄
146
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
The “programmable” circuit for any 4-input Boolean function is obviously M U X4 :
out = f (in0, in1, in2, in3)
where f is “programmed” by a 16-bit word applied on the selected inputs of the multiplexer.
The bad news is: we can not go to far on this way because the size of the resulting universal
circuits increases exponentially. The good news is: usually we do not need universal but particular solution. The circuits are, in most of cases, specific not universal. They “execute” a specific
“program”. But, when a specific binary word is applied on the selected inputs of a multiplexer,
the actual circuit is minimized using the following removing rules and reduction rules.
An EMUX defined by:
out = x ? in1 : in0;
can be removed, if on its selected inputs specific 2-bit binary words are applied, according to the
following rules:
• if {in1, in0} = 00 then out = 0
• if {in1, in0} = 01 then out = x′
• if {in1, in0} = 10 then out = x
• if {in1, in0} = 11 then out = 1
or, if the same variable, y, is applied on both selected inputs:
• if {in1, in0} = yy then out = y
An EMUX can be reduced, if on one its selected inputs a 1-bit binary word are applied, being
substituted with a simpler circuit according to the following rules:
• if {in1, in0} = y0 then out = xy
• if {in1, in0} = y1 then out = y + x′
• if {in1, in0} = 0y then out = x′ y
• if {in1, in0} = 1y then out = x + y
Results: the first level of 2n−1 EMUXs of a M U Xn is reduced, and on the inputs of the
second level (of nn−2 EMUXs) is applied a word containing binary values (0s and 1s) and binary
variables (x0 s and x′0 s). For the next levels the removing rules or the reducing rules are applied.
Example 6.11 Let us solve the problem of majority function for three Boolean variable. The
function maj(x2 , x1 , x0 ) returns 1 if the majority of inputs are 1, and 0 if not. In Figure 6.42
a “programmable” circuit is used to solve the problem.
Because we intend to use the circuit only for the function maj(x2 , x1 , x0 ) the first layer of
EMUXs can be removed resulting the circuit represented in Figure 6.43a.
On the resulting circuit the reduction rules are applied. The result is presented in Figure
6.43b. ⋄
The next examples refer to big n, but “program” containing repetitive patterns.
6.2. COMPLEX, RANDOMLY DEFINED CIRCUITS
the “program”
input variables
✙
✢
1 1
x0
✲
1 0
❄❄
1 0
emux
✲
✲
1 0
emux
x1
x2
147
1 0
❄❄
1 0
emux
✲
❄❄
1 0
emux
✲
✲
1 0
emux
❄❄
✲
0 0
❄❄
1 0
emux
❄❄
❄❄
1 0
emux
❄
m3 (x2 , x1 , x0 )
Figure 6.42: The majority function. The majority function for 3 binary variables is solved by a
3-level binary tree of EMUXs. The actual “program” applied on the “leafs” will allow to minimize the
tree.
Example 6.12 If the ”program” is 128-bit string i127 . . . i0 = (10)64 , it corresponds to a function
of form:
f (x6 , . . . , x0 )
where: the first bit, i0 is the value associated to the input configuration x6 , . . . , x0 = 0000000 and
the last bit i127 is the value associated to input configuration x6 , . . . , x0 = 1111111 (according
with the representation from Figure 6.11 which is equivalent with Figure 6.40).
The obvious regularities of the “program” leads our mind to see what happened with the
resulting tree of EMUXs. Indeed, the structure collapses under this specific “program”. The
upper layer of 64 EMUXs are selected by x0 and each have on their inputs i0 = 1 and i1 = 0,
generating x′0 on their outputs. Therefore, the second layer of EMUXs receive on all selected
inputs the value x′0 , and so on until the output generates x′0 . Therefore, the circuit performs the
function f (x0 ) = x′0 and the structure is reduced to a simple inverter.
In the same way the “program” (0011)32 programs the 7-input MUX to perform the function
f (x1 ) = x1 and the structure of EMUXs disappears.
For the function f (x1 , x0 ) = x1 x′0 the “program” is (0010)32 .
For a 7-input AND the“program” is 0127 1, and the tree of MUXs degenerates in 7 EMUXs
serially connected each having the input i0 connected to 0. Therefore, each EMUX become an
AN D2 and applying the associativity principle results an AN D7 .
In a similar way, the same structure become an OR7 if it is “programmed” with 01127 . ⋄
It is obvious that if the “program” has some uniformities, these uniformities allow to minimize
the size of the circuit in polynomial limits using removing and reduction rules. The simple
“programs” lead toward small circuits.
6.2.3
The many-output random circuit: Read Only Memory
The simple solution for the following many-output random circuits having the same inputs:
f (xn−1 , . . . x0 )
g(xn−1 , . . . x0 )
...
148
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
x0
x0
1
x1
✲
x2
0
❄❄
✲
1 0
emux
✲
x1
❄❄
1 0
emux
❄❄
1 0
emux
x2
❄
1 0
emux
❄
m3 (x2 , x1 , x0 )
a.
✲
m3 (x2 , x1 , x0 )
b.
Figure 6.43: The reduction process. a. For any function the first level of EMUSs is reduced to a
binary vector ((1, 0) in this example). b. For the actual “program” of the 3-input majority function the
second level is supplementary reduced to simple gates (an AN D2 and an OR2 ).
s(xn−1 , . . . x0 )
is to connect in parallel many one-output circuits. The inefficiency of the solution become
obvious when the structure of the MUX presented in Figure 6.9 is considered. Indeed, if we
implement many MUXs with the same selection inputs, then the decoder DCDn is replicated
many time. One DCD is enough for many MUXs if the structure from Figure 6.44a is adopted.
The DCD circuit is shared for implementing the functions f, g, . . . s. The shared DCD is used
to compute all possible minterms (see Appendix C.4) needed to compute an n-variable Boolean
function.
Figure 6.44b is an example of using the generic structure from Figure 6.44a to implement a
specific many-output function. Each output is defined by a different binary string. A 0 removes
the associated AND, connecting the corresponding OR input to 0, and an 1 connects to the
corresponding i-th input of each OR to the i-th DCD output. The equivalent resulting circuit is
represented in Figure 6.44c, where some OR inputs are connected to ground and other directly
to the DCD’s output. Therefore, we use a technology allowing us to make “programmable”
connections of some wires to other (each vertical line must be connected to one horizontal line).
The uniform structure is “programmed” with a more or less random distribution of connections.
If De Morgan transformation is applied, the circuit from Figure 6.44c is transformed in the
circuit represented in Figure 6.45a, where instead of an active high outputs DCD an active low
outputs DCD is considered and the OR gates are substituted with NAND gates. The DCD’s
outputs are generated using NAND gates to decode the input binary word, the same as the gates
used to encode the output binary word. Thus, a multi-output Boolean function works like a
trans-coder. A trans-coder works translating all the binary input words into output binary
words. The list of input words can by represented as an ordered list of sorted binary numbers
starting with 0 and ending with 2n − 1. The table from Figure 6.46 represents the truth table
for the multi-output function used to exemplify our approach. The left column contains all
binary numbers from 0 (on the first line) until 2n − 1 = 11 . . . 1 (on the last line). In the right
column the desired function is defined associating to each input an output. If the left column is
an ordered list, the right column has a more or less random content (preferably more random
for this type of solution).
The trans-coder circuit can be interpreted as a fix content memory. Indeed, it works like a
memory containing at the location 00...00 the word 11...0, ... at the location 11...10 the word
6.2. COMPLEX, RANDOMLY DEFINED CIRCUITS
fm−1 fm−2
f0
gm−1 gm−2
149
g0
sm−1 sm−2
s0
O0
DCD
Om−2
Om−1
n
✻
xn−1 , . . . x0
f (xn−1 , . . . x0 )
a.
0
1
1
g(xn−1 , . . . x0 )
1
0
1
s(xn−1 , . . . x0 )
1
0
0
O0
DCD
Om−2
Om−1
n
✻
xn−1 , . . . x0
b.
f (xn−1 , . . . x0 )
g(xn−1 , . . . x0 )
s(xn−1 , . . . x0 )
f (xn−1 , . . . x0 )
g(xn−1 , . . . x0 )
s(xn−1 , . . . x0 )
O0
DCD
Om−2
Om−1
✻
n
xn−1 , . . . x0
c.
Figure 6.44: Many-output random circuit. a. One DCD and many AND-OR circuits. b. An
example. c. The version using programmable connections.
10...0, and at the last location the word 01...1. The name of this kind of programmable device
is read only memory, ROM.
Example 6.13 The trans-coder from the binary coded decimal numbers to 7 segments display is
a combinational circuit with 4 inputs, a, b, c, d, and 7 outputs A, B, C, D, E, F, G, each associated
to one of the seven segments. Therefore we have to solve 7 functions of 4 variables (see the truth
table from Figure 6.48).
The Verilog code describing the circuit is:
module seven_segments( output reg [6:0]
out ,
input
[3:0]
in );
always @(in) case(in)
4’b0000: out = 7’b0000001;
4’b0001: out = 7’b1001111;
4’b0010: out = 7’b0010010;
150
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
VDD
′
O0
′
O0
DCD
DCD
′
Om−1
′
Om−1
′
Om−1
′
Om−1
✻
✻
xn−1 , . . . x0
xn−1 , . . . x0
a.
g
f
s
❄❄
f
g
❄
s
b.
Figure 6.45: The internal structure of a Read Only Memory used as trans-coder. a.
The internal structure. b. The simplified logic symbol where a thick vertical line is used to represent an
m-input NAND gate.
Input
00 ... 00
...
11 ... 10
11 ... 11
Output
11 ... 0
...
10 ... 0
01 ... 1
Figure 6.46: The truth table for a multi-output Boolean function. The input columns can
be seen as addresses, from 00 . . . 0 to 11 . . . 1 and the output columns as the content stored at the
corresponding addresses.
4’b0011:
4’b0100:
4’b0101:
4’b0110:
4’b0111:
4’b1000:
4’b1001:
default
endcase
out
out
out
out
out
out
out
out
=
=
=
=
=
=
=
=
7’b0000110;
7’b1001100;
7’b0100100;
7’b0100000;
7’b0001111;
7’b0000000;
7’b0000100;
7’bxxxxxxx;
endmodule
The first solution is a 16-location of 7-bit words ROM (see Figure 6.47a. If inverted outputs
are needed results the circuit from Figure 6.47b.
⋄
∗ Programmable Logic Array In the previous example each output of DCD compute the inverted
value of a minterm. But our applications do not need all the possible minterms, for two reasons:
• the function is not defined for all possible input binary configurations (only the input codes representing numbers from 0 to 9 define the output behavior of the circuit)
• in the version with inverted outputs the minterm corresponding for the input 1000 (the number 8)
is not used.
A more flexible solution is needed. ROM consists in two arrays of NANDs, one fix and another configurable (programmable). What if also the first array is configurable, i.e., the DCD circuit is programmed
to compute only some minterms? More, what if instead of computing only minterms (logic products
6.2. COMPLEX, RANDOMLY DEFINED CIRCUITS
151
0
0
DCD4
DCD4
15
x3 x2 x1 x0
✻✻✻✻
✻b c d
a
15
x 3 x2 x 1 x0
✻✻✻✻
❄❄❄❄❄❄
✻b c d A’❄
B’ C’D’ E’ F’ G’
❄❄❄❄❄❄❄
a
A B C D E F G
a.
b.
Figure 6.47: The CLC as trans-coder designed serially connecting a DCD with an
encoder. Example: BCD to 7-segment trans-coder. a. The solution for non-inverting functions. b.
The solution for inverting functions.
containing all the input variable) we compute also some, or only, terms (logic products containing only a
part of input variable)? As we know a term corresponds to a logic sum of minterms. Computing a term
in a programmable array of NANDs two or more NANDs with n inputs are substituted with a NAND
with n − 1 or less inputs. Applying these ideas results another frequently used programmable circuit: the
famous PLA (programmable logic array).
Example 6.14 Let’s revisit the problem of 7 segment trans-coder. The solution is to use a PLA. Because
now the minimal form of equations is important the version with inverted outputs is considered. Results
the circuit represented in Figure 6.49, where a similar convention for representing NANDs as a thick line
is used. ⋄
When PLA are used as hardware support the minimized form of Boolean functions (see Appendix C.4
for a short refresh) are needed. In the previous example for each inverted output its minimized Boolean
expression was computed.
The main effect of substituting, whenever is possible, ROMs with PLAs are:
• the number of decoder outputs decreases
• the size of circuits that implements the decoder decreases (some or all minterms are substituted
with less terms)
• the number of inputs of the NANds on the output level also decreases.
There are applications supposing ROMs with a very random content, so as the equivalent PLA has
the same ar almost the same size and the effort to translate the ROM into a PLA does not deserve. A
typical case is when we “store” into a ROM a program executed by a computing machine. No regularities
are expected in such an applications.
It is also surprising the efficiency of a PLA in solving pure Boolean problems which occur in current
digital designs. A standard PLA circuit, with 16 inputs, 8 outputs and only 48 programmable (min)terms
on the first decoding level, is able to solve a huge amount of pure logic (non-arithmetic) applications. A
full decoder in a ROM circuit computes 66536 minterms, and the previous PLA is designed to support
no more than 48 terms!
Warning! For arithmetic applications PLA are extremely inefficient. Short explanation: a 16-input
XOR supposes 32768 minterms to be implemented, but a 16-input AND can be computed using one
minterm. The behavioral diversity to the output of an adder is similar with the behavioral diversity on
its inputs. But the diversity on the output of 8-input NAND is almost null. The probability of 1 on the
output of 16-input AND is 1/216 = 0.000015.
152
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
abcd
0000
0001
0010
0011
0100
0101
0110
0111
1000
1001
1010
....
1111
ABCDEFG
1111110
0110000
1101101
1111001
0110011
1011011
1011111
1110000
1111111
1111011
------.......
-------
Figure 6.48: The truth table for the 7 segment trans-coder. Each binary represented decimal
(in the left columns of inputs) has associated a 7-bit command (in the right columns of outputs) for the
segments used for display. For unused input codes the output is “don’t care”.
a
b
c
d
✲ A’
✲ B’
✲ C’
✲ D’
✲ E’
✲ F’
✲ G’
bc’d
bc’d’
d
bcd’
a’b’c’d
bcd
cd
b’cd’
a’b’c’
Figure 6.49: Programmable Logic Array (PLA). Example: BCD to 7-segment trans-coder.
Both, decoder and encoders are programmable structures.
6.3
Concluding about combinational circuits
The goal of this chapter was to introduce the main type of combinational circuits. Each presented
circuit is important first, for its specific function and second, as a suggestion for how to build
similar ones. There are a lot of important circuits undiscussed in this chapter. Some of them
are introduced as problems at the end of this chapter.
Simple circuits vs. complex circuits Two very distinct class of combinational circuits
are emphasized. The first contains simple circuits, the second contains complex circuits. The
complexity of a circuit is distinct from the size of a circuit. Complexity of a circuit is given by
the size of the definition used to specify that circuit. Simple circuits can achieve big sizes because
they are defined using a repetitive pattern. A complex circuit can not be very big because its
definition is dimensioned related with its size.
Simple circuits have recursive definitions Each simple circuit is defined initially as an
elementary module performing the needed function on the smallest input. Follows a recursive
6.4. PROBLEMS
153
definition about how can be used the elementary circuit to define a circuit working for any
input dimension. Therefore, any big simple circuit is a network of elementary modules which
expands according to a specific rule. Unfortunately, the actual HDL, Verilog included, are not
able to manage without (strong) restrictions recursive definitions neither in simulation nor in
synthesis. The recursiveness is a property of simple circuits to be fully used only for our mental
experiences.
Speeding circuits means increase their size Depth and size evolve in opposite directions.
If the speed increases, the pay is done in size, which also increases. We agree to pay, but in
digital systems the pay is not fair. We conjecture the bigger is performance the bigger is the
unit price. Therefore, the pay increases more than the units we buy. It is like paying urgency
tax. If the speed increases n times, then the size of the circuit increases more than n times,
which is not fair but it is real life and we must obey.
Big sized complex circuits require programmable circuits There are software tolls for
simulating and synthesizing complex circuits, but the control on what they generate is very low.
A higher level of control we have using programmable circuits such as ROMs or PLA. PLA are
efficient only if non-arithmetic functions are implemented. For arithmetic functions there are a
lot of simple circuits to be used. ROM are efficient only if the randomness of the function is
very high.
Circuits represent a strong but ineffective computational model Combinational circuits represent a theoretical solution for any Boolean function, but not an effective one. Circuits
can do more than algorithms can describe. The price for their universal completeness is their
ineffectiveness. In the general case, both the needed physical structure (a tree of EMUXs) and
the symbolic specification (a binary string) increase exponentially with n (the number of binary
input variables). More, in the general case only a family of circuits represents the solution.
To provide an effective computational tool new features must be added to a digital machine and some restrictions must be imposed on what is to be computable. The next chapters
will propose improvements induced by successively closing appropriate loops inside the digital
systems.
6.4
Problems
Gates
Problem 6.1 Determine the relation between the total number, N , of n-input m-output Boolean
functions (f : {0, 1}n → {0, 1}m ) and the numbers n and m.
Problem 6.2 Let be a circuit implemented using 32 3-input AND gates. Using the appendix
evaluate the area if 3-input gates are used and compare with a solution using 2-input gates.
Analyze two cases: (1) the fan-out of each gate is 1, (2) the fan-out of each gate is 4.
Decoders
Problem 6.3 Draw DCD4 according to Definition 2.9. Evaluate the area of the circuit, using
the cell library from Appendis E, with the placement efficiency8 70%. Estimate the maximum
8
For various reason the area used to place gates on Silicon can not completely used. Some unused spaces
remain between gates. Area efficiency measures the degree of area use.
154
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
propagation time. The wires are considered enough short to be ignored their contribution in
delaying signals.
Problem 6.4 Design a constant depth DCD4 . Draw it. Evaluate the area and the maximum
propagation time using the cell library from Appendix E. Compare the results with the results of
the previous problem.
Problem 6.5 Propose a recursive definition for DCDn using EDMUXs. Evaluate the size and
the depth of the resulting structure.
Multiplexors
Problem 6.6 Draw M U X4 using EMUX. Make the structural Verilog design for the resulting
circuit. Organize the Verilog modules as hierarchical as possible. Design a tester and use it to
test the circuit.
Problem 6.7 Define the 2-input XOR circuit using an EDCD and an EMUX.
Problem 6.8 Make the Verilog behavioral description for a constant depth left shifter by maximum m − 1 positions for m-bit numbers, where m = 2n . The “header” of the project is:
module left_shift( output
input
input
...
endmodule
[2m-2:0]
[m-1:0]
[n-1:0]
out
in
shift
,
,
);
Problem 6.9 Make the Verilog structural description of a log-depth (the depth is log2 16 = 4)
left shifter by 16 positions for 16-bit numbers. Draw the resulting circuit. Estimate the size and
the depth comparing the results with a similar shifter designed using the solution of the previous
problem.
Problem 6.10 Draw the circuit described by the Verilog module leftRotate in the subsection
Shifters.
Problem 6.11 A barrel shifter for m-bit numbers is a circuit which rotate the bits the input
word a number of positions indicated by the shift code. The “header” of the project is:
module barrel_shift(
output
input
input
[m-1:0] out
[m-1:0] in
[n-1:0] shift
,
,
);
...
endmodule
Write a behavioral code and a minimal structural version in Verilog.
Prefix network
Problem 6.12 A prefix network for a certain associative function f ,
Pf (x0 , x1 , . . . xn−1 ) = {y0 , y1 , . . . yn−1 }
receives n inputs and generate n outputs defined as follows:
y0 = f (x0 )
y1 = f (x0 , x1 )
6.4. PROBLEMS
155
y2 = f (x0 , x1 , x2 )
...
yn−1 = f (x0 , x1 , . . . xn−1 ).
Design the circuit POR (n) for n = 16 in two versions: (1) with the smallest size, (2) with
the smallest depth.
Problem 6.13 Design POR (n) for n = 8 and the best product size × depth.
Problem 6.14 Design Paddition (n) for n = 4. The inputs are 8-bit numbers. The addition is a
mod256 addition.
Recursive circuits
Problem 6.15 A comparator is circuit designed to compare two n-bit positive integers. Its
definition is:
module comparator( input
input
output
output
output
...
endmodule
[n-1:0] in1 ,
[n-1:0] in2 ,
eq ,
lt ,
gt );
//
//
//
//
//
first operand
second operand
in1 = in2
in1 < in2
in1 > in2
1. write the behavioral description in Verilog
2. write a structural description optimized for size
3. design a tester which compare the results of the simulations of the two descriptions: the
behavioral description and the structural description
4. design a version optimized for depth
5. define an expandable structure to be used in designing comparators for bigger numbers in
two versions: (1) optimized for depth, (2) optimized for size.
Problem 6.16 Design a comparator for signed integers in two versions: (1) for negative numbers represented in 2s complement, (2) for negative numbers represented a sign and number.
Problem 6.17 Design an expandable priority encoder with minimal size starting from an elementary priority encoder, EPE, defined for n = 2. Evaluate its depth.
Problem 6.18 Design an expandable priority encoder, P E(n), with minimal depth.
Problem 6.19 What is the numerical function executed by a priority encoder circuit if the input
is interpreted as an n-bit integer, the output is an m-bit integer and n ones is a specific warning
signal?
Problem 6.20 Design the Verilog structural descriptions for an 8-input adder in two versions:
(1) using 8 FAs and a ripple carry connection, (2) using 8 HAs and a carry look ahead circuit.
Evaluate both solutions using the cell library from Appendix E.
Problem 6.21 Design an expandable carry look-ahead adder starting from an elementary circuit.
156
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Problem 6.22 Design an enabled incrementer/decrementer circuit for n-bit numbers. If en =
1, then the circuit increments the input value if inc = 1 or decrements the input value if inc = 0,
else, if en = 0, the output value is equal with the input value.
Problem 6.23 Design an expandable adder/subtracter circuit for 16-bit numbers. The circuit
has a carry input and a carry output to allow expandability. The 1-bit command input is sub. For
sub = 0 the circuit performs addition, else it subtracts. Evaluate the area and the propagation
time of the resulting circuit using the cell library from Appendix E.
Problem 6.24 Provide a “divide et impera” solution for the circuit performing firstDetect
function.
Random circuits
Problem 6.25 The Gray counting means to count, starting from 0, so as at each step only one
bit is changed. Example: the three-bit counting means 000, 001, 011, 010, 110, 111, 101, 100,
000, ... Design a circuit to convert the binary counting into the Gray counting for 8-bit numbers.
Problem 6.26 Design a converter from Gray counting to binary counting for n-bit numbers.
Problem 6.27 Write a Verilog structural description for ALU described in Example 2.3. Identify the longest path in the resulting circuit. Draw the circuit for n = 8.
Problem 6.28 Design a 8-bit combinational multiplier for a7 , . . . a0 and b7 , . . . b0 , using as basic
“brick” the following elementary multiplier, containing a FA and an AND:
module em(carry_out, sum_out, a, b, carry, sum);
input
a, b, carry, sum;
output carry_out, sum_out;
assign {carry_out, sum_out} = (a & b) + sum + carry;
endmodule
Problem 6.29 Design an adder for 32 1-bit numbers using the carry save adder approach.
Hint: instead of using the direct solution of a binary tree of adders a more efficient way (from
the point of view of both size and depth) is to use circuits to “compact” the numbers. The
first step is presented in Figure 6.50, where 4 1-bit numbers are transformed in two numbers,
a 1-bit number and a 2-bit number. The process is similar in Figure 6.51 where 4 numbers, 2
1-bit numbers and 2 2-bit numbers are compacted as 2 numbers, one 2-bit number and one 3-bit
number. The result is a smaller and a faster circuit than a circuit realized using adders.
in[3:0]
❄❄❄
3
❄❄❄❄
210
FA
3210
csa4
xx
0x
❄❄ ❄❄
a.
❄❄ ❄❄
out1[1:0]out0[1:0]
b.
Figure 6.50: 4-bit compacter.
Compare the size and depth of the resulting circuit with a version using adders.
6.4. PROBLEMS
157
4
❄
csa4
0x xx
4
7 654
3 210
FA
FA
HA
FA
❄
csa4
0x xx
HA
✲
0 (3)
✲
x
✲
✲ x0 csa4 ✰
0 0 xx
✲
x (3)
✲
x
✲
✲ xx csa4
0x xx
❄
out0
3
a.
xxx
❄
out0
3
0xx
out1
❄
❄
out1
b.
Figure 6.51: 8-bit compacter.
Problem 6.30 Design in Verilog the behavioral and the structural description of a multiply
and accumulate circuit, MACC, performing the function: (a × b) + c, where a and b are 16-bit
numbers and c is a 24-bit number.
Problem 6.31 Design the combinational circuit for computing
c=
7
∑
ai × bi
i=0
where: ai , bi are 16-bit numbers. Optimize the size and the depth of the 8-number adder using a
technique learned in one of the previous problem.
Problem 6.32 Exemplify the serial composition, the parallel composition and the serial-parallel
composition in 0 order systems.
Problem 6.33 Write the logic equations for the BCD to 7-segment trans-coder circuit in both
high active outputs version and low active outputs version. Minimize each of them individually.
Minimize all of them globally.
Problem 6.34 Applying removing rules and reduction rules find the functions performed by
5-level universal circuit programmed by the following binary strings:
1. (0100)8
2. (01000010)4
3. (0100001011001010)2
4. 024 (01000010)
5. 00000001001001001111000011000011
Problem 6.35 Compute the biggest size and the biggest depth of an n-input, 1-output circuit
implemented using the universal circuit.
Problem 6.36 Provide the prof for Zero-One Principle.
158
6.5
CHAPTER 6. GATES:
ZERO ORDER, NO-LOOP DIGITAL SYSTEMS
Projects
Project 6.1 Finalize Project 1.1 using the knowledge acquired about the combinational structures in this chapter.
Project 6.2 Design a combinational floating point single precision (32 bit) multiplier according
to the ANSI/IEEE Standard 754-1985, Standard for Binary Floating Point Arithmetic.
Chapter 7
MEMORIES:
First order, 1-loop digital systems
In the previous chapter
the main combinational, no-loop circuits were presented with emphasis on
• the simple, pattern-based basic combinational circuits performing functions like: decode using demultiplexors, selection using multiplexors, increment, add, various selectable functions using arithmetic
and logic units, compare, shift, priority encoding, ...
• the difference between the simple circuits, which grow according to recursive rules, and the complex,
pattern-less circuits whose complexity must be kept at lowest possible level
• the compromise between area and speed, i.e., how to save area accepting to give up the speed, or
how can be increased the speed accepting to enlarge the circuit’s area.
In this chapter
the first order, one-loop circuits are introduced studying
• how to close the first loop inside a combinational circuit in order to obtain a stable
and useful behavior
• the elementary storage support – the latch – and the way to expand it using the
serial, parallel, and serial-parallel compositions leading to the basic memory circuits,
such as: the master-slave flip-flop, the random access memory and the register
• how to use first order circuits to design basic circuits for real applications, such as
register files, content addressable memories, associative memories or systolic systems.
In the next chapter
the second order, automata circuits are described. While the first order circuits have the smallest degree of
autonomy – they are able only to maintain a certain state – the second order circuits have an autonomous
behavior induced by the loop just added. The following circuits will be described:
• the simplest and smallest elementary, two-state automata: the T flip-flop and JK flip-flop, which
besides the storing function allow an autonomous behavior under a less restrictive external command
• simple automata performing recursive functions, generated by expanding the function of the simplest
two-state automata (example: n-bit counters)
• the complex, finite automata used for control or for recognition and generation of regular streams of
symbols.
159
160
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
The magic images were placed on the wheel of the memory system to which correspondent other wheels on which
were remembered all the physical contents of the terrestrial
world – elements, stones, metals, herbs, and plants, animals, birds, and so on – and the whole sum of the human
knowledge accumulated through the centuries through the
images of one hundred and fifty great men and inventors.
The possessor of this system thus rose above time and reflected the whole universe of nature and of man in his mind.
Frances A. Yates1
A true memory is an associative one. Please do not confuse
the physical support – the random access memory – with the
function – the associative memory.
According to the mechanisms described in Chapter 3 of this book, the step toward a new
class of circuits means to close a new loop. This will be the first loop which closed over the
combinational circuits already presented. Thus, a first degree of autonomy will be reached in
digital systems: the autonomy of the state of the circuit. Indeed, the state of the circuit will be
partially independent by the input signals, i.e., the output of the circuits do not depend on or
not respond to certain input switching.
In this chapter we introduce some of the most important circuits used for building digital
systems. The basic function in which they are involved is the memory function. Some events
on the input of a memory circuit are significant for the state of the circuits and some are not.
Thus, the circuit “memorizes”, by the state it reaches, the significant events and “ignores” the
rest. The possibility to have an “attitude” against the input signals is given to the circuit by the
autonomy induced by its internal loop. In fact, this first loop closed over a simple combinational
circuit makes insignificant some input signals because the circuit is able to compensate their
effect using the signals received back from its output.
The main circuits with one internal loop are:
• the elementary latch - the basic circuit in 1-OS, containing two appropriately loopcoupled gates; the circuit has two stable states being able to store 1 bit of information
• the clocked latch - the first digital circuit which accepts the clock signal as an input
distinct from data inputs; the clock signal determines by its active level when the latch is
triggered, while the data input determines how the latch switches
• the master-slave flip-flop - the serial composition in 1-OS, built by two clocked latches
serially connected; results a circuit triggered by the active transition of clock
• the random access memory (RAM) - the parallel composition in 1-OS, containing a
set of n clocked elementary latches accessed with a DM U Xlog2 n and a M U Xlog2 n
• the register - the serial-parallel composition in 1-OS, made by parallel connecting masterslave flip-flops.
These first order circuits don’t have a direct computational functionality, but are involved
in supporting the following main processes in a computational machine:
1
She was Reader in the History of the Renaissance at the University of London. The quote is from Giordano
Bruno and the Hermetic Tradition. Her other books include The Art of Memory.
7.1. STABLE/UNSTABLE LOOPS
161
• offer the storage support for implementing various memory functions (register files, stacks,
queues, content addressable memories, associative memories, ...)
• are used for synchronizing different subsystems in a complex system (supports the pipeline
mechanism, implements delay lines, stores the state of automata circuits).
7.1
Stable/Unstable Loops
There are two main types of loops closed over a combinational logic circuit: loops generating a
stable behavior and loops generating an unstable behavior. We are interested in the first kind
of loop that generates a stable state inside the circuit. The other loop cannot be used to build
anything useful for computational purposes, except some low performance signal generators.
The distinction between the two types of loops is easy exemplified closing loops over the
simplest circuit presented in the previous chapter, the elementary decoder (see Figure 7.1a).
The unstable loop is closed connecting the output y0 of the elementary decoder to its input
x0 (see Figure 7.1b). Suppose that y0 = 0 = x0. After the time interval equal with tpLH 2 the
output y0 becomes 1. After another time interval equal with tpHL the output y0 becomes again
0. And so on, the two outputs of the decoder are unstable oscillating between 0 and 1 with a
period of time Tosc = tpLH + tpHL , or the frequency fosc = 1/(tpLH + tpHL ).
y1
x0
y0
a.
y1
x0
out1
y1
x0
EDCD
EDCD
y0
b.
y0
out2
c.
Figure 7.1: The two loops closed over an elementary decoder. a. The simplest combinational circuit: the one-input, elementary decoder. b. The unstable, inverting loop containing one (odd)
inverting logic level(s). c. The stable, non-inverting loop containing two (even) inverting levels.
The stable loop is obtained connecting the output y1 of the elementary decoder to the input
x0 (see Figure 7.1c). If y1 = 0 = x0, then y0 = 1 fixing again the value 0 to the output y1.
If y1 = 1 = x0, then y0 = 0 fixing again the value 1 to the output y1. Therefore, the circuit
has two stable states. (For the moment we don’t know how to switch from one state to another
state, because the circuit has no input to command the switching from 0 to 1 or conversely. The
solution comes soon.)
What is the main structural distinction between the two loops?
• The unstable loop has an odd number of inverting levels, thus the signal comes back to
the output having the complementary value.
• The stable loop has an even number of inverting levels, thus the signal comes back to the
output having the same value.
2
the propagation time through the inverter when the output switches from the low logic level to the high level.
162
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
Example 7.1 Let be the circuit from Figure 7.2a, with 3 inverting levels on its internal loop.
If the command input C is 0, then the loop is “opened”, i.e., the flow of the signal through the
circular way is interrupted. If C switches in 1, then the behavior of the circuit is described by the
wave forms represented in Figure 7.2b. The circuit generates a periodic signal with the period
Tosc = 3(tpLH + tpHL ) and frequency fosc = 1/3(tpLH + tpHL ). (To keep the example simple we
consider that tpLH and tpHL have the same value for the three circuits.)⋄
C
✻
C
✲
t
l1
✻
l1
l2
✲ ✛ tpHL
✻
out
✲ ✛
✻
l2
out
a.
✲
t
✲
t
tpLH
✲
b.
t
Figure 7.2: The unstable loop. The circuit version used for a low-cost and low-performance clock
generator. a. The circuit with a three (odd) inverting circuits loop coupled. b. The wave forms drawn
takeing into account the propagation times associated to the low-high transitions (tpLH ) and to the
high-low transitions (tpHL ).
In order to be useful in digital applications, a loop closed over a combinational logic circuit
must contain an even number of inverting levels for all binary combinations applied to its inputs.
Else, for certain or for all input binary configurations, the circuit becomes unstable, unuseful
for implementing computational functions. In the following, only even (in most of cases two)
number of inverting levels are used for building the circuits belonging to 1-OS.
7.2
Elementary Structures
7.3
The Serial Composition: the Edge Triggered Flip-Flop
The first composition in 1-order systems is the serial composition, represented mainly by:
• the master-slave structure as the main mechanism that avoids the transparency of the
storage structures
• the delay flip-flop, the basic storage circuit that allows to close the second loop in the
synchronous digital systems
• the serial register, the fist big and simple memory circuit having a recursive definition.
This class of circuits allows us to design synchronous digital systems. Starting from this
point the inputs in a digital system are divided in two categories:
• clock inputs for synchronizing different parts of a digital system
• data and control inputs that receive the “informational” flow inside a digital system.
7.4. THE PARALLEL COMPOSITION: THE RANDOM ACCESS MEMORY
7.3.1
163
The Serial Register
Starting from the delay function of the last presented circuit (see Figure 2.14) a very important
function and the associated structure can be defined: the serial register. It is very easy to give
a recursive definition to this simple circuit.
Definition 7.1 An n-bit serial register, SRn , is made by serially connecting a D flip-flop with
an SRn−1 . SR1 is a D flip-flop. ⋄
In Figure 7.3 is shown a SRn . It is obvious that SRn introduces a n clock cycle delay between
its input and its output. The current application is for building digital controlled “delay lines”.
IN
✲D
Q
DF-F
Q’
✲
S
RSF-F
✲R
CK
Q
Q’
✲
✲
...
✲
S
Q
✲
OUT
RSF-F
...
✲R
Q’
...
Figure 7.3: The n-bit serial register (SRn ). Triggered by the active edge of the clock, the content
of each RSF-F is loaded with the content of the previous RSF-F.
We hope that now it is very clear what is the role of the master-slave structure. Let us
imagine a “serial register built with D latches”! The transparency of each element generates
the strange situation in which at each clock cycle the input is loaded in a number of latches
that depends by the length of the active level of the clock signal and by the propagation time
through each latch. Results an uncontrolled system, useless for any application. Therefore, for
controlling the propagation with the clock signal we must use the master-slave, non-transparent
structure of D flip-flop that switches on the positive or negative edge of clock.
VeriSim 7.1 The functional description currently used for an n-bit serial register active on the
positive edge of clock is:
module serial_register #(parameter
reg[0:n-1]
n = 1024)(output out
,
input in, enable, clock);
serial_reg;
assign out = serial_reg[n-1];
always @(posedge clock)
if (enable) serial_reg = {in, serial_reg[0:n-2]};
endmodule
⋄
7.4
The Parallel Composition: the Random Access Memory
The parallel composition in 1-OS provides the random access memory (RAM), which is the main
storage support in digital systems. Both, data and programs are stored on this physical support
in different forms. Usually we call these circuits improperly memories, even if the memory
function is something more complex, which suppose besides a storage device a specific access
mechanism for the stored information. A true memory is, for example, an associative memory
(see the next subchapters about applications), or a stack memory (see next chapter).
This subchapter introduces two structures:
164
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
• a trivial composition, but a very useful circuit: the n-bit latch
• the asynchronous random access memory (RAM),
both involved in building big but simple recursive structures.
7.4.1
The n-Bit Latch
The n-bit latch, Ln , is made by parallel connecting n data latches clocked by the same CK. The
system has n inputs and n outputs and stores an n-bit word. Ln is a transparent structure on
the active level of the CK signal. The n-bit latch must be distinguished by the n-bit register
(see the next section) that switches on the edge of the clock. In a synchronous digital system is
forbidden to close a combinational loop over Ln .
VeriSim 7.2 A 16-bit latch is described in Verilog as follows:
module
n_latch #(parameter n = 16)(output reg [n-1:0] out
input
[n-1:0] in
input
clock
always @(in or clock)
if (clock == 1)
// the active-high clock version
//if (clock == 0)
// the active-low clock version
out = in;
endmodule
,
,
);
⋄
The n-bit latch works like a memory, storing n bits. The only deficiency of this circuit is
due to the access mechanism. We must control the value applied on all n inputs when the latch
changes its content. More, we can not use selectively the content of the latch. The two problems
are solved adding some combinational circuits to limit both the changes and the use of the stored
bits.
7.4.2
Asynchronous Random Access Memory
Adding combinational circuits for accessing in a more flexible way an m-bit latch for write and
read operations, results one of the most important circuits in digital systems: the random
access memory. This circuit is the biggest and simplest digital circuit. And we can say it can
be the biggest because it is the simplest.
Definition 7.2 The m-bit random access memory, RAMm , is a linear collection of m D (data)
latches parallel connected, with the 1-bit common data inputs, DIN. Each latch receives the clock
signal distributed by a DM U Xlog2 m . Each latch is accessed for reading through a M U Xlog2 m .
The selection code is common for DMUX and MUX and is represented by the p-bit address code:
Ap−1 , . . . , A0 , where p = log2 m. ⋄
The logic diagram associated with the previous definition is shown in Figure 7.4. Because
no one of the input signal is clock related, this version of RAM is considered an asynchronous
one. The signal W E ′ is the low-active write enable signal. For W E ′ = 0 the write operation is
performed in the memory cell selected by the address An−1 , . . . , A0 .3 The wave forme describing
the relation between the input and output signals of a RAM are represented in Figure 7.5, where
the main time restrictions are the followings:
3
The actual implementation of this system uses optimized circuits for each 1-bit storage element and for the
access circuits. See Appendix C for more details.)
7.4. THE PARALLEL COMPOSITION: THE RANDOM ACCESS MEMORY
✲
WE’
165
′
Om−1
E’
.
DM U Xp
.
.
′
O1
′
O0
Ap−1 . . . A0
✻
i0
i1
✛
CK
D
DL
❄
DOUT
❄ ❄
M U Xp
Q
❄ ❄
CK
...
D
DL
Q
DIN
❄ ❄
CK
...
D
DL
Q
✛
✛
.
.
.
im−1
✛
Figure 7.4: The principle of the random access memory (RAM). The clock is distributed
by a DMUX to one of m = 2p DLs, and the data is selected by a MUX from one of the m DLs. Both,
DMUX and MUX use as selection code a p-bit address. The one-bit data DIN can be stored in the
clocked DL.
• tACC : access time - the propagation time from address input to data output when the
read operation is performed; it is defined as a minimal value
• tW : write signal width - the length of active level of the write enable signal; it is defined
as the shortest time interval for a secure writing
• tASU : address set-up time related to the occurrence of the write enable signal; it is defined
as a minimal value for avoiding to disturb the content of other than the storing cell selected
by the current address applied on the address inputs
• tAH : address hold time related to the end transition of the write enable signal; it is defined
as a minimal value for similar reasons
• tDSU : data set-up time related to the end transition of the write enable signal; it is defined
as a minimal value that ensure a proper writing
• tDH : data hold time related to the end transition of the write enable signal; it is defined
as a minimal value for similar reasons.
The just described version of a RAM represents only the asynchronous core of a memory
subsystem, which must have a synchronous behavior in order to be easy integrated in a robust
design. In Figure 7.4 there is no clock signal applied to the inputs of the RAM. In order to
synchronize the behavior of this circuit with the external world, additional circuits must be
added (see the first application in the next subchapter: Synchronous RAM).
The actual organization of an asynchronous RAM is more elaborated in order to provide the
storage support for a big number of m-bit words.
VeriSim 7.3 The functional description of a asynchronous n = 2p m-bit words RAM follows:
module ram(input
input
input
[m-1:0]
[p-1:0]
din ,
addr,
we ,
// data input
// address
// write enable
166
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
An−1 . . . A0
✻
a1
a2
✲
a3
t
✛ ✲
✛ ✲✛✲
tASU
W E′
✻
tDSU
tDH
✲
✛
tW
tAH
✲✛
✲
t
DIN
✻
data in
✲
t
DOU T
✻
tACC
✛
✲
✛
✲
d(a1)
d(a2)
d(a3)
✛✲
data in
✲
t
Figure 7.5: Read and write cycles for an asynchronous RAM. Reading is a combinational
process of selecting. The access time, tACC , is given by the propagation through a big MUX. The write
enable signal must be strictly included in the time interval when the address is stable (see tASU and tAH ).
Data must be stable related to the positive transition of W E ′ (see tDSU and tDH ).
reg
output [m-1:0]
dout); // data out
[m-1:0]
mem[(1’b1<<p)-1:0];
// the memory
assign
dout = mem[addr];
// reading memory
always @(din or addr or we) if (we) mem[addr] = din;
endmodule
// writing memory
⋄
The real structural version of the storage array will be presented in two stages. First the
number of bits per word will be expanded, then the e solution for a big number of words number
of words will be presented.
Expanding the number of bits per word
The pure logic description offered in Figure 7.4 must be reconsidered in order (1) to optimize it
and (2) to show how the principle it describe can be used for designing a many-bit word RAM.
The circuit structure from Figure 7.6 represents the m-bit word RAM. The circuit is organized
in m columns, one for each bit of the m-bit word. The DMUX structure is shared all by the m
columns, while each column has it own MUX structure. Let us remember that both, the DMUX
and MUX circuits are structured around a DCD. See Figure 6.6 and 6.9, where the first level
in both circuits is a decoder, followed by a linear network of 2-input ANDs for DMUX, and by
an AND-OR circuit for MUX. Then, only one decoder, DCDp , must be provided for the entire
memory. It is shared by the demultiplexing function and by the m multiplexors. Indeed, the
outputs of the decoder, LINEn−1 , ... LINE1 , LINE0 , are used to drive:
• one AND2 gate associate cu each line in the array, whose output clocks the DL latches
associated to one word; with these gates the decoder forms the demultimplexing circuit
used to clock, when WE = 1, the latches selected (addressed) by the current value of the
address: Ap−1 , . . . A0
7.4. THE PARALLEL COMPOSITION: THE RANDOM ACCESS MEMORY
167
• m AND2 gates, one in each column, selecting the read word to be ORed to the outputs
DOUTm−1 , DOUTm−2 , ... DOUT0 ; with the AND-OR circuit from each COLUMN the
decoder forms the multiplexor circuit associated to each output bit of the memory.
The array of lathes is organized in n and m columns. Each line is driven for write by the
output of a demultiplexer, while for the read function the addressed line (word) is selected by
the output of a decoder. The output value is gathered from the array using m multiplexors.
The reading process is a pure combinational one, while the writing mechanism is an asynchronous sequential one. The relation between the WE signal and the address bits is very
sensitive. Due to the combinational hazard to the output of DCD, the WE’ signal must be
activated only when the DCD’s outputs are stabilized to the final value, i.e., tASU before the
fall edge of WE’ or tH after the rise edge of WE’.
On−1
m-COLUMN BLOCK
LINEn−1
CK
DL
Q
D
COLU M Nm−1
DCDp
O1
O0
ADDR
COLU M Nm−2
COLU M N0
LINE1
CK
DL
Q
D
LINE0
✻
CK
DL
Q
D
Ap−1 . . . A0
...
✻
WE’
DINm−1
DOU Tm−1
DINm−2 DOU Tm−2
✻
DIN0
DOU T0
Figure 7.6: The asynchronous m-bit word RAM. Expanding the number of bits per word means
to connect in parallel one-bit word memories which share the same decoder. Each COLUMN contains
the storing latches and the AND-OR circuits for one bit.
Expanding the number of words by two dimension addressing
The factor form on silicon of the memory described in Figure 7.6 is very unbalanced for n >>>
m. Expanding the number of words for the a RAM in the previous, one block version is not
efficient because request a complex lay-out involving very long wires. We are looking for a more
“squarish” version of the lay-out for a big memory. The solution is to connect in parallel many
m-column blocks, thus defining a many-word from which to select one word using another level
168
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
of multiplexing. The reading process selects the many-word containing the requested word from
which the requested word is selected.
The internal organization of memory is now a two dimension array of rows and columns.
Each row contains a many-word of 2q words. Each column contains a number of 2r words. The
memory is addressed using the (p = r + q)-bit address:
addr[p-1:0] = {rowAddr[r-1:0], colAddr[q-1:0]}
The row address rowAddr[r-1:0] selects a many-word, while from the selected many-word,
the column address colAddr[q-1:0] selects the word addressed by the address addr[p-1:0].
Playing with the values of r and q an appropriate lay-out of the memory array can be designed.
In Figure 7.7 the block schematic for the resulting memory is presented. The second decoder
– COLUMN DECODE – selects from the s m-bit words provided by the s COLUMN BLOCKs
the word addressed by addr[p-1:0].
While the size decoder for a one block memory version is in the same order with the number
of words (SDCDp ∈ 2p ), the sum of the sizes of the two decoders in the two dimension version
is much smaller, because usually 2p >> 2r + 2q , for p = r + q. Thus, the area of the memory
circuit is dominated only by the storage elements.
The second level of selection is based also on a shared decoder – COLUMN DECODER. It
forms, with the s two-input ANDs a DM U Xq – the q-input DMUX in Figure 7.7 – which
distributes the write enable signals, we, to the selected m-column block. The same decoder is
shared by the m s-input MUXs used to select the output word from the many-word selected by
ROW DECODE.
The well known principle of ”divide et impera” (divide and conquer) is applied when the
address is divided in two parts, one for selecting a row and another for selecting a column. The
access circuits is thus minimized.
Unfortunately, RAM has not the function of memorizing. It is only a storage support.
Indeed, if we want to “memorize” the number 13, for example, we must store it to the address
131313, for example, and to keep in mind (to memorize) the value 131313, the place where the
number is stored. And than, what’s the help provided us by a the famous RAM memory? No
one. Because RAM is not a memory, it becomes a memory only if the associated processor runs
an appropriate procedure which allows us to forget about the address 131313. Another solution
is provided by additional circuits used to improve the functionality (see the subsection about
Associative Memories.)
7.5
Applications
Composing basic memory circuits with combinational structures result typical system configurations or typical functions to be used in structuring digital machines. The pipeline connection, for
example, is a system configuration for speeding up a digital system using a sort of parallelism.
This mechanism is already described in the subsections 2.5.1 Pipelined connections, and 3.3.2
Pipeline structures. Few other applications of the circuits belonging to 1-OS are described in this
section. The first is a frequent application of 1-OS: the synchronous memory, obtained adding
clock triggered structures to an asynchronous memory. The next is the file register – a typical
storage subsystem used in the kernel of the almost all computational structures. The basic
building block in one of the most popular digital device, the Field Programmable Gate Array,
is also SRAM based structure. Follows the content addressable memory which is a hardware
mechanism useful in controlling complex digital systems or for designing genuine memory
structures: the associative memories.
7.5. APPLICATIONS
169
✲
✲
✲
✲
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
ROW
DECODE
m-COLUMN
BLOCK (s-1)
✲
m-COLUMN
BLOCK (s-2)
.
.
.
.
.
.
.
.
.
.
.
.
.
✲
WE DOUT DIN
✻
✻
✲
✲
m-COLUMN
BLOCK (0)
✲
WE DOUT DIN
✻
WE DOUT DIN
✻
rowAddr[r-1:0]
din[m-1:0]
.........
s-input DMUX
we
COLUMN
DECODE
.
.
.
.
.
.
.
✻
colAddr[q-1:0]
m 2-input
ANDs
m 2-input
ANDs
m 2-input
ANDs
.......
m s-input
ORs
m q-input MUXs
addr[n-1:0]
dout[m-1;0]
Figure 7.7: RAM version with two dimension storage array. A number of m-bit blocks are
parallel connected and driven by the same row decoder. The column decoder selects to outoput an m-bit
word from the (s × m)-bit row.
170
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
CLOCK
✻
✍
✍
✍
✍
✍
✲
t
✲
✛
An−1 . . . A0
✻
a1
a2
a3
✲
a4
t
✲
✛ ✲
✛
W E′
✻
✲
t
✛
✲ ✛
✲
DIN
✻
✲
data in
t
DOU T
✻
✛ ✲
d(a1)
d(a2)
d(a4)
data in
✲
t
Figure 7.8: Read and write cycles for SRAM. For the flow-through version of a SRAM the
time behavior is similar to a register. The set-up and hold time are defined related to the active edge of
clock for all the input connections: data, write-enable, and address. The data output is also related to
the same edge.
7.5.1
Synchronous RAM
It is very hard to consider the time restriction imposed by the wave forms presented in Figure
7.5 when the system is requested to work at high speed. The system designer will be more comfortable with a memory circuit having all the time restrictions defined related only to the active
edge of the system clock. The synchronous RAM (SRAM) is conceived to have all time relations
defined related to the active edge of the clock signal. SRAM is the preferred embodiment of a
storage circuit in the contemporary designs. It performs write and read operations synchronized
with the active edge of the clock signal (see Figure 7.8).
VeriSim 7.4 The functional description of a synchronous RAM (0.5K of 64-bit words) follows:
module sram(
input
input
input
output
[63:0]
[8:0]
reg [63:0]
din
addr
we, clk
dout
,
,
,
);
//
//
//
//
data input
address
write enable & clock signal
data out
reg
[63:0]
mem[511:0];
always
@(posedge clk)
if (we) dout <= din
else
dout <= mem[addr]
;
;
always
endmodule
@(posedge clk)
if (we) mem[addr] <= din
;
⋄
The previously described SRAM is the flow-through version of a SRAM. A pipelined version
is also possible. It introduces another clock cycle delay for the output data.
7.5. APPLICATIONS
7.5.2
171
Register File
The most accessible data in a computational system is stored in a small and fast memory whose
locations are usually called machine registers or simply registers. In most usual embodiment
they have actually the physical structure of a register. The machine registers of a computational
(processing) element are organized in what is called register file. Because computation supposes
two operands and one result in most of cases, two read ports and one write port are currently
provided to the small memory used as register file (see Figure 7.9).
write enable
clock
❄
right addr[n-1:0]
✲
dest addr[n-1:0]
✲
result[m-1:0]
✲
left addr[n-1:0]
✲
✲
right operand[m-1:0]
✲
left operand[m-1:0]
register file m n
Figure 7.9: Register file. In this example it contains 2n m-bit registers. In each clock cycle any two
registers can be read and writing can be performed in anyone.
VeriSim 7.5 Follows the Verilog description of a register file containing 32 32-bit registers. In
each clock cycle any two pair of registers can be accessed to be used as operands and a result can
be stored in any one register.
module register_file(
reg [31:0]
assign
output
output
input
input
input
input
input
input
file[0:31];
left_operand
right_operand
[31:0]
[31:0]
[31:0]
[4:0]
[4:0]
[4:0]
left_operand
right_operand
result
left_addr
right_addr
dest_addr
write_enable
clock
= file[left_addr]
= file[right_addr]
,
,
,
,
,
,
,
);
,
;
always @(posedge clock) if (write_enable) file[dest_addr] <= result;
endmodule
⋄
The internal structure of a register file can be optimized using m × 2n 1-bit clocked latches
to store data and 2 m-bit clocked latches to implement the master-slave mechanism.
7.5.3
Field Programmable Gate Array – FPGA
Few decades ago the prototype of a digital system was realized in a technology very similar with
the one used for the final form of the product. Different types of standard integrated circuits
where connected according to the design on boards using a more or less flexible interconnection
172
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
technique. Now we do not have anymore standard integrated circuits, and making an Application
Specific Integrated Circuit (ASIC) is a very expensive adventure. Fortunately, now there is a
wonderful technology for prototyping (which can be used also for small production chains). It
is based on a one-chip system called Field Programmable Gate Array – FPGA. The name
comes from its flexibility to be configured by the user after manufacturing, i.e., “in the field”.
This generic circuit can be programmed to perform any digital function.
In this subsection the basic configuration of an FPGA circuit will be described4 . The internal
cellular structure of the system is described for the simplest implementation, letting aside details
and improvements used by different producer on this very diverse market (each new generation
of FPGA integrates different usual digital blocks in order to help efficient implementations; for
example: multipliers, block RAMs, ...; learn more about this from the on-line documentation
provided by the FPGA producers).
I/O
I/O
I/O
Switch matrix
❂
I/O
CLB
Long connection
CLB
✙
I/O
CLB
CLB
Local connection
✮
I/O
Figure 7.10: Top level organization of FPGA.
The system level organization of an FPGA
The FPGA chip has a cellular structure with three main programmable components, whose
function is defined by setting on 0 or on 1 control bits stored in memory elements. An FPGA
can be seen as a big structured memory containing million of bits used to control the state of
million of switches. The main type of cells are:
• Configurable Logic Blocks (CLB) used to perform a programmable combinational
and/or sequential function
4
The terminology introduced in this section follows the Xlilinx style in order to support the associated lab
work.
7.5. APPLICATIONS
173
• Switch Nodes which interconnect in the most flexible way the CLB modules and some
of them to the IO pins, using a matrix of programmed switches
• Input-Output Interfaces are two-direction programmable interfaces, each one associated with an IO pin.
Figure 7.10 provides a simplified representation of the internal structure of an FPGA at the top
level. The area of the chip is filled up with two interleaved arrays. One of the CLBs and another
of the Switch Nodes. The chip is boarded by IO interfaces.
The entire functionality of the system can be programmed by an appropriate binary configuration distributed in all the cells. For each IO pin is enough one bit to define if the pin is an
input or an output. For a Switch Node more bits are needed because each switch asks for 6 bits
to be configured. But, most of bits (in some implementations more than 100 per CLB) are used
to program the functions of the combinational and sequential circuits in each node containing a
CLB.
The IO interface
Each signal pin of the FPGA chip can be assigned to be an input or an output. The simplest
form of the interface associated to each IO pin is presented in Figure 7.11, where:
• D-FF0: is the D master-slave flip-flop which synchronously receives the value of the I/O
pin through the associated input non-inverting buffer
• m: the storage element which contains the 1-bit program for the input interface used to
command the tristate buffer; if m = 1 then the tristate buffer is enabled and interface is
in the output mode, else the tristate buffer is disabled and interface is in the input mode
• D-FF1: is the flip-flop loaded synchronously with the output bit to be sent to the I/O
pin if m = 1.
Programmable memory element
✰
m
✲D
Q
D-FF1
clock
✛✲
■
✛
D-FF0
Q
D
I/O pin
Figure 7.11: Input-Output interface.
The storage element m is part of the big distributed RAM containing all the storage elements
used to program the FPGA.
174
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
The switch node
The switch node (Figure 7.12a) consists of a number of programmable switches (4 in our description). Each switch (Figure 7.12b) manages 4 wires, connection them in different configurations
using 6 nMOS transistors, each commanded by the state of 1-bit memory (Figure 7.12c). If mi
= 1 then the associated nMOS transistor is on and between its drain end source the resistor
has a small value. If mi = 0 then the associated nMOS transistor is off and the two ends of the
switch are not connected.
✻ ✻ ✻ ✻
✲
m2
3
2
q
0
1
✛
✛
✶
4
1
0
5
3
2
✲
✲
❄ ❄ ❄
a.
b.
c.
d.
Figure 7.12: The structure of a Switch Node. a. A Switch Node with 4 switches. b. The
organization of a switch. c. A line switch. d. An example of actual connections.
For example, the configuration shown in Figure 7.12d is programmed as follows:
switch 0 : {m0, m1, m2, m3, m4, m5} = 011010;
switch 1 : {m0, m1, m2, m3, m4, m5} = 101000;
switch 2 : {m0, m1, m2, m3, m4, m5} = 000001;
switch 3 : {m0, m1, m2, m3, m4, m5} = 010000;
Any connection is a two-direction connection.
The basic building block
Because any digital circuit can be composed by properly interconnected gates and flip-flops, each
CLB contains a number of basic building blocks, called bit slices (BSs), each able to provide
at least an n-input, 1-output programmable combinational circuit and an 1-bit register.
In the previous chapter was presented an Universal combinational circuit: the n-input multiplexer able to perform any n-variable Boolean function. It was programmed applying on its
selected inputs an m-bit binary configuration (where m = 2n ). Thereby, an M U Xn and a memory for storing the m-bit program provide the structure able to be programmed to perform any
n-input 1-output combinational circuit. In Figure 7.13 it is represented, for n = 4, by the multiplexer MUX and the 16 memory elements m0, m1, ... m15. The entire sub-module is called
LUT (from look-up table). The memory elements m0, m1, ... m15, being part of the big
distributed RAM of the FPGA chip, can be loaded with any out of 65536 binary configuration
used to define the same number of 4-input Boolean function.
Because the arithmetic operations are very frequently used the BS contains a specific circuit
for any arithmetic operation: the circuit computing the value of the carry signal. The module
carry Figure 7.13 has also its specific propagation path defined by a specific input, carryIn,
and a specific output carryOut.
7.5. APPLICATIONS
175
carryOut
✻
mc
md
LUT
m15
✲ 15
✲
✲
MUX
m1
✲
1
❄
m0
✲
✲
❄
✲D
0
EMUX
✲
❄
0
EMUX
✲
out
1
Q
D-FF
1
Carry
0
✻✻✻✻
✻ ✻
in[3:0]
carryIn
clock
Figure 7.13: The basic building block: the bit slice (BS).
The BS module contains also the one-bit register D-FF. Its contribution can be considered
in the current design if the memory element md is programmed appropriately. Indeed, if md =
1, then the output of the BS comes form the output of D-FF, else the output of the BS is a
combinational one, the flip-flop being shortcut.
The memory element mc is used to program the selection of the LUT output or of the Carry
output to be considered as the programmable combinational function of this BS.
The total number of bits used to program the function of the BS previously described is 18.
Real FPGA circuits are now featured with much more complex BSs (please search on their web
pages for details).
There are two kinds of BS: logic type and memory type. The logic type uses LUT to
implement combinational functions. The memory type uses LUT for implementing both, combinatorial functions and memory function (RAM or serial shift register).
The configurable logic block
The main cell used to build an FPGA, CLB (see Figure 7.10) contains many BSs organized in
slices. The most frequent organization is of 2 slices, each having 4 BSs (see Figure 7.14). There
are slices containing logic type BSs (usually called SLICEL), or slices containing memory type
BSs (usually called SLICEM). Some CLBs are composed by two SLICEL, others are composed
by one SLICEL and one SLICEM.
A slice has some fix connections between its BSs. In our simple description, the fix connections refers to the carry chain connections. Obviously, we can afford to make fix connections for
circuits having specific function.
7.5.4
∗ Content Addressable Memory
A normal way to “question” a memory circuit is to ask for:
Q1: the value of the property A of the object B
For example: How old is George? The age is the property and the object is George. The first step to
design an appropriate device to be questioned as previously is exemplified is to define a circuit able to
176
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
Carry chain
Configurable Logic Block
Slice 2
✛
BS3
❲
✛
BS2
✛
BS1
✛
BS0
✛
Slice 1
✛
BS3
✛
BS2
✛
BS1
✛
BS0
✛
Figure 7.14: Configurable logic block.
answer the question:
Q2: where is the object B?
with two possible answers:
1. the object B is not in the searched space
2. the object B is stored in the cell indexed by X.
The circuit for answering Q2-type questions is called Content Addressable Memory, shortly: CAM. (About
the question Q1 in the next subsection.)
The basic cell of a CAM is consists of:
• the storage elements for binary objects
• the “questioning” circuits for searching the value applied to the input of the cell.
In Figure 7.15 there are 4 D latches as storage elements and four XORs connected to a 4-input NAND
used as comparator. The cell has two functions:
• to store: the active level of the clock modify the content of the cell storing the 4-bit input data
into the four D latches
• to search: the input data is continuously compared with the content of the cell generating the
signal AO′ = 0 if the input matches the content.
The cell is written as an m-bit latch and is continuously interrogated using a combinational circuit
as comparator. The resulting circuit is an 1-OS because results serially connecting a memory, one-loop
circuit with a combinational, no-loop circuit. No additional loop is involved.
An n-word CAM contains n CAM cells and some additional combinational circuits for distributing
the clock to the selected cell and for generating the global signal M, activated for signaling a successful
match between the input value and one or more cell contents. In Figure 7.16a a 4-word of 4 bits each is
represented. The write enable, W E, signal is demultiplexed as clock to the appropriate cell, according to
the address codded by A1 A0 . The 4-input NAND generate the signal M . If, at least one address output,
AOi′ is zero, indicating match in the corresponding cell, then M = 1 indicating a successful search.
The input address Ap−1 , . . . , A0 is binary codded on p = log2 n bits. The output address
AOn−1 , . . . , AO0 is an unary code indicating the place or the places where the data input Dm−1 , . . . , D0
matches the content of the cell. The output address must be unary codded because there is the possibility
of match in more than one cell.
Figure 7.16b represents the logic symbol for a CAM with n m-bit words. The input W E indicate the
function performed by CAM. Be very careful with the set-up time and hold time of data related to the
W E signal!
The CAM device is used to locate an object (to answer the question Q2). Dealing with the properties
of an object (answering Q1-type questions) means to use o more complex devices which associate one or
7.5. APPLICATIONS
177
D3
clock
❄ ❄
D✻
❄ ❄
CK
Q
D2
CK
′
D
Q
D1
❄ ❄
CK
′
D
Q
D0
❄ ❄
CK
′
D
Q′
AO ′
a.
D3 , . . . , D0
clock
✲
✲ CK
cam cell
✲
AO ′
b.
Figure 7.15: The Content Addressable Cell. a. The structure: data latches whose content is
compared against the input data using 4 XORs and one NAND. Write is performed applying the clock
with stable data input. b. The logic symbol.
more properties to an object. Thus, the associative memory will be introduced adding some circuits to
CAM.
7.5.5
∗ An Associative Memory
A partially used RAM can be an associative memory, but a very inefficient one. Indeed, let be a RAM
addressed by An−1 . . . A0 containing 2-field words {V, Dm−1 . . . D0 }. The objects are codded using the
address, the values of the unique property P are codded by the data field Dm−1 . . . D0 . The one-bit field
V is used as a validation flag. If V = 1 in a certain location, then there is a match between the object
designated by the corresponding address and the value of property P designated by the associated data
field.
Example 7.2 Let be the 1Mword RAM addressed by A19 . . . A0 containing 2-field 17-bit words
{V, D15 . . . D0 }. The set of objects, OBJ, are codded using 20-bit words, the property P associated
to OBJ is codded using 16-bit words. If
RAM [11110000111100001111] = 1 0011001111110000
RAM [11110000111100001010] = 0 0011001111110000
then:
• for the object 11110000111100001111 the property P is defined (V = 1) and has the value
0011001111110000
• for the object 11110000111100001010 the property P is not defined (V = 0) and the data field is
meaningless.
Now, let us consider the 20-bit address codes four-letter names using for each letter a 5-bit code. How
many locations in this memory will contain the field V instantiated to 1? Unfortunately, only extremely
few of them, because:
• only 24 from 32 binary configurations of 5 bits will be used to code the 24 letters of Latin alphabet
(244 < 220 )
• but more important: how many different name expressed by 4 letters can be involved in a real
application? Usually no more than few hundred, meaning almost nothing related to 220 .
178
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
A1 A0
WE
❄❄❄
DMUX
3 2 1 0
D 3 , . . . , D0
✲
✲ CK cam cell 0
✲
✲ CK cam cell 1
✲
✲ CK cam cell 2
✲
✲ CK cam cell 3
a.
✲
✲
✲
✲
′
AO0
Dm−1 , . . . , D0
WE
′
AO1
Ap−1 , . . . , A0
❄
✲
✲
✲
CAMm×n
′
′
, . . . , AO0
AOn−1
′
AO2
❄
M
′
AO3
M
b.
Figure 7.16: The Content Addressable Memory (CAM). a. A 4-word CAM is built using 4
content addressable cells, a demultiplexor to distribute the write enable (WE) signal, and a N AN D4 to
generate the match signal (M). b. The logic symbol.
⋄
The previous example teaches us that a RAM used as associative memory is a very inefficient solution.
In real applications are used names codded very inefficiently:
number of possible names >>> number of actual names.
In fact, the natural memory function means almost the same: to remember about something immersed
in a huge set of possibilities.
One way to implement an efficient associative memory is to take a CAM and to use it as a programmable decoder for a RAM. The (extremely) limited subset of the actual objects are stored into a
CAM, and the address outputs of the CAM are used instead of the output of a combinational decoder to
select the accessed location of a RAM containing the value of the property P . In Figure 7.17 this version
of an associative memory is presented. CAMm×n is usually dimensioned with 2m >>> n working as a
decoder programmed to decode any very small subset of n addresses expressed by m bits.
wa’
sel
✲
✲
p
address
din
m
❄
q
❄
we’
AOn−1
D
✲
we’
✛
wd’
A
CAMm×n
Array of Lq
(prog dcd)
M
AO0
❄
valide dout
✲
q
❄
dout
Figure 7.17: An associative memory (AM). The structure of an AM can be seen as a RAM with
a programmable decoder implemented with a CAM. The decoder is programmed loading CAM with the
considered addresses.
Here are the three working mode of the previously described associative memory:
7.5. APPLICATIONS
179
define object : write the name of an object to the selected location in CAM
wa’ = 0, address = name of object, sel = cam address
wd’ = 1, din = don’t care
associate value : write the associated value in the randomly accessed array to the location selected by
the active address output of CAM
wa’ = 1, address = name of object, sel = don’t care
wd’ = 0, din = value
search : search for the value associated with the name of the object applied to the address input
wa’ = 1, address = name of object, sel = don’t care
wd’ = 1, din = don’t care
dout is valid only if valide dout = 1.
This associative memory will be dimensioned according to the dimension of the actual subset of
names, which is significantly smaller than the virtual set of the possible names (2p <<< 2m ). Thus, for
a searching space with the size in O(2m ) a device having the size in O(2p ) is used.
7.5.6
∗ Beneš-Waxman Permutation Network
In 1968, even though it was defined by others before, Václav E. Beneš promoted a permutation network
[Benes ’68] and Abraham Waxman published an optimized version [Waksman ’68] of it.
A permutation circuit is a network of programmable switching circuits which receives the sequence
x1 , . . . , xn and can be programmed to to provide on its outputs any of the n! possible permutations.
The two-input programmable switching circuit is represented in Figure 7.18 a. It consists of a D-FF
to store the programming bit and two 2-input multiplexors to perform the programmed switch. The
circuit works as follows:
D-FF = 0 : x′1 = x1 and x′2 = x2
D-FF = 1 : x′1 = x2 and x′2 = x1
The input enable allows to load D-FF with the programming bit. The storage element is a master-slave
structure because in a complex network the D flip-flops are chained because they are loaded with the
programming bits by shifting. The logic symbol for this elementary circuit is represented in Figure 7.18
b (the clock input and enable input are omitted for simplicity).
Beneš-Waxman permutation network (we must recognize credit for this circuit, at least, for both,
Beneš and Waxman) with n inputs has the recursive definition presented in Figure 7.18 c. (The only
difference between the definition provided by Beneš and the optimization done by Waxman refers to the
number of switches on the output layer: in Waxman’s approach there are only n/2 − 1 switches, inetead
of n for the version presented by Beneš.)
Theorem 7.1 The switches of Beneš-Waxman permutation network can be set to realize any permutation.
⋄
Proof. For n = 2 the permutation network is a programmable switch circuit. For n > 2 we consider,
for simplicity, n as a power of 2.
If the two networks Lef tPn/2 and RightPn/2 are permutation networks with n/2 inputs, then we will
prove that it is possible to program the input switches so as each sequence of two successive value on
the outputs contains values that reach the output switch going through different permutation network.
If the “local” order, on the output pair, is not the desired one, then the output switch is used to fix the
problem by an appropriate programming.
The following steps can be followed to establish the programming Boolean sequence –
{qIn1 , . . . , qInn/2 } – for the n/2 input switches and the programming Boolean sequence –
{qOut2 , . . . , qOutn/2 } – for the n/2 − 1 output switches:
1. because y1 ← xi , the input switch where xi is connected is programmed to qIn⌈i/2⌉ = i+1−2×⌈i/2⌉
in order to let xi to be applied on the Lef tPn/2 permutation network (if i is an odd number, then
qIn⌈i/2⌉ = 0, else qIn⌈i/2⌉ = 1); the xi−(−1)i input is the “companion” of xi on the same switch; it
is consequently routed to the input of RightPn/2
180
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
2. the output switch ⌈j/2⌉ is identified using the correspondence yj ← xi−(−1)i ; the state of the switch
is set to sOut⌈j/2⌉ = j − 2 × ⌊j/2⌋ (if j is an odd number, then qOut⌈j/2⌉ = 1, else qOut⌈j/2⌉ = 0)
3. for yj−(−1)j ← xk , the “companion” of yj on the j/2-th output switch, we go back to the step 1
until in the step 2 we reach the second output, y2 ; for the first two outputs there is no need of a
switch because a partial or the total programming cycle ends when y2 receives its value from the
output of the RightPn/2 permutation network
4. if all the output switches are programmed the process stops, else, we start again from the left
output of an un-programmed output switch.
Any step 1, let us call it up-step, programs an input switch, while any step 2, let us call it down-step,
programs an output switch. Any up-step, which solves the connection yi ← xj , is feasible because the
source xj is always connected to a still un-programmed switch. Similarly, any down-step is also feasible.
⋄
The size and depth of the permutation network Pn is computed using the relations:
SP (n) = (2 × SP (n/2) + n − 1) × SP (2)
SP (2) ∈ O(1)
DP (n) = DP (n) + 2 × DP (2)
DP (2) ∈ O(1)
Results:
SP (n) = (n × log2 n − n + 1) × SP (2) ∈ O(nlog n)
DP (n) = −1 + 2 × log2 n ∈ O(log n)
Example 7.3 Let be an 8-input permutation network which must be programmed to perform the following
permutation:
{x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 } → {x8 , x6 , x3 , x1 , x4 , x7 , x5 , x2 } = {y1 , . . . , y8 }
The permutation network with 8 inputs is represented in Figure 7.19. It is designed starting from the
recursive definition.
The programming bits for each switch are established according to the algorithm described in the
previous proof. In the first stage the programming bits for the first layer – {p11 , p12 , p13 , p14 } – and last
layer – {p52 , p53 , p54 } – are established, as follows:
1. (y1 = x8 ) ⇒ (x8 → lef tP4 ) ⇒ (p14 = 1)
2. (p14 = 1) ⇒ (x7 → rightP4 ) ⇒ (p53 = 0)
3. (y5 = x4 ) ⇒ (x4 → lef tP4 ) ⇒ (p12 = 1)
4. (p12 = 1) ⇒ (x3 → rightP4 ) ⇒ (p52 = 1)
5. (y4 = x1 ) ⇒ (x1 → lef tP4 ) ⇒ (p11 = 0)
6. (p11 = 0) ⇒ (x2 → rightP4 ) ⇒ (p54 = 0)
7. (y7 = x5 ) ⇒ (x5 → lef tP4 ) ⇒ (p13 = 0)
8. (p13 = 0) ⇒ (x6 → rightP4 ) ⇒ the first stage of programming closes successfully.
In the second stage there are two P4 permutation networks to be programmed: lef tP4 and rightP4 .
From the first stage of programming resulted the following permutations to be performed:
for lef tP4 : {x1 , x4 , x5 , x8 } → {x8 , x1 , x4 , x5 }
for rightP4 : {x2 , x3 , x6 , x7 } → {x6 , x3 , x7 , x2 }
The same procedure is applied now twice providing the programming bits for the second and the fourth
layers of switches.
The last step generate the programming bits for the third layer of switches.
The programming sequences for the five layers of switches are:
7.5. APPLICATIONS
181
x1
x2
❄❄
0 1
mux
s
p
❄❄
0 1
mux
✛
s
❄
✛
clock
D-FF
e
x1 x2 p
✛
❄ ❄❄
enable
❄
❄
y1
y2
❄ ❄❄
❄
y1 y2 q
q
a.
p1
x1
x2
p2
x3
x4
pn/2−1
xn−3
xn−2
pn/2
xn−1
xn
b.
✲
✲
✲
qIn1
✲
✲
✲
qIn2
✲
✲
✲
qInn/2−1
✲
✲
✲
qInn/2
❄ ❄❄
❄
❄❄
LeftPn/2
❄❄
❄ ❄ ❄❄
RightPn/2
Pn
✲
✲
✲
qOut1
y1
✲
✲
✲
✲
✲
✲
qOut2
y3
✲
✲
✲
✲qOutn/2−1
✲ yn−3
✲ yn−2
✲
✲
✲
✲ qOutn/2
✲ yn−1
✲ yn
y2
y4
c.
Figure 7.18: Beneš-Waxman permutation network. a. The programmable switching circuit.
b. The logic symbol for the programmable switching circuit. c. The recursive definition of a n-input
Beneš-Waxman permutation network, Pn .
182
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
p1
x1
p2
x2
x3
x4
p3
x5
x6
p4
x7
x8
P8
❄
❄❄
❄
0
❄❄
1
❄
❄❄
0
❄
❄❄
1
✰
P4
x1
x4
❄
x5
❄❄
x8
❄
1
x2
❄❄
1
❄
x3
x6
❄❄
0
❄
x7
❄❄
0
✰
P2
❄
x8
❄❄
1
x1
❄
y1 = x8
❄
❄❄
❄
❄❄
x4
❄
y2 = x 6
0
❄
❄❄
1
0
x5
x6
x3
❄
❄❄
❄
❄❄
x7
0
1
x2
❄
❄❄
❄
❄❄
❄
❄❄
❄
❄
❄
❄
❄
❄
y3 = x3
1
y4 = x 1
y 5 = x4
0
y6 = x7
y7 = x5
✙
0
y8 = x2
Figure 7.19: .
prog1
prog2
prog3
prog4
prog5
=
=
=
=
=
{0
{1
{1
{{-
1
1
0
0
1
0
0
1
0
1}
0}
0}
1}
0}
To insert in five clock cycles the programming sequences into the permutation network on the inputs
{p1 , p2 , p3 , p4 } (see Figure 7.19) are successively applied the following 4-bit words: 0100, 0001, 1010,
1100, 0101. During the insertion the input enable on each switch is activated.
⋄
7.5.7
∗ First-Order Systolic Systems
When a very intense computational function is requested for an Application Specific Integrated Circuit
(ASIC) systolic systems represent an appropriate solution. In a systolic system data are inserted and/or
extracted rhythmically in/from a uniform modular structure. H. T. Kung and Charles E. Leiserson
published the first paper describing a systolic system in 1978 [Kung ’79] (however, the first machine
known to use a a systolic approach was the Colossus Mark II in 1944). The following example of systolic
system is taken from this paper.
7.5. APPLICATIONS
183
Let us design the circuit which multiplies a
a11 a12 0
0
a21 a22 a23 0
a31 a32 a33 a34
0 a42 a43 a44
..
..
.
.
0
0
..
.
0
0
0
band matrix with a vector as follows:
0 ···
y1
x1
0 · · · x2 y2
0 ···
y
x3
= 3
a45 · · · ×
x
y
4
4
..
..
x
y5
5
.
.
..
..
..
..
.
.
.
.
The main operation executed for matrix-vector operations is multiply and accumulate (MACC):
Z =A×B+C
for which a specific combinational module is designed. Interleaving MACCs with memory circuits is
provided a structure able to compute and to control the flow of data in the same time. The systolic
vector-matrix multiplier is represented in Figure 7.20.
The systolic module is represented in Figure 7.20a, where a combinational multiplier (M = A × B)
is serially connected with an combinational adder (M + C). The result of MACC operation is latched in
the output latch which latches besides the result of the computation, the two input value A and B. The
latch is transparent on the high level of the clock. It is used to buffer intermediary results and to control
the data propagation through the system.
The system is configured using pairs of modules to generate a master-slave structures, where one
module receives ck and another ck’. The resulting structure is a non-transparent one ready to be used
in a pipelined connection.
For a band matrix having the width 4, two non-transparent structures are used (see Figure 7.20c).
Data is inserted in each phase of the clock (correlate data insertion with the phase of clock represented
in Figure 7.20b) as follows:
The result of the computation is generated sequentially to the output yi of the circuit from Figure
7.20c, as follows:
y1 = a11 x1 + a12 x2
y2 = a21 x1 + a22 x2 + a23 x3
y3 = a31 x1 + a32 x2 + a33 x3 + a34 x4
y4 = a42 x2 + a43 x3 + a44 x4 + a45 x5
y5 = . . .
...
The output X of the module is not used in this application (it is considered for matrix matrix
multiplication only). The state of the system in each phase of the clock (see Figure 7.20b) is represented
by two quadruples:
(Y1 , Y2 , Y3 , Y4 )
(Z1 , Z2 , Z3 , Z4 )
If the initial state of the system is unknown,
(Y1 , Y2 , Y3 , Y4 ) = (−, −, −, −)
(Z1 , Z2 , Z3 , Z4 ) = (−, −, −, −)
then the state of the system in the first 10 phases of the clock, numbered in Figure 7.20c, are the following:
Phase: (1)
(Y1 , Y2 , Y3 , Y4 ) = (−, −, −, −)
(Z1 , Z2 , Z3 , Z4 ) = (−, −, −, 0)
Phase: (2)
(Y1 , Y2 , Y3 , Y4 ) = (0, −, −, −)
(Z1 , Z2 , Z3 , Z4 ) = (−, −, 0, 0)
Phase: (3)
(Y1 , Y2 , Y3 , Y4 ) = (0, 0, −, −)
(Z1 , Z2 , Z3 , Z4 ) = (−, 0, 0, 0)
Phase: (4)
(Y1 , Y2 , Y3 , Y4 ) = (x1 , 0, 0, −)
184
CHAPTER 7. MEMORIES:
A
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
B
C
❄
✲
×
✛
✲
+
ck
✻
1
✲ ck
ck
❄
❄
3
4
5
6
7
8
9 10
✲
t
❄
Latch
❄
X
❄
Y
❄
Z
a.
b.
...
...
...
...
(9)
-
a33
-
a42
(8)
a23
-
a32
-
(7)
-
a22
-
a31
(6)
a12
-
a21
-
(5)
-
a11
-
0
0
0
0
0
❄ (1 - 4)
(1) (2) ✛
...
2
-
-
0
0 y1 y1 y2 y2 y3
x4 x 3 x 3 x2 x2 x1 x1
(10)
❄
(10)
0
0
0
❄
❄
❄
A1
A2
A3
A4
ck
ck
ck
ck
✛ Z1 C 1 ✛ Z2 C 2 ✛ Z3 C 3 ✛ Z4 C 4 ✛
✲ B1 1 Y 1 ✲ B2 2 Y 2 ✲ B3 3 Y 3 ✲ B4 4 Y 4
✲ (2) (1)
ck
✻
✻
✻
✻
ck’
c.
Figure 7.20: Systolic vector-matrix multiplier. a. The module. b. The clock signal with
indexed half periods. c. How the modular structure is fed with the data in each half period of the clock
signal.
7.6. CONCLUDING ABOUT MEMORY CIRCUITS
185
(Z1 , Z2 , Z3 , Z4 ) = (0, 0, 0, 0)
(5)
(Y1 , Y2 , Y3 , Y4 ) = (x1 , x1 , 0, 0)
(Z1 , Z2 , Z3 , Z4 ) = (0, a11 x1 , 0, 0)
(6)
(Y1 , Y2 , Y3 , Y4 ) = (x2 , x1 , x1 , 0)
(Z1 , Z2 , Z3 , Z4 ) = (a11 x1 + a12 x2 , a11 x1 , a21 x1 , 0)
(7)
(Y1 , Y2 , Y3 , Y4 ) = (x2 , x2 , x1 , x1 )
(Z1 , Z2 , Z3 , Z4 ) = (y1 , a21 x1 + a22 x2 , a21 x1 , a31 x1 )
(8)
(Y1 , Y2 , Y3 , Y4 ) = (x3 , x2 , x2 , x1 )
(Z1 , Z2 , Z3 , Z4 ) = (a21 x1 + a22 x2 + a23 x3 , a21 x1 + a22 x2 , a31 x1 + a32 x2 , a31 x1 )
(9)
(Y1 , Y2 , Y3 , Y4 ) = (x3 , x3 , x2 , x2 )
(Z1 , Z2 , Z3 , Z4 ) = (y2 , a31 x1 + a32 x2 + a33 x3 , a31 x1 + a32 x2 , a42 x2 )
(10)
(Y1 , Y2 , Y3 , Y4 ) = (x4 , x3 , x3 , x2 )
(Z1 , Z2 , Z3 , Z4 ) = (a31 x1 + a32 x2 + a33 x3 + a34 x4 , a31 x1 + a32 x2 + a33 x3 , a42 x2 + a43 x3 , a42 x2 )
(11)
(Y1 , Y2 , Y3 , Y4 ) = (x4 , x4 , x3 , x3 )
(Z1 , Z2 , Z3 , Z4 ) = (y3 , . . .)
...
In each clock cycle 4 multiplications and 4 additions are performed. The pipeline connections allow the
synchronous insertion and extraction of data. The maximum width of the matrix band determines the
number of modules used to design the systolic system.
7.6
Concluding About Memory Circuits
For the first time, in this chapter, both composition and loop are used to construct digital
systems. The loop adds a new feature and the composition expands it. The chapter introduced
only the basic concepts and the main ways to use them in implementing actual digital systems.
The first closed loop in digital circuits latches events Closing properly simple loops in
small combinational circuits vey useful effects are obtained. The most useful is the “latch effect”
allowing to store certain temporal events. An internal loop is able to determine an internal
state of the circuit which is independent in some extent from the input signals (the circuit
controls a part of its inputs using its own outputs). Associating different internal states to
different input events the circuit is able to store the input event in its internal states. The first
loop introduces the first degree of autonomy in a digital system: the autonomy of the internal
state. The resulting basic circuit for building memory systems is the elementary latch.
Meaningful circuits occur by composing latches The elementary latches are composed
in different modes to obtain the main memory systems. The serial composition generates the
master-slave flip-flop which is triggered by the active edge of the clock signal. The parallel composition introduces the concept of random access memory. The serial-parallel composition
defines the concept of register.
Distinguishing between “how?” and “when?” At the level of the first order systems
occurs a very special signal called clock. The clock signal becomes responsible for the history
sensitive processes in a digital system. Each “clocked” system has inputs receiving information
about “how” to switch and another special input – the clock input acting on one of its edge called
186
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
the active edge of clock – and another special input indicating “when” the system switches. We
call this kind of digital systems synchronous systems, because any change inside the system is
triggered synchronously by the same edge (positive or negative) of the clock signal.
Registers and RAMs are basic structures First order systems provide few of the most
important type of digital circuits used to support the future developments when new loops will
be closed. The register is a synchronous subsystem which, because of its non-transparency,
allows closing the next loop leading to the second order digital systems. Registers are used also
for accelerating the processing by designing pipelined systems. The random access memory
will be used as storage element in developing systems for processing a big amount of data or
systems performing very complex computations. Both, data and programs are stored in RAMs.
RAM is not a memory, it is only a physical support Unfortunately RAM has not the
function of memorizing. It is only a storage element. Indeed, when the word W is stored at the
address A we must memorize the address A in order to be able to retrieve the word W . Thus,
instead of memorizing W we must memorize A, or, as usual, we must have a mechanism to
regenerate the address A. In conjunction with other circuits RAM can be used to build systems
having the function of memorizing. Any memory system contains a RAM but not only a RAM,
because memorizing means more than storing.
Memorizing means to associate Memorizing means both to store data and to retrieve it.
The most “natural” way to design a memory system is to provide a mechanism able to associate
the stored data with its location. In an associative memory to read means to find, and to write
means to find a free location. The associative memory is the most perfect way of designing
a memory, even if it is not always the most optimal as area (price), time and power.
To solve ambiguities a new loop is needed At the level of the first order systems the
second latch problem can not be solved. The system must be more “intelligent” to solve the
ambiguity of receiving synchronously contradictory commands. The system must know more
about itself in order to be “able” to behave under ambiguous circumstances. Only a new loop
will help the system to behave coherently. The next chapter, dealing with the second level of
loops, will offer a robust solution to the second latch problem.
The storing and memory functions, typical for the first order systems, are not true computational features. We will see that they are only useful ingredients allowing to make digital
computational systems efficient.
7.7
Problems
Stable/unstable loops
Problem 7.1 Simulate in Verilog the unstable circuit described in Example 3.1. Use 2 unit
time (#2) delay for each circuit and measure the frequency of the output signal.
Problem 7.2 Draw the circuits described by the following expressions and analyze their stability
taking into account all the possible combinations applied on their inputs:
d = b(ad)′ + c
d = (b(ad)′ + c)′
7.7. PROBLEMS
187
c = (ac′ + bc)′
c = (a ⊕ c) ⊕ b.
Simple latches
Problem 7.3 Illustrate the second latch problem with a Verilog simulation. Use also versions
of the elementary latch with the two gates having distinct propagation times.
Problem 7.4 Design and simulate an elementary clocked latch using a NOR latch as elementary
latch.
Problem 7.5 Let be the circuit from Figure 7.21. Indicate the functionality and explain it.
Hint: emphasize the structure of an elementary multiplexer.
c
c’
c’
q
d
c
c’
c
ck
Figure 7.21: ?
Problem 7.6 Explain how it works and find an application for the circuit represented in Figure
7.22.
Hint: Imagine the tristate drivers are parts of two big multiplexors.
out1
out2
in
Figure 7.22: ?
Master-slave flip-flops
Problem 7.7 Design an asynchronously presetable master-slave flip-flop.
Hint: to the slave latch must be added asynchronous set and reset inputs (S’ and R’ in the
NAND latch version, or S and R in the NOR latch version).
Problem 7.8 Design and simulate in Verilog a positive edge triggered master-slave structure.
188
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
Problem 7.9 Design a positive edge triggered master slave structure without the clock inverter.
Hint: use an appropriate combination of latches, one transparent on the low level of the clock
and another transparent on the high level of the clock.
Problem 7.10 Design the simulation environment for illustrating the master-slave principle
with emphasis on the set-up time and the hold time.
Problem 7.11 Let be the circuit from Figure 7.23. Indicate the functionality and explain it.
Modify the circuit to be triggered by the other edge of the clock.
Hint: emphasize the structures of two clocked latches and explain how they interact.
c
c’
c
c
c’
c’
q
d
c’
c’
c
c
ck
Figure 7.23: ?
Problem 7.12 Let be the circuit from Figure 7.24. Indicate the functionality and explain it.
Assign a name for the questioned input. What happens if the NANDs are substituted with NORs.
Rename the questioned input. Combine both functionality designing a more complex structure.
Hint: go back to Figure 2.6c.
c
c’
c
c’
c
d
c’
q
c
c’
?
c’
ck
c
Figure 7.24: ?
Enabled circuits
Problem 7.13 An n-bit latch stores the n-bit value applied on its inputs. It is transparent on
the low level of the clock. Design an enabled n-bit latch which stores only in the clock cycle in
which the enable input, en, take the value 1 synchronized with the positive edge of the clock.
Define the set-up time and the hold time related to the appropriate clock edge for data input and
for the enable signal.
Problem 7.14 Provide a recursive Verilog description for an n-bit enabled latch.
7.7. PROBLEMS
189
RAMs
Problem 7.15 Explain the reason for tASU and for tAH in terms of the combinational hazard.
Problem 7.16 Explain the reason for tDSU and for tDH .
Problem 7.17 Provide a structural description of the RAM circuit represented in Figure 7.4
for m = 256. Compute the size of the circuit emphasizing both the weight of storing circuits and
the weight of the access circuits.
Problem 7.18 Design a 256-bit RAM using a two-dimensional array of 16×16 latches in order
to balance the weight of the storing circuits with the weight of the accessing circuits.
Problem 7.19 Design the flow-through version of SRAM defined in Figure 7.8.
Hint: use additional storage circuits for address and input data, and relate the W E ′ signal with
the clock signal.
Problem 7.20 Design the register to latch version of SRAM defined in Figure 7.25.
Hint: the write process is identical with the flow-through version.
CLOCK
✻
✍
✍
✍ ✲
t
ADDRESS
✻
✲
addr
t
DOU T
Register to Latch
✻
✲
data(addr)
✲
t
DOU T
Pipeline
✻
✲
data(addr)
✲
t
Figure 7.25: Read cycles. Read cycle for the register to latch version and for the pipeline version of
SRAM .
Problem 7.21 Design the pipeline version of SRAM defined in Figure 7.25.
Hint: only the output storage device must be adapted.
Registers
Problem 7.22 Provide a recursive description of an n-bit register. Prove that the (algorithmic)
complexity of the concept of register is in O(n) and the complexity of a ceratin register is in
O(log n).
190
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
Problem 7.23 Draw the schematic for an 8-bit enabled and resetable register. Provide the
Verilog environment for testing the resulting circuit. Main restriction: the clock signal must be
applied only directly to each D flip-flop.
Hint: an enabled device performs its function only if the enable signal is active; to reset a
register means to load it with the value 0.
Problem 7.24 Add to the register designed in the previous problem the following feature: the
content of the register is shifted one binary position right (the content is divided by two neglecting
the reminder) and on most significant bit (MSB) position is loaded the value of the one input
bit called SI (serial input). The resulting circuit will be commanded with a 2-bit code having the
following meanings:
nop : the content of the register remains unchanged (the circuit is disabled)
reset : the content of the register becomes zero
load : the register takes the value applied on its data inputs
shift : the content of the register is shifted.
Problem 7.25 Design a serial-parallel register which shifts 16 16-bit numbers.
Definition 7.3 The serial-parallel register, SP Rn×m , is made by a SP R(n−1)×m serial connected with a Rm . The SP R1×m is Rm . ⋄
Hint: the serial-parallel register, SP Rn×m can be seen in two manners. SP Rn×m consists in
m parallel connected serial registers SRn , or SP Rn×m consists in n serially connected registers
Rm . We prefer usually the second approach. In Figure 7.26 is shown the serial-parallel SP Rn×m .
✲
IN
✲
Rm
Rm
✲
CK
✲
Rm
OUT
✲
...
a.
✲
IN
b.
...
SP Rn×m
OUT
✲
CK
Figure 7.26: The serial-parallel register. a. The structure. b. The logic symbol.
Problem 7.26 Let be tSU , tH , tp , for a register and tpCLC the propagation time associated with
the CLC loop connected with the register. The maximal and minimal value of each is provided.
Write the relations governing these time intervals which must be fulfilled for a proper functioning
of the loop.
Pipeline systems
Problem 7.27 Explain what is wrong in the following always construct used to describe a
pipelined system.
7.8. PROJECTS
module pipeline
191
#(parameter
n = 8, m = 16, p = 20)
(output
reg[m-1:]
output_reg,
input
wire[n-1:0] in,
clock);
reg[n-1:0]
input_reg;
reg[p-1:0]
pipeline_reg;
wire[p-1:0]
out1;
wire[m-1:0]
out2;
clc1
first_clc(out1, input_reg);
clc2
second_clc(out2, pipeline_reg);
always @(posedge clock) begin
input_reg = in;
pipeline_reg = out1;
output_reg = out2;
end
endmodule
module clc1(out1, in1);
// ...
endmodule
module clc2(out2, in2);
// ...
endmodule
Hint: revisit the explanation about blocking and nonblocking evaluation in Verilog.
Register file
Problem 7.28 Draw register file 16 4 at the level of registers, multiplexors and decoders.
Problem 7.29 Evaluate for register file 32 5 minimum input arrival time before clock
(tin reg ), minimum period of clock (Tmin ), maximum combinational path delay (tin out ) and
maximum output required time after clock (treg out ) using circuit timing from Appendix Standard cell libraries.
CAMs
Problem 7.30 Design a CAM with binary codded output address, which provides as output
address the first location containing the searched binary configuration, if any.
Problem 7.31 Design an associative memory, AM , implemented as a maskable and readable
CAM. A CAM is maskable if any of the m input bits can be masked using an m-bit mask word.
The masked bit is ignored during the comparison process. A CAM is readable if the full content
of the first matched location in sent to the data output.
Problem 7.32 Find examples for the inequality
number of possible names >>> number of actual names
which justify the use of the associative memory concept in digital systems.
7.8
Projects
Project 7.1 Let be the module system containing system1 and system2 interconnected through
the two-direction memory buffer module bufferMemory. The signal mode controls the sense of
192
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
the transfer: for mode = 0 system1 is in read mode and system2 in write mode, while for
mode = 1 system2 is in read mode and system1 in write mode. The module library provide the
memory block described by the module memory.
module system( input
[m-1:0] in1
,
input
[n-1:0] in2
,
output [p-1:0] out1
,
output [q-1:0] out2
,
input
clock
);
wire
[63:0] memOut1 ;
wire
[63:0] memIn1 ;
wire
[13:0]] addr1
;
wire
we1
;
wire
[255:0] memOut2 ;
wire
[255:0] memIn2 ;
wire
[11:0] addr2
;
wire
we2
;
wire
mode
; // mode = 0: system1 reads, system2 writes
// mode = 1: system2 reads, system1 writes
wire
[1:0]
com12, com21
;
system1 system1(in1, out1, com12, com21,
memOut1 ,
memIn1 ,
addr1
,
we1
,
mode
,
clock
);
system2 system2(in2, out2, com12, com21,
memOut2 ,
memIn2 ,
addr2
,
we2
,
clock
);
bufferMemory
bufferMemory(
memOut1 ,
memIn1 ,
addr1
,
we1
,
memOut2 ,
memIn2 ,
addr2
,
we2
,
mode
,
clock
);
endmodule
module memory #(parameter n=32, m=10)
(
output reg [n-1:0] dataOut
input
[n-1:0] dataIn
input
[m-1:0] readAddr
input
[m-1:0] writeAddr
input
we
input
enable
input
clock
,
,
,
,
,
,
);
reg [n-1:0] memory[0:(1 << m)-1];
always @(posedge clock) if (enable) begin
//
//
//
//
//
//
data output
data input
read address
write address
write enable
module enable
7.8. PROJECTS
193
if (we) memory[writeAddr] <= dataIn ;
dataOut <= memory[readAddr]
;
end
endmodule
Design the module bufferMemory.
Project 7.2 Design a systolic system for multiplying a band matrix of maximum width 16 with
a vector. The operands are stored in serial registers.
194
CHAPTER 7. MEMORIES:
FIRST ORDER, 1-LOOP DIGITAL SYSTEMS
Chapter 8
AUTOMATA:
Second order, 2-loop digital systems
In the previous chapter
the memory circuit were described discussing about
• how is built an elementary memory cell
• how applying all type of compositions the basic memory structures (flip-flops, registers, RAMs) can be obtained
• how the basic memory structures are in used real applications
In this chapter
the second order, two-loop circuits are presented with emphasis on
• defining what is an automaton
• the smallest 2-state automata, such as T flip-flop and JK flip-flop
• big and simple automata exemplified by the binary counters
• small and complex finite automata exemplified by the control automata
In the next chapter
the third order, three-loop systems are described taking into account the type of system
through which the third loop is closed:
• combinational circuit - resulting optimized design procedures for automata
• memory systems - supposing simplified control
• automata - with the processor as typical structure.
195
196
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
The Tao of heaven is impartial.
If you perpetuate it, it perpetuates you.
Lao Tzu1
Perpetuating the inner behavior is the
magic of the second loop.
The next step in building digital systems is to add a new loop over systems containing 1-OS.
This new loop must be introduced carefully so as the system remains stable and controllable.
One of the most reliable ways is to build synchronous structures, that means to close the loop
through a way containing a register. The non-transparency of registers allows us to separate
with great accuracy the current state of the machine from the next state of the same machine.
This second loop increases the autonomous behavior of the system including it. As we shall
see, in 2-OS each system has the autonomy of evolving in the state space, partially independent
from the input dynamics, rather than in 1-OS in which the system has only the autonomy of
preserving a certain state.
The basic structure in 2-OS is the automaton, a digital system with outputs evolving according to two variables: the input variable and a “hidden” internal variable named the internal
state variable, simply the em state. The autonomy is given by the internal effect of the state.
The behavior of the circuit output can not be explained only by the evolution of the input,
the circuit has an internal autonomous evolution that “memorizes” previous events. Thus the
response of the circuit to the actual input takes into account the more or less recent history. The
state space is the space of the internal state and its dimension is responsible for the behavioral
complexity. Thus, the degree of autonomy depends on the dimension of the state space.
register
❄
Uncloked Lathes
Unclocked Latch
Cloked Lathes
clock
1-OS
1-OS
1-OS
❄
✻
❄
clock
✻
1-OS
CLC
Cloked Lathes
clock
1-OS
✻
❄
CLC
0-OS
✻
a.
0-OS
✻
0/1-OS
❄
b.
1-OS
❄
Figure 8.1: The two type of 2-OS. a. The asynchronous automata with a hazardous loop over
a transparent latch. b. The synchronous automata with a edge clock controlled loop closed over a
non-transparent register.
An automaton is built closing a loop over a 1-OS represented by a collection of latches. The
loop can be structured using the previous two type of systems. Thus, there are two type of
automata:
• asynchronous automata, for which the loop is closed over unclocked latches, through
combinational circuit and/or unclocked latches as in Figure 8.1a
1
Quote from Tao Te King of Lao Tzu translated by Brian Browne Walker.
8.1. OPTIMIZING DFF WITH AN ASYNCHRONOUS AUTOMATON
197
• synchronous automata, having the loop closed through an 1-OS and all latches are clocked
latches connected on the loop in master-slave configurations (see Figure 8.1b).
Our approach will be focused on the synchronous automata, after considering only in the first
subchapter an asynchronous automaton used to optimize the internal structure of the widely
used flip-flop: DFF.
Definition 8.1 Formally, a finite automaton is an automaton (see Appendix G) having a finite
set of states. ⋄
Actually, all implementable automata are finite. Traditionally, this term is used to distinguish
a subset of automata whose behavior is described using a constant number of states. Even if
the input string is infinite, the behavior of the automaton is limited to a trajectory traversing a
constant (finite) number of states. A finite automaton will be an automaton having a random
combinational function for computing the next state. Therefore, a finite automaton is a complex
structure.
A “non-finite” automaton that is an automaton designed to evolve in a state space proportional with the length of the input string. Now, if the input string is infinite the number of
states must be also infinite. Such an automaton can be defined only if its transition function is
simple. Its combinational loop is a simple circuit even if it can be a big one. The “non-finite”
automaton has a number of states that does not affect the definition (see the following examples
of counters, for sum prefix automaton, ...). We classify the automata in two categories:
• “non-finite”, recursive defined, simple automata, called functional automata, or simply
automata
• non-recursive defined, complex automata, called finite automata.
We start this chapter with an example of asynchronous circuit, because of its utility and
because we intend to show how complex is the management of its behavior. We will continue
presenting only synchronous automata, starting with small automata having only two states
(the smallest state space). We will continue with simple, recursive defined automata and we will
end with finite automata, that are the most complex automata.
8.1
Optimizing DFF with an asynchronous automaton
The very important feature added by the master-slave configuration – that of edge triggering
the flip-flop – was paid by increasing two times the size of the structure. An improvement is
possible for DFF (the master-slave D flip-flop) using the structure presented in Figure 8.2, where
instead of 8 2-input NANDs and 2 invertors only 6 2-input gates are used. The circuit contains
three elementary unclocked latches: the output latch, with the inputs R’ and S’ commanded by
the outputs of the other two latches, L1 and L2. L1 and L2 are loop connected building up a
very simple asynchronous automaton with two inputs – D and CK – and two outputs – R’ and
S’.
The explanation of how this DFF, designed as a 2-OS, works uses the static values on the
inputs of the latches. For describing the process of switching in 1 the triplets (x,y,z) are used,
while for switching in 0 are used [x,y,z], where:
x : is the stable value in the set-up time interval (in a time interval, equal with tsu , before
the positive transition of CK)
y : is the stable value in the hold time interval (in a time interval of th , after the positive
transition of CK; the transition time, t+ is considered very small and is neglected)
198
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
z : is a possible value after the hold time interval (after th measured from the positive
transition of CK)
For the process of transition in 1 we follow the triplets (x,y,z) in Figure 8.2:
in set-up time interval : CK = 0 forces the values R’ and S’ to 1, does not matter what is
the value on D. Thus, the output latch receives passive values on both of its inputs.
in hold time interval : CK = 1 frees L1 and L2 to follow the signals they receive on their
inputs. The first order and the second order loops are now closed. L2 switches to S’ =
0, because of the 0 received from L1, which maintains its state because D and CK have
passive values and the output S’ of L2 reinforces its state to R’ = 1. The output latch is
then set because of S’ = 0.
after the hold time interval : the possible transition in 0 of D after the hold time does not
affect the output of the circuit, because the second loop, from L2 to L1, forces the output
R’ to 1, while L2 is not affected by the transition of its input to the passive value because
of D = 0. Now, the second loop allow the system to “ignore” the switch of D after the hold
time.
(0,0,1)
(1,1,1)
[1,1,1]
[0,0,0]
Async. Automaton
(1,1,0)
[0,0,1]
L1
D
L2
(0,1,1)
[0,1,1]
CK
(1,1,1) R’
S’ (1,0,0)
[1,0,0]
[1,1,1]
Q’
Q
Figure 8.2: The D flip-flop implemented as a 2-OS system. The asynchronous automaton
built up loop connecting two unclocked latches allows to trigger the output latch according to the input
data value available at the positive transition of clock.
For the process of transition in 0 we follow the triplets [x,y,z] in Figure 8.2:
in set-up time interval : CK = 0 forces the values R’ and S’ to 1, does not matter what is
the value on D. Thus, the output latch receives passive values on both of its inputs. The
output of L1 applied to L2 is also forced to 1, because of the input D = 0.
in hold time interval : CK = 1 frees L1 and L2 to follow the signals they receive on their
inputs. The first order and the second order loops are now closed. L1 switches to R’ = 0,
because of the 0 maintained on D. L2 does not change its state because the input received
from L1 has the passive value and the CK input switches also in the passive value. The
output latch is then reset because of R’ = 0.
8.2. TWO STATES AUTOMATA
199
after the hold time interval : the possible transition in 1 of D after the hold time does not
affect the state of the circuit, because 1 is a passive value for a NAND elementary latch.
The effect of the second order loop is to “inform” the circuit that the set signal was, and still
is, activated by the positive transition of CK and any possible transition on the input D must be
ignored. The asynchronous automaton L1 & L2 behaves as an autonomous agent who “knows”
what to do in the critical situation when the input D takes an active value in an unappropriate
time interval.
8.2
Two States Automata
The smallest two-state half-automata can be explored almost systematically. Indeed, there are
only 16 one-input two-state half-automata and 256 with two inputs. We choose only two of
them: the T flip-flop, the JK flip-flop, which are automata with Q = Y and f = g. For simple
2-operand computations 2-input automata can be used. One of them is the adder automaton.
This section ends with a small and simple universal automaton having 2 inputs and 2 states.
8.2.1
The Smallest Automaton: the T Flip-Flop
The size and the complexity of an automaton depends at least on the dimension of the sets
defining it. Thus, the smallest (and also the simplest) automaton has two states, Q = {0, 1}
(represented with one bit), one-bit input, T = {0, 1}, and Q = Y . The associated structure in
represented in Figure 8.3, where is represented a circuit with one-bit input, T, having a one-bit
register, a D flip-flop, for storing the 1-bit coded state, and a combinational logic circuit, CLC,
for computing the function f .
What can be the meaning of an one-bit “message”, received on the input T, by a machine
having only two states? We can “express” with the two values of T only the following things:
no op : T = 0 - the state of the automaton remains the same
switch : T = 1 - the state of the automaton switches.
T
T
❄❄
CLC
CK
CK
❄
❄
D
D
DF-F
T
DF-F
Q
a.
CK
✻
❄
Q
TF-F
Q
b.
❄
Q
Q
❄
c.
Figure 8.3: The T flip-flop. a. It is the simplest automaton because: has 1-bit state register (a
DF-F), a 2-input loop circuit (one as automaton input and another to close the loop), and direct output
from the state register. b. The structure of the T flip-flop: the XOR2 circuits complements the state is
T = 1. c. The logic symbol.
The resulting automaton is the well known T flip-flop. The actual structure of a T flip-flop
is obtained connecting on the loop a commanded invertor, i.e., a XOR gate (see Figure 8.3b).
The command input is T and the value to be inverted is Q, the state and the output of the
circuit.
200
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
This small and simple circuit can be seen as a 2-modulo counter because for T = 1 the output
“says”: 01010101... Another interpretation of this circuit is: the T flip-flop is a frequency divider.
Indeed, if the clock frequency is fCK , then the frequency of the signal received to the output Q
is fCK /2 (after each clock cycle the circuit comes back in the same state).
8.2.2
The JK Automaton: the Greatest Flip-Flop
The “next” automaton in an imaginary hierarchy is one having two inputs. Let’s call them J
and K. Thus, we can define the famous JK flip-flop. Also, the function of this automaton results
univocally. For an automaton having only two states the four input messages coded with J and
K will be compulsory:
no op : J = K = 0 - the flip-flop output does not change (the same as T = 0 for T flip-flop)
reset : J = 0, K = 1 - the flip-flop output takes the value 0 (specific for D flip-flop)
set : J = 1, K = 0 - the flip-flop output takes the value 1 (specific for D flip-flop)
switch : J = K = 1 - the flip-flop output switches in the complementary state (the same as
T = 1 for T flip-flop)
Only for the last function the loop acts specific for a second order circuit. The flip-flop
must “tell to itself” what is its own state in order “to knows” how to switch in the other state.
Executing this command the circuit asserts its own autonomy. The vagueness of the command
“switch” imposes a sort of autonomy to determine a precise behavior. The loop that assures
this needed autonomy is closed through two AND gates (see Figure 8.4a).
J
K
❄
CK
✻
❄
J
S
R
Q
❄
RSF-F
Q’
a.
K
JKF-F
Q’
❄
Q
❄
❄
Q’
Q
b.
Figure 8.4: The JK flip-flop. It is the simplest two-input automaton. a. The structure: the loop is
closed over a master-slave RSF-F using only two AN D2 . b. The logic symbol.
Finally, we solved the second latch problem. We have a two state machine with two command
inputs and for each input configuration the circuit has a predictable behavior. The JK flip-flop is
the best flip-flop ever defined. All the previous ones can be reduced to this circuit with minimal
modifications (J = K = T for T flip-flop or K ′ = J = D for D flip-flop).
8.2.3
∗ Serial Arithmetic
As we know the ripple carry adder has the size in O(n) and the depth also in O(n) (remember Figure
6.18). If we agree with the time in this magnitude order, then there is a better solution where a second
order circuit is used.
The best solution for the n-bit adder is a solution involving a small and simple automaton. Instead
of storing the two numbers to be added in (parallel) registers, as in the pure combinational solution, the
sequential solutions needs serial registers for storing the operands. The system is presented in Figure 8.5,
containing three serial registers (two for the operands and one for the result) and the adder automaton.
8.2. TWO STATES AUTOMATA
✲
IN
201
SRn
✲A
✲B
✲C
OUT
CK
✲
S
FA
C+
Q
IN
SRn
OUT
D
✛
DF-F
Adder automaton
✛
OUT
SRn
IN
✛
Figure 8.5: Serial n-bit adder. The state of the adder automaton has the value of the carry generated
adding the previous 2 bits received from the output of the two serial registers containing the operands.
The adder automaton is a two states automaton having in the loop the carry circuit of a full adder
(FA). The one-bit state register contains the carry bit from the previous cycle. The inputs A and B of
FA receive synchronously, at each clock cycle, bits having the same binary range from the serial registers.
First, LSBs are read from the serial registers. Initially, the automaton is in the state 0, that means
CR = 0. The output S is stored bit by bit in the third serial register during n clock cycles. The final
(n + 1)-bit result is contained in the output serial register and in the state register.
The operation time remains in the order of O(n), but the structure involved in computation becomes
the constant structure of the adder automaton. The product of the size, SADD (n), into the time, TADD (n)
is in O(n) for this sequential solution. Again, Conjecture 2.1 acts emphasizing the slowest solution as
optimal. Let us remember that for a carry-look-ahead adder, the fastest O(1) variant, the same product
was in O(n3 ). The price for the constant execution time is, in this example, in O(n2 ). I believe it is too
much. We will prefer architectural solutions which allow us to avoid the structural necessity to perform
the addition in constant time.
8.2.4
∗ Universal 2-input and 2-state automaton
Any binary (two-operand) simple operation on n-bit operands can be performed serially using a 2state automaton. The internal state of the automaton stores the “carry” information from one stage of
processing to another. In the adder automaton, just presented, the internal state is used to store the
carry bit generated adding the i-th bits of a number. It is used in the next stage for adding the (i + 1)-th
bits. This mechanism can be generalized, resulting an universal 2-input (for binary operation), 2-state
automaton (for “carry” bit).
Definition 8.2 An Universal 2-input (in1, in2), 2-state (codded by state[0]) automaton is a programmable structure using a 16-bit program word, {next state f unc[7 : 0], out f unc[7 : 0]}. It is defined
by the following Verilog code:
module univ_aut(
output
input
input
input
reg
assign
[7:0]
out
in1
in2
next_state_func
out_func
reset
clock
state;
out = out_func[{state, in2, in1}];
always @(posedge clock)
,
,
,
,
,
,
);
//
//
//
//
//
the
the
the
the
the
output
first operand
second operand
program for loop
program for output
202
CHAPTER 8. AUTOMATA:
if (reset)
else
endmodule
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
state <= 0;
state <= next_state_func[{state, in2, in1}];
⋄
The universal programmable automaton is implemented using two 3-input universal combinational
circuits, one for the output function and another for the loop function. The total number of automata
can be programmed on this structure is 216 (the total number of 16-bit “programs”). Most of them are
meaningless, but the simplicity of solution deserves our attention.2
8.3
Functional Automata: the Simple Automata
The smallest automata before presented are used in recursively extended configuration to perform similar functions for any n. From this category of circuits we will present in this section
only the binary counters. The next circuit will be also a simple one, having the definition independent by size. It is a sum-prefix automaton. The last subject will be a multiply-accumulate
circuit built with two simple automata serially connected.
8.3.1
Counters
The first simple automaton is a composition starting from one of the function of T flip-flop: the
counting. If one T flip-flop counts modulo-21 , maybe two T flip-flops will count modulo-22 and
so on. Seems to be right, but we must find the way for connecting many T flip-flops to perform
the counter function.
For the synchronous counter3 built with n T flip-flops, Tn−1 , . . . , T0 , the formal rule is very
simple: if IN C0 , then the first flip-flop, T0 , switches, and the i-th flip-flop, for i = 1, . . . , n − 1,
switches only if all the previous flip-flops are in the state 1. Therefore, in order to detect the
switch condition for i-th flip-flop an AN Di+1 must be used.
Definition 8.3 The n-bit synchronous counter, COU N Tn , has a clock input, CK, a command
input, IN C0 , an n-bit data output, Qn−1 , . . . Q0 , and an expansion output, IN Cn . If IN C0 = 1,
the active edge of clock increments the value on the data output (see Figure 8.6). ⋄
There is also a recursive, constructive, definition for COU N Tn .
Definition 8.4 An n-bit synchronous counter, COU N Tn is made by expanding a COU N Tn−1
with a T flip-flop with the output Qn−1 , and an AN Dn+1 , with the inputs IN C0 , Qn−1 , . . . , Q0 ,
which computes IN Cn (see Figure 8.6). COU N T1 is a T flip-flop and an AN D2 with the inputs
Q0 and IN C0 which generates IN C1 . ⋄
Example 8.1
∗
The Verilog description of a synchronous counter follows:
module sync_counter #(parameter n = 8)(output
output
input
t_reg
2
3
t_reg(
.out
.in
[n-1:0] out
inc_n
inc_0
reset
clock
(out)
,
(prefix_out[n-1:0]) ,
,
,
,
,
);
The processing element in Connection Machine [Hillis ’85] used a similar simple and small machine.
There exist also asinchronous counters. They are simpler but less performant.
8.3. FUNCTIONAL AUTOMATA: THE SIMPLE AUTOMATA
203
CK
❄
❄
IN C0
T
Tn−1
IN Cn−1
COU N Tn−1
Qn−2
Q
...
Q0
...
...
IN C0
IN Cn
❄
Qn−1
❄
Qn−2
❄
Q0
...
Figure 8.6: The synchronous counter. The recursive definition of a synchronous counter has
SCOU N T (n) ∈ O(n2 ) and TCOU N T (n) ∈ O(log n), because for the i-th range one TF-F and one AN Di
are added.
.reset
.clock
and_prefix
assign
endmodule
(reset)
(clock)
and_prefix( .out
.in
,
);
(prefix_out)
({out, inc_0})
,
);
inc_n = prefix_out[n];
module t_reg #(parameter n = 8)(
output
input
input
always @(posedge clock) if (reset)
else
endmodule
reg [n-1:0] out
[n-1:0] in
reset
clock
out <= 0;
out <= out ^ in;
,
,
,
);
The reset input is added because it is used in real applications. Also, a reset input is good in simulation
because makes the simulation possible allowing an initial value for the flip-flops (reg[n-1:0] out in
module t reg) used in design. ⋄
It is obvious that CCOU N T (n) ∈ O(1) because the definition for any n has the same, constant
size (in number of symbols used to write the Verilog description for it or in the area occupied
by the drawing of COU N Tn ). The size of COU N Tn , according to the Definition 4.4, can be
computed starting from the following iterative form:
SCOU N T (n) = SCOU N T (n − 1) + (n + 1) + ST
and results:
SCOU N T (n) ∈ O(n2 )
because of the AND gates network used to command the T flip-flop. The counting time is the
clock period. The minimal clock period is limited by the propagation time inside the structure.
It is computed as follows:
TCOU N T (n) = tpT + tpAN Dn + tSU ∈ O(log n)
where: tpT ∈ O(1) is the propagation time through the T flip-flop, tpAN Dn ∈ O(log n) is the
propagation time through the AN Dn (in the fastest version it is implemented using a tree of
AN D2 gates) gate and tSU ∈ O(1) is the set-up time at the input of T flip-flop.
204
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
In order to reduce the size of the counter we must find another way to solve the function
performed by the network of ANDs. Obviously, the network of ANDs is an AND prefix-network.
Thus, the problem could be reduced to the problem of the general form of prefix-network. The
optimal solution exists and has the size in O(n) and the time in O(log n) (see in this respect
the section 8.2).
Finishing this short discussion about counters must be emphasized the autonomy of this
circuit which consists in switching in the next state according to the current state. We “tell”
simply to the circuit “please count”, and the circuit know what to do. The loop allow “him to
know” how to behave.
Real applications uses more complex counters able to be initialized in any states or the count
in both ways, up and down. Such a counter is described by the following code:
module full_counter #(parameter n = 4)(output
input
input
always @(posedge clock)
if (reset)
else if (load)
else if (count) if (down)
else
else
endmodule
reg [n-1:0] out
[n-1:0] in
reset
load
down
count
clock
out
out
out
out
out
<=
<=
<=
<=
<=
0
in
out - 1
out + 1
out
,
,
,
,
,
,
);
;
;
;
;
;
The reset operation has the highest priority, and the counting operations have the lowest
priority.
8.3.2
∗ Accumulator Automaton
The accumulator automaton is a generalization of the counter automaton. A counter can add 1 to the
value of its state in each clock cycle. An accumulator automaton can add in each clock cycle any value
applied on its inputs.
Many applications require the accumulator function performed by a system which adds a string of
numbers returning the final sum and all partial results – the prefixes. Let be p numbers x1 , . . . , xp . The
sum-prefixes are:
y1 = x1
y2 = x1 + x2
y3 = x1 + x2 + x3
...
yp = x1 + x2 + . . . + xp .
This example of arithmetic automata generates at each clock cycle one prefix starting with y1 . The
initial value in the register Rm+n is zero. The structure is presented in Figure 8.7 and consists in an
adder, ADDm+n , two multiplexors and a state register, Rm+n . This automaton has 2m+n states and
computes sum prefixes for p = 2m numbers, each represented with n bits. The supplementary m bits are
needed because in the worst case adding two numbers of n bits results a number of n + 1 bits, and so
on, ... adding 2m n-bit numbers results, in the worst case, a n + m-bit number. The automaton must
be dimensioned such as in the worst case the resulting prefix can be stored in the state register. The
two multiplexors are used to initialize the system clearing the register (for acc = 0 and clear = 1), to
maintain unchanged the accumulated value (for acc = 0 and clear = 0), or to accumulate the n-bit
input value (for acc = 1 and clear = 0). It is obvious the accumulate function has priority: for acc =
1 and clear = 1 the automaton accumulates ignoring the clear command.
8.3. FUNCTIONAL AUTOMATA: THE SIMPLE AUTOMATA
205
in
❄
m
Am+n−1
...
n
❄ ❄
m+n
...
Bm+n−1
A0
B0
ADDm+n
Sm+n−1
...
S0
❄
1
❄
❄
1
❄
0
(m + n) × EM U X
✛
0
(m + n) × EM U X
✛
clear
acc
❄
Rm+n
clock
❄
out
Figure 8.7: Accumulator automaton. It can be used as sum prefix automaton because in each
clock cycle outputs a new value as a result of a sequential addition of a stream of signed integers.
The size of the systems depends on the speed of adder and can be found between O(m + n) (for ripple
carry adder) and O((m + n)3 ) (for carry-look-ahead adder).
It is evident that this automaton is a simple one, having a constant sized definition. The four
components are all simple recursive defined circuits. This automaton can be build for any number of
states using the same definition. In this respect this automaton is a “non-finite”, functional automaton.
8.3.3
∗ Sequential multiplication
Multiplication is performed sequentially by repeated additions and shifts. The n-bit multiplier is inspected
and the multiplicand is accumulated shifted according to the position of the inspected bit or bits. If in
each cycle one bit is inspected (radix-2 multiplication), then the multiplication is is performed in n cycles.
If 2 bits are inspected (radix-2 multiplication) in each cycle, then the operation is performed in n/2 cycles,
and so on.
∗ Radix-2 multiplication
The generic three-register structure for radix-2 multiplication is presented in the following Verilog module.
op
com
❄
✲
❄
op2
Combinatorial Logic
❄ ✻
❄
en
prod
❄ ✻
en
❄
ptoduct[2*n-1:0]
❄
op1
❄ ✻
en
❄
product[n-1:0]
Figure 8.8: Radix-2 sequential multiplier.
module rad2mult #(parameter n = 8)(output
input
input
[2*n-1:0]
[n-1:0]
[2:0]
product , // result output
op
, // operands input
com
, // command input
206
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
input
reg
wire
[n-1:0] op2, op1, prod
[n:0]
sum ;
assign
sum = prod + op2;
clock
);
; // multiplicand, multiplier, upper result
always @(posedge clock) if (com[2])
case(com[1:0])
2’b00: prod
<= 0
;
2’b01: op1
<= op
;
2’b10: op2
<= op
;
2’b11: {prod, op1} <= (op1[0] == 1) ?
// clear prod
// load multiplier
// load multiplicand
{sum, op1[n-1:1]} :
{prod, op1} >> 1 ; // multiplication step
endcase
assign product = {prod, op1};
endmodule
The sequence of commands applied to the previous module is:
initial begin
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
#2
com =
com =
op =
com =
op =
com =
com =
com =
com =
com =
com =
com =
com =
com =
$stop
3’b100
3’b101
8’b0000_1001
3’b110
8’b0000_1100
3’b111
3’b111
3’b111
3’b111
3’b111
3’b111
3’b111
3’b111
3’b000
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
end
In real application the first three steps can be merged in one, depending on the way the multiplier is
connected. The effective number of clock cycles for multiplication is n.
∗ Radix-4 multiplication
The time performance for the sequential multiplication is improved if in each clock cycle 2 bits are
considered instead of one. In order to keep simple the operation performed in each cycle a small and
simple two-state automaton is included in design.
op
com
❄
✲
❄
op2
❄
❄
Combinatorial Logic
❄ ✻
en
❄
prod
❄
❄ ✻
en
ptoduct[2*n-1:0]
CLC
❄
op1
❄
❄ ✻
✻
en
❄
state
op1[1:0]
product[n-1:0]
Figure 8.9: Radix-4 sequential multiplier.
8.3. FUNCTIONAL AUTOMATA: THE SIMPLE AUTOMATA
207
Let us consider positive integer multiplication. The design inspects by turn each 2-bit group of the
multiplier, m[i+1:i] for i = 0, 2, ... n-2, doing the following simple actions:
m[i+1:i] = 00 : adds 0 to the result and multiply by 4 the multiplicand
m[i+1:i] = 01 : adds the multiplicand to the result and multiply by 4 the multiplicand
m[i+1:i] = 10 : adds twice the multiplicand to the result and multiply by 4 the multiplicand
m[i+1:i] = 11 : subtract the multiplicand from the results, multiply by 4 the multiplicand, and sends
to the next cycle the information that the current value of multiplicand must be added
4 times to the result
The inter-cycle message is stored in the state of the automaton. In the initial cycle the state of the
automaton is ignored, while in the next stages it is added to the value of m[i+1:i].
module rad4mult #(parameter n = 8)(output
input
input
input
[2*n-1:0]
[n-1:0]
[2:0]
product
op
com
clock
, // result output
, // operands input
, // command input
);
reg
reg
reg
[n-1:0]
[n+1:0]
op2, op1
prod
state
; // multiplicand, mutiplier, upper part of result
; // upper part of result
; // the state register of a two-state automaton
reg
reg
[n+1:0]
nextProd
nextState
;
;
wire
[2*n+1:0]
next
;
/*
com = 3’b000 // nop
com = 3’b001 // clear prod register
com = 3’b010 // load op1
com = 3’b011 // load op2
com = 3’b101 // mult
com = 3’b110 // lastStep
*/
always @(posedge clock)
case(com)
3’b001: begin
prod
state
end
3’b010: op1
3’b011: op2
3’b101: begin
{prod, op1}
state
end
3’b110: if (state) prod
default prod
endcase
& initialize automaton
<= 0
<= 0
;
;
<=
<=
<=
<=
;
;
;
;
op
op
next
nextState
<= prod + op2
<= prod
;
;
assign next = {{2{nextProd[n+1]}}, nextProd, op1[n-1:2]};
// begin algorithm
always @(*) if (state) case(op1[1:0])
2’b00: begin
nextProd
=
nextState
=
end
2’b01: begin
nextProd
=
nextState
=
end
2’b10: begin
nextProd
=
nextState
=
end
prod + op2
0
;
;
prod + (op2 << 1) ;
0
;
prod - op2
1
;
;
208
CHAPTER 8. AUTOMATA:
2’b11:
else
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
begin
end
endcase
case(op1[1:0])
2’b00: begin
2’b01:
end
begin
2’b10:
end
begin
2’b11:
end
begin
nextProd
nextState
= prod
= 1
;
;
nextProd
nextState
= prod
= 0
;
;
nextProd
nextState
= prod + op2
= 0
;
;
nextProd
nextState
= prod + (op2 << 1) ;
= 0
;
nextProd
nextState
= prod - op2
= 1
;
;
end
endcase
// end algorithm
assign product = {prod[n-1:0], op1};
endmodule
The sequence of commands for n = 8 is:
initial begin
#2
#2
#2
#2
#2
#2
#2
#2
#2
com =
com =
op =
com =
op =
com =
com =
com =
com =
com =
com =
$stop
3’b001
3’b010
8’b1100_1111
3’b011
8’b1111_1111
3’b101
3’b101
3’b101
3’b101
3’b110
3’b000
;
;
;
;
;
;
;
;
;
;
;
;
end
The effective number of clock cycles for positive integer radix-4 multiplication is n/2 + 1.
Let’s now solve the problem multiplying signed integers. The Verilog description of the circuit is:
module signedRad4mult #(parameter n = 8)(
reg
reg
reg
reg
[n-1:0] op1
[n:0]
op2
[n:0]
prod
state
;
;
;
;
reg
reg
[n:0]
;
;
wire
/*
com =
com =
com =
com =
com =
[2*n:0] next
3’b000
3’b001
3’b010
3’b011
3’b100
nextProd
nextState
//
//
//
//
//
//
//
//
//
output
input
input
input
[2*n-1:0]
[n-1:0]
[2:0]
product
op
com
clock
, // result output
, // operands input
, // command input
);
signed mutiplier
signed multiplicand
upper part of the signed result
the state register of a two-state automaton
;
nop
clear prod register & initialize automaton
load op1
load op2 with one bit sign expansion
mult
8.3. FUNCTIONAL AUTOMATA: THE SIMPLE AUTOMATA
209
*/
always @(posedge clock)
case(com)
3’b001: begin
prod
<= 0
;
state
<= 0
;
end
3’b010: op1
<= op
;
3’b011: op2
<= {op[n-1], op}; // to allow left shift
3’b100: begin
{prod, op1} <= next
;
state
<= nextState
;
end
default prod
<= prod
;
endcase
assign next = {{2{nextProd[n]}}, nextProd, op1[n-1:2]};
// begin algorithm
always @(*) if (state) case(op1[1:0])
2’b00: begin
nextProd
=
nextState
=
end
2’b01: begin
nextProd
=
nextState
=
end
2’b10: begin
nextProd
=
nextState
=
end
2’b11: begin
nextProd
=
nextState
=
end
endcase
else
case(op1[1:0])
2’b00: begin
nextProd
=
nextState
=
end
2’b01: begin
nextProd
=
nextState
=
end
2’b10: begin
nextProd
=
nextState
=
end
2’b11: begin
nextProd
=
nextState
=
end
endcase
// end algorithm
assign product = {prod[n-1:0], op1};
endmodule
The sequence of commands for n = 8 is:
initial begin
#2
#2
#2
#2
#2
#2
#2
#2
end
com =
com =
op =
com =
op =
com =
com =
com =
com =
com =
$stop
3’b001
3’b010
8’b1111_0001
3’b011
8’b1111_0001
3’b100
3’b100
3’b100
3’b100
3’b000
;
;
;
;
;
;
;
;
;
;
;
prod + op2
0
;
;
prod + (op2 << 1) ;
0
;
prod - op2
1
;
;
prod
1
;
;
prod
0
;
;
prod + op2
0
;
;
prod - (op2 << 1) ;
1
;
prod - op2
1
;
;
210
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
The effective number of clock cycles for signed integer radix-4 multiplication is n/2.
8.3.4
∗ “Bit-eater” automaton
A very useful function is to search the bits of a binary word in order to find the positions occupied by the
1s. For example, inspecting the number 00100100 we find in 2 steps a 1 on the 5-th position and another
on the 2-nd position. A simple automaton does this operation in a number of clock cycles equal with
the number of 1s contained in its initial state. In Figure 8.10 is represented The “bit-eater” automaton
which is a simple machine containing:
in
n
✲
load
S0
next state
❄
❄
1
0
n × EM U X
n × XOR
✻
❄
✻
2
Rn
clock
state
✲
eat
❄
enable
out
DM U Xlog2 n
in
P En
out
zero
E’
✻
log2 n
❄
out
in
✻
❄
n bit
Figure 8.10: “Bit-eater” automaton. Priority encoder outputs the index of the most significant 1
in register, and the loop switches it into 0 using the demultiplexer to “point” it and a XOR to invert it.
• an n-bit state register
• a multiplexer used to initialize the automaton with the number to be inspected, if load = 1, then
the register takes the input value, else the automaton’s loop is closed
• a priority encoder circuit which computes the index of the most significant bit of the state and
activates the demultiplexer (E ′ = 0), if its enable input is activated, (eat = 1)
• an enabled decoder (or a demultiplexor) which decodes the value generated by the priority encoder
applying 1 only to the input of one XOR circuit if zero = 0 indicating at least one bit of the state
word is 1
• n 2-input XORs used to complement the most significant 1 of the state.
In each clock cycle after the initialization cycle the output takes the value of the index of the most
significant bit of state having the value 1, and the next state is computed clearing the pointed bit.
Another, numerical interpretation is: while state ̸= 0, the output of the automaton takes the integer
value of the base 2 logarithm of the state value, |log2 (state)|, and the next state will be next state =
state − |log2 (state)|. If state = 0, then n bit = 1, the output takes the value 0, and the state remains in
0.
8.3.5
∗ Sequential divisor
The sequential divisor circuit receives two n-bit positive integers, the dividend and the divisor, and returns
other n-bit positive integers: the quotient and the remainder. The sequential algorithm to compute:
dividend/divisor = quotient + remainder
8.3. FUNCTIONAL AUTOMATA: THE SIMPLE AUTOMATA
211
a sort of “trial & error” algorithm which computes the n bits of the quotient starting with quotient[n-1].
Then in the first step remainder = dividend - (divisor << (n-1)) is computed and if the result is
a positive number the most significant bit of the quotient is 1 and the operation is validated as the new
state of the circuit, else the most significant bit of the quotient is 0. Next step we try with divisor <<
(n-2) computing quotient[n-2], and so on until the quotient[o] bit is determined. The structure of
the circuit is represented in figure 8.11.
divisor
dividend
quotient[n-1]
❄❄
❄ ❄
subtract
sub[n]
sub[n:0]
✲
❄
❄
mux
1
com[0]
✲
❄
0
❄
0
com[1]
mux
0
❄❄
❄
✛
1
✲
❄
1
❄
mux
❄
mux
0
✲
1
❄
quotient[n-2:0]
❄
❄
❄
0
mux
1
❄
quotient
remainder
❄
❄
Figure 8.11: Sequential divisor.
The Verilod description of the circuit is:
module divisor #(parameter n = 8)( output
output
output
input
input
input
input
parameter
nop = 2’b00,
ld = 2’b01,
div = 2’b10;
wire
[n:0]
reg [n-1:0] quotient
reg [n-1:0] remainder
error
[n-1:0] dividend
[n-1:0] divisor
[1:0]
com
clk
sub;
assign error
= (divisor == 0) & (com == div)
assign sub = {remainder, quotient[n-1]} - {1’b0, divisor}
always @(posedge clk)
if (com == ld)
else if (com == div)
begin
end
begin
end
endmodule
,
,
,
,
,
,
);
;
;
quotient
remainder
<= dividend ;
<= 0
;
quotient
remainder
<= {quotient[n-2:0], ~sub[n]}
;
<= sub[n] ?
{remainder[n-2:0], quotient[n-1]} :
sub[n-1:0]
;
212
CHAPTER 8. AUTOMATA:
8.4
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
∗ Composing with simple automata
Using previously defined simple automata some very useful subsystem can be designed. In this section are
presented some subsystems currently used to provide solutions for real applications: Last-In First-Out
memory (LIFO), First-In-First-Out memory (FIFO), and a version of the multiply accumulate circuit
(MACC). All are simple circuits because result as simple compositions of simple circuits, and all are
expandable for any n . . . m, where n . . . m are a parameters defining different part of the circuit. For
example, the memory size and the word size are independent parameters is a FIFO implementation.
8.4.1
∗ LIFO memory
The LIFO memory or the stack memory has many applications in structuring the processing systems. It
is used both for building the control part of the system, or for designing the data section of a processing
system.
Definition 8.5 LIFO memory implements a data structure which consists of a string S = <s0, s1,
s2, ...> of maximum 2m n-bit recordings accessed for write, called push, and read, called pop, at the
same end, s0, called top of stack (TOS). ⋄
Example 8.2 Let be the stack S = <s0, s1, s2, ...>. It evolve as follows under the sequence of five
commands:
push a --> S = <a, s0, s1, s2, ...>
push b --> S = <b, a, s0, s1, s2, ...>
pop
--> S = <a, s0, s1, s2, ...>
pop
--> S = <s0, s1, s2, ...>
pop
--> S = <s1, s2, ...>
⋄
Real applications request additional functions for a LIFO used for expression evaluation. An minimally expanded set of functions for the LIFO S = <s0, s1, s2, ...> contains the following operations:
• nop: no operation; the contents of S in untouched
• write a: write a in TOS
<s0, s1, s2, ...> --> <a, s1, s2, ...>
used for unary operations; for example:
a = s0 + 1
the TOS is incremented and write back in TOS (pop, push = write)
• pop:
<s0, s1, s2, ...> --> <s1, s2, ...>
• popwr a: pop & write
<s0, s1, s2, ...> --> <a, s2, ...>
used for binary operations; for example:
a = s0 + s1
the first two positions in LIFO are popped, added and the result is pushed back into the LIFO
memory (pop, pop, push = pop, write)
• push a:
<s0, s1, s2, ...> --> <a, s0, s1, s2, ...>
A possible implementation of such a LIFO is presented in Figure 11.1, where:
• Register File is organized, using the logic surrounding it, as a 2m n-bit stream of words accessed
at TOS
• LeftAddrReg is a m-bit register containing the pointer to TOS = s0
• Dec is the decrement circuit pointing to s1
8.4. ∗ COMPOSING WITH SIMPLE AUTOMATA
213
• IncDec is the circuit which increment, decrement or do not touch the content of the register
LeftAddrReg as follows:
– increment for push
– decrement for pop or popwr
– keeps unchanged for nop or write
Its output is used to select the destination in Register File and to up date, in each clock cycle,
the content of the register LeftAddrReg.
❄
writeEnable
✲
com
✲
Dec
IncDec
✻
in
✲
rightAddr
rightOp
✲
destAddr
✲
result
✲
✲
stack1
✲
stack0
Register File
leftOp
leftAddr
LeftAddrReg
✻ ✻
reset
Figure 8.12: LIFO memory. The LeftAddrReg register is, in conjunction with IncDec commbinational circuit, an up/downn counter used as stack pointer to organize in Register File an expression
evaluation stack.
A Verilog description of the previously defined LIFO (stack) is:
module lifo #(‘include "0_parameters.v")
(
output [m-1:0] stack0, stack1,
input [m-1:0] in
,
input [2:0]
com
,
input
reset , clock );
/* The command codes
nop
= 3’b000, //
write = 3’b001, // we
pop
= 3’b010, // dec
popwr = 3’b011, // dec, we
push = 3’b101; // inc, we */
reg
wire
// The
assign
// The
always
[n-1:0] leftAddr; // the main pointer
[n-1:0] nextAddr;
increment/decrement circuit
nextAddr = com[2] ? (leftAddr + 1’b1) : (com[1] ? (leftAddr - 1’b1) : leftAddr);
address register for TOS
@(posedge clock) if (reset) leftAddr <= 0
;
else
leftAddr <= nextAddr;
// The register file
reg [m-1:0] file[0:(1’b1 << n)-1];
assign stack0 = file[leftAddr]
,
stack1 = file[leftAddr - 1’b1];
always @(posedge clock) if (com[0]) file[nextAddr] <= in;
endmodule
214
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
Faster implementations can be done using registers instead of different kind of RAMs and counters
(see the chapter SELF-ORGANIZING STRUCTURES: N-th order digital systems). For big stacks,
optimized solutions are obtained combining a small register implemented stack with a big RAM based
implementation.
8.4.2
∗ FIFO memory
The FIFO memory, or the queue memory is used to interconnect subsystems working logical, or both
logical and electrical, asynchronously.x
Definition 8.6 FIFO memory implements a data structure which consists in a string of maximum 2m nbit recordings accessed for write and read, at its two ends. Full and empty signals are provided indicating
the write operation or the read operation are not allowed. ⋄
✲
in
we
write counter
up
rst
✲
out
out
in
RAM
✲ w addr[n-1:0]
r addr[n-1:0]
✛
✻
clock
read counter
up
rst
✻
reset
write
✲ w addr[n-1:0] eq r addr[n-1:0]✛
w addr[n]
full
read
r addr[n]
empty
Figure 8.13: FIFO memory. Two pointers, evolving in the same direction, and a two-port RAM
implement a LIFO (queue) memory. The limit flags are computed combinational from the addresses used
to write and to read the memory.
A FIFO is considered synchronous if both read and write signals are synchronized with the same clock
signal. If the two commands, read and write, are synchronized with different clock signals, then the FIFO
memory is called asynchronous.
In Figure 8.13 is presented a solution for the synchronous version, where:
• RAM is a 2n m-bit words two-port asynchronous random access memory, one port for write to
the address w addr and another for read form the address r addr
• write counter is an (n + 1)-bit resetable counter incremented each time a write is executed; its
output is w addr[n:0], initially it is reset
• read counter is an (n + 1)-bit resetable counter incremented each time a read is executed; its
output is r addr[n:0], initially it is reset
• eq is a comparator activating its output when the least significant n bits of the two counters are
identical.
FIFO works like a circular memory addressed by two pointers (w addr[n-1:0] and r addr[n-1:0])
running on the same direction. If the write pointer after a write operation becomes equal with the read
8.4. ∗ COMPOSING WITH SIMPLE AUTOMATA
215
pointer, then the memory is full and the full signal is 1. If the read pointer after a read operation becomes
equal with the write pointer, then the memory is empty and the empty signal is 1. The n + 1-th bit in
each counter is used to differentiate between empty and full when w addr[n-1:0] and r addr[n-1:0]
are the same. If w addr[n] and r addr[n] are different, then w addr[n-1:0] = r addr[n-1:0] means
full, else it means empty.
The circuit used to compare the two addresses is a combinational one. Therefore, its output has a
hazardous behavior which affects the outputs full and empty. These two outputs must be used carefully
in designing the system which includes this FIFO memory. The problem can be managed because the
system works in the same clock domain (clock is the same for both ends of FIFO and for the entire
system). We call this kind of FIFO synchronous FIFO.
VeriSim 8.1 A Verilog synthesisable description of a synchronous FIFO follows:
module simple_fifo(output [31:0] out
output
empty
output
full
input
[31:0] in
input
write
input
read
input
reset
input
clock
wire
[9:0]
write_addr, read_addr;
counter write_counter(
read_counter(
.out
.reset
.count_up
.clock
.out
.reset
.count_up
.clock
dual_ram memory(.out
.in
.read_addr
.write_addr
.we
.clock
assign
eq
phase
empty
full
=
=
=
=
,
,
,
,
,
,
,
);
(write_addr
(reset
(write
(clock
(read_addr
(reset
(read
(clock
),
),
),
)),
),
),
),
));
(out
),
(in
),
(read_addr[8:0] ),
(write_addr[8:0]),
(write
),
(clock
));
read_addr[8:0] == write_addr[8:0] ,
~(read_addr[9] == write_addr[9]) ,
eq & phase
,
eq & ~phase
;
endmodule
module counter(output
input
input
input
reg [9:0]
out
,
reset
,
count_up,
clock
);
always @(posedge clock) if (reset)
out <= 0;
else if (count_up) out <= out + 1;
endmodule
module dual_ram(
reg
[63:0]
output [31:0] out
input
[31:0]
input
[8:0]
input
[8:0]
input
input
mem[511:0];
,
in
read_addr
write_addr
we
clock
,
,
,
,
);
216
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
assign
out = mem[read_addr]
;
always @(posedge clock) if (we) mem[write_addr] <= in
endmodule
;
⋄
An asynchronous FIFO uses two independent clocks, one for write counter and another for read
counter. This type of FIFO is used to interconnect subsystems working in different clock domains. The
previously described circuit is unable to work as an asynchronous FIFO. The signals empty and full
are meaningless, being generated in two clock domains. Indeed, write counter and read counter are
triggered by different clocks generating the signal eq with hazardous transitions related to two different
clocks: write clock and read clock. This signal can not be used neither in the system working with write
clock nor in the system working with read clock. Read clock is unable to avoid the hazard generated by
write clock, and write clock is unable to avoid the hazard generated by read clock. Special tricks must be
used.
8.4.3
∗ The Multiply-Accumulate Circuit
The functional automata can be composed in order to perform useful functions in a digital system.
Otherwise, we can say that a function can be decomposed in many functional units, some of them
being functional automata, in order to implement it efficiently. Let’s take the example of the MultiplyAccumulate Circuit (MACC) and implement it in few versions. It is mainly used to implement one of the
most important numerical functions performed in our digital machines, the scalar product of two vectors:
a1 × b1 + . . . + an × bn .
in1
in0
❄❄
load
❄❄
reg1
reg0
selects the number with less 1s
✠
❄
1
❄
mux
0
❄
✛
1
❄
mux
0
❄
✛
REDA1
❄
REDA0
❄
❄
COMP
❄
❄
✛
SHIFTER
✛
BIT EATER
start
AUTOMATON
❄
❄
ADDER
❄
outReg
✛
✛
✲
done
readOut
❄out
Figure 8.14:
We will offer in the following a solution involving two serially connected functional automata: an
accumulator automaton and “bits eater” automaton.
The starting idea is that the multiplication is also an accumulation. Thus we use an accumulator
automaton for implementing both operations, the multiplication and the sum of products, without any
loss in the execution time.
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
217
The structure of the multiply-accumulate circuit is presented in Figure 8.15 and consists in:
“bits eater” automaton – used to indicates successively the positions of the bits having the value 1
from the first operand ai ; it also points out the end of the multiplication (see Figure 8.10)
combinational shifter – shifts the second operand, bi , with a number of positions indicated by the
previous automaton
accumulate automaton – performs the partial sums for each multiplication step and accumulate the
sum of products, if it is not cleared after each multiplication (see Figure 8.7).
in
clock
load
n
❄
in
❄
❄
load
nop
“Bit-Eater” Automaton
data in
SHIFTER shift
out
✛
out
eat
n bit
✲
2n
❄
done
in
Accumulate Automaton
✲
clear
acc
clear
out
2n+m
❄
out
Figure 8.15: A multiply-accumulate circuit (MAC). A sequential version of MAC results
serially connecting an automaton, generating by turn the indexes of the binary ranges equal with 1 in
multiplier, with a combinational shifter and an accumulator.
In order to execute a multiplication only we must execute the following steps:
• load the “beat-eater” automaton with the first operand and clear the content of the output register
in accumulator automaton
• select to the input the second operand which remains applied to the input of the shifter circuit
during the operation
• wait for the end of operation indicated by the done output.
The operation is performed in a number of clock cycles equal with the number of 1s of the first operand.
Thus, the mean execution time is proportional with n/2. To understand better how this machine works
the next example will be an automaton which controls it.
If a MACC function is performed the clear of the state register of the accumulator automaton is
avoided after each multiplication. Thus, the register accumulates the results of the successive multiplications.
8.5
Finite Automata: the Complex Automata
After presenting the elementary small automata and the large and simple functional automata
it is the time to discuss about the complex automata. The main property of these automata
218
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
is to use a random combinational circuit, CLC, for computing the state transition function and
the output transition function. Designing a finite automaton means mainly to design two CLC:
the loop CLC (associated to the state transition function f ) and the output CLC (associated to
the output transition function g).
8.5.1
Basic Configurations
The half automaton is an automaton without the output function (see Figure 8.16a,b) defined
only for theoretical reasons (see Appendix G). The utility of this concept is due to the fact
that many optimization techniques are related only with the loop circuits of the automaton.
The main feature of the automaton is the autonomy and the half-automaton concept describes
especially this type of behavior.
X
❄ ❄
Q
loopCLC
❄
CK
✲
✲
✲
❄
halfAut
CK
X
✲
X
outCLC
halfAut
CK
X
Y
✲
halfAut
CK
❄
outCLC
outCLC
❄
❄
outReg
outReg
❄
e.
❄
d.
❄
✲
CK
outCLC
Y
✲
halfAut
❄
❄
c.
CK
b.
❄
✲
halfAut
stateReg
a.
X
X
Y
❄
f.
Y
Figure 8.16: Automata types. a. The structure of the half-automaton (A1/2 ), the no-output
automaton: the state is generated by the previous state and the previous input. b. The logic symbol
of half-automaton. c. Immediate Mealy automaton: the output is generated by the current state and
the current input. d. Immediate Moore automaton: the output is generated by the current state. e.
Delayed Mealy automaton: the output is generated by the previous state and the previous input. f.
Delayed Moore automaton: the output is generated by the previous state.
All kind of automata can be described starting from a half-automaton, adding only combinational (no loops) circuits and/or memory (one loop) circuits. In Figure 8.16 are presented all
the four types of automata:
Mealy automaton : results connecting to the “output” of an A1/2 the output CLC that re-
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
219
ceives also the input X (Figure 8.16c) and computes the output function g; a combinational
way occurs between the input and the output of this automaton allowing a fast response,
in the same clock cycle, to the input variation
Moore automaton : results connecting to the “output” of an A1/2 the output CLC (Figure
8.16d) that computes the output function g; this automaton reacts to the input signal in
the next clock cycle
delayed Mealy automaton : results serially connecting a register, R, to the output of the
Mealy automaton (Figure 8.16e); this automaton reacts also to the input signal in the
next clock cycle, but the output is hazard free because it is registered
delayed Moore automaton : results serially connecting a register, R, to the output of the
Moore automaton (Figure 8.16f); this automaton reacts to the input signal with a two
clock cycles delay.
Real applications use all the previous type of automata, because they react with different
delay to the input change. The registered outputs are preferred if possible.
8.5.2
Designing Finite Automata
The behavior of a finite automaton can be defined in many ways. Graphs, transition tables, flowcharts, transition V/K diagrams or HDL description are very good for defining the transition
functions f and g. All this forms provide non-recursive definitions. Thus, the resulting automata
has the size of the definition in the same order with the size of the structure. Therefore, the
finite automata are complex structures even when they have small size.
In order to exemplify the design procedure for a finite automaton let be two examples, one
dealing with a 1-bit input string and another related with a system built around the multiplyaccumulate circuit (MAC) previously described.
Example 8.3 The binary strings 1n 0m , for n ≥ 1 and m ≥ 1, are recognized by a finite halfautomaton by its internal states. Let’s define and design it. The transition diagram defining the
behavior of the half-automaton is presented in Figure 8.17, where:
reset
✇
q0
1
[10]
❫ ☛
1
q1
0
[11]
0
✲
❯ ☛
q3
[00]
✗
[01]
1
q2
⑥
0
Figure 8.17: Transition diagram. The transition diagram for the half-automaton which recognizes
strings of form 1n 0m , for n ≥ 1 and m ≥ 1. Each circle represent a state, each (marked) arrow represent
a (conditioned) transition.
• q0 - is the initial state in which 1 must be received, if not the the half-automaton switches
in q3 , the error state
220
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
• q1 - in this state at least one 1 was received and the first 0 will switch the machine in q2
• q2 - this state acknowledges a well formed string: one or more 1s and at least one 0 are
already received
• q3 - the error state: an incorrect string was received.
Q1
Q0
Q1
1
1
0
1
1
0
0
0
Q0
Q1 , Q +
0
{Q+
, Q+
}
1
0
Q1
X0
X0
0
✯+
f (Q1 , Q0 , X0 ) =
Q0
0 X0′
X0 X0 0
Q1 , Q 0
a.
X0 1
Q1
Q0
0
0
Q+
1
X0′
0
Q+
0
c.
b.
1
X0
Figure 8.18: VK transition maps. The VK transition map for the half-automaton used to recognize
1n 0m , for n ≥ 1 and m ≥ 1. a. The state transition function f . b. The VK diagram for the next most
significant state bit, extracted from the previous full diagram. c. The VK diagram for the next least
significant state bit.
clock
X0 reset
Q+
0
Q+
1
R
D
D-FF1
Q’
S
Q
Q1
✛✲ R
D
D-FF0
Q’
S
Q
Q0
Figure 8.19: A 4-state finite half-automaton. The structure of the finite half-automaton used
to recognize binary string belonging to the 1n 0m set of strings.
The first step in implementing the structure of the just defined half-automaton is to assign
binary codes to each state.
In this stage we have the absolute freedom. Any assignment can be used. The only difference
will be in the resulting structure but not in the resulting behavior.
For a first version let be the codes assigned int square brackets in Figure 8.17. Results the
transition diagram presented in Figure 8.18. The resulting transition functions are:
′ ′
Q+
1 = Q1 · X0 = ((Q1 · X0 ) )
′
′
′ ′
Q+
0 = Q1 · X0 + Q0 · X0 = ((Q1 · X0 ) · (Q0 · X0 ))
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
221
(The 1 from q0+ map is double covered. Therefore, it is taken into consideration as a “don’t
care”.) The circuit is represented in Figure 8.19 in a version using inverted gated only. The
2-bit state register is designed by 2 D flip-flops. The reset input is applied on the set input of
D-FF1 and on the reset input of D-FF0.
The Verilog behavioral description of the automaton is:
module rec_aut( output reg [1:0]
state
,
input
in
,
input
reset
,
input
clock
);
always @(posedge clock)
if (reset) state <= 2’b10;
else
case(state)
2’b00: state <= 2’b00
2’b01: state <= {1’b0, ~in}
2’b10: state <= {in, in}
2’b11: state <= {in, 1’b1}
endcase
endmodule
;
;
;
;
⋄
Example 8.4 ∗ The execution time of the MAC circuit is data dependent, depends on how many 1s contains the multiplicand. Therefore, the data flow through it has no a fix rate. The best way to interconnect
this version of MAC circuit supposes two FIFOs, one to its input and another to its output. Thus, a
flexible buffered way to interconnect MAC is provided.
clock
reset
IN FIFO
read
empty
(end)
✛
✛
❄
in
MAC
reset
clear
load
nop
done
✛
✛
✛
❄
✲
❄
Finite Automaton
reset
✛
✻
out
❄
full
write
✛
OUT FIFO
reset
✛
Figure 8.20: The Multiply-Accumulate System. The system consists in a multiply-accumulate
circuit (MAC), two FIFOs and a finite automaton (FA) controlling all of them.
A complex finite automaton must be added to manage the signals and the commands associated
with the three simple subsystems: IN FIFO, OUT FIFO, and MAC (see Figure 8.20). The flow-chart
describing the version for performing multiplications is presented in Figure 8.21, where:
q0 : wait first state – the system waits to have at least one operand in IN FIFO, clearing in the same
time the output register of the accumulator automaton, when empty = 0 reads the first operand
from IN FIFO and loads it in MAC
q1 : wait second state – if IN FIFO is empty, the system waits for the second operand
q2 : multiply state – the system perform multiplication while done = 0
222
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
q3 : write state – the system writes the result in OUT FIFO and read the second operand from IN FIFO
if full = 0 to access the first operand for the next operation, else waits while full = 1.
reset
q0
00
(00)
clear, nop
1
empty
0
load, read
q1
1
empty
01
(01)
0
nop
q2
1
q3
1
full
done
11
(10)
0
10
(11)
0
write, read
Figure 8.21: Flow chart describing a Mealy finite automaton. The flow-chart describes the
finite automaton FA from Figure 8.20, which controls MAC and the two FIFOs in MAC system. (The
state coding shown in parenthesis will be used in the next chapter.)
The flow chart can be translated into VK transition maps (see Figure 8.22) or in a Verilog description.
From the VK transition maps result the following equations describing the combinational circuits for the
loop (q1+ , q0+ ) and for the outputs.
′
Q+
1 = Q1 · Q0 + Q0 · empty + Q1 · f ull
′
′
′
′
Q+
0 = Q1 · Q0 + Q0 · done + Q1 · empty
clear = Q′1 · Q′0
nop = Q′1 · Q′0 + Q′1 · empty
load = Q′1 · Q′0 · empty ′
read = Q1 · Q′0 · f ull′ + Q′1 · Q′0 · empty ′
write = Q1 · Q′0 · f ull′
The resulting circuit is represented in Figure 8.23, where the state register is implemented using 2 D
flip-flops and the combinational circuits are implemented using a PLA.
If we intend to use a software tool to implement the circuit the following Verilog description is a must.
module macc_control(read
write
load
clear
nop
,
,
,
,
,
//
//
//
//
//
read from IN FIFO
write in OUT FIFO
load the multiplier in MAC
reset the output of MAC
stops the multiplication
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
Q1
Q0
Q1
11
01
10
00
Q0
0
Q1
Q0
empty’
done’
Q1
Q1
Q0
1
1
1
Q+
0
nop
clear
Q1
Q1
Q0
Q1
Q0
Q0
empty’
full’ empty’
load
empty
Q0
empty’
full
Q+
1
1
empty’
0
Q+
, Q+
1
0
Q1
1
done’ empty’
1
full
reference
Q0
223
full’
write
read
Figure 8.22: Veitch-Karnaugh transition diagrams. The transition VK diagrams for FA (see
+
Figure 8.20). The reference diagram has a box for each state. The state transition diagram, Q+
1 Q1 ,
contains in the same positions the description of the next state. For each output a diagram describe the
output’s behavior in the corresponding state.
empty
full
done
reset
clock
input
output
,
,
,
,
);
// IN FIFO is empty
// OUT FIFO is full
// the multiplication is concluded
empty, full, done, reset, clock;
read, write, load, clear, nop;
reg [1:0]
reg
state;
read, write, load, clear, nop;
parameter
wait_first
wait_second
multiply
write_result
=
=
=
=
// as variables
2’b00,
2’b01,
2’b11,
2’b10;
// THE STATE TRANSITION FUNCTION
always @(posedge clock)
if (reset) state <= wait_first;
else
case(state)
wait_first : if (empty)
else
wait_second : if (empty)
else
multiply
: if (done)
else
write_result: if (full)
else
endcase
state
state
state
state
state
state
state
state
<=
<=
<=
<=
<=
<=
<=
<=
wait_first;
wait_second;
wait_second;
multiply;
write_result;
multiply;
write_result;
wait_first;
// THE OUTPUT TRANSITION FUNCTION (MEALY IMMEDIATE)
always @(state or empty or full or done)
case(state)
wait_first :
if (empty)
{read, write, load, clear, nop} = 5’b00011;
224
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
empty
full
done
Q’
q0
D
+
q0
DF-F
read
Q
write
load
Q’
q1
D
+
q1
clear
nop
DF-F
Q
Half-automaton
clock
Figure 8.23: FA’s structure. The FA is implemented with a two-bit register and a PLA with 5 input
variables (2 for state bits, and 3 for the input sibnals), 7 outputs and 10 products.
wait_second :
multiply
:
write_result:
else
{read,
if (empty)
{read,
else
{read,
{read,
if (full)
{read,
else
{read,
write, load, clear, nop} = 5’b10111;
write, load, clear, nop} = 5’b00001;
write, load, clear, nop} = 5’b00000;
write, load, clear, nop} = 5’b00000;
write, load, clear, nop} = 5’b00000;
write, load, clear, nop} = 5’b11000;
endcase
endmodule
The resulting circuit will depend by the synthesis tool used because the previous description is “too”
behavioral. There are tools which will synthesize the circuit codding the four states using four bits ....!!!!!.
If we intend to impose a certain solution, then a more structural description is needed. For example,
the following “very” structural code which translate directly the transition equations extracted from VK
transition maps.
module macc_control(read
write
load
clear
nop
empty
full
done
reset
clock
input
,
,
,
,
,
,
,
,
,
);
//
//
//
//
//
//
//
//
read from IN FIFO
write in OUT FIFO
load the multiplier in MAC
reset the output of MAC
stops the multiplication
IN FIFO is empty
OUT FIFO is full
the multiplication is concluded
empty, full, done, reset, clock;
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
output
225
read, write, load, clear, nop;
reg [1:0]
state;
// as an actual register
// THE STATE TRANSITION FUNCTION
always @(posedge clock)
if (reset) state <= 2’b00;
else state <= {(state[1] & state[0] | state[0] & ~empty |
state[1] & full),
(~state[1] & state[0] | state[0] & ~done |
~state[1] & ~empty)};
assign
read
write
load
clear
nop
= ~state[1] & ~state[0] & ~empty |
state[1] & ~state[0] & ~full,
= state[1] & ~state[0] & ~full,
= ~state[1] & ~state[0] & ~empty,
= ~state[1] & ~state[0],
= ~state[1] & ~state[0] | ~state[1] & empty;
endmodule
The resulting circuit will be eventually an optimized form of the version represented in Figure 8.23
because instead a PLA, the current tools use an minimized network of gates. ⋄
The finite automaton has two distinct parts:
• the simple, recursive defined part, that consists in the state register; it can be minimized
only by minimizing the definition of the automaton
• the complex part, that consists in the PLA that computes functions f and g and this is
the part submitted to the main minimization process.
Our main goal in designing finite automaton is to reduce the random part of the automaton,
even if the price is to enlarge the recursive defined part. In the current VLSI technologies we
prefer big size instead of big complexity. A big sized circuit has now a technological solution,
but for describing very complex circuits we have not yet efficient solutions (maybe never).
8.5.3
∗ Control Automata: the First “Turning Point”
A very important class of finite automata is the class of control automata. A control automaton is
embedded in a system using three main connections (see Figure 8.24):
• the p-bit input operation[p-1:0] selects the control sequence to be executed by the control
automaton (it receives the information about “what to do”); it is used to part the ROM in 2p
parts, each having the same dimension; in each part a sequence of maximum 2n operation can be
“stored” for execution
• the m-bit command output, command[m-1:0], the control automaton uses to generate “the command” toward the controlled subsystem
• the n-bit input flags[q-1:0] the control automaton uses to receive information, represented by
some independent bits, about “what happens” in the controlled subsystems commanded by the
output command[m-1:0].
The size and the complexity of the control sequence asks the replacement of the PLA with a ROM,
at least for the designing and testing stages in implementing the application. The size of the ROM has
the magnitude order:
SROM (n, p, q) ∈ O(2n+p+q ).
In order to reduce the ROM’s size we start from the actual applications which emphasize two very
important facts:
226
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
✛
flags[q-1:0]
operation[p-1:0]
”What to do”
✲
CLC(ROM)
”What happens”
✛
❄
❄Q
state[n-1:0]
+
R
✛
command[m-1:0]
”The command”
Figure 8.24: Control Automaton. The functional definition of control automaton. Control means
to issue commands and to receive back signals (flags) characterizing the effect of the command.
1. the automaton can “store” the information about “what to do” in the state space, i.e., each current
state belongs to a path through the state space, started in one initial state given by the code used
to specify the operation
2. in most of the states the automaton tests only one bit from the flags[q-1:0] input and if not, a
few additional states in the flow-chart solve the problem in most of the cases.
✛
n
CLC(ROM)
✛
0
✛
nEM U X
S0
operation[n-1:0]
✻
✛
❄
1
1
R
✛
command[m-1:0]
t
✲
✲
✲
”0”
flags[q-1:0]
.
.
.
❄
M U Xq
✲
✲
MOD
TEST
NEXT
T
”1”
Figure 8.25: Optimized version of control automata. The flags received from the controlled
system have independent meaning considered in distinct cycles. The flag selected by the code TEST, T,
decides from what half of ROM the next state and output will be read.
Starting from these remarks the structure of the control automaton can be modified (see Figure 8.25).
Because the sequence is only initialized using the code operation[n-1:0], this code is used only for
addressing the first command line from ROM in a single state in which M OD = 1. For this feature we
must add n EMUXs and a new output to the ROM to generate the signal M OD. This change allows
us to use in a more flexible way the “storing space” of ROM. Because a control sequence can have the
dimension very different from the dimension of the other control sequence it is not efficient to allocate fix
size part of ROM for each sequence as in we did in the initial solution. The version presented in Figure
8.25 uses for each control sequence only as much of space as needed to store all lines of command.
The second modification refers to the input flags[q-1:0]. Because the bits associated with this
input are tested in different states, M U Xq selects in each state the appropriate bit using the t-bit field
TEST. Thus, the q − 1 bits associated to the input flags[q-1:0] are removed from the input of the
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
227
ROM, adding only t output bits to ROM. Instead of around q bits we connect only one, T, to the input
of ROM.
This new structure works almost the same as the initial structure but the size of ROM is very strongly
minimized. Now the size of the ROM is estimated as being:
SROM (2n ).
Working with the control automaton in this new version we will make another remark: the most part
of the sequence generated is organized in a linear sequence. Therefore, the commands associated to the
linear sequences can be stored in ROM at the successive addresses, i.e., the next address for ROM can be
obtained incrementing the current address stored in the register R. The structure represented in Figure
8.26 results. What is new in this structure is an increment circuit connected to the output of the state
register and a small combinational circuit that transcodes the bits M1 , M0 , T into S1 and S0 . There are
4 transition modes coded by M1 , M0 :
3
✛
n
CLC(ROM)
R
✛
nM U X4
S1 S0
1
0
✛
✛
✛
✛
operation[n-1:0]
✻✻
✛
t
command[m-1:0]
n
2
.
.
.
✲
✲
❄
MUXT
✲
✲
INC
MOD
TEST
flags[q-1:0]
2
JUMP
✲
T
❄
TC
Figure 8.26: The simplest Controller with ROM (CROM). The Moore form of control
automaton is optimized using an incremented circuit (INC) to compute the most frequent next address
for ROM.
• inc, codded by M1 M0 = 00: the next address for ROM results by incrementing the current address;
the selection code must be S1 S0 = 00
• jmp, codded by M1 , M0 = 01: the next address for ROM is given by the content of the one field to
the output of ROM; the selection code must be S1 S0 = 01
• cjmp, codded by M1 , M0 = 10: if the value of the selected flag, T, is 1, then the next address
for ROM is given by the content of the one field to the output of ROM, else the next address for
ROM results by incrementing the current address; the selection code must be S1 S0 = 0T
• init, codded by M1 , M0 = 11: the next address for ROM is selected by nM U X4 from the initialization input operation; the selection code must be S1 S0 = 1−
Results the following logic functions for the transcoder TC: S1 = M1 M0 , S0 = M1 T + M0 .
The output of ROM can be seen as a microinstruction defined as follows:
<microinstruction>::= <setLabel> <Command> <Mod> <Test> <useLabel>;
<command>::= <to be defined when use>;
<mod>::= jmp | cjmp | init | inc ;
<test>::= <to be defined when use>;
<setLabel>::= setLabel(<number>);
<useLabel>::= useLabel(<number>);
<number>::= 0 | 1 | ... | 9 | <number><number>;
228
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
This last version will be called CROM (Controller with ROM) and will be considered, in the present
approach, to have enough functional features to be used as controller for the most complex structures
described in this book.
Very important comment! The previous version of the control automaton’s structure is characterized by two processes:
• the first is the increasing of the structural complexity.
• the second is the decreasing of the dimension and of the complexity of the binary configuration
“stored” in ROM.
In this third step both, the size and the complexity of the system grows without any functional improvement. The only effect is reducing the (algorithmic) complexity of ROM’s content.
We are in a very important moment of digital system development, in which the physical complexity
starts to compensate the “symbolic” complexity of ROM’s content. Both, circuits and symbols, are
structures but there is a big difference between them. The physical structures have simple recursive
definitions. The symbolic content of ROM is (almost) random and has no simple definition.
We agree to grow a little the complexity of the physical structure, even the size, in order to create the
condition to reduce the effort to set up the complex symbolic content of ROM.
This is the first main “turning point” in the development of digital systems. We have here the first
sign about the higher complexity of symbolic structures. Using recursive defined objects the physical
structures are maintained at smaller complexity, rather than the symbolic structures, that must assume
the complexity of the actual problems to be solved with the digital machines. The previous defined
CROM structure is so thought as the content of ROM to be easy designed, easy tested and easy maintained
because it is complex. This is the first moment, in our approach, when the symbolic structure has more
importance than the physical structure of a digital machine.
Example 8.5 Let’s revisit the automaton used to control the MAC system. Now, because a more powerful
tool is available, the control automaton will perform three functions, multiply, multiply and accumulate,
no operation, codded as follows:
mult: op = 01,
macc: op = 11,
noop: op = 00.
The CROM circuit is actualized in Figure 8.27 with the word of ROM organized as follows:
<microinstruction>::= <setLabel><Command> <Mod> <Test> <useLabel>
<command>::= <c1> <c2> <c3> <c4> <c5>
<c1> ::= nop | <c2> ::= clear | <c3> ::= load | <c4> ::= read | <c5> ::= write | <mod>::= jmp | cjmp | init | inc
<test>::= empty | full | done | stop | n_done | n_empty
<setLabel>::= setLabel(<number>);
<useLabel>::= useLabel(<number>);
<number>::= 0 | 1 | ... | 9 | <number><number>;
The fields <c1> ... <c5> are one-bit fields takeing the value 0 for “-”. When nothing is specified,
then in the corresponding position is 0. The bit end is used to end the accumulation. If stop = 0 the macc
operation does not end, the system waits for a new pairs of numbers to be multiplied and accumulated.
The result is sent out only when stop = 1.
The function mult is defined in the flowchart from Figure 8.23 as a Mealy automaton. Because the
CROM automaton is defined as a Moore automaton the code sequence will be defined takeing into account
the Moore version of the control multiply-accumulate automaton. The function macc is defined in Figure
8.28 as a Moore automaton. The function nop consist in looping in the reset state waiting for a command
different from nop. The content of ROM has the following symbolic definition:
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
229
clock
3✛
✛
ROM
R
✛
MUX
S1 S0
nop
clear
load
read
write
✛
✛
✛
✛
✛
3
✲
1✛
0✛
✻ ✻
5
4
3 MUXT
2
1
0
reset
INC
5
jumpAddr
❄
✲
✲
✲
✲
✲
✲
2✛
mode
test
n empty
n done
empty
full
done
stop
✻
5
{0, 0, op[1], op[1], op[0]}
❄2
✲ TC
T
Figure 8.27: Using a CROM. A more complex control can be done for Multiply Accumulate System
using a CROM instead of a standard finite automaton.
// no operation
setLabel(0)
// multiplication
setLabel(1)
init;
nop
nop
setLabel(2)
nop
setLabel(3)
cjmp
setLabel(4)
cjmp
read
// multiply and accumulate
setLabel(5)
nop
setLabel(8)
nop
setLabel(6)
nop
setLabel(7)
cjmp
read
cjmp
cjmp
setLabel(9)
cjmp
jmp
setLabel(10)
cjmp
write
// 00000
clear
load
cjmp
done
full
write
cjmp
read
empty
setLabel(3);
setLabel(4);
jmp
empty
setLabel(1);
inc;
setLabel(2);
clear
load
cjmp
n_done
inc;
n_empty
stop
empty
loop8;
full
jmp
cjmp
read
empty
setLabel(7);
empty
setLabel(5);
inc;
setLabel(6);
setLabel(8);
setLabel(10);
setLabel(9);
setLabel(10);
setLabel(0);
setLabel(0);
//
//
//
//
//
//
//
//
//
//
//
q0
q1
q2
q3
q4
q5
q6
q9
q10!
q7
q8
The binary sequence is stored in ROM starting from the address zero with the line labelled as
setLabel(0). The sequence associated to the function mult has 6 lines because a Moore automaton
has usually more states when the equivalent Mealy version. For macc function the correspondence with
the state are included in commentaries on each line. An additional state (q10) occurs also here, because
this version of CROM can not consider jump addresses depending on the tested bits; only one jump
address per line is available.
230
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
init
clear, nop
1
empty
q0
0000
0
load, read, nop
nop
1
q1
0001
q2
0010
empty
0
q3
0011
1
0
done
q4
0100
read
q5
0101
1
empty
0
q6
0110
1
stop
0
q7
0111
1
full
0
q9
1001
1
write
empty
0
q8
1000
Figure 8.28: Control flowchart. The control flowchart for the function macc.
The binary image of the previous code asks codes for the fields acting on the loop. ⋄
Verilog descriptions for CROM
The most complex part in defining a CROM unit is the specification of the ROM’s content. There are
few versions to be used. One is to provide the bits using a binary file, another is to generate the bits
using a Verilog program. Let us start with the first version.
The description of the unit CROM implies the specification of the following parameters used for
dimensioning the ROM:
• comDim: the number of bits used to encode the command(s) generated by the control automaton;
it depends by the system under control
• adrDim: the number of bits used to encode the address for ROM; it depends on the number of
state of the control automaton
• testDim: the number of bits used to select one of the flags coming back from the controlled system;
it depends by the functionality performed by the entire system.
The following description refers to the CROM represented in Figure 8.26. It is dimensioned to generate
a 5-bit command, to have maximum 32 internal states, and to evolve according to maximum 8 flags
(the dimensioning fits with the simple application presented in the previous example). Adjusting these
parameters, the same design can by reused in different projects. Depending on the resulting size of the
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
231
ROM, its content is specified in various ways. For small sizes the ROM content can be specified by a
hand written file of bits, while for big sizes it must be generated automatically starting from a “friendly”
definition.
A generic Verilog description of the simple CROM already introduced follows:
module crom #(‘include "0_parameter.v")(output [comDim - 1:0]
command ,
input [addrDim - 1:0]
operation,
input [(1 << testDim) - 1:0] flags
,
input
reset
,
input
clock
);
reg [addrDim - 1:0] stateRegister;
wire
wire
wire
wire
wire
wire
[comDim + testDim + addrDim + 1:0] romOut ;
flag
;
[testDim - 1:0]
test
;
[1:0]
mode
;
[addrDim - 1:0]
nextAddr;
[addrDim - 1:0]
jumpAddr;
rom rom(.address (stateRegister),
.data
(romOut
));
assign {command, test, jumpAddr, mode} = romOut,
flag = flags[test];
mux4 addrSelMux(.out(nextAddr
),
.in0(stateRegister + 1
),
.in1(jumpAddr
),
.in2(operation
),
.in3(operation
),
.sel({&mode, (mode[1] & flag | mode[0])}));
always @(posedge clock) if (reset) stateRegister <= 0
;
else
stateRegister <= nextAddr;
endmodule
The simple uniform part of the previous module consists in two multiplexer, an increment circuit,
and a register. The complex part of the module is formed by a very small one (the transcoder) and a big
one the ROM.
From the simple only the addrSelMux multiplexor asks for a distinct module. It follows:
module mux4 #(‘include "0_parameter.v")(out, in0, in1, in2, in3, sel);
input [1:0]
sel
;
input [addrDim - 1:0] in0, in1, in2, in3;
output [addrDim - 1:0] out
;
reg
[addrDim - 1:0] out
;
always @(in0 or in1 or in2 or in3 or sel)
case(sel)
2’b00: out = in0;
2’b01: out = in1;
2’b10: out = in2;
2’b11: out = in3;
endcase
endmodule
The big complex part has a first version described by the following Verilog module:
232
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
module rom #(‘include "0_parameter.v")
(input [addrDim - 1:0]
address,
output [comDim + testDim + addrDim + 1:0] data
);
reg [comDim + testDim + addrDim + 1:0] mem [0:(1 << addrDim) - 1];
initial $readmemb("0_romContent.v", mem); // the fix content of the memory
assign data = mem[address]; // it is a read only memory
endmodule
The file 0 parameter.v defines the dimensions used in the project crom. It must be placed in the
same folder with the rest of the files defining the project. For our example its content is:
parameter comDim = 5,
addrDim = 5,
testDim = 3
The initial line loads in background, in a transparent mode, the memory module mem. The module
rom does not have explicit writing capabilities, behaving like a “read only” device. The synthesis tools
are able to infer from the previous description that it is about a ROM combinational circuit.
The content of the file 0 romContent.v is filled up according to the micro-code generated in Example
7.6. Obviously, after the first 4 line the our drive to continue is completely lost.
/*
/*
/*
/*
//
/*
/*
00
01
02
03
*/
*/
*/
*/
30 */
31 */
00000_000_00000_11
11000_011_00001_10
10100_000_00000_00
10000_011_00011_10
...
00000_000_00000_00
00000_000_00000_00
Obviously, after filling up the first 4 lines our internal drive to continue is completely lost. The full
solution asks for 270 bits free of error bits. Another way to generate them must be found!
Binary code generator
Instead of defining and writing bit by bit the content of the ROM, using a hand written file (in our
example 0 romContent.v), is easiest to design a Verilog description for a “machine” which takes a file
containing lines of microinstructions and translate it into the corresponding binary representation. Then,
in the module rom the line initial ...; must be substituted with the following line:
‘include "codeGenerator.v" // generates ROM’s content using to ’theDefinition.v’
which will act by including the the description of a loading mechanism for the memory mem. The
code generating machine is a program which has as input a file describing the behavior of the automaton.
Considering the same example of the control automaton for MAC system, the file theDefinition is a
possible input for our code generator. It has the following form:
// MAC control automaton
setLabel(0); init;
// multiplication
setLabel(1); nop;
clear;
nop;
load;
setLabel(2); nop;
cjmp;
setLabel(3); cjpm; done;
setLabel(4); cjmp; full;
read; write;
// multiply & accumulate
setLabel(5); nop;
clear;
cjmp;
read;
empty;
useLabel(3);
useLabel(4);
jmp;
empty;
useLabel(1);
inc;
useLabel(2);
cjmp;
empty;
useLabel(0);
useLabel(5);
8.5. FINITE AUTOMATA: THE COMPLEX AUTOMATA
setLabel(8);
setLabel(7);
nop;
cjmp;
read;
cjmp;
cjmp;
setLabel(9); cjmp;
jmp;
setLabel(10); cjmp;
write;
load;
notDone;
inc;
notEmpty;
stop;
empty
useLabel(8);
full;
jmp;
233
read;
inc;
useLabel(7);
useLabel(8);
useLabel(10);
useLabel(9);
useLabel(10);
useLabel(0);
The theDefinition file consist in a stream of Verilog tasks. The execution of these tasks generate
the ROM’a content.
The file codeGenerator.v “understand” and use the file theDefinition, whose content follows:
// Generate the binary content of the ROM
reg
nopReg
;
reg
clearReg
;
reg
loadReg
;
reg
readReg
;
reg
writeReg
;
reg [1:0] mode
;
reg [2:0] test
;
reg [4:0] address
;
reg [4:0] counter
;
reg [4:0] labelTab[0:31];
task endLine;
begin
mem[counter] =
{nopReg, clearReg, loadReg, readReg, writeReg, mode, test, address};
nopReg
= 1’b0;
clearReg = 1’b0;
loadReg = 1’b0;
readReg = 1’b0;
writeReg = 1’b0;
counter = counter + 1;
end
endtask
// sets labelTab in the first pass associating ’counter’ with ’labelIndex’
task setLabel; input [4:0] labelIndex; labelTab[labelIndex] = counter; endtask
// uses the content of labelTab in the second pass
task useLabel; input [4:0] labelIndex; begin address = labelTab[labelIndex];
endLine;
end
endtask
// external commands
task nop ; nopReg
= 1’b1; endtask
task clear; clearReg = 1’b1; endtask
task load ; loadReg = 1’b1; endtask
task read ; readReg = 1’b1; endtask
task write; writeReg = 1’b1; endtask
// transition mode
task inc ; begin mode = 2’b00; endLine; end endtask
task jmp ; mode = 2’b01; endtask
task cjmp; mode = 2’b10; endtask
task init; begin mode = 2’b11; endLine; end endtask
234
CHAPTER 8. AUTOMATA:
// flag selection
task empty ; test
task full
; test
task done
; test
task stop
; test
task notDone; test
task notEmpt; test
=
=
=
=
=
=
3’b000;
3’b001;
3’b010;
3’b011;
3’b100;
3’b101;
initial begin counter
nopReg
clearReg
loadReg
readReg
writeReg
‘include
‘include
end
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
endtask
endtask
endtask
endtask
endtask
endtask
= 0;
= 0;
= 0;
= 0;
= 0;
= 0;
"theDefinition.v"; // first pass
"theDefinition.v"; // second pass
The file theDefinition is included twice because if a label is used before it is defined, only at the
second pass in the memory labelTab the right value of a label will be found when the task useLabel is
executad.
8.6
∗ Automata vs. Combinational Circuits
As we saw, both combinational circuits (0-OS) and automata (2-OS) execute digital functions. Indeed,
there are combinational circuits performing addition or multiplication, but there are also sequential circuits performing the same functions. What is the correlation between a gates network and an automaton
executing the same function? What are the conditions in which we can transform a combinational circuit
in an automaton or conversely? The answer to this question will be given in this last section.
Let be a Mealy automaton, his two CLCs (LOOP CLC and OU T CLC), the initial state of the
automaton, q(t0 ) and the input sequence for the first n clock cycle: x(t0 ), . . . , x(tn−1 ). The combinational
circuit that generates the corresponding output sequence y(t0 ), . . . , y(tn−1 ) is represented in Figure 8.29.
Indeed, the first pair LOOP CLC, OU T CLC computes the first output, y(t0 ), and the next state, q(t1 )
to be used by the second pair of CLCs to compute the second output and the next state, and so on.
Example 8.6 The ripple carry adder (Figure 6.18) has as correspondent automaton the adder automaton
from the serial adder (Figure 8.5). ⋄
Should be very interesting to see how a complex problem having associated a finite automaton can
be solved starting from a combinational circuit and reducing it to a finite automaton. Let us revisit in
the next example the problem of recognizing strings from the set 1a 0b , for a, b > 0.
Example 8.7 The universal combinational circuit (see 2.3.1) is used to recognize all the strings having
the form:
x0 , x1 , . . . xi , . . . xn−1 ∈ 1a 0b
for a, b > 0, and a + b = n. The function performed by the circuit will be:
f (xn−1 , . . . , x0 )
which takes value 1 for the following inputs:
xn−1 , . . . , x0 = 0000 . . . 01
xn−1 , . . . , x0 = 000 . . . 011
xn−1 , . . . , x0 = 00 . . . 0111
...
xn−1 , . . . , x0 = 00011 . . . 1
xn−1 , . . . , x0 = 0011 . . . 11
xn−1 , . . . , x0 = 011 . . . 111
8.6. ∗ AUTOMATA VS. COMBINATIONAL CIRCUITS
235
initial state
q(t0 )
x(t0 )
✲
LOOP CLC
q(t1 )
x(t1 )
✲
x(t2 )
❄
LOOP CLC
q(t2 )
✲
❄
❄
LOOP CLC
❄
❄
OUT CLC
❄
y(t0 )
❄
OUT CLC
❄
✲
✲
y(t1 )
❄
OUT CLC
✲
y(t2 )
❄
.
.
.
q(tn )
x(tn )
✲
❄
LOOP CLC
❄
❄
OUT CLC
✲
y(tn )
✻
❄
Figure 8.29: Converting an automata into a combinational circuit. The conversion rule
from the finite (Mealy) automaton into a combinational logic circuit means to use a pair of circuits (LOOP
CLC, OUTPUT CLC) for each clock cycle. The time dimension is transformed in space dimension.
Any function f (xn−1 , . . . , x0 ) of n variables can be expressed using certain minterms from the set of
2n minterms of n variables. Our functions uses only n − 1 minterms from the total number of 2n . They
are:
m2i −1
for i = 1, . . . (n − 1), i.e., the functions takes the value 1 for m1 or m3 or m7 or m15 or . . ..
Figure 8.30 represents the universal circuits receiving as ”program” the string:
. . . 001000000010001010
where 1s corresponds to minterms having the value 1, and 0s to the minterms having the value 0.
Initially the size of the resulting circuit is too big. For an n-bit input string from x0 to xn−1 the circuits
contains 2n − 1 elementary multiplexors. But, a lot of EMUXs have applied 0 on both selected inputs.
They will generate 0 on their outputs. If the multiplexors generating 0 are removed and substituted with
connections to 0, then the resulting circuit containing only n(n − 1)/2 EMUXs is represented in Figure
8.31a.
The circuit can be more reduced if we take into account that some of them are identical. Indeed, on
the first line all EMUXs are identical an the third (from left to right) can do the ”job” of the first tree
circuits. Therefore, the output of the third circuit from the first line will be connected to the input of all
the circuits from the second line. Similarly, on the second line we will maintain only two EMUXs, and
so on on each line. Results the circuit from Figure 8.31b containing (2n − 1) EMUXs.
This last form consists in a serial composition made using the same combinational circuit: an EMUX
and an 2-input AND (the EMUX with the input 0 connected to 0). Each stage of the circuit receives one
input value starting with x0 . The initial circuit receives on the selected inputs a fix binary configuration
(see Figure 8.31b). It can be considered as the initial state of the automaton. Now we are in the position
to transform the circuit in a finite half-automaton connecting the emphasized module in the loop with a
2-bit state register (see Figure 8.32a).
The resulting half-automaton can be compared with the half-automaton from Figure 8.19, reproduced
in Figure 8.32b. Not-surprisingly they are identical. ⋄
To transform a combinational circuit in a (finite) automaton the associated tree (or trees) of EMUXs
must degenerate into a linear graph of identical modules. An interesting problem is: how many of
236
CHAPTER 8. AUTOMATA:
x0
x2
00
10
00
00
00
10
00
10
10
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
x1
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
x3
x4
s1w0
s1w0
xi
Figure 8.30: The Universal Circuit “programmed” to recognize 1a 0b . A full tree of 2n
EMUXs are used to recognize the strings belonging to 1a 0b . The “program” is applied on the selected
inputs of the first level of EMUXs.
x0
10
10
10
s1w0
s1w0
s1w0
0
s1w0
x1
0
s1w0
0
s1w0
x2
s1w0
s1w0
s1w0
x3
x4
10
s1w0
s1w0
xi
a.
x0
0
s1w0
x1
x2
x3
x4
0
s1w0
0
s1w0
10
10
s1w0
s1w0
s1w0
s1w0
s1w0
s1w0
xi
b.
Figure 8.31: Minimizing the Universal Circuit “programmed to recognize 1a 0b . a. The
first step of minimizing the full tree of 2n − 1 EMUXs to a tree containing 0.5n(n + 1) EMUXs. Each
EMUX selecting between 0 and 0 is substituted with a connection to 0. b. The minimal combinational
network of EMUXs obtained removing the duplicated circuits. The resulting network is a linear stream
of identical CLCs.
8.7. ∗ THE CIRCUIT COMPLEXITY OF A BINARY STRING
237
xi reset
clock
reset
xi
✲
clock
a.
❄0
s
1 0
w
✲
❄
D ✛ ✲
S
R
❄❄
s
1 0
w
D
D-FF1
D-FF0
Q
Q
q1
q0
+
q0
+
q1
❄
R
D
D-FF1
Q’
b.
S
✛✲ R
Q
D
D-FF0
Q’
q1
S
Q
q0
Figure 8.32: From a big and simple CLC to a small and complex finite automata.
a. The resulting half-automaton obtained collapsing the stream of identical circuits. b. Minimizing the
structure of the two EMUXs results a circuit identical with the solution provided in Figure 8.19 for the
same problem.
“programs”, P = mp−1 , mp−2 , . . . m0 , applied as “leaves” of Universal Circuit allows the tree of EMUXs
to be reduced to a linear graph of identical modules?
8.7
∗ The Circuit Complexity of a Binary String
Greg Chaitin taught us that simplicity means the possibility to compress. He expressed the complexity of
a binary string as being the length of the shortest program used to generate that string. An alternative
form to express the complexity of a binary string is to use the size of the smallest circuit used to generate
it.
Definition 8.7 The circuit complexity of a binary string P of length p, CCP (p), is the size of the
minimized circuit used to generate it. ⋄
Definition 8.8 The universal circuit used to generate any p-bit sting, pU-Generator, consists in a nUCircuit programmed with the string to be generated and triggered by a resetable counter (see Figure 8.33).
⋄
According to the actual content of the “program” P = mp−1 . . . m0 the nU-Circuit can be reduced to
a minimal size using techniques previously described in the section 2.3. The minimal size of the counter
is in O(log p) (the “first” proposal for an actual value is 11(1 + log2 p) + 5). Therefore, the minimal size
of pU-Generator, used to generate an actual string of p bits is the very precisely defined number CCP (p).
Example 8.8 Let us compute the circuit size of the following 16-bit strings:
P 1 = 0000000000000000
P 2 = 1111111111111111
P 3 = 0101000001010000
P 4 = 0110100110110001
For both, P1 and P2 the nU-Circuit is reduced to circuits containing no gates. Therefore, CC(P 1) =
CC(P 2) = 11(1 + log2 16) + 5 + 0 = 60.
For P3, applying the removing rules the first level of EMUXs in nU-Circuitis is removed and to the
inputs of the second level the following string is applied:
x′0 , x′0 , 0, 0, x′0 , x′0 , 0, 0
238
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
mp−1 mp−2
reset
start
✲
✲
m0
❄ ❄
❄
counter
clock
nU-Circuit
✲
n = log2 p
pU-Generator
❄
valid out
❄
out
Figure 8.33: The universal string generator. The counter, starting from zero, selects to the
output out the bits of the “program” one by one starting with m0 .
We continue applying removing and reducing rules. Results the inputs of the third level:
x′0 x2 , x′0 x2
The last level is removed because its inputs are identic. The resulting circuit is: x′0 x2 . It has the size 3.
Therefore CC(P 3) = 60 + 3 = 63.
For P4, applying the removing rules results the following string for the second level of EMUXs:
x′0 , x0 , x0 , x′0 , x0 , 1, 0, x′0
No removing or reducing rule apply for the next level. Therefore, the size of the resulting circuit is:
CC(P 4) = 1 + 7SEM U X + 88 = 103. ⋄
The main problem in computing the circuit complexity of a string is to find the minimal form
of a Boolean function. Fortunately, there are rigorous formal procedures to minimize logic functions
(see Appendix C.4 for some of them). (Important note: the entire structure of pU-Generator can be
designed composing and closing loops in a structure containing only elementary multiplexors and inverters.
In the langauge of the partial recursive functions these circuits perform the elementary selection and the
elementary increment. “Programming” uses only the function zero and the elementary increment. No
restrictions imposed by primitive recursiveness or minimalization are applied!)
An important problem rises: how many of the n-bit variable function are simple? The answer comes
from the next theorem.
Theorem 8.1 The weight, w, of Turing-computable functions, of n binary variables, in the set of the
formal functions decreases twice exponentially with n. ⋄
n
Proof Let be a given n. The number of formal n-input function is N = 22 , because the definition
are expressed with 2n bits. Some of this functions are Turing-computable. Let be these functions defined
by the compressed m-bit strings. The value of m depends on the actual function, but is realized the
condition that max(m) < 2n and m does not depends by n. Each compressed form of m bits corresponds
only to one 2n -bit uncompressed form. Thus, the ratio between the Turing-computable function of and
the formal function, both of n variables, is smaller than
n
max(w) = 2−(2
−max(m))
.
And, because max(m) does not depends by n, the ratio has the same form for no matter how big becomes
n. Results:
n
max(w) = const/22 .
8.8. CONCLUDING ABOUT AUTOMATA
239
⋄
A big question arises: how could be combinational circuits useful with this huge ratio between complex
circuits and simple circuits? An answer could be: potentially this ratio is very high, but actually, in the
real world of problems this ratio is very small. It is small because we do not need to compute too many
complex functions. Our mind is usually attracted by simple functions in a strange manner for which we
do not have (yet?) a simple explanation.
The Turing machine is limited to perform only partial recursive functions (see Chapter 9 in this
book). The halting problem is an example of a problem that has no solutions on a Turing machine (see
subsection 9.3.5???? in this book). Circuits are more powerful but they are not so easy“programmed”
as the Turing Machine, and the related systems. We are in a paradoxical situation: the circuit does not
need algorithms and Turing Machine is limited only to the problems that have an algorithm. But without
algorithms many solutions exist and we do not know the way to find them. The complexity of the
way to find of a solution becomes more and more important.
The working hypothesis will be that at the level of combinational (without autonomy) circuits the
segregation between simple circuits and complex programs is not productive. In most of cases the digital
system grows toward higher orders where the autonomy of the structures allow an efficient segregation
between simple and complex.
8.8
Concluding about automata
A new step is made in this chapter in order to increase the autonomous behavior of digital
systems. The second loop looks justified by new useful behaviors.
Synchronous automata need non-transparent state registers The first loop, closed for
gain the storing function, is applied carefully to obtain stable circuits. Tough restrictions can be
applied (even number of inverting levels on the loop) because of the functional simplicity. The
functional complexity of automata rejects any functional restrictions applied for the transfer
function associated to loop circuits. The unstable behavior is avoided using non-transparent
memories (registers) to store the state4 . Thus, the state switches synchronized by clock. The
output switches synchronously for delayed version of the implementation. The output is asynchronous for the immediate versions.
The second loop means the behavior’s autonomy Using the first loop to store the state
and the second to compute any transition function, a half-automaton is able to evolve in the
state space. The evolution depends by state and by input. The state dependence allows an
evolution even if the input is constant. Therefore, the automaton manifests its autonomy being
able to behave, evolving in the state space, under constant input. An automaton can be used
as “pure” generator of more or less complex sequence of binary configuration. the complexity
of the sequence depends by the complexity of the state transition function. A simple function
on the second loop determine a simple behavior (a simple increment circuit on the second loop
transforms a register in a counter which generate the simple sequence of numbers in the strict
increasing order).
Simple automata can have n states When we say n states, this means n can be very big,
it is not limited by our ability to define the automaton, it is limited only by the possibility to
implement it using the accessible technologies. A simple automata can have n states because the
state register contains log n flip-flops, and its second loop contains a simple (constant defined)
circuit having the size in O(f (log n)). The simple automata can be big because they can be
specified easy, and they can be generated automatically using the current software tools.
4
Asynchronous automata are possible but their design is restricted by to complex additional criteria. Therefore,
asynchronous design is avoided until stronger reason will force us to use it.
240
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
Complex automata have only finite number of states Finite number of states means:
a number of states unrelated with the length (theoretically accepted as infinite) of the input
sequence, i.e., the number of states is constant. The definition must describe the specific behavior
of the automaton in each state. Therefore, the definition is complex having the size (at least)
linearly related with the number of states. Complex automata must be small because they
suppose combinational loops closed through complex circuits having the description in the same
magnitude order with their size.
Control automata suggest the third loop Control automata evolve according to their
state and they take into account the signals received from the controlled system. Because the
controlled system receives commands from the same control automaton a third loop prefigures.
Usually finite automata are used as control automata. Only the simple automata are involved
directly in processing data.
An important final question: adding new loops the functional power of digital systems is
expanded or only helpful features are added? And, if indeed new helpful features occur, who is
helped by these additional features?
8.9
Problems
Problem 8.1 Draw the JK flip-flop structure (see Figure 8.4) at the gate level. Analyze the
set-up time related to both edges of the clock.
Problem 8.2 Design a JK FF using a D flip-flop by closing the appropriate combinational loop.
Compare the set-up time of this implementation with the set-up time of the version resulting in
the previous problem.
Problem 8.3 Design the sequential version for the circuit which computes the n-bit AND prefixes. Follow the approach used to design the serial n-bit adder (see Figure 8.5).
Problem 8.4 Write the Verilog structural description for the universal 2-input, 2-state programmable automaton.
Problem 8.5 Draw at the gate level the universal 2-input, 2-state programmable automaton.
Problem 8.6 Use the universal 2-input, 2-state automaton to implement the following circuits:
• n-bit serial adder
• n-bit serial subtractor
• n-bit serial comparator for equality
• n-bit serial comparator for inequality
• n-bit serial parity generator (returns 1 if odd)
Problem 8.7 Define the synchronous n-bit counter as a simple n-bit Increment Automaton.
Problem 8.8 Design a Verilog tester for the resetable synchronous counter from Example 4.1.
Problem 8.9 Evaluate the size and the speed of the counter defined in Example 4.1.
8.9. PROBLEMS
241
Problem 8.10 Improve the speed of the counter designed in Example 4.1 designing an improved
version for the module and prefix.
Problem 8.11 Design a reversible counter defined as follows:
module smartest_counter
(
output [n-1:0]
input
[n-1:0]
input
input
input
input
input
endmodule
#(parameter
out
,
in
,
reset
,
load
,
down
,
count
,
clock
);
n = 16)
//
//
//
//
//
preset value
reset counter to zero
load counter with ’in’
counts down if (count)
counts up or down
Problem 8.12 Simulate a 3-bit counter with different delay on its outputs. It is the case in
real world because the flop-flops can not be identical and their load could be different. Use it as
input for a three input decoder implemented in two versions. One without delays and another
assigning delays to the inverters and the the gates used to implement the decoder. Visualize the
outputs of the decoder in both cases and interpret what you will find.
Solution:
module dec_spyke;
reg
clock,
enable;
reg [2:0]
counter;
wire
out0, out1, out2, out3, out4, out5, out6, out7;
initial begin
clock = 0;
enable = 1;
counter = 0;
forever #20 clock = ~clock;
end
initial #400 $stop;
always @(posedge clock)
begin
if (counter[0])
if (&counter[1:0])
end
dmux dmux(
.out0
.out1
.out2
.out3
.out4
.out5
.out6
.out7
.in
.enable
(out0)
(out1)
(out2)
(out3)
(out4)
(out5)
(out6)
(out7)
(counter)
(enable)
counter[0] <= #3 ~counter[0];
counter[1] <= #4 ~counter[1];
counter[2] <= #5 ~counter[2];
,
,
,
,
,
,
,
,
,
);
242
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
initial $vw_dumpvars;
endmodule
module dmux(out0, out1, out2, out3, out4, out5, out6, out7, in, enable);
input
input
output
[2:0]
enable;
in;
out0, out1, out2, out3, out4, out5, out6, out7;
assign {out0, out1, out2, out3, out4, out5, out6, out7} = 1’b1 << in;
/*
not
#1 not0(nin2, in[2]);
not
#1 not1(nin1, in[1]);
not
#1 not2(nin0, in[0]);
not
#1 not3(in2, nin2);
not
#1 not4(in1, nin1);
not
#1 not5(in0, nin0);
nand
nand
nand
nand
nand
nand
nand
nand
*/
endmodule
#2
#2
#2
#2
#2
#2
#2
#2
nand0(out0,
nand1(out1,
nand2(out2,
nand3(out3,
nand4(out4,
nand5(out5,
nand6(out6,
nand7(out7,
nin2,
nin2,
nin2,
nin2,
in2,
in2,
in2,
in2,
nin1,
nin1,
in1,
in1,
nin1,
nin1,
in1,
in1,
nin0,
in0,
nin0,
in0,
nin0,
in0,
nin0,
in0,
enable);
enable);
enable);
enable);
enable);
enable);
enable);
enable);
Problem 8.13 Justify the reason for which the LIFO circuit works properly without a reset
input, i.e., the initial state of the address counter does not matter.
Problem 8.14 How behaves simple stack .
Problem 8.15 Design a LIFO memory using a synchronous RAM (SRAM) instead of an asynchronous one as in the embodiment represented in Figure 11.1.
Problem 8.16 Some applications ask the access to the last two data stored into the LIFO. Call
them tos, for the last pushed data, and prev tos for the previously pushed data. Both accessed
data can be popped from stack. Double push is allowed. The accessed data can be rearranged
swapping their position. Both, tos and prev tos can be pushed again in the top of stack. Design
such a LIFO defined as follows:
module two_head_lifo(
output [31:0] tos
,
output [31:0] prev_tos
,
input
[31:0] in
,
input
[31:0] second_in
,
input
[2:0]
com
, // the operation
input
clock
);
// the semantics of ’com’
parameter
nop
= 3’b000, // no operation
swap
= 3’b001, // swap the first two
8.9. PROBLEMS
243
pop
pop2
push
push2
push_tos
push_prev
=
=
=
=
=
=
3’b010,
3’b011,
3’b100,
3’b101,
3’110b,
3’b111;
//
//
//
//
//
//
pop tos
pop tos and prev_tos
push in as new tos
push ’in’ and ’second_in’
push ’tos’ (double tos)
push ’prev_tos’
endmodule
Problem 8.17 Write the Verilog description of the FIFO memory represented in Figure 8.13.
Problem 8.18 Redesign the FIFO memory represented in Figure 8.13 using a synchronous
RAM (SRAM) instead of the asynchronous RAM.
Problem 8.19 There are application asking for a warning signal before the FIFO memory is
full or empty. Sometimes full and empty come to late for the system using the FIFO memory.
For example, no more then 3 write operation are allowed, or no more than 7 read operation
are allowed are very useful in systems designed using pipeline techniques. The threshold for this
warning signals is good to be programmable. Design a 256 8-bit entries FIFO with warnings
activated using a programmable threshold. The interconnection of this design are:
module th_fifo(output
input
input
input
input
input
output
output
output
output
input
input
endmodule
[7:0]
[7:0]
[3:0]
[3:0]
out
,
in
,
write_th, //
read_th , //
write
,
read
,
w_warn , //
r_warn , //
full
,
empty
,
reset
,
clock
);
write threshold
read threshold
write warning
read warning
Problem 8.20 A synchronous FIFO memory is written or read using the same clock signal.
There are many applications which use a FIFO to interconnect two subsystems working with
different clock signals. In this cases the FIFO memory has an additional role: to cross from
the clock domain clock in into another clock domain, clock out. Design an asynchronous
FIFO using a synchronous RAM.
Problem 8.21 A serial memory implements the data structure of a fix length circular list. The
first location is accessed, for write or read operation, activating the input init. Each read or
write operation move the access point one position right. Design an 8-bit word serial memory
using a synchronous RAM as follows:
module serial_memory(
output
input
input
input
input
input
[7:0]
[7:0]
out
in
init
write
read
clock
,
,
,
,
,
);
endmodule
Problem 8.22 A list memory is a circuit in which a list can be constructed by insert, can be
accessed by read forward, read back, and modified by insert, delete. Design such a circuit
using two LIFOs.
244
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
Problem 8.23 Design a sequential multiplier using as combinational resources only an adder,
a multiplexors.
Problem 8.24 Write the behavioral and the structural Verilog description for the MAC circuit
represented in Figure 8.15. Test it using a special test module.
Problem 8.25 Redesign the MAC circuit represented in Figure 8.15 adding pipeline register(s)
to improve the execution time. Evaluate the resulting speed performance using the parameters
form Appendix E.
Problem 8.26 How many 2-bit code assignment for the half-automaton from Example 4.2 exist? Revisit the implementation of the half-automaton for four of them different from the one
already used. Compare the resulting circuits and try to explain the differences.
Problem 8.27 Ad to the definition of the half-automaton from Example 4.2 the output circuits
for: (1) error, a bit indicating the detection of an incorrectly formed string, (2)ack, another
bit indicating the acknowledge of a well formed sting.
Problem 8.28 Multiplier control automaton can be defined testing more than one input variable
in some states. The number of states will be reduced and the behavior of the entire system will
change. Design this version of the multiply automaton and compare it with the circuit resulted
in Example 4.3. Reevaluate also the execution time for the multiply operation.
Problem 8.29 Revisit the system described in Example 4.3 and design the finite automaton for
multiply and accumulate (MACC) function. The system perform MACC until the input FIFO
is empty and end = 1.
Problem 8.30 Design the structure of TC in the CROM defined in 4.4.3 (see Figure 8.26).
Define the codes associated to the four modes of transition (jmp, cjmp, init, inc) so as to
minimize the number of gates.
Problem 8.31 Design an easy to actualize Verilog description for the CROM unit represented
in Figure 8.26.
Problem 8.32 Generate the binary code for the ROM described using the symbolic definition
in Example 4.4.
Problem 8.33 Design a fast multiplier converting a sequential multiplier into a combinational
circuit.
8.10
Projects
Project 8.1 Finalize Project 1.2 using the knowledge acquired about the combinational and
sequential structures in this chapter and in the previous two.
Project 8.2 The idea of simple FIFO presented in this chapter can be used to design an actual
block having the following additional features:
• fully buffered inputs and outputs
• programmable thresholds for generating the empty and full signals
8.10. PROJECTS
245
• asynchronous clock signals for input and for output (the design must take into consideration
that the two clocks – clockIn, clockOut – are considered completely asynchronous)
• the read or write commands are executed only if the it is possible (reads only if not-empty,
or writes only if not-full).
The module header is the following:
module asyncFIFO #(‘include "fifoParameters.v")
(
output reg [n-1:0] out
,
output reg
empty
,
output reg
full
,
input
[n-1:0] in
,
input
write
,
input
read
,
input
[m-1:0] inTh
, // input threshold
input
[m-1:0] outTh
, // output threshold
input
reset
,
input
clockIn ,
input
clockOut);
// ...
endmodule
The file fifoParameters.v has the content:
parameter
n = 16
m = 8
,
// word size
// number of levels
Project 8.3 Design a stack execution unit with a 32-bit ALU. The stack is 16-level depth
(stack0, stack1, ... stack15) with stack0 assigned as the top of stack. ALU has the
following functions:
• add: addition
{stack0, stack1, stack2, ...} <= {(stack0 + stack1), stack2, stack3,...}
• sub: subtract
{stack0, stack1, stack2, ...} <= {(stack0 - stack1), stack2, stack3,...}
• inc: increment
{stack0, stack1, stack2, ...} <= {(stack0 + 1), stack1, stack2, ...}
• dec: decrement
{stack0, stack1, stack2, ...} <= {(stack0 - 1), stack1, stack2, ...},
• and: bitwise AND
{stack0, stack1, stack2, ...} <= {(stack0 & stack1), stack2, stack3,...}
• or: bitwise OR
{stack0, stack1, stack2, ...} <= {(stack0 | stack1), stack2, stack3,...}
• xor: bitwise XOR
{stack0, stack1, stack2, ...} <= {(stack0 ⊕ stack1), stack2, stack3,...}
• not: bitwise NOT
{stack0, stack1, stack2, ...} <= {(∼stack0), stack1, stack2, ...}
246
CHAPTER 8. AUTOMATA:
SECOND ORDER, 2-LOOP DIGITAL SYSTEMS
• over:
{stack0, stack1, stack2, ...} <= {stack1, stack0, stack1, stack2, ...}
• dup: duplicate
{stack0, stack1, stack2, ...} <= {stack0, stack0, stack1, stack2, ...}
• rightShift: right shift one position (integer division)
{stack0, stack1, ...} <= {({1’b0, stack0[31:1]}), stack1, ...}
• arithShift: arithmetic right shift one position
{stack0, stack1, ...} <= {({stack0[31], stack0[31:1]}), stack1, ...}
• get: push dataIn in top of stack
{stack0, stack1, stack2, ...} <= {dataIn, stack0, stack1, ...},
• acc: accumulate dataIn
{stack0, stack1, stack2, ...} <= {(stack0 + dataIn), stack1, stack2, ...},
• swp: swap the last two recordings in stack
{stack0, stack1, stack2, ...} <= {stack1, stack0, stack2, ...}
• nop: no operation
{stack0, stack1, stack2, ...} <= {stack0, stack1, stack2, ...}.
All the register buffered external connections are the following:
• dataIn[31:0] : data input provided by the external subsystem
• dataOut[31:0] : data output sent from the top of stack to the external subsystem
• aluCom[3:0] : command code executed by the unit
• carryIn : carry input
• carryOut : carry output
• eqFlag : is one if (stack0 == stack1)
• ltFlag : is one if (stack0 ¡ stack1)
• zeroFlag : is one if (stack0 == 0)
Project 8.4
Chapter 9
PROCESSORS:
Third order, 3-loop digital systems
In the previous chapter
the circuits having an autonomous behavior were introduced pointing on
• how the increased autonomy adds new functional features in digital systems
• the distinction between finite automata and uniform automata
• the segregation mechanism used to reduce the complexity
In this chapter
the third order, three-loop systems are studied presenting
• how a “smart register” can reduce the complexity of a finite automaton
• how an additional memory helps for designing easy controllable systems
• how the general processing functions can be performed loop connecting two appropriate automata forming a processor
In the next chapter
the fourth order, four-loop systems are suggested with emphasis on
• the four types of loops used for generating different kind of computational structures
• the strongest segregation which occurs between the simple circuits and the complex
programs
247
248
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
The soft overcomes the hard in the world
as a gentle rider controls a galloping horse.
Lao Tzu1
The third loop allows the softness of symbols to act
imposing the system’s function.
In order to add more autonomy in digital systems the third loop must be closed. Thus,
new effects of the autonomy are used in order to reduce the complexity of the system. One of
them will allow us to reduce the apparent complexity of an automaton, another, to reduce the
complexity of the sequence of commands, but, the main form of manifesting of this third loop
will be the control process.
❄
❄
❄
Automaton
a.
0-OS
✻
❄
❄
❄
2-OS
1-OS
✻
❄
❄
Automaton
2-OS
❄
automaton
❄
❄
Automaton
c.
easier to control
Memory
❄
automaton
❄
❄
Automaton
b.
simpler (& smalleer)
CLC
2-OS
Processor
2-OS
✻
❄
Figure 9.1: The three types of 3-OS machines. a. The third loop is closed through a combinational circuit resulting less complex, sometimes smaller, finite automaton. b. The third loop is closed
through memories allowing a simplest control. c. The third loop is closed through another automaton
resulting the Processor: the most complex and powerful circuit.
The third loop can be closed in three manners, using the three types of circuits presented in
the previous chapters.
• The first 3-OS type system is a system having the third loop closed through a combinational
circuit, i.e., over an automaton or a network of automata the loop is closed through a 0-OS
(see Figure 9.1a).
• The second type (see Figure 9.1b) has on the loop a memory circuit (1-OS).
• The third type connects in a loop two automata (see Figure 9.1c). This last type is typical
for 3-OS, having the processor as the main component.
1
Quote from Tao Te King of Lao Tzu translated by Brian Browne Walker.
9.1. IMPLEMENTING FINITE AUTOMATA WITH ”INTELLIGENT REGISTERS”
249
All these types of loops will be exemplified emphasizing a new and very important process
appearing at the level of the third order system: the segregation of the simple from the
complex in order to reduce the global (apparent) complexity.
9.1
Implementing finite automata with ”intelligent registers”
The automaton function rises at the second order level, but this function can be better implemented using the facilities offered by the systems having a higher order. Thus, in this section
we resume a previous example using the feature offered by 3-OS. The main effect of these new
approaches: the ratio between the simple circuits and the complex circuits grows, without spectacular changes in the size of circuits. The main conclusion of this section: more autonomy
means less complexity.
9.1.1
Automata with JK “registers”
In the first example we will substitute the state register with a more autonomous device: a
“register” made by JK flip-flops. The “JK register” is not a register, it is a network of parallel
connected simple automata. We shall prove that, using this more complicated flip-flop, the
random part of the system will be reduced and in most of big sized cases the entire size of the
system could be also reduced. Thus, both the size and the complexity diminishes when we work
with autonomous (“smart”) components.
But let’s start to disclose the promised magic method which, using flip-flops having two
inputs instead of one, offers a minimized solution for the combinational circuit performing the
loop’s function f . The main step is to offer a simple rule to substitute a D flip-flop with a JK
flip-flop in the structure of the automaton.
The JK flip-flop has more autonomy than the D flip-flop. The first is an automaton and
the second is only a storage element used to delay. The JK flip-flop has one more loop than
the D flip-flop. Therefore, for switching from a state to another the input signals of a JK flipflop accepts more “ambiguity” than the signal to the input of a D flip-flop. The JK flip-flop
transition can be commanded as follows:
• for 0 → 0 transition, JK can be 00 or 01, i.e., JK=0– (“–” means “don’t care” value)
• for 0 → 1 transition, JK can be 11 or 10, i.e., JK=1–
• for 1 → 0 transition, JK can be 11 or 01, i.e., JK=–1
• for 1 → 1 transition, JK can be 00 or 10, i.e., JK=–0
From the previous rule results the following rule:
• for 0 → A, JK=A–
• for 1 → A, JK=–A’.
Using these rules, each transition diagram for Q+
i can be translated in two transition diagrams
for Ji and Ki . Results: twice numbers of equations. But surprisingly, the entire size of the
random circuit which computes the state transition will diminish.
Example 9.1 The half-automaton designed in Example 8.4 is reconsidered in order to be designed using JK flip-flops instead of D flip-flops. The transition map from Figure 8.18 (reproduces in Figure 9.2a) is translated in JK transition maps in Figure 9.2b. The resulting circuit
is represented in Figure 9.2c.
250
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
Q1
Q1
Q0
1
1
0
1
1
0
0
0
X0 1
Q0
0 X0′
X0 X0 0
X0
0
Q+
, Q+
1
0
Q1 , Q0
a.
clock
Q0
Q1
Q1
–
0
–
0
X0′
X0′
Q0
–
✲
–
S
K1 = X0′
J1 = 0
Q1
Q0
reset
J
K
JK-FF1
Q
R
✛✲
S
J
K
JK-FF0
Q
Q’
R
✛
Q’
Q1
–
–
X0
0
Q0
J0 = Q1 X0 = (Q′1 + X0′ )′
0
X0
–
–
❄
❄
Q1
Q0
K0 = Q′1 X0 = (Q1 + X0′ )′
c.
b.
Figure 9.2: Translating D transition diagrams in the corresponding JK transition
diagrams. The transition VK diagrams for the JK implementation of the finite half-automaton used
to recognize binary string belonging to the 1n 0m set of strings.
The size of the random circuit which computes the state transition function is now smaller
(from the size 8 for D–FF to size 5 for JK–FF). The increased autonomy of the now used
flip-flops allows a smaller “effort” for the same functionality. ⋄
Example 9.2 ∗ Let’s revisit also Example 8.5. Applying the transformation rules results the VK diagrams from Figure 9.3 from which we extract:
Q1
Q0
Q1
11
01
10
00
Q0
Q1
-
Q0
J1
Q+
0
Q1
empty’
full’
K1
1
done’
empty’
Q+
1
Q1
-
Q1
Q0
empty’
full
reference
Q0
1
-
Q0
-
Q1
-
Q0
done
-
empty’
J0
-
K0
Figure 9.3: Translating D transition diagrams in the corresponding JK transition
diagrams. The transition VK diagrams for the JK implementation of the finite automaton used to
control MAC circuit (see Example 4.3).
J1 = Q0 · empty ′
K1 = Q′0 · f ull′
J0 = Q′1 · empty ′
K0 = Q1 · done
If we compare with the previous D flip-flop solution where the loop circuit is defined by
′
Q+
1 = Q1 · Q0 + Q0 · empty + Q1 · f ull
9.1. IMPLEMENTING FINITE AUTOMATA WITH ”INTELLIGENT REGISTERS”
251
′
′
′
′
Q+
0 = Q1 · Q0 + Q0 · done + Q1 · empty
results a big reduction of complexity. ⋄
empty
full
done
write
Q’
Q0
J
JKF-F
Q
read
K
load
clear
Q’
Q1
J
nop
JKF-F
Q
K
Half-automaton
clock
Figure 9.4: Finite automaton with smart “JK register”. The new implementation of FA
from Figure 8.20 using a ”JK register” as a state register. The associated half-automaton is simpler (the
corresponding PLA is smaller).
In this new approach, using a “smart register”, a part of loopCLC from the automaton built
with a true register was segregated in the uniform structure of the “JK register”. Indeed, the
size of loopCLC decreases, but the size of each flip-flop increases with 3 units (instead of an
inverter between S and R in D flip-flop, there are two AN D2 in JK flip-flop). Thus, in this new
variant the size of loopCLC decreases on the account of the size of the “JK register”.
This method acts as a mechanism that emphasizes more uniformities in the designing process
and allows to build for the same function a less complex and, only sometimes, a smaller circuit.
The efficiency of this method increases with the complexity and the size of the system.
We can say that loopCLC of the first versions has only an apparent complexity, because of
a certain quantity of “order” distributed, maybe hidden, among the effective random parts of
it. Because the “order” sunken in “disorder” can not be easy recognized we say that “disorder
+ order” means “disorder”. In this respect, the apparent complexity must be defined. The
apparent complexity of a circuit is reduced segregating the “hidden order”, until the circuit
remains really random. The first step is done. The next step, in the following subsection.
What is the explanation for this segregation that implies the above presented minimization
in the random part of the system? Shortly: because the “JK register” is a “smart register”
having more autonomy than the true register built by D flip-flops. A D flip-flop has only the
partial autonomy of staying in a certain state, instead of the JK flip-flop that has the autonomy
to evolve in the state space. Indeed, for a D flip-flop we must all the time “say” on the input
what will be the next state, 0 or 1, but for a JK flip-flop we have the vague, almost “evasive”,
command J = K = 1 that says: “switch in the other state”, without indicating precisely, as
for D flip-flop, the next state, because the JK “knows”, aided by the second loop, what is its
present state.
252
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
Because of the second loop, that informs the JK flip-flop about its own state,
the expressions for Ji and Ki do not depend by Qi , rather than Q+
i that depends on Qi .
Thus, Ji and Ki are simplified. More autonomy means less control. For this reason the PLA
that closes the third loop over a “JK register” is smaller than a PLA that closes the second loop
over a true register.
9.1.2
∗ Automata using counters as registers
Are there ways to “extract” more “simplicity” by segregation from the PLA associated to an automaton?
For some particular problems there is at least one more solution: to use a synchronous setable counter,
SCOU N Tn . The synchronous setable counter is a circuit that combines two functions, it is a register
(loaded on the command L) and in the same time it is a counter (counting up under the command U).
The load has priority before the count.
Instead of using few one-bit counters, i.e. JK flip-flops, one few-bit counter is used to store the state
and to simplify, if possible, the control of the state transition. The coding style used is the incremental
encoding (see E.4.3), which provides the possibility that some state transitions to be performed by
counting (increment).
Warning: using setable counters is not always an efficient solution!
Follows two example. One is extremely encouraging, and another is more realistic.
Example 9.3 The half-automaton associated to the codes assignment written in parenthesis in Figure
8.21 is implemented using an SCOU N Tn with n = 2. Because the states are codded using increment
encoding, the state transitions in the flow-chart can be interpreted as follows:
• in the state q0 if empty = 0, then the state code is incremented, else it remains the same
• in the state q1 if empty = 0, then the state code is incremented, else it remains the same
• in the state q2 if done = 1, then the state code is incremented, else it remains the same
• in the state q3 if f ull = 0, then the state code is incremented, else it remains the same
Half-automaton
empty’
load
full’
✲
✲
✲
✲
❄
0
1
2
MUX w
3
S0 S1
✻✻
✲
❄
I1
up
load
Q1
I0
SCOU N T2
O1
O0
reset
✛
reseet
clock
Q0
Figure 9.5: Finite half-automaton implemented with a setable counter. The last implementation of the half-automaton associated with FA from Figure 8.20 (with the function defined in Figure
8.21 where the states coded in parenthesis). A synchronous two-bit counter is used as state register. The
simple four-input MUX commands the counter.
Results the very simple (not necessarily very small) implementation represented in Figure 9.5, where
a 4-input multiplexer selects according to the state the way the counter switches: by increment (up = 1)
or by loading (load = 1).
Comparing with the half-automaton part in the circuit represented in Figure 9.4, the version with
counter is simpler, eventually smaller. But, the most important effect is the reducing complexity. ⋄
Example 9.4 This example is also a remake. The half-automaton of the automaton which controls the
operation macc in Example 4.6 will be implemented using a presetable counter as register. See Figure
9.1. IMPLEMENTING FINITE AUTOMATA WITH ”INTELLIGENT REGISTERS”
253
empty’
done
empty
end
full’
✲0
✲1
✲2
✲3
✲4
✲5
✲6
✲7
✲8
✲9
✲ 10
✲ 11
✲ 12
✲ 13
✲ 14
✲ 15
M U X4
W
U
L
I3 I2 I1 I0
SCOU N T4
R
Q3 Q2 Q1 Q0
clock
✛
reset
S3 S 2 S 1 S0
1
✻✻✻✻
empty
Figure 9.6: Finite half-automaton for controlling the function macc. The function was
previously implemented using a CROM in Example 4.6.
8.28 for the state encoding. The idea is to have in the flow-chart as many as possible transitions by
incrementing.
Building the solution starts from a SCOU N T4 and a M U X4 connected as in Figure 9.6. The multiplexer selects the counter’s operation (load or up-increment) in each state according to the flow-chart
description. For example in the state 0000 the transition is made by counting if empty = 0, else the state
remains the same. Therefore, the multiplexer selects the value of empty ′ to the input U of the counter.
The main idea is that the loading inputs I3 , I2 , I1 and I0 must have correct values only if in the
current state the transition can be made by loading a certain value in the counter. Thus, in the definition
of the logical functions associated with these inputs we have many “don’t care”s. Results the circuit
represented in Figure 9.6. The random part of the circuit is designed using the transition diagrams from
Figure 9.7.
The resulting structure has a minimized random part. We assumed even the risk of increasing the
recursive defined part of the circuit in order to reduce the random part of it. ⋄
Now, the autonomous device that allows reducing the randomness is the counter used as state register.
An adequate state assignment implies many transitions by incrementing the state code. Thus, the basic
function of the counter is many times involved in the state transition. Therefore, the second loop of the
system, the simple defined “loop that counts”, is frequently used by the third loop, the random loop.
The simple command UP, on the third loop, is like a complex “macro” executed by the second loop using
simple circuits. This hierarchy of autonomies simplifies the system, because at the higher level the loop
uses simple commands for complex actions. Let us remember:
• the loop over a true register (in 2-OS) uses the simple commands for the simplest actions: load 0
in D flip-flop and load 1 in D flip-flop
• the loop over a “JK register” (in 3-OS) uses beside the previous commands the following: no op
(remain in the same state!) and switch (switch in the complementary state!)
• the loop over a SCOU N Tn substitutes the command switch with the same simple expressed, but
more powerful, command increment.
254
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
q3
q2
- - - -
- - - -
1001
- - - -
- - - -
- - - -
0111
0001
001
- - - -
0011
- - - -
0000
- - - -
0010
0000
✲
empty
q1
+
+
q3
, . . . q0
q3
q3
-
-
-
-
❫
-
q2
1
-
q2
q0
+
q3
q1
+
q2
-
q0
q3
-
1
-
q1
q2
q0
+
q1
-
q3
-
1
1
1
q1
-
q2
q0
-
-
1
-
-
-
1
1
1
-
1
-
q0
+
q0
q1
Figure 9.7: Transition diagrams for the presetable counter used as state register.
The complex (random) part of the automaton is represented by the loop closed to the load input of the
presetable counter.
The “architecture” used on the third loop is more powerful than the two previous. Therefore, the effort
of this loop to implement the same function is smaller, having the simpler expression: a reduced random
circuit.
The segregation process is more deep, thus we imply in the designing process more simple, recursive
defined, circuits. The apparent complexity of the previous solution is reduced towards, maybe on, the
actual complexity. The complexity of the simple part is a little increased in order to “pay the price” for a
strong minimization of the random part of the system. The quantitative aspects of our small example are
not very significant. Only the design of the actual large systems offers a meaningful example concerning
the quantitative effects.
9.2
Loops closed through memories
Because the storage elements do not perform logical or arithmetical functions - they only store
- a loop closed through the 1-OS seems to be unuseful or at least strange. But a selective
memorizing action is used sometimes to optimize the computational process! The key is to
know what can be useful in the next steps.
The previous two examples of the third order systems belongs to the subclass having a
combinational loop. The function performed remains the same, only the efficiency is affected.
In this section, because automata having the loop closed through a memory is presented, we
expect the occurrence of some supplementary effects.
In order to exemplify how a trough memory loop works an Arithmetic & Logic Automaton
– ALA – will be used (see Figure 9.8a). This circuit performs logic and arithmetic functions
on data stored in its own state register called accumulator – ACC –, used as left operand and
on the data received on its input in, used as right operand. A first version uses a control
automaton to send commands to ALA, receiving back one flag: crout.
A second version of the system contains an additional D flip-flop used to store the value of
the CRout signal, in each clock cycle when it is enabled (E = 1), in order to be applied on the
CRin input of ALU. The control automaton is now substituted with a command automaton,
used only to issue commands, without receiving back any flag.
Follow two example of using this ALA, one without an additional loop and another with the
third loop closed trough a simple D flip-flop.
9.2. LOOPS CLOSED THROUGH MEMORIES
255
out
in
✻
The second
loop
✮
❄
❄
ALU
CRout
ACC
clock
✛
CRin ✛
Right
Left
✻
<func>
<carry>
crout
Arithmetic & Logic Automaton
CONTROL
AUTOMATON
✲
a.
out
in
✻
The second
loop
✮
❄
❄
ALU
CRout
ACC
clock
✛
CRin ✛
Right
Left
<func>
<carry>
COMMAND
AUTOMATON
✻
✶
Arithmetic & Logic Automaton
✲D
Q
D-FF
The third loop
clock
R E
✻✻<enable>
<reset>
b.
Figure 9.8: The third loop closed over an arithmetic and logic automaton. a. The basic
structure: a simple automaton (its loop is closed through a simple combinational circuit: ALU) working
under the supervision of a control automaton. b. The improved version, with an additional 1-bit state
register to store the carry signal. The control is simpler if the third loop “tells” back to the arithmetic
automaton the value of the carry signal in the previous cycle.
Version 1: the controlled Arithmetic & Logic Automaton
In the first case ALA is controlled (see Figure 9.8a) using the following definition for the
undefined fields of < microinstruction> specified in 8.4.3:
<command> ::= <func> <carry>;
<func> ::= and | or | xor | add | sub | inc | shl | right;
<test> ::= crout | -;
Let be the sequence of commands that controls the increment of a double-length number:
bubu
cucu
inc cjmp crout bubu // ACC = in + 1
right jmp cucu
// ACC = in
inc
// ACC = in + 1
...
The first increment command is followed by different operarion according to the value of
crout. If crout = 1 then the next command is an increment, else the next command is a
simple load of the upper bits of the double-length operand into the accumulator. The control
automaton decides according to the result of the first increment and behaves accordingly.
Version 2: the commanded Arithmetic & Logic Automaton
The second version of Arithmetic & Logic Automaton is a 3-OS because of the additional loop
closed through the D flip-flop. The role of this new loop is to reduce, to simplify and to speed up
the routine that performs the same operation. Now the microinstruction is actualized differently:
256
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
<command> ::= <func>;
<func> ::= and | or | xor | add | sub | inc | addcr | shl;
<test> ::= - ;
The field <test> is not used, and the control automaton can be substituted by a command
automaton. The field <func> is codded so as one of its bit is 1 for all arithmetic functions. This
bit is used to enable the switch of D-FF. A new function is added: addcr. The instruction add
the value of carry to the input in.
The resulting difference in how the system works is that in each clock cycle CRin is given by
the content of the D flip-flop. Thus, the sequence of commands that performs the same action
becomes:
inc
// ACC = in + 1
addcr // ACC = in + Q
In the two previous use of the arithmetic and logic automaton the execution time remains
the same, but the expression used to command the structure in the second version is shorter
and simpler. The explanation for this effect is the improved autonomy of the second version of
the ALA. The first version was a 2-OS but the second version is a 3-OS. A significant part of
the random content of the ROM from CROM can be removed by this simple new loop. Again,
more autonomy means less control. A small circuit added as a new loop can save much
from the random part of the structure. Therefore, this kind of loop acts as a segregation method.
Specific for this type of loop is that adding simple circuits we save random, i.e., complex,
structured symbolic structures. The circuits grow by simple physical structure and the complex
symbolic structures are partially avoided.
In the first version the sequence of commands are executed by the automaton all the time
in the same manner. In the second version, a simpler sequence of commands are executed
different, according to the processed data that impose different values in the carry flop-flop.
This “different execution” can be thought as an “interpretation”.
In fact, the execution is substituted by the interpretation, so as the apparent complexity of the
symbolic structure is reduced based on the additional autonomy due to the third structural loop.
The autonomy introduced by the new loop through the D flip-flop allowed the interpretation of
the commands received from the sequencer, according to the value of CR.
The third loop allows the simplest form of interpretation, we will call it static interpretation.
The fourth loop allows a dynamic interpretation, as we will see in the next chapter.
9.3
Loop coupled automata: the second ”turning point”
This last step in building 3-OS stresses specifically on the maximal segregation between the
simple physical structure and the complex symbolic structures. The third loop allows
us to make a deeper segregation between simple and complex.
We are in the point where the process of segregation between simple and complex physical
structures ends. The physical structures reach the stage from which the evolution can be done
only coupled with the symbolic structures. From this point a machine means: circuits that
execute or interpret bit configurations structured under restrictions imposed by the formal
languages used to describe the functionality to be performed.
9.3.1
∗ Push-down automata
The first example of loop coupled automata uses a finite automaton and a functional automaton: the stack
(LIFO memory). A finite complex structure is interconnected with an “infinite” but simple structure.
The simple and the complex are thus perfectly segregated. This approach has the role of minimizing
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
257
the size of the random part. More, this loop affects the magnitude order of the randomness, instead of
the previous examples (Arithmetic & Logic Automaton) in which the size of randomness is reduced only
by a constant. The proposed structure is a well known system having many theoretical and practical
applications: the push-down automaton.
X
❄
X
FA
✲
DIN
LIFO
✛
X
DOUT
✻
❄
{P U SH, P OP, −}
Y
Figure 9.9: The push-down automaton (PDA). A finite (random) automaton loop-coupled with
an “infinite” stack (a simple automaton) is an enhanced toll for dealing with formal languages.
Definition 9.1 The push-down automaton, PDA, (see Figure 9.9) built by a finite automaton loop
connected with a push-down stack (LIFO), is defined by the six-tuple:
P DA = (X × X ′ , Y × Y ′ × X, Q, f, g, z0 )
where:
X : is the finite alphabet of the machine; the input string is in X ∗
X’ : is the finite alphabet of the stack, X ′ = X ′ ∪ {z0 }
Y : is the finite output set of the machine
Y’ : is the set of commands issued by the finite automaton toward LIFO, {P U SH, P OP, −}
Q : is the finite set of the automaton states (i.e., |Q| ̸= h(max l(s)), where s ∈ X ∗ is received on the
input of the machine)
f : is the state transition function of the machine
f : X × X′ × Q → Q × X × Y ′
(i.e., depending on the received symbol, by the value of the top of stack (TOS) and by the automaton’s state, the automaton switches in a new state, a new value can be sent to the stack and the
stack receives a new command (PUSH, POP or NOP))
g : is the output transition function of the automaton - g : Q → Y
z0 : is the initial value of TOS. ⋄
Example 9.5 The problem to be solved is designing a machine that recognizes strings having the form
$x&y$, where $, & ∈ X and x, y ∈ X ∗ , X being a finite alphabet and y is the antisymmetric version of x.
The solution is to use a PDA with f and g described by the flow-chart given in Figure 9.10. Results
a five state, initial (in q0 ) automaton, each state having the following meaning and role:
q0 : is the initial state in which the machine is waiting for the first $
q1 : in this state the received symbols are pushed into the stack, excepting & that switches the automaton
in the next state
q2 : in this state, each received symbol is compared with T OS, that is poped on, while the received symbol
is not $; when the input is $ and T OS = z0 the automaton switches in q3 , else, if the received
symbols do not correspond with the successive value of the T OS or the final value of TOS differs
from z0 , the automaton switches in q4
258
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
q0
0
X =$
1
q1
1
X =&
PUSH X
q2
0
1
X = T OS
X =$
0
0
0
1
T OS = Z0
1
POP
q3
q4
OK
WRONG
Figure 9.10: Defining the behavior of a PDA. The algorithm detecting the antisymmetrical
sequences of symbols.
q3 : if the automaton is in this state the received string was recognized as a well formed string
q4 : if the automaton is in this state the received string was wrong. ⋄
The reader can try to solve the problem using only an automaton. For a given X set, especially for
a small set, the solution is possible and small, but the LOOP PLA of the resulting automaton will be a
circuit with the size and the form depending by the dimension and by the content of the set X. If only
one symbol is added or at least is changed, then the entire design process must be restarted from scratch.
The automaton imposes a solution in which the simple, recursive part of the solution is mixed up with
the random part, thus all the system has a very large apparent complexity. The automaton must store
in the state space what PDA stores in stack. You imagine how huge become the state set in a such crazy
solution. Both, the size and the complexity of the solution become unacceptable.
The solution with PDA, just presented, does not depend by the content and by the dimension of the
set X. In this solution the simple is well segregated from the complex. The simple part is the “infinite”
stack and the complex part is a small, five-state finite automaton.
9.3.2
The elementary processor
The most representative circuit in the class of 3-OS is the processor. The processor is maybe the
most important digital circuit because of its flexibility to compute any computable function.
Definition 9.2 The processor, P, is a circuit realized loop connecting a functional automaton
with a finite (control) automaton. ⋄
The function of a processor P is specified by the sequences of commands “stored” in the
loopCLC of the finite automaton used for control. (In a microprogrammed processor each sequence represents a microprogram. A microprogram consists in a sequence of microinstructions
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
259
each containing the commands executed by the functional automaton and fields that allow to
select the next microinstruction.)
In order to understand the main mechanisms involved by the third loop closed in digital
systems we will present initially only how an elementary processor works.
Definition 9.3 The elementary processor, EP, is a processor executing only one control sequence, i.e., the associated finite automaton is a strict initial automaton. ⋄
An EP performs only one function. It is a structure having a fix, nonprogrammable function.
The two parts of an EP are very different. One, the control automaton, is a complex structure,
while another, the functional automaton, is a simple circuit assembled from few recursively defined circuits (registers, ALU, file registers, multiplexors, and the kind). This strong segregation
between the simple part and the complex part of a circuit is the key idea on which the efficiency
of this approach is based.
Even on this basic level the main aspect of computation manifest. It is about control and
execution. The finite automaton performs the control, while the functional automaton executes
the logic or arithmetic operations on data. The control depends on the function to be computed
(the 2nd level loop at the level of the automaton) and on the actual data received by the system
(the 3rd level loop at the system level).
Example 9.6 Let’s revisit Example 5.2 in order to implement the function interpol using an
EP. The organization of the EP intepolEP is presented in Figure 9.11.
reset
✛
✛
asndAck
send
✲getAck
DFF
DFF
DFF
✲
✲
in
✲
DFF
✛
The functional automaton RALU
inR
❄
❄
8xEMUX
✲
✛
out
❄❄❄
✛
Regiser File
FA
✻
ALU
✲
✻
❄
❄
outR
get
❄
✛
✛
✛
2nd
Loop
3rd Loop
✛
✛
2nd
Loop
zero
interpolEP
Figure 9.11: The elementary processor interpolEP.
The functional automaton consists of a register file, an Arithmetic and Logic Unit and a 2way multiplexer. Such a simple functional automaton can be called RALU (Registers & ALU).
In each clock cycle two operands are read from the register file, they are operated in ALU, and
the result is stored back at destination register in the register file. The multiplexor is used to
260
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
load the register file with data. The loop closed from the ALU’s output to the MUX’s input is a
2nd level loop, because each register in the file register contains a first level loop.
The system has fully buffered connections. Synchronization signals (send, get, sendAck,
getAck) are connected through D–FFs (one-bit registers) and data through two 8-bit registers:
inR and outR.
The control of the system is performed by the finite automaton FA. It is initialized by the
reset signal, and evolve by testing three independent 1-bit signals: send (the sending external
subsystem provides a new input byte), get (the receiving external subsystem is getting the data
provided by the EP), zero (means the current output of ALU has the value 0). The last 1-bit
signal closes the third loop of the system. The transition function is described in the following
lines:
STATE
FUNCTION
TEST
EXT. SIGNAL
NEXT STATE
waitSend
reg0 <= inReg,
sendAck,
test
reg1 <= reg1,
waitGet
outReg <= reg1,
if (send)
else
if (zero)
else
if (get)
else
move1
move0
add
divide
reg2
reg1
reg1
reg1
next
next
next
next
next
next
next
next
next
next
<=
<=
<=
<=
getAck,
reg1,
reg0,
reg0 + reg2,
reg1 >> 1,
=
=
=
=
=
=
=
=
=
=
test;
waitSend;
add;
waitGet;
move1;
waitGet;
move0;
waitSend;
divide;
waitGet;
The outputs of the automaton provide the command for the acknowledge signals for the
external subsystems, and the internal command signals for RALU and output register outR. ⋄
Example 9.7 ∗ The EP structure is exemplified framed inside the simple system represented in Figure
9.12, where:
reset
✲
read
inFIFO
empty
stack com
✲
in2
✲
❄
in1
❄
mux
Elementary Processor
❄
❄
right
left
LIFO
✛
alu
❄
✛
✻
✻
❄
acc reg
out
✲
✻
✕
❄
write outFIFO
❄
Control
Automaton
full
the third loop
Figure 9.12: An example of elementary processor (EP). The third loop is closed between a
simple execution automaton (alu & acc reg) and a complex control automaton used to generate the
sequence of operations to be performed by alu and to control the data flow between EP and the associated
memory resources: LIFO, inFIFO, outFIFO.
inFIFO : provides the input data for EP when read = 1 if empty = 0
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
261
outFIFO : receives the output data generated by EP when write = 1 if full = 0
LIFO : stores intermediary data for EP if push = 1 and send back the last sent data if pop
Elementary Processor : is one of the simplest embodiment of an EP containing:
Control Automaton : a strict initial control automaton (see CROM from Figure 8.26)
alu : an Arithmeetic & Logic Unit
acc reg : an accumulator register, used as state register for Arithmetic & Logic Automaton which
is a functional automaton
mux : is the multiplexer for select the left operand from inFIFO or from LIFO.
The control automaton is a one function CROM that commands the functional automaton, receiving from
it only the carry output, cr, of the adder embedded in ALU.
The description of PE must be supplemented with the associated microprogramming language, as
follows:
<microinstruction> ::= <label> <command> <mod> <test> <next>;
<label> ::= <any string having maximum 6 symbols>;
<command> ::= <func> <inout>;
<mod> ::= jmp | cjmp | - ;
<test> ::= zero | notzero | cr | notcr | empty | nempty | full | nfull;
<next> ::= <label>;
<func> ::= left | add | half0 | half1 | - ;
<inout> ::= read | write | push | pop ;
where:
notcr: inverted cr
nempty: inverted empty
nfull: inverted full
left: acc_reg <= left
add: acc_reg <= left + acc_reg
half0: acc_reg <= {0, acc_reg[n-1:1]}
half1: acc_reg <= {1, acc_reg[n-1:1]}
left = read ? out(inFIFO) : out(LIFO)
and by default command are:
inc for <mode>
right: acc_reg <= acc_reg
The only microprogram executed by the previous described EP receives a string of numbers and generates another string of numbers representing the mean values of the successive two received numbers.
The numbers are positive integers. Using the previous defined microprogramming language results the
following microprogram:
microprogram mean;
bubu
read,
cucu
cjmp,
read,
half0;
out
write,
jmp,
one
half1,
endmicroprogram
cjmp,
empty,
add,
empty,
cucu;
cjmp,
bubu,
left;
cr,
one;
cjmp,
bubu;
jmp,
full,
out;
out;
262
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
On the first line PE waits for non-empty inFIFO; when empty becomes inactive the last left command
puts in the accumulator register the correct value. The second microinstruction PE waits for the second
number, when the number arrives the microprogram goes to the next line. The third line adds the content
of the register with the just read number from inFIFO. If cr = 1, the next microinstruction will be one,
else the next will be the following microinstruction. The fourth and the last microinstructions performs
the right shift setting the most significant bit on 0, i.e., the division for finishing to compute the mean
between the two received numbers. The line out send out the result when full = 0. The jump to bubu
restart again the procedure, and so on unending. The line one performs a right shift setting the most
significant bit on 1. ⋄
The entire physical structure of EP is not relevant for the actual function it performs. The
function is defined only by the loopCLC of the finite automaton. The control performed by the
finite automaton combines the simple functional facilities of the functional automaton that is
a simple logic-arithmetic automaton. The randomness is now concentrated in the structure of
loopCLC which is the single complex structure in the system. If loopCLC is implemented as a
ROM, then its internal structure is a symbolic one. As we said at the beginning of this section,
at the level of 3-OS the complexity is segregated in the symbolic domain. The complexity is
driven away from the circuits being lodged inside the symbolic structures supported by ROM.
The complexity can not be avoided, it can be only transferred in the more controllable space of
the symbolic structures.
9.3.3
Executing instructions vs. interpreting instructions
A processor is a machine which composes & loops functions performed by elementary
processors. Let us call them elementary computations or, simply, instructions. But now it
is not about composing circuits. The big difference from a physical composition or a physical
looping, already discussed, is that now the composition and looping are done ”in the symbolic
domain”.
As we know, an EP computes a function of variables received from an external sub-system
(in the previous example from inFIFO), and sends the result to an external sub-system (in the
previous example to outFIFO). Besides input variables a processor receives also functions. The
results are stored sometimes internally or in specific external resources (for example a LIFO
memory), and only at the end of a complex computation a result or a partial result is outputed.
The ”symbolic composition” is performed applying the computation g on the results of computations hm , . . . h0 . Let’s call now g, hi , or other similar simple computations, instructions.
The ”symbolic looping” means to apply the same string of instructions to the same variables
as many time as needed.
Any processor is characterized by its instruction set architecture (ISA). As we mentioned,
an instruction is equivalent with an elementary computation performed by an EP, and its code
is used to specify:
• the operation to be performed (<op code>)
• sometimes an immediate operand, i.e., a value known at the moment the computation is
defined (<value>),
therefore, in the simplest cases instruction ::= <op code> <value>
A program is a sequence of instructions allowing to compose and to loop more or less complex
computations.
There are two ways to perform an instruction:
• to execute it: to transcode op code in one or many elementary operations executed in one
clock cycle
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
263
• to interpret it: to expand op code is a sequence of operations performed in many clock
cycles.
Accordingly, two kind of processors are defined:
• executing processors
• interpreting processors.
elementary functions
✇
storage elements
VARIABLES
composing & looping
✢
PROCESSOR
storage elements
PROGRAM
COUNTER
DATA
&
PROGRAMS
ALU
NEXT PC
EXECUTION UNIT
or
INTERPRETATION UNIT
❨
decoding or µ-composing & µ-looping
Figure 9.13: The processor (P) in its environment. P works loop connected with an external
memory containing data and programs. Inside P elementary function, applied to a small set of very
accessible variables, are composed in linear or looped sequences. The instructions read from the external
memory are executed in one (constant) clock cycle(s) or they are interpreted by a sequence of elementary
functions.
In Figure 9.13 the processing module is framed in a typical context. The data to be computed
and the instructions to be used perform the computation are stored in a RAM module (see in
Figure 9.13 DATA & PROGRAMS). PROCESSOR is a separate unit used to compose and to
loop strings of instructions. The internal resources of a processor consists, usually, in:
• a block to perform elementary computations, containing:
– an ALU performing at least simple arithmetic operations and the basic logic operations
– a memory support for storing the most used variable
• the block used to transform each instruction in an executable internal mico-code, with two
possible versions:
– a simple decoder allowing the execution of each instruction in one clock cycle
– a microprogrammed unit used to ”expand” each instruction in a microprogram, thus
allowing the interpretation of each instruction in a sequence of actions
• the block used to compose and to loop by:
– reading the successive instructions organized as a program (by incrementing the PROGRAM COUNTER register) from the external memory devices, here grouped under
the name DATA & PROGRAMS
264
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
– jumping in the program space (by adding signed value to PROGRAM COUNTER)
In this section we introduce only the executing processors (in Chapter 11 the interpreting
processor will be used to exemplify how the functional information works).
Informally, the processor architecture consists in two main components:
• the internal organization of the processor at the top level used to specify:
– how are interconnected the top levels blocks of processor
– the micro-architecture: the set of operations performed by each top level block
• the instruction set architecture (ISA) associated to the top level internal organization.
Von Neumann architecture / Harvard architecture
When the instruction must be executed (in one clock cycle) two distinct memories are mandatory,
one for programs and one for data, because in each cycle a new instruction must be fetched and
sometimes data must be exchanged between the external memory and the processor. But,
when an instructions is interpreted in many clock cycles it is possible to have only one external
memory, because, if a data transfer is needed, then it can be performed adding one or few extra
cycles to the process of interpretation.
DATA
PROGRAM
DATA & PROGRAM
MEMORY
MEMORY
MEMORY
✻
❄
✻
❄
✻
❄
PROCESSOR
a.
PROCESSOR
b.
Figure 9.14: The two main computer architectures. a. Harvard Architecture: data and
programs are stored in two different memories. b. Von Neumann Architecture: both data and programs
are stored in the same memory.
Two kind of computer architecture where imposed from the beginning of the history of
computers:
• Harvard architecture with two external memories, one for data and another for programs (see Figure 10.6a)
• von Neumann architecture with only one external memory used for storing both data
and programs (see Figure 10.6b).
The preferred embodiment for an executing processor is a Hardvare architecture, and the
preferred embodiment for an interpreting processor is a von Neumann architecture. For technological reasons in the first few decades of development of computing the von Neumann architecture was more taken into account. Now the technology being freed by a lot of restriction, we
pay attention to both kind of architectures.
In the next two subsections both, executing processor (commercially called Reduced Instruction Set Computer – RISC – processors) and interpreting processor (commercially called
Complex Instruction Set Computer – CISC – processors) are exemplified by implementing
very simple versions.
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
9.3.4
265
An executing processor
The executing processor is simpler than an interpreting processor. The complexity of computation moves almost completely from the physical structure of the processor into the programs
executed by the processor, because a RISC processor has an organization containing mainly
simple, recursively defined circuits.
The organization
The Harvard architecture of a RISC executing machine (see Figure 10.6a) determine the internal
structure of the processor to have mechanisms allowing in each clock cycle cu address both, the
program memory and the data memory. Thus, the RALU-type functional automaton, directly
interfaced with the data memory, is loop-connected with a control automaton designed to fetch
in each clock cycle a new instruction from the program memory. The control automaton does
not “know” the function to be performed, as it does for the elementary processor, rather he
“knows” how to “fetch the function” from an external storage support, the program memory2 .
Program Memory
addr
Data Memory
instruction
dataOut
✻
dataIn
addr
✻
toyRISC processor
jmpAddr
❄
✻
value
Decode
✛
❄
PC
pc + 1
2nd
Loop
✲
❄
Next PC
✻
❘
✠
✲ MUX
✻
3rd Loop
✻
RALU
✲
File Register
ALU
✻
❄
2nd
Loop
✻
leftOp
rightOp
Control
Figure 9.15: The organization of toyRISC processor.
The organization of the simple executive processortoyRISC is given in Figure 9.15, where
the RALU subsystem is connected with the Control subsystem, thus closing a 3rd loop.
Control section is simple functional automaton whose state, stored in the register called Program Counter (PC), is used to compute in each clock cycle the address from where the next
instruction is fetched. There are two modes to compute the next address: incrementing, with
1 or signed number the current address. The next address can be set, independently from the
current value of PC, using a value fetched from an internal register or a value generated by
the currently executed instruction. The way the address is computed can be determined by the
value, 0 or different from 0, of a selected register. More, the current pc+1 can be stored in an
internal register when the control of the program call a new function and a return is needed.
For all the previously described behaviors the combinational circuit NextPC is designed. It
contains outCLC and loopCLC of the automaton whose state is stored in PC.
2
The relation between an elementary processor and a processor is somehow similar with the relation between
a Turing Machine and an Universal Turing Machine.
266
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
RALU section accepts data coming form data memory, from the currently executed instruction, or from the Control automaton, thus closing the 3dr loop.
Both, the Control automaton and the RALU automaton are simple, recursively defined
automata. The computational complexity is completely moved in the code stored inside the
program memory.
The instruction set architecture
The architecture of toyRISC processor is described in Figure 9.16.
/**********************************************************************************
INSTRUCTION SET ARCHITECTURE
reg
[15:0] pc; // program counter
reg
[31:0] programMemory[0:65535];
reg
[31:0] dataMemory[0:n-1];
instruction[31:0] = {opCode[5:0], dest[4:0], left[4:0], value[15:0]} |
{opCode[5:0], dest[4:0], left[4:0], right[4:0], noUse[10:0]};
**********************************************************************************/
parameter
// CONTROL
nop
= 6’b00_0000,
// no operation: pc = pc+1;
zjpm
= 6’b00_0010,
// cond jump: pc = (rf[left] = 0) ? pc + value : pc+1
nzjmp = 6’b00_0011,
// !cond jump: pc = !(rf[left] = 0) ? pc + value : pc+1
rjmp
= 6’b00_0100,
// relative jump: pc = pc + value;
ret
= 6’b00_0101,
// return from subroutine: pc = rf[left][15:0];
ajmp
= 6’b00_0110,
// absolute jump: pc = value;
call
= 6’b00_0111,
// subroutine call: pc = value; rf[dest] = pc+1;
// ARITHMETIC & LOGIC
inc
= 6’b11_0000,
// rf[dest] = rf[left] + 1; pc = pc+1;
dec
= 6’b11_0001,
// rf[dest] = rf[left] - 1; pc = pc+1;
add
= 6’b11_0010,
// rf[dest] = rf[left] + rf[right]; pc = pc+1;
sub
= 6’b11_0011,
// rf[dest] = rf[left] - rf[right]; pc = pc+1;
inccr = 6’b11_0100,
// rf[dest] = (rf[left] + 1)[32]; pc = pc+1;
deccr = 6’b11_0101,
// rf[dest] = (rf[left] - 1)[32]; pc = pc+1;
addcr = 6’b11_0110,
// rf[dest] = (rf[left] + rf[right])[32]; pc = pc+1;
subcr = 6’b11_0111,
// rf[dest] = (rf[left] - rf[right])[32]; pc = pc+1;
lsh
= 6’b11_1000,
// rf[dest] = rf[left] >> 1; pc = pc+1;
ash
= 6’b11_1001,
// rf[dest] = {rf[left][31], rf[left][31:1]}; pc = pc+1;
move
= 6’b11_1010,
// rf[dest] = rf[left]; pc = pc+1;
swap
= 6’b11_1011,
// rf[dest] = {rf[left][15:0], rf[left][31:16]} pc = pc+1;
neg
= 6’b11_1100,
// rf[dest] = ~rf[left]; pc = pc+1;
bwand = 6’b11_1101,
// rf[dest] = rf[left] & rf[right]; pc = pc+1;
bwor
= 6’b11_1110,
// rf[dest] = rf[left] | rf[right]; pc = pc+1;
bwxor = 6’b11_1111,
// rf[dest] = rf[left] ^ rf[right]; pc = pc+1;
// MEMORY
read
= 6’b10_0000,
// read from dataMemory[rf[right]]; pc = pc+1;
load
= 6’b10_0111,
// rf[dest] = dataOut; pc = pc+1;
store = 6’b10_1000,
// dataMemory[rf[right]] = rf[left]; pc = pc+1;
val
= 6’b01_0111;
// rf[dest] = {{16*{value[15]}}, value}; pc = pc+1;
Figure 9.16: The architecture of toyRISC processor.
The 32-bit instruction has two forms: (1) control form, and (2) arithmetic-logic & memory
form. The first field, opCode, is used to determine what is the form of the current instruction.
Each instruction is executed in one clock cycle.
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
module toyRISC(output
input
output
output
input
output
input
input
reg
wire
wire
wire
wire
wire
[15:0]
[31:0]
[31:0]
[31:0]
[31:0]
[15:0]
[15:0]
[31:0]
[31:0]
[31:0]
[31:0]
pc
aluOut
leftOp
rightOp
regFileIn
incPc
instrAddr
instruction
dataAddr
dataOut
dataIn
we
reset
clock
, //
, //
, //
, //
, //
, //
,
);
267
program memory address
instruction from program memory
data memory address
data send to data memory
data received from data memory
write enable for data memory
;
;
;
;
;
;
assign we = instruction[31:26] == 6’b101000 ;
assign dataAddr = rightOp
assign dataOut = leftOp
;
;
always @(posedge clock) if (reset) pc <=0; else pc <= instrAddr;
nextPc nextPc(
.addr
.incPc
.pc
.jmpVal
.leftOp
.enable
(instrAddr
(incPc
(pc
(instruction[15:0]
(leftOp
(~|instruction[31:30]
),
),
),
),
),
));
fileReg fileReg(.leftOut
(leftOp
.rightOut
(rightOp
.in
(regFileIn
.leftAddr
(instruction[15:11]
.rightAddr (instruction[20:16]
.destAddr
(instruction[25:21]
.writeEnable(&instruction[31:30] | &instruction[28:26]
),
),
),
),
),
),
));
mux4_32 mux(.out(regFileIn
),
.in0({16’b0, incPc}
),
.in1({{16{instruction[15]}}, instruction[15:0]} ),
.in2(dataIn
),
.in3(aluOut
),
.sel(instruction[31:30]
));
alu alu(.out
(aluOut
.leftIn (leftOp
.rightIn(rightOp
.func
(instruction[29:26]
),
),
),
));
endmodule
Figure 9.17: The top module of toyRISC processor. The four module of the design are all
simple circuits with no meaningful random glue logic used to interconnect them.
268
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
Implementing toyRISC
The structure of toyRISC will be implemented as part of a bigger project realized for a SoC,
where the program memory and data memory are on the same chip, tightly coupled with our
design. Therefore, the connections of the module are not very rigorously buffered.
The Figure 9.17 describe the structure of the top level of our design, which is composed by
simple modules. Because the glue logic is minimal (see lines 17, 29 and 37), the overall structure
could be considered also a very simple one.
The time performance
The longest combinational path in a system using our toyRISC, which imposes the minimum
clock period, is:
Tclock = tclock
to instruction
+ tlef tAddr
to lef tOp
+ tthroughALU + tthroughM U X + tf ileRegSU
Because the system is not buffered the clock frequency depends also by the time behavior of
the system directly connected with toyRISC. In this case tclock to instruction – the access time
of the program memory, related to the active edge of the clock – is an extra-system parameter
limiting the speed of our design. The internal propagation time to be considered are: the read
time from the file register (tlef tAddr to lef tOp or trightAddr to rightOp ), the maximum propagation
time through ALU (dominated by the time for an 32-bit arithmetic operation), the propagation
time through a 4-way 32-bit multiplexer, and the set-up time on the file register’s data inputs.
The way from the output of the file register through Next PC circuit is “shorter” because it
contains a 16-bit adder, comparing with the 32-bit one of the ALU.
9.3.5
∗ An interpreting processor
The interpreting processor are known also as processors having a Complex Instruction Set Computer
(CISC) architecture, or simply as CISC Processors. The interpreting approach allows us to design complex
instructions which are transformed at the hardware level in a sequence of operations. Lets remember
that an executing (RISC) processor has almost all instructions implemented in one clock cycle. It is not
decided what style of designing an architecture is the best. Depending on the application sometimes a
RISC approach is mode efficient, sometimes a CISC approach is preferred.
The organization
Our CISC Processor is a machine characterized by using a register file to store the internal (the most
frequently used) variables. The top level view of this version of processor is represented in Figure 9.18.
It contains the following blocks:
• REGISTER & ALU – RALU – 32 32-bit registers organized in a register file, and an ALU; the
registers are used also for control purposes (program counter, return address, stack pointer in the
external memory, ...)
• INPUT & OUTPUT BUFFER REGISTERS used to provide full synchronous connections with the
external “world”, minimizing tin reg , treg out , maximizing fmax , and avoiding tin out (see subsection
1.1.5); the registers are the following:
COM REG : sends out the 2-bit read or write command for the external data & program memory
ADDR REG : sends out the 32-bit address for the external data & program memory
OUT REG : sends out the 32-bit data for the external memory
DATA REG : receives back, with one clock cycle delay related to the command loaded in COM
REG, 32-bit data from the external data & program memory
INST REG : receives back, with one clock cycle delay related to the command loaded in COM
REG, 32-bit instruction from the external data & program memory
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
269
in
❄
data
mem addr
instruction
❄
REG
✲ enINST
inst
DATA REG
clock
❄
✛
addr
ADDR
✛
❲
inst[15:0]
✛ write enable
✛ dest addr
✛ left addr
✛ right addr
result
REG
REGISTER FILE
inst[31:11]
❄
CONTROL
right out
left out
AUTOMATON
(CROM)
✛
out
OUT
✛
data out
REG
data in
❄
❄
1
0
LEFT MUX
✛
com
COM
✻
value
✛
❄❄❄❄
left sel
✛
0
1
2
3
RIGHT MUX
✻
reset
✛right sel
command
✛
mem com
❄
REG
❄
right
left
ALU
✛
alu com
✕ flags
alu out
■
the second loop
RALU
the third loop
Figure 9.18: An interpreting processor. The organization is simpler because only one external
memory is used.
• CONTROL AUTOMATON used to control the fetch and the interpretation of the instructions
stored in the external memory; it is an initial automaton initialized, for each new instruction, by
the operation code (inst[31:26] received from INST REG)
The instruction of our CISC Processor has two formats. The first format is:
{
opcode[5:0]
dest_addr[4:0]
left_addr[4:0]
right_addr[4:0]
rel_addr[10:0]
= instr[31:0];
,
,
,
,
}
//
//
//
//
//
operation code
selects the destination
selects the left operand
selects the right operand
small signed jump for program address
The relative address allows a positive or a negative jump of 1023 instructions in the program space. It is
sufficient for almost all jumps in a program. If not, special absolute jump instruction can solve this very
rare cases.
The second format is used when the right operand is a constant value generated at the compiling
time in the instruction body. It is:
{
opcode[5:0]
dest_addr[4:0]
,
,
270
CHAPTER 9. PROCESSORS:
left_addr[4:0]
value[15:0]
= instr[31:0];
,
}
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
// signed integer
When the instruction is fetched from the external memory it is memorized in INST REG because its
content will be used in different stages of the interpretation, as follows:
• inst[31:26] = opcode[5:0] to initialize CONTROL AUTOMATON in the state from which
flows the sequence of commands used to interpret the current instruction
• inst[29:26] = opcode[3:0] to command the function performed by ALU in the step associated
to perform the main operation associated with the current instruction (for example, if the instruction is add 12, 3, 7, then the bits opcode[3:0] are used to command the ALU to do the addition
of registers 3 and 7 in the appropriate step of interpretation)
• inst[25:11] = {dest addr, left addr, right addr} is used to address the REGISTER FILE
unit when the main operation associated with the current instruction is performed
• inst[15:0] = value is selected to form the right operand when an instruction operating with
immediate value is interpreted
• inst[10:0] = rel addr is used in jump instructions, in the appropriate clock cycle, to compute
the next program address.
The REGISTER FILE unit contains 32 32-bit registers. In each clock cycle, any ordered pair of
registers can be selected as operands, and the result can be stored back in any of them. They have the
following use:
• r0, r1, ...
r29 are general purpose registers;
• r31 is used as Program Counter (PC);
• r30 is used to store the Return Address (RA) when the call instruction is interpreted (no embedded
calls are allowed for this simple processor3 ).
CONTROL AUTOMATON has the structure presented in Figure 9.19. In the fetch cycle init =
1 allows the automaton to jump into the state codded by opcode, from which a sequence of operations
flows with init = 0, ignoring the initialization input. This is the simplest way to associate for each
instruction the interpreting sequence of elementary operation.
The output of CONTROL AUTOMATON commands all the top level blocks of our CISC Processor
using the following fields:
{en_inst
,
write_enable
,
dest_addr[4:0] ,
left_addr[4:0] ,
alu_com[3:0]
,
right_addr[4:0],
left_sel
,
right_sel[1:0] ,
mem_com[1:0]
}
= command
//
//
//
//
//
//
//
//
//
write enable for the instruction register
write back enable for the register file
destination address in file register
left operand address in file register
alu functions
right operand address in file register
left operand selection
right operand selection
memory command
The fields dest addr, left addr, right addr, alu com are sometimes selected from INST REG
(see ADDR MUX and FUNC MUX in Figure 9.19) and sometimes their value is generated by CONTROL
AUTOMATON according to the operation to be executed in the current clock cycle. The other command
fields are generated by CONTROL AUTOMATON in each clock cycle.
3
If embedded calls are needed, then this register contains the stack pointer into a stack organized in the
external memory. We are not interested in adding the feature of embedded calls, because in this digital system
lessons we intend to keep the examples small and simple.
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
inst[31:11]
271
inst[31:26]
❄
inst[29:26]
inst[25:11]
❄
1
0
INIT MUX
❄
reset
✛init
next state
❄
STATE REG
clock
state
❄
❄
❄
0
1
ADDR MUX
❄
✛
❄
0
1
FUNC MUX
“RANDOM” CLC
✛
✻
❄
❄
alu com
{dest addr, left addr, right addr}
flag
❄
{en instr, write enable, left sel, right sel, mem com}
Figure 9.19: The control automaton for our CISC Processor. It is a more compact version
of CROM (see Figure 8.26). Instead of a CLC used for the complex part of executing processor, for an
interpreting processor a sequential machine is used to solve the problem of complexity.
CONTROL AUTOMATON receives back from ALU only one flag: the least significant bit of ALU,
alu out[0]; thus closing the third loop4 .
In each clock cycle the content of two registers can be operated in ALU and the result stored in a
third register.
The left operand can be sometimes data in if left sel = 1. It must be addressed two clock cycles
before use, because the external memory is supposed to be a synchronous one, and the input register
introduces another one cycle delay. The sequence generated by CONTROL AUTOMATON takes care
by this synchronization.
The right operand can be sometimes value = instr reg[15:0] if right sel = 2’b1x. If right sel
= 2’b01 the right operand is the 11-bit signed integer rel addr = instr reg[10:0]
The external memory is addressed with a delay of one clock cycle using the value of left out. We
are not very happy about this additional delay, but this is the price for a robust design. What we loose
in number of clock cycles used to perform some instructions is, at least partially, recuperated by the
possibility to increase the frequency of the system clock.
Data to be written in the external memory is loaded into OUT REG from the right output of FILE
REG. It is synchronous with the address.
The command for the external memory is also delayed one cycle by the synchronization register COM
REG. It is generated by CONTROL AUTOMATON.
Data and instructions are received back from the external memory with two clock cycle delay, one
because of we have an external synchronous memory, and another because of the input re-synchronization
done by DATA REG and INST REG.
The structural Verilog description of the top level of our CISC Processor is in Figure 9.20.
4
The second loop is closed once in the big & simple automaton RALU, and another in the complex finite
automaton CONTROL AUTOMATON. The first loop is closed in each flip-flop used to build the registers.
272
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
module cisc_processor(input
clock
,
input
reset
,
output
reg [31:0] addr_reg,
// memory address
output
reg [1:0]
com_reg ,
// memory command
output
reg [31:0] out_reg ,
// data output
input
[31:0] in
); // data/inst input
// INTERNAL CONNECTIONS
wire
[25:0] command;
wire
flag;
wire
[31:0] alu_out, left, right, left_out, right_out;
// INPUT & OUTPUT BUFFER REGISTERS
reg [31:0] data_reg, inst_reg;
always @(posedge clock) begin
if (command[25]) inst_reg <= in ;
data_reg <= in
;
addr_reg <= left_out
;
out_reg <= right_out
;
com_reg <= command[1:0]
;
end
// CONTROL AUTOMATON
control_automaton control_automaton(.clock (clock
),
.reset (reset
),
.inst
(inst_reg[31:11]),
.command(command
),
.flag
(alu_out[0]
));
// REGISTER FILE
register_file register_file(.left_out
(left_out
),
.right_out
(right_out
),
.result
(alu_out
),
.left_addr
(command[18:14] ),
.right_addr
(command[9:5]
),
.dest_addr
(command[23:19] ),
.write_enable
(command[24]
),
.clock
(clock
));
// MULTIPLEXERS
mux2 left_mux( .out(left
),
.in0(left_out
),
.in1(data_reg
),
.sel(command[4]));
mux4 right_mux( .out(right
),
.in0(right_out
),
.in1({{21{inst_reg[10]}}, inst_reg[10:0]}
),
.in2({16’b0, inst_reg[15:0]}
),
.in3({inst_reg[15:0], 16’b0}
),
.sel(command[3:2]
));
// ARITHMETIC & LOGIC UNIT
cisc_alu alu(
.alu_out(alu_out
),
.left
(left
),
.right (right
),
.alu_com(command[13:10] ));
endmodule
Figure 9.20: The top module of our CISC Processor.
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
273
Microarchitecture
The complex part of our CISC Processor is located in the block called CONTROL AUTOMATON. More
precisely, the only complex circuit in this design is the loop of the automaton called “RANDOM CLC”
(see Figure 9.19). The Verilog module describing CONTROL AUTOMATON is represented in Figure
9.21.
The micro-architecture defines all the fields used to command the simple parts of this processor.
Some of them are used inside the control automaton module, while others command the top modules
of the processor.
The inside used fields command are the following:
init : allows the jump of the automaton into the initial state associated with each instruction when
init = new seq
addr sel : the three 5-bit addresses for FILE REGISTER are considered only if the field addr sel takes
the value from inst, else three special combinations of addresses are generated by the control
automaton
func sel : the field alu com is considered only if the field func sel takes the value from out, else the
code opcode[3:0] selects the ALU’s function
The rest of fields command the function performed in each clock cycle by the top modules of our
CISC Processor. They are:
en inst : enables the load of data received from the external memory only when it represents the next
instruction to be interpreted
write enable : enables write back into FILE REGISTER the result from the output of ALU
alu com : is a 4-bit field used to command ALU’s function for the specific purpose of the interpretation
process (it is considered only if func sel = from aut)
left sel : is the selection code for LEFT MUX (see Figure 9.18)
right sel : is the selection code for RIGHT MUX (see Figure 9.18)
mem com : generated the commands for the external memory containing both data and programs.
The micro-architecture (see Figure 9.22) is subject of possible changes during the definition of the
transition function of CONTROL AUTOMATON.
Instruction set architecture (ISA)
There is a big flexibility in defining the ISA for a CISC machine, because we accepted to interpret
each instruction using a sequence of micro-operations. The control automaton is used as sequencer for
implementing instructions beyond what can be simply envisaged inspecting the organization of our simple
CISC processor.
An executing (RISC) processor displays its architecture in its organization, because the control is very
simple (the decoder is a combinational circuit used to trans-code only). The complexity of the control
of an interpreting processor hides the architecture in the complex definition of the control automaton
(which can have a strong generative power).
In Figure 9.23 is sketched a possible instruction set for our CISC processor. There are at least the
following classes of instructions:
• Arithmetic & Logic Instructions: the destination register takes the value resulted from operating
any two registers (unary operations, such as increment, are also allowed)
• Data Move Instructions: data exchange between the external memory and the file register are
performed
• Control Instructions: the flow of instruction is controlled according to the fix or data dependent
patterns
• ...
We limit our discussion to few and small classes of instructions because our goal is to offer only a structural
image about what an interpretative processor is. An exhaustive approach is an architectural one, which
is far beyond out intention in these lessons about digital systems, not about computational systems.
274
CHAPTER 9. PROCESSORS:
module control_automaton(
input
input
input
output
input
[20:0]
[25:0]
[3:0]
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
clock
reset
inst
command
flags
,
,
,
,
);
‘include "micro_architecture.v"
‘include "instruction_set_architecture.v"
// THE STRUCTURE OF ’inst’
wire
[5:0]
opcode ;
//
wire
[4:0]
dest
,
//
left_op ,
//
right_op;
//
assign {opcode, dest, left_op,
operation code
selects destination register
selects left operand register
selects right operand register
right_op} = inst;
// THE STRUCTURE OF ’command’
reg
en_inst
; // enable load a new instruction in inst_reg
reg
write_enable; // writes the output of alu at dest_addr
reg
[4:0]
dest_addr
; // selects the destination register
reg
[4:0]
left_addr
; // selects the left operand in file register
reg
[3:0]
alu_com
; // selects the operation performed by the alu
reg
[4:0]
right_addr ; // selects the right operand in file register
reg
left_sel
; // selects the source of the left operand
reg
[1:0]
right_sel
; // selects the source of the right operand
reg
[1:0]
mem_com
; // generates the command for memory
assign command = {en_inst, write_enable, dest_addr, left_addr, alu_com,
right_addr, left_sel, right_sel, mem_com};
// THE STATE REGISTER
reg
[5:0]
state_reg
; // the state register
reg
[5:0]
next_state ; // a "register" used as variable
always @(posedge clock) if (reset) state_reg <= 0
;
else
state_reg <= next_state ;
‘include "the_control_automaton’s_loop.v"
endmodule
Figure 9.21: Verilog code for control automaton.
Implementing ISA
Implementing a certain Instruction Set Architecture means to define the transition functions of the control
automaton:
• the output transition function, in our case to specify for each state the value of the command code
(see Figure 9.21)
• the state transition function, which specifies the value of next state
The content of the file the control automaton’s loop.v contains the description of the combinational
circuit associated to the control automaton. It generates both the 26-bit command code and the 6-bit
next state code. The following Verilog code is the most compact way to explain how the control
automaton works. Please read the next “always” as the single way to explain rigorously how out CISC
machine works.
// THE CONTROL AUTOMATON’S LOOP
always @(state_reg or opcode or dest or left_op or right_op or flag)
begin
en_inst
= 1’bx
;
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
// MICRO-ARCHITECTURE
// en_inst
parameter
no_load
load_inst
// write_enable
parameter
no_write
write_back
// alu_func
parameter
alu_left
alu_right
alu_inc
alu_dec
alu_add
alu_sub
alu_shl
alu_half
alu_zero
alu_equal
alu_less
alu_carry
alu_borrow
alu_and
alu_or
alu_xor
// left_sel
parameter
left_out
from_mem
// right_sel
parameter
right_out
jmp_addr
low_value
high_value
// mem_com
parameter
mem_nop
read
write
= 1’b0, // disable instruction register
= 1’b1; // enable instruction register
= 1’b0,
= 1’b1; // write back the current ALU output
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
4’b0000,
4’b0001,
4’b0010,
4’b0011,
4’b0100,
4’b0101,
4’b0110,
4’b0111,
4’b1000,
4’b1001,
4’b1010,
4’b1011,
4’b1100,
4’b1101,
4’b1110,
4’b1111,
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
left
right
left + 1
left - 1
left + right
left - right
{1’b0, left[31:1]}
{left[31], left[31:1]}
{31’b0, (left == 0)}
{31’b0, (left == right)}
{31’b0, (left < right)}
{31’b0, add[32]}
{31’b0, sub[32]}
left & right
left | right
left ^ right
= 1’b0, // left out of the reg file as left op
= 1’b1; // data from memory as left op
=
=
=
=
2’b00,
2’b01,
2’b10,
2’b11;
//
//
//
//
right
right
right
right
out of the reg file as right op
op = {{22{inst[10]}}, inst[9:0]}
op = {{16{inst[15]}}, inst[15:0]}
op = {inst[15:0], 16’b0}
= 2’b00,
= 2’b10, // read from memory
= 2’b11; // write to memory
Figure 9.22: The micro-architecture of our CISC Processor.
write_enable = 1’bx
;
dest_addr
= 5’bxxxxx ;
left_addr
= 5’bxxxxx ;
alu_com
= 4’bxxxx ;
right_addr
= 5’bxxxxx ;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= 2’bxx
;
next_state
= 6’bxxxxxx;
// INITIALIZE THE PROCESSOR
if (state_reg == 6’b00_0000)
// pc = 0
begin
en_inst
= no_load
write_enable = write_back
dest_addr
= 5’b11111
left_addr
= 5’b11111
;
;
;
;
275
276
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
// INSTRUCTION SET ARCHITECTURE (only samples)
// arithmetic & logic instructions & pc = pc + 1
parameter
move
= 6’b10_0000, // dest_reg = left_out
inc
= 6’b10_0010, // dest_reg = left_out + 1
dec
= 6’b10_0011, // dest_reg = left_out - 1
add
= 6’b10_0100, // dest_reg = left_out + right_out
sub
= 6’b10_0101, // dest_reg = left_out - right_out
bwxor
= 6’b10_1111; // dest_reg = left_out ^ right_out
// ...
// data move instructions & pc = pc + 1
parameter
read
= 6’b01_0000, // dest_reg = mem(left_out)
rdinc
= 6’b01_0001, // dest_reg = mem(left_out + value)
write
= 6’b01_1000, // mem(left_out) = right_out
wrinc
= 6’b01_1001; // mem(left_out + value) = right_out
// ...
// control instructions
parameter
nop
= 6’b11_0000, // pc = pc + 1
jmp
= 6’b11_0001, // pc = pc + value
call
= 6’b11_0010, // pc = value, ra = pc + 1
ret
= 6’b11_0011, // pc = ra
jzero
= 6’b11_0100, // if (left_out = 0) pc = pc + value;
// else pc = pc + 1
jnzero = 6’b11_0101; // if (left_out != 0) pc = pc + value;
// else pc = pc + 1
// ...
Figure 9.23: The instruction set architecture of our CISC Processor. The partial definition
of the file instruction set architecture.v included in the conttrol automaton.v file.
alu_com
right_addr
left_sel
right_sel
mem_com
next_state
=
=
=
=
=
=
alu_xor
;
5’b11111
;
left_out
;
right_out
;
mem_nop
;
state_reg + 1;
end
// INSTRUCTION FETCH
if (state_reg == 6’b00_0001)
// rquest for a new instruction & increment pc
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_inc
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_read
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0010)
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
// wait for memory to read doing nothing (synchronous memory)
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0011)
// load the new instruction in instr_reg
begin
en_inst
= load_inst
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0100)
// initialize the control automaton
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= opcode[5:0] ;
end
// EXECUTE THE ONE CYCLE FUNCTIONAL INSTRUCTIONS
if (state_reg[5:4] == 2’b10)
// dest = left_op OPERATION right_op
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= dest
;
left_addr
= left_op
;
alu_com
= opcode[3:0] ;
right_addr
= right_op
;
left_sel
= left_out
;
right_sel
= right_out
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
277
278
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
// EXECUTE MEMORY READ INSTRUCTIONS
if (state_reg == 6’b01_0000)
// read from left_reg in dest_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_read
;
next_state
= 6’b01_0010
;
end
if (state_reg == 6’b01_0001)
// read from left_reg + <value> in dest_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_read
;
next_state
= 6’b01_0010
;
end
if (state_reg == 6’b01_0010)
// wait for memory to read doing nothing
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b01_0011)
// the data from memory is loaded in data_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
end
if (state_reg == 6’b01_0100)
// data_reg is loaded in dest_reg & go to fetch
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= dest
;
left_addr
= 5’bxxxxx
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= from_mem
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
// EXECUTE MEMORY WRITE INSTRUCTIONS
if (state_reg == 6’b01_1000)
// write right_op to left_op & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_left
;
right_addr
= right_op
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_write
;
next_state
= 6’b00_0001
;
end
if (state_reg == 6’b01_1000)
// write right_op to left_op + <value> & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_add
;
right_addr
= right_op
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_write
;
next_state
= 6’b00_0001
;
end
// CONTROL INSTRUCTIONS
if (state_reg == 6’b11_0000)
// no operation & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
279
280
CHAPTER 9. PROCESSORS:
mem_com
next_state
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
= mem_nop
= 6’b00_0001
end
if (state_reg == 6’b11_0001)
// jump to (pc + <value>) & go to fetch
begin
en_inst
= no_load
write_enable = write_back
dest_addr
= 5’b11111
left_addr
= 5’b11111
alu_com
= alu_add
right_addr
= 5’bxxxxx
left_sel
= left_out
right_sel
= low_value
mem_com
= mem_nop
next_state
= 6’b00_0001
end
if (state_reg == 6’b11_0010)
// call: first step: ra = pc + 1
begin
en_inst
= no_load
write_enable = write_back
dest_addr
= 5’b11110
left_addr
= 5’b11111
alu_com
= alu_left
right_addr
= 5’bxxxxx
left_sel
= left_out
right_sel
= 2’bxx
mem_com
= mem_nop
next_state
= 6’b11_0110;
end
if (state_reg == 8’b0011_0110)
// call: second step: pc = value
begin
en_inst
= no_load
write_enable = write_back
dest_addr
= 5’b11111
left_addr
= 5’bxxxxx
alu_com
= alu_right
right_addr
= 5’bxxxxx
left_sel
= 1’bx
right_sel
= jmp_addr
mem_com
= mem_nop
next_state
= 6’b00_0001
end
if (state_reg == 6’b11_0011)
// ret: pc = ra
begin
en_inst
= no_load
write_enable = write_back
dest_addr
= 5’b11111
left_addr
= 5’b11110
alu_com
= alu_left
right_addr
= 5’bxxxxx
left_sel
= left_out
right_sel
= 2’bxx
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
9.3. LOOP COUPLED AUTOMATA: THE SECOND ”TURNING POINT”
mem_com
next_state
= mem_nop
= 6’b00_0001
;
;
end
if ((state_reg == 6’b11_0100) && flag)
// jzero: if (left_out = 0) pc = pc + value;
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && ~flag)
// jzero: if (left_out = 1) pc = pc + 1;
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && ~flag)
// jnzero: if (left_out = 1) pc = pc + value;
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && flag)
// jnzero: if (left_out = 0) pc = pc + 1;
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
281
282
CHAPTER 9. PROCESSORS:
mem_com
next_state
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
= mem_nop
= 6’b00_0001
;
;
end
end
The automaton described by the previous code has 36 states for the 25 instructions implemented
(see Figure 9.23). More instructions can added if new state are described in the previous “always”.
Obviously, the most complex part of the processor is this combinational circuit associated to the control
automaton.
Time performance
The representation from Figure 9.24 is used to evaluate the time restrictions imposed by our CISC
processor.
from the external memory
clock
❄
INPUT REGISTERS
❄
❄
✛
CONTROL
RALU
✲
AUTOMATON
❄
❄
OUTPUT REGISTERS
❄
to the external memory
Figure 9.24: The simple block diagram of our CISC processor. The fully buffered solution
imposed for designing this interpretative processor minimizes the depth of signal path entering and
emerging in/from the circuit, and avoid a going through combinational path.
The full registered external connections of the circuit allows us to provide the smallest possible values
for minimum input arrival time before clock, tin reg , maximum output required time after clock, treg out ,
and no path for maximum combinational path delay, tin out . The maximum clock frequency is fully
determined by the internal structure of the processor, by the path on the loop closed inside RALU or
between RALU and CONTROL AUTOMATON. The actual time characterization is:
• tin
reg
• treg
= tsu – the set-up time for the input registers
out
= treg – the propagation time for the output registers
• Tmin = max(tRALU
tRALU
tprocessor
loop
loop , tprocessor loop ),
loop
= tstate
= tstate
reg
reg
+ taut
where:
+ taut
out clc
out clc
+ treg
+ treg
f ile
f ile
+ tmux + talu + treg
+ tmux + talu
f lag
+ taut
f ile su
in clc
+ tstate
reg su
9.4. ∗ THE ASSEMBLY LANGUAGE: THE LOWEST PROGRAMMING LEVEL
283
This well packed version of a simple processor is very well characterized as time behavior. The
price for this is the increasing number of clock cycle used for executing an instruction. The effect of
the increased number of clock cycles is sometimes compensated by the possibility to use a higher clock
frequency. But, all the time the modularity is the main benefit.
Concluding about our CISC processor
A CISC processor is more complex than a stack processor because for each instruction the operands must
be selected from the file register. The architecture is more flexible, but the loop closed in RALU is longer
than the loop closed in SALU.
A CISC approach allows more complex operations performed during an instruction because it is
interpreted, not simply executed in one clock cycle.
Interpretation allows a single memory for both data and programs with all the resulting advantages
and disadvantages.
An interpreting processor contains a simple automaton – RALU – and a complex one – Control
Automaton – because its complex behavior.
Both, Stack Processor and CISC Processor are only simple exercises designed for presenting the
circuit aspects of the closing of hte third loop. The real and complex architectural aspects are minimally
presented because this text book is about circuits not abut computation.
9.4
∗ The assembly language: the lowest programming level
The instruction set represent the machine language: the lowest programming level in a computation
machine. The programming is very difficult at this level because of the concreteness of the process.
Too many details must be known by the programmer. The main improvement added by a higher level
language is the level of abstraction used to present the computational resources. Writing a program
in machine language we must have in mind a lot of physical details of the machine. Therefore, a real
application must be developed in a higher level language.
The machine language can be used only for some very critical section of the algorithms. The automatic
translation done by a compiler from a high level language into the machine language is some times
unsatisfactory for high performance application. Only in this cases small part of the code must be
generated “manually” using the machine language.
9.5
Concluding about the third loop
The third loop is closed through simple automata avoiding the fast increasing of the
complexity in digital circuit domain. It allows the autonomy of the control mechanism.
”Intelligent registers” ask less structural control maintaining the complexity of a finite
automaton at the smallest possible level. Intelligent, loop driven circuits can be controlled using
smaller complex circuits.
The loop through a storage element ask less symbolic control at the microarchitectural level. Less symbols are used to determine the same behavior because the local
loop through a memory element generates additional information about the recent history.
Looping through a memory circuit allows a more complex “understanding” because
the controlled circuits “knows” more about its behavior in the previous clock cycle. The circuit is
somehow “conscious” about what it did before, thus being more “responsible” for the operation
it performs now.
284
CHAPTER 9. PROCESSORS:
THIRD ORDER, 3-LOOP DIGITAL SYSTEMS
Looping through an automaton allows any effective computation. Using the theory
of computation (see chapter Recursive Functions & Loops in this book) can be proved that any
effective computation can be done using a three loop digital system. More than three loops are
needed only for improving the efficiency of the computational structures.
The third loop allows the symbolic functional control using the arbitrary meaning
associated to the binary codes embodied in instructions or micro-instructions. Both, the coding
and the decoding process being controlled at the design level, the binary symbols act actualizing
the potential structure of a programmable machine.
Real processors use circuit level parallelism discussed in the first chapter of this book.
They are: data parallelism, time parallelism and speculative parallelism. How all these kind of
parallelism are used is a computer architecture topic, beyond the goal of these lecture notes.
9.6
Problems
Problem 9.1 Interrupt automaton with asynchronous input.
Problem 9.2 Solving the second degree equations with an elementary processor.
Problem 9.3 Compute y if x, m and n is given with an elementary processor..
Problem 9.4 Modify the unending loop of the processor to avoid spending time in testing if a
new instruction is in inFIFO when it is there.
Problem 9.5 Define an instruction set for the processor described in this chapter using its
microarchitecture.
Problem 9.6 Is it closed another loop in our Stack Processor connecting tos to the input of
DECODE unit?
Problem 9.7 Our CISC Processor: how must be codded the instruction set to avoid FUNC
MUX?
9.7
Projects
Project 9.1 Design a specialized elementary processor for rasterization function.
Project 9.2 Design a system integrating in a parallel computational structure 8 rasterization
processors designed in the previous project.
Project 9.3 Design a floating point arithmetic coprocessor.
Project 9.4 Design the RISC processor defined by the following Verilog behavioral description:
module risc_processor(
);
endmodule
Project 9.5 Design a version of Stack Processor modifying SALU as follows: move MUX4 to
the output of ALU and the input of STACK.
Chapter 10
COMPUTING MACHINES:
≥4–loop digital systems
In the previous chapter
was introduced the main digital system - the processor - and we discussed how works the
third loop in a digital system emphasizing
• effects on the size of digital circuits
• effects on the complexity of digital systems
• how the apparent complexity can be reduced to the actual complexity in a digital
system
In this chapter
a very short introduction in the systems having more than three internal loops is provided,
talking abut
• how are defined the basic computational structures: microcontrollers, computers,
stack machines, co-processors
• how the classification in orders starts to become obsolete with the fourth order systems
• the concept of embedded computation
In the next chapter
some futuristic systems are described as N-th order systems having the following features:
• they can behave as self-organizing systems
• they are cellular systems easy to be expanded in very large and simple powerful
computational systems
285
286
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
Software is getting slower more rapidly than hardware becomes faster.
Wirth’s law1
To compensate the effects of the bad behavior of software guys, besides the job done by the Moore law a
lot of architectural work must be added.
The last examples of the previous chapter emphasized a process that appears as a ”turning
point” in 3-OS: the function of the system becomes lesser and lesser dependent on the physical
structure and the function is more and more assumed by a symbolic structure (the program
or the microprogram). The physical structure (the circuit) remains simple, rather than the
symbolic structure, “stored” in program memory of in a ROM, that establishes the functional
complexity. The fourth loop creates the condition for a total functional dependence on the
symbolic structure. By the rule, at this level an universal circuit - the processor - executes (in
RISC machines) or interprets (in CISC machines) symbolic structures stored in an additional
device: the program memory.
10.1
Types of fourth order systems
There are four main types of fourth order systems (see Figure 10.1) depending on the order of
the system through which the loop is closed:
1. P & ROM is a 4-OS with loop closed through a 0-OS - in Figure 10.1a the combinational
circuit is a ROM containing only the programs executed or interpreted by the processor
2. P & RAM is a 4-OS with loop closed through a 1-OS - is the computer, the most
representative structure in this order, having on the loop a RAM (see Figure 10.1b) that
stores both data and programs
3. P & LIFO is a 4-OS with loop closed through a 2-OS - in Figure 10.1c the automaton is
represented by a push-down stack containing, by the rule, data (or sequences in which the
distinction between data and programs does not make sense, as in the Lisp programming
language, for example)
4. P & CO-P is a 4-OS with loop closed through a 3-OS - in Figure 10.1d COPROCESSOR
is also a processor but a specialized one executing efficiently critical functions in the system
(in most of cases the coprocessor is a floating point arithmetic processor).
The representative system in the class of P & ROM is the microcontroller the most successful circuit in 4-OS. The microcontroller is a “best seller” circuit realized as a one-chip computer.
The core of a microcontroller is a processor executing/interpreting the programs stored in a
ROM.
The representative structure in the class of P & RAM is the computer. More precisely, the
structure Processor - Channel - Memory represents the physical support for the well known von
Neumann architecture. Almost all present-day computers are based on this architecture.
1
Niklaus Wirth is an already legendary Swiss born computer scientist with many contributions in developing
various programming languages. The best known is Pascal. Wirth’s law is a sentence which Wirth made popular,
but he attributed it to Martin Reiser.
10.1. TYPES OF FOURTH ORDER SYSTEMS
287
The third type of system seems to be strange, but a recent developed architecture is a stack
oriented architecture defined for the successful Java language. Naturally, a real Java machine is
endowed also with the program memory.
The third and the fourth types are machines in which the segregation process emphasized
physical structures, a stack or a coprocessor. In both cases the segregated structures are also
simple. The consequence is that the whole system is also a simple system. But, the first two
systems are very complex systems in which the simple is net segregated by the random. The
support of the random part is the ROM physical structure in the first case and the symbolic
content of the RAM memory in the second.
Programs
✮
PROCESSOR
ROM
0-OS
3-OS
✻
✻
✻✻
DATA
ADDRESS
READ
WAIT
a.
Data & Programs
✮
PROCESSOR
RAM
1-OS
3-OS
✻
✻
✻
DATA
✻✻
2
ADDRESS
{READ, W RIT E, −}
WAIT
b.
Data
✮
LIFO
PROCESSOR
2-OS
3-OS
✻
✻
✻
DATA
✻
2
{P U SH, P OP, −}
WAIT
c.
Functions
✮
CO-PROCESSOR
PROCESSOR
3-OS
3-OS
✻
✻
DATA
✻
✻
COMMANDS
d.
WAIT
Figure 10.1: The four types of 4-OS machines. a. Fix program computers usual in embedded
computation. b. General purpose computer. c. Specialized computer working working on a restricted
data structure. d. Accelerated computation supported by a specialized co-processor.
The actual computing machines have currently more than order 4, because the processors
involved in the applications have additional features. Many of these features are introduced by
new loops that increase the autonomy of certain subsystems. But theoretically, the computer
function asks at least four loops.
288
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
10.1.1
The computer – support for the strongest segregation
The ROM content is defined symbolically and after that it is converted in the actual physical
structure of ROM. Instead, the RAM content remains in symbolic form and has, in consequence, more flexibility. This is the main reason for considering the PROCESSOR & RAM =
COMPUTER as the most representative in 4-OS.
The computer is not a circuit. It is a new entity with a special functional definition, currently
called computer architecture. Mainly, the computer architecture is given by the machine language. A program written in this language is interpreted or executed by the processor. The
program is stored in the RAM memory. In the same subsystem are stored data on which the
program “acts”. Each architecture can have many associated computer structures (organizations).
Starting from the level of four order systems the behavior of the system is controlled mainly
by the symbolic structure of programs. The architectural approach settles the distinction between the physical structures and the symbolic structures. Therefore, any computing machine
supposes the following triadic definition (suggested by [”Milutinovic” ’89]):
• the machine language (usually called architecture)
• the storage containing programs written in the machine language
• the machine that interprets the programs, containing:
– the machine language ...
– the storage ...
– the machine ... containing:
∗ ...
and so on until the machine executes the programs.
Does it make any sense to add new loops? Yes, but not too much! It can be justified to
add loops inside the processor structure to improve its capacity to interpret fast the machine
language by using simple circuits. Another way is to see PROCESSOR & COPROCESSOR or
PROCESSOR & LIFO as performant processors and to add over them the loop through RAM.
But, mainly these machines remain structures having the computer function. The computer
needs at least four loops to be competent, but currently it is implemented on system having
more loops in order to become performant.
10.2
∗ The stack processor – a processor as 4-OS
The best way to explain how to use the concept of architecture to design an executive processor is to
use an example having an appropriate complexity. One of the simplest model of computing machine is
the stack machine. A stack machine finds always its operands in the first two stages of a stack (LIFO)
memory. The last two pushed data are the operands involved in the current operation. The computation
must be managed to have accessible the current operand(s) in the data stack. The stack used in a stack
processor have some additional features allowing an efficient data management. For example: double
pop, swap, . . ..
The high level description of a stack processor follows. The purpose of this description is to offer an
example of how starts the design of a processor. Once the functionality of the machine is established at
the higher level of the architecture, there are many ways to implement it.
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
10.2.1
289
∗ The organization
Our Stack Processor is a sort of simple processing element characterized by using a stack memory
(LIFO) for storing the internal variables. The top level internal organization of a version of Stack
Processor (see Figure 10.2) contains the following blocks:
• STACK & ALU – SALU – is the unit performing the elementary computations; it contains:
– a two-output stack; the top of stack (stack0 or tos) and the previous recording (stack1)
are accessible
– an ALU with the operands from the top of stack (left op = stack0 and right io =
stack1)
– a selector for the input of stack grabbing data from: (0) the output of ALU, (1) external data
memory, (2) the value provided by the instruction, or (3) the value of pc +1 to be used as
return address
• PROGRAM FETCH – a unit used to generate in each clock cycle a new address for fetching from
the external program memory the next instruction to be executed
• DECODER – is a combinational circuit used to trans-code the operation code – op code – into
commands executed by each internal block or sub-block.
the fourth loop
✛
✛
data out
❘
data addr
✲
instr addr
ret addr
✛
SALU
✲
STACK
✲
❄
stack com
pc com
✻
MUX4
2
3
0
✻ ✻ ✻ ✻
data in
NEXT PC
stack1
stack0
✛data sel
1
❄
left
❄
PC
✛
ALU
right
✻
✛
PROGRAM
FETCH
✎
<value>
✛
alu com
✲
data mem com
mem ready
instruction
DECODER
✛
<op code>
✲
Figure 10.2: An executing Stack Processor. Elementary functions are performed by ALU
on variables stored in a stack (LIFO) memory. The decoder supports the one-cycle execution of the
instructions fetched from the external memory.
Figure 10.3 represents the Verilog top module for our Stack Processor (stack processor).
The two loop connected automata are SALU and PROGRAM FETCH. Both are simple, recursive
defined structures. The complexity of the Stack Processor is given by the DECODE unit: a combinational circuit used to trans-code op code providing 5 small command words to specify how behaves
290
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
module stack_processor(input
clock
,
input
reset
,
output [31:0] instr_addr ,
// instruction address
output [31:0] data_addr
,
// data address
output [1:0]
data_mem_com,
// data memory command
output [31:0] data_out
,
// data output
input
[23:0] instruction ,
// instruction input
input
[31:0] data_in
,
// data input
input
mem_ready
); // data memory is ready
// INTERNAL CONNECTIONS
wire
[2:0]
stack_com
;
// stack command
wire
[3:0]
alu_com
;
// alu command
wire
[1:0]
data_sel
;
// data selection for SALU
wire
[2:0]
pc_com
;
// program counter command
wire
[31:0] tos
,
// top of stack
ret_addr
;
// return from subroutine address
// INSTRUCTION DECODER UNIT
decode decode( .op_code
(instruction[23:16]) ,
.test_in
(tos)
,
.mem_ready
(mem_ready)
,
.stack_com
(stack_com)
,
.alu_com
(alu_com)
,
.data_sel
(data_sel)
,
.pc_com
(pc_com)
,
.data_mem_com
(data_mem_com)
);
// SALU: STACK WITH ARITHMETIC & LOGIC UNIT
salu
salu(
.stack0
(tos)
,
.stack1
(data_out)
,
.in1
(data_in)
,
.in2
({16’b0, instruction[15:0]}),
.in3
(ret_addr)
,
.s_com
(stack_com)
,
.data_sel
(data_sel)
,
.alu_com
(alu_com)
,
.reset
(reset)
,
.clock
(clock)
);
assign data_addr = tos;
// PROGRAM COUNTER UNIT
program_counter pc( .clock
(clock)
,
.reset
(reset)
,
.addr
(instr_addr)
,
.inc_pc
(ret_addr)
,
.value
(instruction[15:0]) ,
.tos
(tos)
,
.pc_com
(pc_com)
);
endmodule
Figure 10.3: The top level structural description of a Stack Processor. The Verilog code
associated to the circuit represented in Figure 10.2.
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
291
each component of the system. The Verilog decode module uses test in = tos and mem ready to make
decisions. The value of tos can be tested (if it is zero or not, for example) to decide a conditional jump
in program (on this way only PROGRAM FETCH module is affected). The mem ready input received
from data memory allows the processor to adapt itself to external memories having different access time.
The external data and program memories are both synchronous: the content addressed in the current
clock cycle is received back in the next clock cycle. Therefore, instruction received in each clock cycle
corresponds to instr addr generated in the previous cycle. Thus, the fetch mechanism fits perfect with
the behavior of the synchronous memory. For data memory mem ready flag is used to “inform” the decode
module to delay one clock cycle the use of the data received from the external data memory.
In each clock cycle ALU unit from SALU receives on its data inputs the two outputs of the stack,
and generates the result of the operation selected by the alu com code. If MUX4 has the input 0 selected
by the data sel code, then the result is applied to the input of stack. The result is written back in
tos if a unary operation (increment, for example) is performed (write the result of increment in tos is
equivalent with the sequence pop, increment & push). If a binary operation (addition, for example) is
performed, then the first operand is popped from stack and the result is written back over the the new
tos (double pop & push involved in a binary operation is equivalent with pop & write).
MUX4 selects for the stack input, according to the command data sel, besides the output of ALU,
data received back from the external data memory, the value carried by the currently executed instruction,
or the value pc+1 (to be used as return address).
module decode( input
input
input
output
output
output
output
output
[7:0]
[31:0]
[2:0]
[3:0]
[1:0]
[2:0]
[1:0]
op_code
,
test_in
,
mem_ready
,
stack_com
,
alu_com
,
data_sel
,
pc_com
,
data_mem_com);
‘include "micro_architecture.v"
‘include "instruction_set_architecture.v"
‘include "decoder_implementation.v"
endmodule
Figure 10.4: The decode module. It contains the three complex components of the description of
Stack Processor.
The unit PC generates in each clock cycle the address for program memory. It uses mainly the
value from the register PC, which contains the last used address, to fetch an instruction. The content of
tos or the value contained in the current instruction are also used to compute different conditioned or
unconditioned jumps.
To keep this example simple, the program memory is a synchronous one and it contains anytime the
addressed instruction (no misses in this memory).
Because our Stack Processor is designed to be an executing machine, besides the block associated
with the elementary functions (SALU) and the block used to compose & and loop them (PC) there is only
a decoder used as execution unit (see Figure 9.13. The decoder module – decode – is the most complex
module of Stack Processor (see Figure 10.4). It contains three sections:
• micro-architecture: it describes the micro-operations performed by each top level block listing
the meaning of all binary codes used to command them
• instruction set architecture: describe each instruction performed by Stack Processor
• decoder implementation: describe how the micro-architecture is used to implement the instruction set architecture.
10.2.2
∗ The micro-architecture
Any architecture can be implemented using various micro-architectures. For our Stack Processor one of
them is presented in Figure 10.5.
292
CHAPTER 10. COMPUTING MACHINES:
// MICROARCHITECTURE
// pc_com
parameter
stop
=
next
=
small_jmp
=
big_jmp
=
abs_jmp
=
ret_jmp
=
// alu_com
parameter
alu_left
=
alu_right
=
alu_inc
=
alu_dec
=
alu_add
=
alu_sub
=
alu_shl
=
alu_half
=
alu_zero
=
alu_equal
=
alu_less
=
alu_carry
=
alu_borrow =
alu_and
=
alu_or
=
alu_xor
=
// data_sel
parameter
alu
=
mem
=
val
=
return
=
// stack_com
parameter
s_nop
=
s_swap
=
s_push
=
s_write
=
s_pop
=
s_popwr
=
s_pop2
=
// data_mem_com
parameter
mem_nop
=
read
=
write
=
Figure 10.5:
3’b000,
3’b001,
3’b010,
3’b011,
3’b100,
3’b101;
//
//
//
//
//
//
pc
pc
pc
pc
pc
pc
=
=
=
=
=
=
≥4–LOOP DIGITAL SYSTEMS
pc
pc + 1
pc + value
pc + tos
value
tos
4’b0000,
4’b0001,
4’b0010,
4’b0011,
4’b0100,
4’b0101,
4’b0110,
4’b0111,
4’b1000,
4’b1001,
4’b1010,
4’b1011,
4’b1100,
4’b1101,
4’b1110,
4’b1111;
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
//
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
alu_out
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
left
right
left + 1
left - 1
left + right = add[32:0]
left - right = sub[32:0]
{1’b0, left[31:1]}
{left[31], left[31:1]}
{31’b0, (left == 0)}
{31’b0, (left == right)}
{31’b0, (left < right)}
{31’b0, add[32]}
{31’b0, sub[32]}
left & right
left | right
left ^ right
2’b00,
2’b01,
2’b10,
2’b11;
//
//
//
//
stack_input
stack_input
stack_input
stack_input
3’b000,
3’b001,
3’b010,
3’b100,
3’b101,
3’b110,
3’b111;
//
//
//
//
//
//
//
no operation
swap the content of the first two
push
write in tos
pop
pop2 & push
pops two values
2’b00,
2’b01,
2’b10;
// no data memory command
// read from data memory
// write to data memory
=
=
=
=
alu_out
data_in
value
ret_addr
The micro-architecture of our Stack Processor. The content of file
micro architecture.v defines each command word generated by the decoder describing the associated
micro-commands and their binary codes.
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
293
The decoder unit generates in each clock cycle a command word having the following 5-field structure:
{alu com, data sel, stack com data mem com, pc com} = command
where:
• alu com: is a 4-bit code used to select the arithmetic or logic operation performed by ALU in the
current cycle; it specifies:
– well known binary operations such as: add, subtract, and, or, xor
– usual unary operations such as: increment, shifts
– test operations indicating by alu out[0] the result of testing, for example: if an input is zero
or if an input is less than another input
• data sel: is a 2-bit code used to select the value applied on the input of the stack for the current
cycle as one from the following:
– the output of ALU
– data received from data memory addressed by tos (with a delay of one clock cycle controlled
by mem ready signal because the external data memory is synchronous)
– the 16-bit integer selected from the current instruction
– pc+1, generated by the PROGRAM FETCH module, to be pushed in stack when the a call
instruction is executed
• stack com: is a 3-bit code used to select the operation performed by the stack unit in the current
cycle (it is correlated with the ALU operation selected by alu com); the following micro-operations
are codded:
– push: it is the well known standard writing operation into a stack memory
– pop: it is the well known standard reading operation into a stack memory
– write: it writes in top of stack, which is equivalent with popping an operand and pushing
back the result of operation performed on it (used mainly in performing unary operations)
– pop & write: it is equivalent with popping two operands from stack and pushing back the
result of operation performed on them (used mainly in performing binary operations)
– double pop: it is equivalent with two successive pops, but is performed in one clock cycle;
some instructions need to remove both the content of stack0 and of stack1 (for example,
after a data write into the external data memory)
– swap: it exchange the content of stack0 and of stack1; it is useful, for example to make a
subtract in the desired order.
• data mem com: is a 2-bit command for the external data memory; it has three instantiations:
– memory nop: keep memory doing nothing is a very important command
– read: commands the read operation from data memory with the address from tos; the data
will be returned in the next clock cycle; in the current cycle mem read is activated to allow
stoping the processor one clock cycle (the associated read instruction will be executed in two
clock cycles)
– write: the data contained in stack1 is written to the address contained in stack0 (both,
address and data will be popped from stack)
• pc com: is a 3-bit code used to command how is computed the address for the fetching of the next
instruction; 6 modes are used:
– stop: program counter is not incremented (the processor halts or is waiting for a condition
to be fulfilled)
– next: it is the most frequent mode to compute the program counter by incrementing it
294
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
– small jump: compute the next program counter adding to it the value contained in the
current instruction (instruction[15:0]) interpreted as a 16-bit signed integer; a relative
jump in program is performed
– big jump: compute the next program counter adding to it the value contained in tos interpreted as a 32-bit signed integer; a relative big jump in program is performed
– absolute jump: the program counter takes the value of instruction[15:0]; thhe processor
performs an absolute jump in program
– return jump: is an absolute jump performed using the content of tos (usually performs a
return from a subroutine, or is used to call a subroutine in a big addressing space)
The 5-field just explained can not be filled up without inter-restrictions imposed by the meaning of the
micro-operations. There exist inter-correlations between the micro-operations assembled in a command.
For example, if ALU performs an addition, then the stack must perform mandatory pop & poop& &
push == pop write. If the ALU operation is increment, then the stack must perform write. Some fields
are sometimes meaningless. For example, when an unconditioned small jump is performed the fields
alu com and data sel can take don’t care values. But, for obvious reasons, no times stack com and
data mem com can take don’t care values.
Each unconditioned instruction has associated one 5-field commands, and each conditioned instructions is defined using two 5-field commands.
10.2.3
∗ The instruction set architecture
Instruction set architecture is the interface between the hardware and the software part of a computing
machine. It grounds the definition of the lowest level programming language: the assembly language.
It is an interface because allows the parallel work of two teams once its definitions is frozen. One is the
hardware team which starts to design the physical structure, and the other is the software team which
starts to grow the symbolic structure of the hierarchy of programs. Each architecture can be embodied
in many forms according to the technological restrictions or to the imposed performances. The main
benefit of this concept is the possibility to change the hardware without throwing out the work done by
the software team.
The implementation of our Stack Processor has, as the majority of the currently produced processors,
an instruction set architecture containing the following class of instructions:
arithmetic and logic instructions having the form:
• [stack0, stack1, s2, ...] = [op(stack0, stack1), s2, ...]
where: stack0 is the top of stack, stack1 is the next recording in stack, and op is an arithmetic
or logic binary operation
• [stack0, stack1, s2, ...]
if the operation op is unary
= [(op(stack0), stack1, s2, ...]
input-output instructions which uses stack0 as data addr and stack1 as data out
stack instructions (only for stack processors) used to immediate load the stack or to change the content
in the first two recordings (stack0 and stack1)
test instructions used to test the content of stack putting the result of the test back into the stack
control instructions used to execute unconditioned or conditioned jumps in the instruction stream by
modifying the variable program counter used to address in the program space.
The instruction set architecture is given as part of the Verilog code describing the module
decode: the content of the file instruction set architecture.v (a more complete stage of this
module in Appendix: Designing a stack processor). Figure 10.6 contains an incipient form of file
instruction set architecture.v. From each class of instructions only few examples are shown. Each
instruction is performed in one clock cycle, except load whose execution can be delayed if data ready
= 0.
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
295
// INSTRUCTION SET ARCHITECTURE
// arithmetic & logic instructions (pc <= pc + 1)
parameter
nop
= 8’b0000_0000, // s0, s1, s2 ... <= s0, s1, s2, ...
add
= 8’b0000_0001, // s0, s1, s2 ... <= s0 + s1, s2, ...
inc
= 8’b0000_0010, // s0, s1, s2 ... <= s0 + 1, s1, s2, ...
half
= 8’b0000_0011; // s0, s1, s2 ... <= s0/2, s1, s2, ...
// ...
// input output instructions (pc <= pc + 1)
parameter
load
= 8’b0001_0000, // s0, s1, s2 ... <= data_mem[s0], s1, s2, ...
store
= 8’b0001_0001; // s0, s1, s2 ... <= s2, s3, ...; data_mem[s0] = s1
// stack instructions (pc <= pc + 1)
parameter
push
= 8’b0010_0000, // s0, s1, s2 ... <= value, s0, s1, ...
pop
= 8’b0010_0010, // s0, s1, s2 ... <= s1, s2, ...
dup
= 8’b0010_0011, // s0, s1, s2 ... <= s0, s0, s1, s2, ...
swap
= 8’b0010_0100, // s0, s1, s2 ... <= s1, s0, s2, ...
over
= 8’b0010_0101; // s0, s1, s2 ... <= s1, s0, s1, s2, ...
// ...
// test instructions (pc <= pc + 1)
parameter
zero
= 8’b0100_0000, // s0, s1, s2 ... <= (s0 == 0), s1, s2, ...
eq
= 8’b0100_0001; // s0, s1, s2 ... <= (s0 == s1), s2, ...
// ...
// control instructions
parameter
jmp
= 8’b0011_0000, // pc = pc + value
call
= 8’b0011_0001, // pc = s0; s0, s1, ... <= pc + 1, s1, ...
cjmpz
= 8’b0011_0010, // if (s0 == 0) pc <= pc + value, else pc <= pc + 1
cjmpnz = 8’b0011_0011, // if (s0 == 0) pc <= pc + 1, else pc <= pc + value
ret
= 8’b0011_0111; // pc = s0; s0, s1, ... <= s1, s2, ...
// ...
Figure 10.6: Instruction set architecture of our Stack Processor. From each subset few
typical example are shown. The content of data stack is represented by: s0, s1, s2, ....
10.2.4
∗ Implementation: from micro-architecture to architecture
Designing a processor (in our case designing the Stack Processor) means to use the micro-architecture
to implement the instruction set architecture. For an executing processor the ”connection” between
micro-architecture and architecture is done by the decoder which is a combinational structure.
The main body of the decode module – decoder implementation.v – contains the description of
the Stack Processor’s instruction set architecture in term of micro-architecture.
The structure of the file decoder implementation.v is suggested in Figure 10.7, where the output
variables are the 5 command fields (declared as registers) and the input variables are: the operation code
from instruction, the value of tos received as test in and the flag received from the external memory:
mem ready.
The main body of this vile consists in a big case structure with an entry for each instruction. In
Figure 10.7 only few instructions are implemented (nop, add, load) to show how an unconditioned
instruction nop, add or a conditioned instruction load is executed.
Instruction nop does not affect the state of stack and PC is incremented. We‘must take care only
about three command fields. PC must be incremented (next, and the fields commanding memory resources (stack, external data memory) must be set on ”no operation” (s nop, mem nop. The operation
performed by ALU and data selected as right operand have no meaning for this instruction.
296
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
Instruction add pops the two last recordings in stack, adds them, pushes back the result in tos, and
increments PC. Meantime the data memory receives no active command.
Instruction load is executed in two clock cycles. In the first cycle, when mem ready = 0, the command read is sent to the external data memory, and the PC is maintained unchanged. The operation
performed by ALU does not matter. The selection code for MUX4 does not matter. In the next clock
cycle data memory sets it flag on 1 (mem ready = 1 means the requested data is available), data selected
is from memory mem), and the output of MUX4 is pushed in stack ((s push).
By default the decoder generates “dont’care” commands. Another possibility is to have nop instruction
the “by default” instruction. Or by default to have a halt instruction which stops the processor. The
first version is good as a final solution because generates a minimal solution. The last version is preferred
in the initial stage of development because provides an easy testing and debugging solution.
Follows the description of some typical instructions from a possible instruction set executed by our
Stack Processor.
Instruction inc increments the top of stack, and increments also PC. The right operand of ALU
does not matter. The code describing this instruction, to be inserted into the big case sketched in Figure
10.7, is the following:
inc
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next
alu_inc
alu
s_write
m_nop
;
;
;
;
;
end
Instruction store stores the value contained in stack1 at the address from stack0 in external data
memory. Both, data and address are popped from stack. The associated code is:
store
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next ;
4’bx ;
2’bx ;
s_pop2;
write ;
end
Instruction push pushes {16’b0, instruction[15:0]} in in stack. The code is:
push
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next
4’bx
val
s_push
m_nop
;
;
;
;
;
end
Instruction dup pushes in stack the top of stack, thus duplicating it. ALU performs alu left, the
right operand does not matter, and in the stack is pusher the output of ALU. The code is:
dup
:
begin
end
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next
alu_left
alu
s_push
m_nop
;
;
;
;
;
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
297
// THE IMPLEMENTATION
reg
reg
reg
[3:0]
[2:0]
[1:0]
alu_com
pc_com, stack_com
data_sel, data_mem_com
;
;
;
always @(op_code or test_in or mem_ready
)
case(op_code)
// arithmetic & logic instructions
nop :
begin
pc_com
= next
;
alu_com
= 4’bx
;
data_sel
= 2’bx
;
stack_com
= s_nop
;
data_mem_com
= mem_nop
;
end
add :
begin
pc_com
= next
;
alu_com
= alu_add
;
data_sel
= alu
;
stack_com
= s_popwr
;
data_mem_com
= mem_nop
;
end
// ...
// input output instructions
load
:
if (mem_ready)
begin
pc_com
= next
alu_com
= 4’bx
data_sel
= mem
stack_com
= s_write
data_mem_com
= mem_nop
end
else
begin
pc_com
= stop ;
alu_com
= 4’bx ;
data_sel
= 2’bx ;
stack_com
= s_nop ;
data_mem_com
= read ;
end
// ...
// ...
default
begin
pc_com
= 3’bx ;
alu_com
= 4’bx ;
data_sel
= 2’bx ;
stack_com
= 3’bx ;
data_mem_com
= 2’bx ;
end
endcase
;
;
;
;
;
Figure 10.7: Sample from the file decoder implementation.v. Implementation consists in a
big case form with an entry for each instruction.
298
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
Instruction over pushes stack1 in stack, thus duplicating the second stage of stack. ALU performs
alu right, and in the stack is pusher the output of ALU.
over
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next
alu_right
alu
s_push
m_nop
;
;
;
;
;
end
The sequence of instructions:
over;
over;
duplicates the first two recordings in stack to be reused later in another stage of computation.
Instruction zero substitute the top of stack with 1, if its content is 0, or with 0 if the content is
different from 0.
zero
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
next
alu_zero
alu
s_write
m_nop
;
;
;
;
;
end
This instruction is used in conjunction with a conditioned jump (cjmpz or cjmpnz) to decide according
to the value of stack0.
Instruction jmp adds to PS the signed value instruction[15:0].
jmp
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
rel_jmp
4’bx
2’bx
s_nop
m_nop
;
;
;
;
;
end
This instruction is expressed as follows:
jmp <value>
where, <value> is expressed sometimes as an explicit signed integer, but usually as a label which takes
a numerical value only when the program is assembled. For example:
jmp loop1;
Instruction
call performs an absolute jump to the subroutine placed at the address
instruction[15:0], and saves in tos the return address (ret addr) which is pc + 1. The address
saved in stack will be used by ret instruction to return the processor from the subroutine into the main
program.
call
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
end
The instruction is used, for example, as follows:
jmp subrt5;
where subrt5 is the label of a certain subroutine.
abs_jmp
4’bx
return
s_write
m_nop
;
;
;
;
;
10.2. ∗ THE STACK PROCESSOR – A PROCESSOR AS 4-OS
299
Instruction cjmpz performs a relative jump if the content of tos is zero; else PC is incremented.
The content of stack is unchanged. (A possible version of this instruction pops the tested value from the
stack.)
cjmpz
:
if (test_in == 32’b0)
begin
pc_com
alu_com
data_sel
stack_com
data_mem
end
else
begin
pc_com
alu_com
data_sel
stack_com
data_mem
end
=
=
=
=
=
small_jmp
4’bx
2’bx
s_nop
m_nop
=
=
=
=
=
next
4’bx
2’bx
s_nop
m_nop
;
;
;
;
;
;
;
;
;
;
The instruction is used, for example, as follows:
jmp george;
where george is a label to be converted in a signed 16-bit integer in the assembly process.
Instruction ret performs a jump from subroutine back into the main program using the address
popped from tos.
ret
:
begin
pc_com
alu_com
data_sel
stack_com
data_mem
=
=
=
=
=
ret_jmp
4’bx
2’bx
s_pop
m_nop
;
;
;
;
;
end
The hardware resources of this Stack Processor permits up to 256 instructions to be defined. For
this simple machine we do not need to define too many instructions. Therefore, a “smart” codding of
instructions will allow minimizing the size of decoder. More, for some critical paths the depth of decoder
can be also minimized, eventually reduced to zero. For example, maybe it is possible to set
alu com = instruction[19:16]
data sel = instruction[21:20]
allowing the critical loop to be closed faster.
10.2.5
∗ Time performances
Evaluating the time behavior of the just designed machine does not make us too happy. The main reason
is provided by the fact that all the external connections are unbuffered.
All the three inputs, instruction, data in, mem ready must be received long time before the
active edge of clock because their combinational path inside the Stack Processor are too deep. More,
these paths are shared partially with the internal loops responsible for the maximum clock frequency.
Therefore, optimizing the clock interferes with optimizing tin reg .
Similar comments apply to the output combinational paths.
The most disturbing propagation path is the combinational path going from inputs to outputs (for
example: from instruction to data mem com). The impossibility to avoid tin out make this design very
unfriendly at the system level. Connecting this module together with a data memory a program memory
and some input-output circuits will generate too many (restrictive) time dependencies.
300
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
This kind of approach can be useful only if it is strongly integrated with the design of the associated
memories and interfaces in a module having all inputs and outputs strictly registered.
The previously described Stack Processor remains to be a very good bad example of a pure functionally
centered design which ignores the basic electrical restrictions.
10.2.6
∗ Concluding about our Stack Processor
The simple processor exemplified by Stack Processor is typical for a computational engine: it contains an
simple working 3loop system – SALU – and another simple automaton – Program Fetch – both driven
by a decoder to execute what is codded in each fetched instruction. Therefore, the resulting system is
a 4th order one. This is not the solution! A lot of improvement are possible, and a lot of new features
can be added. But it is very useful to exemplify one of the main virtue of the fourth loop: the 4-loop
processing. A processor with more than the minimal 3 loops is easiest to be controlled. In our cases the
operands are automatically selected by the stack mechanism. Results a lot of advantages in control and
some performance loss. But, the analysis of pros & cons is not a circuit design problem. It is a topics to
be investigated in the computer architecture domain.
The main advantages of a stack machine is its simplicity. The operands are in each cycle already
selected, because they are the first to recording in the top of stack. Results a simple instruction containing
only two fields: op code[7:0] and value[15:0].
The loop inside SALU is very short allowing a high clock frequency (if other loop do not impose a
smaller one).
The main disadvantage of a stack machine is the stack discipline which sometimes adds new instructions in the code generated by the compiler.
Writing a compiler for this kind of machine is simple because the discipline in selecting the operands
is high. The efficiency of the resulting code is debatable. Sometimes a longer sequence of operation is
compensated by the higher frequency allowed by a stack architecture.
A real machine can adopt a more sophisticated stack in order to remove some limitation imposed by
the restricted access imposed by the discipline.
10.3
Embedded computation
Now we are prepared to revisit the Chapter OUR FINAL TARGET in order to offer an optimal
implementation for the small & simple system toyMachine. The main application for such
a machine is in the domain of the embedded computation. The technology of embedded
computation uses programmable machines of various complexity to implement by programming
functions formerly implemented by big & complex circuits.
Instead of the behavioral description by the module toyMachine (see Figure 5.4) we are
able to provide now a structural description. Even if the behavioral description offered by the
module toyMachine is synthesisable will we see that the following structural version provides a
half sized circuit.
10.3.1
The structural description of toyMachine
A structural description is supposed to be a detailed description which provide a hierarchical
description of the design using on the “leafs of the tree” simple and optimal circuits. A structural
description answers the question of “how”.
The top module
In the top module of the design – toyMachineStructure – there are two structures (see Figure
10.8):
controlSection : manages the instruction flow read from the program memory and executed,
one per clock cycle, by the entire system; the specific control instructions are executed by
10.3. EMBEDDED COMPUTATION
301
this module using data, when needed, provided by the other modules (“dialog” bits for the
stream flow, values from controlSection); the asynchronous inta signal constrains the
specific action of jumping to the instruction addressed with the content of refFile[31]
dataSection : performs the functional aspect of computation, operating on data internally
stored by the register file, or received from the external world; it generate also the output
signals loading the output register outRegister with the results of the internal computation.
toyMachine
reset
int
clock
readyIn
readyOut
instruction
dataIn
inStream
✲
✲
✲
✲
✲
✲ immValue
✲ opCode
✲
✲
✲
✲ opCode
✲ destAddr
✲ leftAddr
✲ rightAddr
✲ immValue
✲
✲
✲
✲
✲
controlSection
✲ inta
✲ progAddr
dataSection
✲ dataAddr
✲ dataOut
✲ outStream
✲ write
✲ readIn
✲ writeOut
Figure 10.8: The top level block schematic of the toyMachine design.
The block dataSection is a third order (3-loop) digital system having the third loop closed
over the regFile through alu with carry (see Figure 10.9). The second loop is closed over alu
and the carry flip-flop. The first loop in this section is closed in the latches used to build the
module regFile and the flip-flop carry.
The block controlSection is a second order (2-loop) digital system, because the first loop
is closed in the master-slave flip-flops of the (programCounter module, the second loop is closed
over programCounter through pcMux inc and add (see Figure 10.9).
Thus, the toyMachine system is a fourth order digital system, the last loop being
closed through dataSection and controlSection. The module controlSection sends to the
dataSection the value progAddr as the return address from the sub-routine associated to the
interruupt. The module dataSection sends back to the module controlSection the absolute
jump address.
See in Figure 10.10 the code describing the top module. Unlike the module toyMachine
(see Chapter OUR FINAL TARGET), which describe on one level design the behavior of toyMachine, the module toyMachineStucture is a pure structural description providing only the
302
CHAPTER 10. COMPUTING MACHINES:
reset
int
inta
≥4–LOOP DIGITAL SYSTEMS
✻
✲
progAddr
❄❄
❄
❄
programCounter
intEnable
✲
✲
ieMux
✻
✻✻✻✻
”1”
✲
✲
✲
inc
✲
adder
nextIESel
✻
contrDecode
✻
pcMux
nextPcSel
instruction
contrloSection
write
✛
✲
✲
❄
❄
❄
✲
result
writeBack
destAddr
regFle
dataDecode
leftOp
dataOut
addrOut
leftAddr
rightAddr
rightOp
✛
✛
✛
✛
✛
arithLogOp
❄❄ ❄
❄❄ ❄
logicModule
arithModule
✛
dataIn
❄❄
❄❄❄❄❄❄❄❄
✲
carry
resultMux
readyOut
readyIn
readIn
alu
✛
✲writeOut
❄
❄❄
inRegister
inStream
✻
outRegister
dataSection
Figure 10.9:
✲outStream
10.3. EMBEDDED COMPUTATION
module toyMachineStructure
(
input
[15:0] inStream
,
input
readyIn
,
output
readIn
,
output [15:0] outStream
,
input
readyOut
,
output
writeOut
,
input
int
,
output
inta
,
output [31:0] dataAddr, dataOut, progAddr ,
output
write
,
input
[31:0] dataIn, instruction
,
input
reset, clock
); // 414 LUTs, 170 MHz
wire
[31:0] leftOp, immValue
;
wire
[5:0]
opCode
;
wire
[4:0]
destAddr, leftAddr, rightAddr
;
assign opCode
= instruction[31:26]
;
assign destAddr
= instruction[25:21]
;
assign leftAddr
= instruction[20:16]
;
assign rightAddr
= instruction[15:11]
;
assign immValue
= {{16{instruction[15]}}, instruction[15:0]};
dataSection dataSection(inStream
,
readyIn
,
readIn
,
outStream
,
readyOut
,
writeOut
,
dataAddr
,
dataOut
,
write
,
dataIn
,
progAddr
,
opCode
,
destAddr
,
leftAddr
,
rightAddr
,
immValue
,
inta
,
clock
);
controlSection controlSection( int
,
inta
,
progAddr,
dataAddr,
opCode ,
immValue,
readyIn ,
readyOut,
reset
,
clock
);
endmodule
Figure 10.10: The top module toyMachineStructure.
303
304
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
module controlSection(
input
int
,
output
inta
,
output [31:0] progAddr
,
input
[31:0] dataAddr
,
input
[5:0]
opCode
,
input
[31:0] immValue
,
input
readyIn
,
input
readyOut
,
input
reset
,
input
clock
);
reg
[31:0] programCounter
;
reg
intEnable
;
wire
[1:0]
nextPcSel, nextIESel;
wire
nextIE
;
assign inta = intEnable & int;
contrDecode
contrDecode(opCode
,
dataAddr
,
inta
,
readyIn
,
readyOut
,
nextPcSel
,
nextIESel
);
always @(posedge clock) if (reset) begin
programCounter
intEnable
end
else
begin
programCounter
intEnable
end
mux4_32 pcMux( .out(progAddr
),
.in0(programCounter
),
.in1(programCounter + 1
),
.in2(programCounter + immValue ),
.in3(dataAddr
),
.sel(nextPcSel
));
mux4_1 ieMux(
.out(nextIE
),
.in0(intEnable ),
.in1(1’b0
),
.in2(1’b1
),
.in3(1’b0
),
.sel(nextIESel ));
endmodule
<= 32’b0;
<= 1’b0 ;
<= progAddr ;
<= nextIE
;
Figure 10.11: The module controlSection.
10.3. EMBEDDED COMPUTATION
305
top level description of the same digital system. It contains two modules, one for each main
sub-system of or design.
The interrupt
There are many ways to solve the problem of the interrupt signal in a computing machine.
The solutions are different depending on the way the signals int and inta are connected to
the external systems. The solution provided here is the simplest one. It is supposed that both
signals are synchronous with the toyMachine structure. This simple solution consists of a
2-state half-automaton (the one-bit register intEnable and the multiplexer ieMux).
Because the input int is considered synchronously generated with the system clock, the
signal inta is combinational generated.
The next subsection provides an enhanced version of this module which is able to manage
asynchronous int signal.
module contrDecode( input
[5:0]
opCode
,
input
[31:0] dataAddr
,
input
inta
,
input
readyIn
,
input
readyOut
,
output reg [1:0]
nextPcSel
,
output reg [1:0]
nextIESel
);
‘include "0_toyMachineArchitecture.v"
always @(*) if (inta)
nextIESel = 2’b10
;
else if (opCode == ei)
nextIESel = 2’b01
;
else if (opCode == di)
nextIESel = 2’b10
;
else
nextIESel = 2’b00
;
always @(*) if (inta)
nextPcSel = 2’b11
else case(opCode)
jmp
:
nextPcSel = 2’b11
zjmp
: if (dataAddr == 0)
nextPcSel = 2’b10
else
nextPcSel = 2’b01
nzjmp
: if (dataAddr !== 0)
nextPcSel = 2’b10
else
nextPcSel = 2’b01
receive : if (readyIn)
nextPcSel = 2’b01
else
nextPcSel = 2’b00
issue
: if (readyOut)
nextPcSel = 2’b01
else
nextPcSel = 2’b00
halt :
nextPcSel = 2’b00
default
nextPcSel = 2’b01
endcase
endmodule
;
;
;
;
;
;
;
;
;
;
;
;
Figure 10.12: The module contrDecode.
The control section
This unit fetches in each clock cycle a new instruction from the program memory. The instruction
is decoded locally for its use and is also sent for the use of the data unit. For each instruction
there is a specific way to use the content of the program counter in order to compute the address
of the next instruction. For data and interrupt instructions (see Figure 5.3) the next instruction
is always fetched form the address programCounter + 1. For the control instructions (see
Figure 5.3) there are different modes for each instruction. The internal structure of the module
306
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
controlSection is designed to provide the specific modes of computing the next value for the
program counter.
The multiplexers pcMux from the control section (see Figure 10.9) is used to select the next
value of the program counter, providing the value of progAddr, as follows:
• program counter keep its own value for the halt instruction or in the wait instructions for
input or output to become ready
• program counter is incremented for the linear part of the program
• program counter is added to the immValue provided by the current instruction
• program counter is set to the value provided by regFile[leftAddr] for unconditioned
jump
The selection bits for pcMux are generated by the contrDecode. It uses for generating
the selections for the multiplexers opCode from instruction, asyncInta, the content of
regFile[leftAddr] and the input signals readyIn, readyOut.
The only complex module in the control section is the combinational circuit described in
contrDecode (see Figure 10.12). The type reg in this description must be understood as a
variable. The actual structure of a register is not generated.
The data section
The module dataSection includes mainly the data storage resources and the combinational
circuits allowing the execution of each data instruction in one clock cycle.
Data is stored in the register file, regFile, which allows to read two variable as operands
for the current instruction, selected by leftAddr and rightAddr, and to store the result of the
current instruction to the location selected by desrAddr (except the case when inta = 1 forces
reading leftOp form the location 30, to be used as absolute jump address, and loading the
location 31 with the current value of programCounter).
The arithmetic-logic unit, alu, operate in each clock cycle on the operands received from the
two outputs of the register file: leftOp and rightOp. The operation code is given directly from
the output of the dataDecode block described by the module dataDecode (see Figure 10.18).
The input of the register file is provided from the alu output and from other four sources:
• inRegister: because the input bits can be submitted to arithmetic and logic processing
only if they are stored in the register file first
• immValue: is used to generate immediate values for the purpose of the program
• dataIn: data provided by the external data memory addressed by leftOp
• programCounter: is saved as the “return” address to be used after running the program
started by the acknowledged interrupt
The first three of these inputs are selected according to the current instructions by the selection
code resultSel generated by the module dataDecode for the multiplexor resultMux. The last
one is forced at the input of the register file by the occurrence of the signal inta.
The register file description uses the code presented in the subsection Register file. Only
the sizes are adapted to our design (see Figure 10.14.
The module called alu (see Figure 10.15) performs the arithmetic-logic functions of our
small instruction set. Because the current synthesis tools are able to synthesize very efficiently
uniform arithmetic and logic circuits, this Verilog contains a behavioral description.
10.3. EMBEDDED COMPUTATION
module dataSection( input
[15:0] inStream
,
input
readyIn
,
output
readIn
,
output [15:0] outStream
,
input
readyOut
,
output
writeOut
,
output [31:0] dataAddr
,
output [31:0] dataOut
,
output
write
,
input
[31:0] dataIn
,
input
[31:0] progAddr
,
input
[5:0]
opCode
,
input
[4:0]
destAddr
,
input
[4:0]
leftAddr
,
input
[4:0]
rightAddr
,
input
[31:0] immValue
,
input
inta
,
input
clock
);
reg
[15:0] inRegister ;
reg
[15:0] outRegister ;
reg
carry
;
wire
[31:0] result
;
wire
carryOut
;
wire
[31:0] leftOp
;
wire
[31:0] rightOp
;
wire
[1:0]
arithLogOp ;
wire
[2:0]
resultSel
;
wire
writeBack
;
wire
inRegEnable ;
wire
outRegEnable;
wire
carryEnable ;
assign dataAddr
= leftOp
;
assign dataOut
= leftOp
;
assign outStream
= outRegister
;
dataDecode dataDecode
(arithLogOp, resultSel, writeBack, inRegEnable, outRegEnable,
carryEnable, readIn, writeOut, write, opCode, readyIn, readyOut, inta
always @(posedge clock) begin
if (inRegEnable )
inRegister <= inStream
if (outRegEnable)
outRegister <= leftOp[15:0]
if (carryEnable)
carry
<= carryOut
end
regFile regFile(.destAddr
(inta ? 5’b11110 : destAddr ),
.writeBack (writeBack
),
.leftAddr
(inta ? 5’b11111 : leftAddr ),
.rightAddr (rightAddr
),
.in
(result
),
.leftOut
(leftOp
),
.rightOut
(rightOp
),
.clock
(clock
));
alu alu(result, carryOut, leftOp, rightOp, carry, inRegister,
progAddr, dataIn, immValue, arithLogOp, resultSel
);
endmodule
Figure 10.13: The module dataSection.
307
);
;
;
;
308
CHAPTER 10. COMPUTING MACHINES:
module regFile( input
input
input
input
input
output
output
input
[4:0]
[4:0]
[4:0]
[31:0]
[31:0]
[31:0]
reg [31:0] regFile[0:31]
destAddr
writeBack
leftAddr
rightAddr
in
leftOut
rightOut
clock
≥4–LOOP DIGITAL SYSTEMS
,
,
,
,
,
,
,
);
;
always @(posedge clock) if (writeBack) regFile[destAddr] <= in
;
assign leftOut = regFile[leftAddr] ;
assign rightOut = regFile[rightAddr];
endmodule
Figure 10.14: The module regFile.
The dataDecode block, described in our design by the Verilog module dataDecoder, takes
the opCode field from instruction, the dialog signals, readyIn and readyOut, and inta and
trans-codes them. This is the only complex module from the data section.
Multiplexors
The design of toyMachine uses a lot of multiplexors. Their description is part of the project.
As for the usual functions from an ALU, or small combinational circuits, for multiplexors
behavioral descriptions work very well because the software synthesis tools are enough “smart”
to “know” how to provide optimal solutions.
Concluding about toyMachine
For the same system – toyMachine – we have now two distinct descriptions: toyMachine,
the initial behavioral description (see Chapter OUR FINAL TARGET), and the structural
description toyMachineStructure just laid down in this subsection. Both descriptions, the
structural and the behavioral, are synthesisable, but the resulting structures are very different.
The synthesis of toyMachine design provides a number of components 5.85 times bigger
than the synthesis of the module toyMachineStructure. A detailed description (about 7105
symbols, without spaces) provided a smallest structure then the structure provided by the
behavioral description (about 3083 symbols, without spaces).
The actual structure generated by the behavioral description is not only bigger, but it is
completely unstructured. The structured version provided by the alternative design is easy to
understand, to debug and to optimize.
10.3.2
Interrupt automaton: the asynchronous version
Sometimes for the interrupt automaton a more rigorous solution is requested. In the already
provided solution the int signal must be stable until inta is activated. In many systems this is
an unacceptable restriction. Another restriction is the synchronous switch of int.
This new version for the interrupt automaton accepts an asynchronous int signal having any
width exceeding the period of the clock. The flow chart describing the automaton is in Figure
10.21. It has 4 states:
10.3. EMBEDDED COMPUTATION
module alu( output
output
input
input
input
input
input
input
input
input
input
wire
wire
[31:0]
[31:0]
[31:0]
[31:0]
[31:0]
[15:0]
[31:0]
[31:0]
[31:0]
[1:0]
[2:0]
309
result
carryOut
leftOp
rightOp
carry
inRegister
progAddr
dataIn
immValue
arithLogOp
resultSel
,
,
,
,
,
,
,
,
,
,
);
logicResult ;
arithResult ;
logicModule
logicModule(leftOp
rightOp
arithLogOp
logicResult
,
,
,
);
arithModule
arithModule(leftOp
rightOp
carry
arithLogOp
arithResult
carryOut
,
,
,
,
,
);
mux8_32 resultMux(
.out(result
.in0(arithResult
.in1(logicResult
.in2({leftOp[31], leftOp[31:1]}
.in3(immValue
.in4({immValue[15:0], leftOp[15:0]}
.in5({{16{inRegister[15]}}, inRegister}
.in6(progAddr
.in7(dataIn
.sel(resultSel
),
),
),
),
),
),
),
),
),
));
endmodule
Figure 10.15: The module alu.
module arithModule( input
input
input
input
output
output
[31:0]
[31:0]
[1:0]
[31:0]
leftOp
rightOp
carry
arithLogOp
arithResult
carryOut
,
,
,
,
,
);
assign {carryOut, arithResult} = leftOp + (rightOp ^ {32{arithLogOp[0]}}) +
(arithLogOp[1] & (arithLogOp[0] ^ carry));
endmodule
Figure 10.16: The module arithModule.
310
CHAPTER 10. COMPUTING MACHINES:
module logicModule( input
input
input
output
[31:0]
[31:0]
[1:0]
reg [31:0]
always @(*) case(arithLogOp)
2’b00: logicResult
2’b01: logicResult
2’b10: logicResult
2’b11: logicResult
endcase
endmodule
leftOp
rightOp
arithLogOp
logicResult
=
=
=
=
≥4–LOOP DIGITAL SYSTEMS
,
,
,
);
~leftOp
leftOp & rightOp
leftOp | rightOp
leftOp ^ rightOp
;
;
;
;
Figure 10.17: The module logicModule.
dis : the initial state of the automaton when the interrupt action is disabled
en : the state when the interrupt action is enabled
mem : is the state memorizing the occurrence of an interrupt when interrupt is disabled
inta : is the acknowledge state.
The input signals are:
int : is the asynchronous interrupt signal
ei : is a synchronous bit resulting from the decode of the instruction ei (enable interrupt)
di : is a synchronous bit resulting from the decode of the instruction di (disable interrupt)
The output signal is asyncInta. It is in fact a synchronous hazardous signal which will be
synchronized using a D–FF.
Because int is asynchronous it must be used to switch the automaton in another state in
which asyncInta will be eventually generated.
The state codding style applied for this automaton is imposed by a asynchronous int signal.
It will be of the reduced dependency by the asynchronous input variable int. Let us try first the
following binary codes (see the codes in square brackets in Figure 10.21) for the four states of
the automaton:
dis : Q1 Q0 = 00
en : Q1 Q0 = 11
mem : Q1 Q0 = 01
inta : Q1 Q0 = 10
The critical transitions are from the states dis and en, where the asynchronous input int is
tested. Therefore, the transitions from these two states takes place as follows:
+
+
+
′
• from state dis = 00: if (ei = 0) then {Q+
1 , Q0 } = {0, int}; else {Q1 , Q0 } = {1, int };
+
+
therefore: {Q1 , Q0 } = {ei, ei ⊕ int}
+
+
+
′
• from state en = 11: if (di = 0) then {Q+
1 , Q0 } = {1, int }; else {Q1 , Q0 } = {0, int};
+
+
′
′
therefore: {Q1 , Q0 } = {di , di ⊕ int}
10.3. EMBEDDED COMPUTATION
module dataDecode(
output
output
output
output
output
output
output
output
output
input
input
input
input
reg [1:0]
reg [2:0]
reg
reg
reg
reg
reg
reg
reg
[5:0]
311
arithLogOp ,
resultSel
,
writeBack
,
inRegEnable ,
outRegEnable,
carryEnable ,
readIn
,
writeOut
,
write
,
opCode
,
readyIn
,
readyOut
,
inta
);
‘include "0_toyMachineArchitecture.v"
always @(*) begin
arithLogOp
resultSel
writeBack
inRegEnable
outRegEnable
carryEnable
readIn
writeOut
write
if (inta)
else
= 2’b00 ;
= 3’b000;
= 1’b0 ;
= 1’b0 ;
= 1’b0 ;
= 1’b0 ;
= 1’b0 ;
= 1’b0 ;
= 1’b0 ;
begin resultSel
= 3’b110;
writeBack
= 1’b1 ;
end
case(opCode)
add
: begin arithLogOp =
resultSel
=
writeBack
=
carryEnable =
end
sub
: begin arithLogOp =
resultSel
=
writeBack
=
carryEnable =
end
addc
: begin arithLogOp =
resultSel
=
writeBack
=
carryEnable =
end
subc
: begin arithLogOp =
resultSel
=
writeBack
=
carryEnable =
end
ashr
: begin resultSel
=
writeBack
=
end
Figure 10.18: The module dataDecode.
2’b00 ;
3’b000;
1’b1 ;
1’b1 ;
2’b01 ;
3’b000;
1’b1 ;
1’b1 ;
2’b10 ;
3’b000;
1’b1 ;
1’b1 ;
2’b11 ;
3’b000;
1’b1 ;
1’b1 ;
3’b010;
1’b1 ;
312
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
neg
: begin arithLogOp = 2’b00 ;
resultSel
= 3’b001;
writeBack
= 1’b1 ;
end
bwand
: begin arithLogOp = 2’b01 ;
resultSel
= 3’b001;
writeBack
= 1’b1 ;
end
bwor
: begin arithLogOp = 2’b10 ;
resultSel
= 3’b001;
writeBack
= 1’b1 ;
end
bwxor
: begin arithLogOp = 2’b11 ;
resultSel
= 3’b001;
writeBack
= 1’b1 ;
end
val
: begin resultSel
= 3’b011;
writeBack
= 1’b1 ;
end
hval
: begin resultSel
= 3’b100;
writeBack
= 1’b1 ;
end
get
: begin resultSel
= 3’b101;
writeBack
= 1’b1 ;
end
send
:
outRegEnable
= 1’b1 ;
receive : if (readyIn)
begin inRegEnable
= 1’b1 ;
readIn
= 1’b1 ;
end
issue
: if (readyOut) writeOut = 1’b1 ;
datawr :
write = 1’b1
;
datard : begin resultSel
= 3’b111;
writeBack
= 1’b1 ;
end
default begin
arithLogOp
= 2’b00 ;
resultSel
= 3’b000;
writeBack
= 1’b0 ;
inRegEnable
= 1’b0 ;
outRegEnable
= 1’b0 ;
carryEnable
= 1’b0 ;
readIn
= 1’b0 ;
writeOut
= 1’b0 ;
write
= 1’b0 ;
end
endcase
end
endmodule
Figure 10.19: The module dataDecode (continuation).
10.3. EMBEDDED COMPUTATION
313
module
mux4_32(output reg [31:0] out
input
[31:0] in0
input
[31:0] in1
input
[31:0] in2
input
[31:0] in3
input
[1:0]
sel
always @(*) case(sel)
2’b00: out = in0;
2’b01: out = in1;
2’b10: out = in2;
2’b11: out = in3;
endcase
endmodule
,
,
,
,
,
);
module
mux4_1( output reg
out
input
in0
input
in1
input
in2
input
in3
input
[1:0]
sel
always @(*) case(sel)
2’b00: out = in0;
2’b01: out = in1;
2’b10: out = in2;
2’b11: out = in3;
endcase
endmodule
,
,
,
,
,
);
module
,
,
,
,
,
,
,
,
,
);
mux8_32(output reg [31:0] out
input
[31:0] in0
input
[31:0] in1
input
[31:0] in2
input
[31:0] in3
input
[31:0] in4
input
[31:0] in5
input
[31:0] in6
input
[31:0] in7
input
[2:0]
sel
always @(*)
case(sel)
3’b000: out = in0;
3’b001: out = in1;
3’b010: out = in2;
3’b011: out = in3;
3’b100: out = in4;
3’b101: out = in5;
3’b110: out = in6;
3’b111: out = in7;
endcase
endmodule
Figure 10.20: The multiplexor modules multiplexors.
314
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
reset
00 [00]
dis
0
1
int∗
0
0
1
1
ei
ei
11 [10]
10 [11]
inta
01 [01]
en
mem
asyncInta
0
1
ei
0
1
int∗
0
0
1
di
1
asyncInta
di
Figure 10.21: Interrupt automaton for a limited width and an asynchronous int signal.
Therefore, the transitions triggered by the asynchronous input int influence always only one
state bit.
For an implementation with registers results the following equations for the state transition
and output functions:
′
′ ′
Q+
1 = Q1 Q0 di + Q1 Q0 ei
′
′
′ ′
Q+
0 = Q1 Q0 (di ⊕ int) + Q1 Q0 ei + Q1 Q0 (ei ⊕ int)
asyncInta = Q1 Q′0 + Q′1 Q0 ei
We are not very happy about the resulting circuits because the size is too big to my taste.
Deserve to try another equivalent state coding, preserving the condition that the transitions
depending on the int input are reduced dependency type. The second coding proposal is (see
the un-bracketed codes in Figure 10.21):
dis : Q1 Q0 = 00
en : Q1 Q0 = 10
mem : Q1 Q0 = 01
inta : Q1 Q0 = 11
The new state transition functions are:
′ ′
′ ′
Q+
1 = Q1 Q0 di + Q1 Q0 ei
′
′
′
Q+
0 = Q0 int + Q1 Q0 ei
The Verilog behavioral description for this version is presented in Figure 10.22.
10.3. EMBEDDED COMPUTATION
module interruptAutomaton(input
input
input
output
input
input
315
int
,
ei
,
di
,
regasyncInt,
reset
,
clock
);
reg [1:0] state
;
reg [1:0] nextState;
always @(posedge clock) if (reset) state <= 0
;
else
state <= nextState;
always @(int or ei or di or state)
case(state)
2’b00: if (int) if (ei) {nextState,
else
{nextState,
else if (ei)
{nextState,
else
{nextState,
2’b01: if (ei)
{nextState,
else
{nextState,
2’b10: if (int) if (di) {nextState,
else
{nextState,
else if (di)
{nextState,
else
{nextState,
2’b11:
{nextState,
endcase
endmodule
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
asyncInt}
=
=
=
=
=
=
=
=
=
=
=
3’b11_0;
3’b01_0;
3’b10_0;
3’b00_0;
3’b00_1;
3’b01_0;
3’b01_0;
3’b11_0;
3’b00_0;
3’b10_0;
3’b00_1;
Figure 10.22: The module interruptAutomaton.
316
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
If we make another step re-designing the loop for an “intelligent” JK register, then results
for the loop the following expressions:
J1 = Q′0 ei
K1 = di + Q0
J0 = int
K0 = Q1 + ei
and for the output transition:
asyncInta = Q0 (Q1 + ei) = Q0 K0
A total of 4 2-input gates for the complex part of the automaton. The final count: 2 JK-FFs, 2
ANDs, 2 ORs. Not bad! The structural description for this version is presented in Figure 10.23
and in Figure 10.24
module interruptAutomaton(input int
,
input ei
, di,
output asyncInta,
input reset
,
input clock
);
wire q1, q0, notq1, notq0;
JKflipFlop ff1(.Q
(q1
),
.notQ (notq1
),
.J
(notq0 & ei),
.K
(di | q0
),
.reset(reset
),
.clock(clock
));
JKflipFlop ff0(.Q(q0),
.notQ (notq0 ),
.J
(int
),
.K
(q1 | ei),
.reset(reset ),
.clock(clock ));
assign asyncInta = q0 & (q1 | ei);
endmodule
Figure 10.23: The structural description of the module interruptAutomaton implemented using JK-FFs..
module JKflipFlop(output reg Q
,
output
notQ ,
input
J
, K,
input
reset,
input
clock);
assign notQ = ~Q;
always @(posedge clock) if (reset) Q <= 0
;
else
Q <= J & notQ | ~K & Q;
endmodule
Figure 10.24: The module JKflipFlop.
The synthesis process will provide a very small circuit with the complex part implemented
using only 4 gates. The module interruptUnit in the toyMachine design must be redesigned
10.4. PROBLEMS
317
including the just presented module interruptAutomaton. The size of the overall project will
increase, but the interrupt mechanism will work with less electrical restrictions imposed to the
external connections.
10.4
Problems
Problem 10.1 Interpretative processor with distinct program counter block.
10.5
Projects
Project 10.1
318
CHAPTER 10. COMPUTING MACHINES:
≥4–LOOP DIGITAL SYSTEMS
Chapter 11
# ∗ SELF-ORGANIZING
STRUCTURES:
N-th order digital systems
In the previous chapter
the concept of computer, as at least four-loop system, was introduced. The basic part of
the section is contained in the Problems section. Adding few loops the functionality of the
system remains the same - basic computation - the only effect is optimizing area, power,
speed.
In this chapter
the self-organizing systems are supported by a cellular approach consisting in n-loop
systems. The main structure discussed are:
• the stack memory, as the simplest n-loop system
• the cellular automata, as the simplest self-organizing system
• fractal structures, as “non-uniform” network of processors
In the next chapter
we make only the first step in closing a new kind of loops over an n-order system, thus
introducing the new category of systems with global loops.
319
320CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
Von Neumann’s architecture is supported by a structure which consists of two distinct subsystems. The first is a processor and the second is a memory that stores the bits to be processed.
In order to be processed a bit must be carried from the memory to the processor and many times
back, from the processor to the memory. Too much time is wasted and many structures are
involved only to move bits inside the computing machine. The functional development in the
physical part of a digital systems stopped when this universal model was adopted. In the same
time the performances of the computation process are theoretically limited. All the sort of
parallelism pay tribute to this style that is sequentially founded. We have only one machine,
each bit must be accessed, processed and after that restored. This “ritual” stopped the growing
process of digital machines around the fifth order. There are a small number of useful systems
having more than five loops.
The number of loops can become very large if we give up this model and we have the nerve
to store “each bit” near its own “processor”. A strange, but maybe a winning solution is to
“interleave” the processing elements with the storage circuits [Moto-Oka ’82]. Many of us believe
that this is a more “natural” solution. Until now this way is only a beautiful promise, but this
way deserves more attention.
Definition 11.1 A digital system, DS, belongs to n-OS if having the size in O(f (n)) contains
a number of internal loop in O(n). ⋄
A paradigmatic example of n-loop digital system is the cellular automaton (CA). Many
applications of CA model self-organizing systems.
For the beginning, as a simple introduction, the stack memory is presented in a version
belonging to n-OS. The next subject will be a new type of memory which tightly interleaves
the storage elements with processing circuits: the Connex memory. This kind of memory
allows fine grain deep parallel processes in computation. We end with the eco-chip, a spectacular application of the two-dimensional cellular automata enhanced with the Connex memory’s
functions.
The systems belonging to n-OS support efficiently different mechanisms related to some
parallel computation models. Thus, there are many chances to ground true parallel computing
architecture using such kind of circuits.
11.1
Push-Down Stack as n-OS
There are only a few “exotic” structures that are implemented as digital systems with a great
number of loops. One of these is the stack function that needs at least two loops to be realized, as
a system in 2-OS (reversible counter & RAM serially composed). There is another, more uniform
solution for implementing the push-down stack function or LIFO (last-in first-out) memory. This
solution uses a simple, i.e., recursive defined, structure.
Definition 11.2 The n-level push-down stack, LIF On , is built serial connecting a LIF On−1
with a LIF O1 as in Figure 11.1. The one level push-down stack is a register, Rm , loop connected
with m × M U X2 , so as:
S1 S0 = 00 means: no op - the content of the register does not change
S1 S0 = 01 means: push - the register is loaded with the input value, IN
S1 S0 = 10 means: pop - the register is loaded with the extension input value, EXTIN. ⋄
It is evident that LIF On is a bi-directional serial-parallel shift register. Because the content
of the serial-parallel register shifts in both directions each Rm is contained in two kind of loops:
11.2. CELLULAR AUTOMATA
321
LIF O1
✲
IN
IN
EXTOUT
LIFO (n-1)
✛
OUT
OUT
EXTIN
✛
✲0
✲ 1 MUX ✲ R m
✲2
S1
✻
✲
S0
✻
2
✲
{P U SH, P OP, −}
Figure 11.1: The LIFO structure as n − OS
• through its own MUX for no op function
• through two successive LIF O1
Thus, LIF O1 is a 2-OS, LIF O2 is a 3-OS, LIF O3 is a 4-OS, . . ., LIF Oi is a (i + 1)OS, . . ..
The push-down stack implemented as a bi-directional serial-parallel register is an example
of digital system having the order related with the size. Indeed: LIF On−1 is a n − OS.
11.2
Cellular automata
A cellular automaton consists of a regular grid of cells. Each cell has a finite number of states.
The grid has a finite number of dimensions, usually no more than three. The transition function
of each cell is defined in a constant neighborhood. Usually, the next state of the cell depends
on its own state and the states of the adjacent cells.
11.2.1
General definitions
The linear cellular automaton
Definition 11.3 The one-dimension cellular automaton is linear array of n identical cells,
where each cell is connected in a constant neighborhood of +/- m cells, see Figure 11.2a for
m = 1. Each cell is a s-state finite automaton.
⋄
Definition 11.4 An elementary cellular automaton is a one-dimension cellular automaton with
m = 1 and s = 2. The transition function of each automaton is a three-input Boolean function
defined by the decimally expressed associated Boolean vector.
⋄
Example 11.1 The Boolean vector of the three-input function
f (x2 , x1 , x0 ) = x2 ⊕ (x1 + x0 )
is:
00011110
and defines the transition rule 30.
⋄
322CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
✲
✛✲
✛✲
✛✲
✛✲
✛✲
✛
a.
✛✲
✛✲
✻
❄
✻
❄
✛✲
✛✲
✻
❄
✻
❄
✻
❄
■
✛✲
❘
✛✲
✻■ ✒
❄✠ ❘
✛✲
✛✲
■✒
✠❘
✛✲
✻■ ✒
❄✠ ❘
✛✲
✻■ ✒
❄✠ ❘
✛✲
✒
✻■ ✒
❄✠ ❘
✻■
❄ ❘
✻
❄
✠
c.
b.
✲
✲
❄
✻
❄
✛✲
✛✲
✻
✻ ✒
❄✠
✛✲
❄
✻
❄
✻
✛
✛
✲
✲
❄
✻
❄
✛✲
✛✲
✻
❄
✻
❄
✛
✛
✻
e.
d.
Figure 11.2: Cellular automaton. a. One-dimension cellular automaton. b. Two-dimension
cellular automaton with von Neumann neighborhood. c. Two-dimension cellular automaton with Moore
neighborhood. d. Two-dimension cellular automaton with toroidal shape. e. Two-dimension cellular
automaton with rotated toroidal shape.
Definition 11.5 The Verilog definition of the elementary cellular automaton is:
module eCellAut #(parameter n = 127)
// n-cell elementary cellular automaton
(
output [n-1:0] out ,
input
[7:0]
func,
// the Boolean vector for the transition rule
input
[n-1:0] init,
// used to initialize the cellular automaton
input
rst ,
// loads the initial state of the cellular automaton
input
clk );
genvar i;
generate for (i=0; i<n;
eCell eCell(.out
.func
.init
.in0
.in1
.rst
.clk
i=i+1) begin: C
(out[i]
(func
(init[i]
((i==0) ? out[n-1] : out[i-1]
((i==n-1) ? out[0] : out[i+1]
(rst
(clk
),
),
),
),
),
),
));
11.2. CELLULAR AUTOMATA
323
end
endgenerate
endmodule
where the elementary cell, eCell, is:
module eCell
(
output
input
input
input
input
input
input
// elementary cell
reg
[7:0]
always @(posedge clk)
out ,
func,
init,
in0 ,
in1 ,
rst ,
clk );
// input form the previous cell
// input from the next cell
if (rst)
else
out <= init
;
out <= func[{in1, out, in0}];
endmodule
⋄
Example 11.2 The elementary cellular automaton characterized by the rule 90 (01011010)
provides, starting from the initial state 1’b1 << n/2, the behavior represented in Figure 11.3,
where the sequence of lines of bits represent the sequence of the states of the cellular automaton
starting from the initial state.
The shape generated by the elementary cellular automaton 90 is the Sierpinski triangle or
the Sierpinski Sieve. It is a fractal named after the Polish mathematician Waclaw Sierpinski
who first described it in 1915.
⋄
Example 11.3 The elementary cellular automaton characterized by the rule 30 (00011110)
provides, starting from the initial state 1’b1 << n/2, the behavior represented in Figure 11.4,
where the sequence of lines of bits represent the sequence of the states of the cellular automaton
starting from the initial state.
⋄
324CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
0000000000000000000000000000000000000000000000000000000000000001000000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000000010100000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000000100010000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000001010101000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000010000000100000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000101000001010000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000001000100010001000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000010101010101010100000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000100000000000000010000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000001010000000000000101000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000010001000000000001000100000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000101010100000000010101010000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000001000000010000000100000001000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000010100000101000001010000010100000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000100010001000100010001000100010000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000001010101010101010101010101010101000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000010000000000000000000000000000000100000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000101000000000000000000000000000001010000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000001000100000000000000000000000000010001000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000010101010000000000000000000000000101010100000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000100000001000000000000000000000001000000010000000000000000000000000000000000000000000
0000000000000000000000000000000000000000001010000010100000000000000000000010100000101000000000000000000000000000000000000000000
0000000000000000000000000000000000000000010001000100010000000000000000000100010001000100000000000000000000000000000000000000000
0000000000000000000000000000000000000000101010101010101000000000000000001010101010101010000000000000000000000000000000000000000
0000000000000000000000000000000000000001000000000000000100000000000000010000000000000001000000000000000000000000000000000000000
0000000000000000000000000000000000000010100000000000001010000000000000101000000000000010100000000000000000000000000000000000000
0000000000000000000000000000000000000100010000000000010001000000000001000100000000000100010000000000000000000000000000000000000
0000000000000000000000000000000000001010101000000000101010100000000010101010000000001010101000000000000000000000000000000000000
0000000000000000000000000000000000010000000100000001000000010000000100000001000000010000000100000000000000000000000000000000000
0000000000000000000000000000000000101000001010000010100000101000001010000010100000101000001010000000000000000000000000000000000
0000000000000000000000000000000001000100010001000100010001000100010001000100010001000100010001000000000000000000000000000000000
0000000000000000000000000000000010101010101010101010101010101010101010101010101010101010101010100000000000000000000000000000000
0000000000000000000000000000000100000000000000000000000000000000000000000000000000000000000000010000000000000000000000000000000
0000000000000000000000000000001010000000000000000000000000000000000000000000000000000000000000101000000000000000000000000000000
0000000000000000000000000000010001000000000000000000000000000000000000000000000000000000000001000100000000000000000000000000000
0000000000000000000000000000101010100000000000000000000000000000000000000000000000000000000010101010000000000000000000000000000
0000000000000000000000000001000000010000000000000000000000000000000000000000000000000000000100000001000000000000000000000000000
0000000000000000000000000010100000101000000000000000000000000000000000000000000000000000001010000010100000000000000000000000000
0000000000000000000000000100010001000100000000000000000000000000000000000000000000000000010001000100010000000000000000000000000
0000000000000000000000001010101010101010000000000000000000000000000000000000000000000000101010101010101000000000000000000000000
0000000000000000000000010000000000000001000000000000000000000000000000000000000000000001000000000000000100000000000000000000000
0000000000000000000000101000000000000010100000000000000000000000000000000000000000000010100000000000001010000000000000000000000
0000000000000000000001000100000000000100010000000000000000000000000000000000000000000100010000000000010001000000000000000000000
0000000000000000000010101010000000001010101000000000000000000000000000000000000000001010101000000000101010100000000000000000000
0000000000000000000100000001000000010000000100000000000000000000000000000000000000010000000100000001000000010000000000000000000
0000000000000000001010000010100000101000001010000000000000000000000000000000000000101000001010000010100000101000000000000000000
0000000000000000010001000100010001000100010001000000000000000000000000000000000001000100010001000100010001000100000000000000000
0000000000000000101010101010101010101010101010100000000000000000000000000000000010101010101010101010101010101010000000000000000
0000000000000001000000000000000000000000000000010000000000000000000000000000000100000000000000000000000000000001000000000000000
0000000000000010100000000000000000000000000000101000000000000000000000000000001010000000000000000000000000000010100000000000000
0000000000000100010000000000000000000000000001000100000000000000000000000000010001000000000000000000000000000100010000000000000
0000000000001010101000000000000000000000000010101010000000000000000000000000101010100000000000000000000000001010101000000000000
0000000000010000000100000000000000000000000100000001000000000000000000000001000000010000000000000000000000010000000100000000000
0000000000101000001010000000000000000000001010000010100000000000000000000010100000101000000000000000000000101000001010000000000
0000000001000100010001000000000000000000010001000100010000000000000000000100010001000100000000000000000001000100010001000000000
0000000010101010101010100000000000000000101010101010101000000000000000001010101010101010000000000000000010101010101010100000000
0000000100000000000000010000000000000001000000000000000100000000000000010000000000000001000000000000000100000000000000010000000
0000001010000000000000101000000000000010100000000000001010000000000000101000000000000010100000000000001010000000000000101000000
0000010001000000000001000100000000000100010000000000010001000000000001000100000000000100010000000000010001000000000001000100000
0000101010100000000010101010000000001010101000000000101010100000000010101010000000001010101000000000101010100000000010101010000
0001000000010000000100000001000000010000000100000001000000010000000100000001000000010000000100000001000000010000000100000001000
0010100000101000001010000010100000101000001010000010100000101000001010000010100000101000001010000010100000101000001010000010100
0100010001000100010001000100010001000100010001000100010001000100010001000100010001000100010001000100010001000100010001000100010
1010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101
0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
Figure 11.3:
11.2. CELLULAR AUTOMATA
325
0000000000000000000000000000000000000000000000000000000000000001000000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000000011100000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000000110010000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000001101111000000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000011001000100000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000000110111101110000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000001100100001001000000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000011011110011111100000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000000110010001110000010000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000001101111011001000111000000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000011001000010111101100100000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000000110111100110100001011110000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000001100100011100110011010001000000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000011011110110011101110011011100000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000000110010000101110001001110010010000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000001101111001101001011111001111111000000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000011001000111001111010000111000000100000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000000110111101100111000011001100100001110000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000001100100001011100100110111011110011001000000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000011011110011010011111100100010001110111100000000000000000000000000000000000000000000
0000000000000000000000000000000000000000000110010001110011110000011110111011000100010000000000000000000000000000000000000000000
0000000000000000000000000000000000000000001101111011001110001000110000100010101110111000000000000000000000000000000000000000000
0000000000000000000000000000000000000000011001000010111001011101101001110110101000100100000000000000000000000000000000000000000
0000000000000000000000000000000000000000110111100110100111010001001111000100101101111110000000000000000000000000000000000000000
0000000000000000000000000000000000000001100100011100111100011011111000101111101001000001000000000000000000000000000000000000000
0000000000000000000000000000000000000011011110110011100010110010000101101000001111100011100000000000000000000000000000000000000
0000000000000000000000000000000000000110010000101110010110101111001101001100011000010110010000000000000000000000000000000000000
0000000000000000000000000000000000001101111001101001110100101000111001111010110100110101111000000000000000000000000000000000000
0000000000000000000000000000000000011001000111001111000111101101100111000010100111100101000100000000000000000000000000000000000
0000000000000000000000000000000000110111101100111000101100001001011100100110111100011101101110000000000000000000000000000000000
0000000000000000000000000000000001100100001011100101101010011111010011111100100010110001001001000000000000000000000000000000000
0000000000000000000000000000000011011110011010011101001011110000011110000011110110101011111111100000000000000000000000000000000
0000000000000000000000000000000110010001110011110001111010001000110001000110000100101010000000010000000000000000000000000000000
0000000000000000000000000000001101111011001110001011000011011101101011101101001111101011000000111000000000000000000000000000000
0000000000000000000000000000011001000010111001011010100110010001001010001001111000001010100001100100000000000000000000000000000
0000000000000000000000000000110111100110100111010010111101111011111011011111000100011010110011011110000000000000000000000000000
0000000000000000000000000001100100011100111100011110100001000010000010010000101110110010101110010001000000000000000000000000000
0000000000000000000000000011011110110011100010110000110011100111000111111001101000101110101001111011100000000000000000000000000
0000000000000000000000000110010000101110010110101001101110011100101100000111001101101000101111000010010000000000000000000000000
0000000000000000000000001101111001101001110100101111001001110011101010001100111001001101101000100111111000000000000000000000000
0000000000000000000000011001000111001111000111101000111111001110001011011011100111111001001101111100000100000000000000000000000
0000000000000000000000110111101100111000101100001101100000111001011010010010011100000111111001000010001110000000000000000000000
0000000000000000000001100100001011100101101010011001010001100111010011111111110010001100000111100111011001000000000000000000000
0000000000000000000011011110011010011101001011110111011011011100011110000000001111011010001100011100010111100000000000000000000
0000000000000000000110010001110011110001111010000100010010010010110001000000011000010011011010110010110100010000000000000000000
0000000000000000001101111011001110001011000011001110111111111110101011100000110100111110010010101110100110111000000000000000000
0000000000000000011001000010111001011010100110111000100000000000101010010001100111100001111110101000111100100100000000000000000
0000000000000000110111100110100111010010111100100101110000000001101011111011011100010011000000101101100011111110000000000000000
0000000000000001100100011100111100011110100011111101001000000011001010000010010010111110100001101001010110000001000000000000000
0000000000000011011110110011100010110000110110000001111100000110111011000111111110100000110011001111010101000011100000000000000
0000000000000110010000101110010110101001100101000011000010001100100010101100000000110001101110111000010101100110010000000000000
0000000000001101111001101001110100101111011101100110100111011011110110101010000001101011001000100100110101011101111000000000000
0000000000011001000111001111000111101000010001011100111100010010000100101011000011001010111101111111100101010001000100000000000
0000000000110111101100111000101100001100111011010011100010111111001111101010100110111010100001000000011101011011101110000000000
0000000001100100001011100101101010011011100010011110010110100000111000001010111100100010110011100000110001010010001001000000000
0000000011011110011010011101001011110010010111110001110100110001100100011010100011110110101110010001101011011111011111100000000
0000000110010001110011110001111010001111110100001011000111101011011110110010110110000100101001111011001010010000010000010000000
0000001101111011001110001011000011011000000110011010101100001010010000101110100101001111101111000010111011111000111000111000000
0000011001000010111001011010100110010100001101110010101010011011111001101000111101111000001000100110100010000101100101100100000
0000110111100110100111010010111101110110011001001110101011110010000111001101100001000100011101111100110111001101011101011110000
0001100100011100111100011110100001000101110111111000101010001111001100111001010011101110110001000011100100111001010001010001000
0011011110110011100010110000110011101101000100000101101011011000111011100111011110001000101011100110011111100111011011011011100
0110010000101110010110101001101110001001101110001101001010010101100010011100010001011101101010011101110000011100010010010010010
1101111001101001110100101111001001011111001001011001111011110101010111110010111011010001001011110001001000110010111111111111111
0001000111001111000111101000111111010000111111010111000010000101010100001110100010011011111010001011111101101110100000000000000
0011101100111000101100001101100000011001100000010100100111001101010110011000110111110010000011011010000001001000110000000000000
0110001011100101101010011001010000110111010000110111111100111001010101110101100100001111000110010011000011111101101000000000000
1101011010011101001011110111011001100100011001100100000011100111010101000101011110011000101101111110100110000001001100000000000
1001010011110001111010000100010111011110110111011110000110011100010101101101010001110101101001000000111101000011111010000000001
0111011110001011000011001110110100010000100100010001001101110010110101001001011011000101001111100001100001100110000011000000011
0100010001011010100110111000100110111001111110111011111001001110100101111111010010101101111000010011010011011101000110100000110
Figure 11.4:
326CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
The two-dimension cellular automaton
Definition 11.6 The two-dimension cellular automaton consists of a two-dimension array of
identical cells, where each cell is connected in a constant neighborhood, see Figure 11.2b (the
von Neumann neighborhood) and 11.2c (the Moore neighborhood). Each cell is a s-state finite
automaton.
⋄
There are also many ways of connecting the border cells. The simplest one is to connect them
to ground. Another is close the array so as the surface takes a toroidal shape (see Figure 11.2d).
A more complex form is possible if we intend to preserve also a linear connection between the
cells. Results a twisted toroidal shape (see Figure 11.2e).
Definition 11.7 The Verilog definition of the two-dimension elementary cellular automaton
with a toroidal shape (Figure 11.2d) is:
module eCellAut4 #(parameter n = 8)
(
output [n*n-1:0]
out ,
input
[31:0]
func,
input
[n*n-1:0]
init,
input
rst ,
input
clk );
// n*n-cell two-dimension cellular automaton
// the Boolean vector for the transition rule
// used to initalize the cellular automaton
// loads the inital state of the cellular automaton
genvar i;
generate for (i=0; i<n*n; i=i+1) begin: C
eCell4 eCell4( .out
(out[i]
.func
(func
.init
(init[i]
.in0
(out[(i/n)*n+(i-((i/n)*n)+n-1)%n]
.in1
(out[(i/n)*n+(i-((i/n)*n)+1)%n]
.in2
(out[(i+n*n-n)%(n*n)]
.in3
(out[(i+n)%(n*n)]
.rst
(rst
.clk
(clk
end
endgenerate
endmodule
),
),
),
),
),
),
),
),
));
//
//
//
//
east
west
south
north
where the elementary cell, eCell4, is:
module eCell4
(
output
input
input
input
input
input
input
input
input
// 4-input elementary cell
reg
[31:0]
always @(posedge clk)
endmodule
⋄
out ,
func,
init,
in0 ,
in1 ,
in2 ,
in3 ,
rst ,
clk );
//
//
//
//
//
if (rst)
else
north connection
east connectioin
south connection
west connectioin
out <= init
out <= func[{in3, in2, out, in1, in0}]
;
;
11.2. CELLULAR AUTOMATA
327
Example 11.4 Let be a 8 × 8 cellular automaton with a von Neumann neighborhood and a
toroidal shape. The cells are 2-state automata. The transition function is a 5-input Boolean
OR, and the initial state is state 1 in the bottom right cell and 0 the the rest of cells. The system
will evolve until all the cells will switch in the state 1. Figure 11.5 represents the 8-step evolution
from the initial state to the final state.
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000001
00000001
00000000
00000000
00000000
00000000
00000000
00000001
10000011
10000011
00000001
00000000
00000000
00000000
00000001
10000011
11000111
11000111
10000011
00000001
00000000
00000001
10000011
11000111
11101111
11101111
11000111
10000011
00000001
10000011
11000111
11101111
11111111
11111111
11101111
11000111
10000011
11000111
11101111
11111111
11111111
11111111
11111111
11101111
11000111
11101111
11111111
11111111
11111111
11111111
11111111
11111111
11101111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
initial
step 1
step 2
step 3
step 4
step 5
step6
step 7
final
Figure 11.5:
⋄
Definition 11.8 The Verilog definition of the two-dimension elementary cellular automaton
with linearly connected cells (Figure 11.2e) is:
module eCellAut4L #(parameter n = 8)
// two-dimension cellular automaton
(
output [n*n-1:0]
out ,
input
[31:0]
func,
// the Boolean vector for the transition rule
input
[n*n-1:0]
init,
// used to initalize the cellular automaton
input
rst ,
// loads the inital state of the cellular automaton
input
clk );
genvar i;
generate for (i=0; i<n*n; i=i+1) begin: C
eCell4 eCell4( .out
(out[i]
.func
(func
.init
(init[i]
.in0
(out[(i+n*n-1)%(n*n)]
.in1
(out[(i+1)%(n*n)]
.in2
(out[(i+n*n-n)%(n*n)]
.in3
(out[(i+n)%(n*n)]
.rst
(rst
.clk
(clk
end
endgenerate
endmodule
),
),
),
),
),
),
),
),
));
//
//
//
//
east
west
south
north
where the elementary cell, eCell4, is the same as in the previous definition.
⋄
Example 11.5 Let us do the same for the two-dimension elementary cellular automaton with
linearly connected cells (Figure 11.2e). The insertion of 1s in all the cells is done now in 7
steps. See Figure 11.6.
Looks like a twisted toroidal shape offers a better neighborhood than a simple toroidal shape.
⋄
11.2.2
Applications
328CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
00000000
00000000
00000000
00000000
00000000
00000000
00000000
00000001
10000001
00000000
00000000
00000000
00000000
00000000
00000001
00000011
11000011
10000001
00000000
00000000
00000000
00000001
00000011
10000111
11100111
11000011
10000001
00000000
00000001
00000011
10000111
11001111
11111111
11100111
11000011
10000001
00000011
10000111
11001111
11111111
11111111
11111111
11100111
11000011
10000111
11001111
11111111
11111111
11111111
11111111
11111111
11100111
11001111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
11111111
initial
step 1
step 2
step 3
step 4
stap 5
step 6
final
Figure 11.6:
11.3
Systolic systems
Leiserson’s systolic sorter. The initial state: in each cell = ∞. For no operation: in1 =
+∞, in2 = −∞. To insert the value v: in1 = v, in2 = −∞. For extract: in1 = in2 = +∞.
A
ck
B
C
✲A
✲B
✲C
✲
Latch
Y
a.
Y
Z
ck
✲
✲
✲
min(A,B,C)
out
med(A,B,C)
in1
max(A,B,C)
in2
✻
❄ ❄ ❄
X
X
✛ X
✲B
✲C
Y
Z
ck
✻
ck’
Z
A
✛ X
✲B
✲C
A
Y
ck
Z
✻
✛ X
✲B
✲C
A
Y
Z
ck
✛ X
✲B
✲C
✻
A
Y
ck
Z
✛
✲
✲
✻
ck
c.
b.
Figure 11.7: Systolic sorter. a. The internal structure of cell. b. The logic symbol of cell. c. The
organization of the systolic sorter.
module systolicSorterCell #(parameter n=8)(input
[n-1:0] a, b, c,
output reg [n-1:0] x, y, z,
input
rst, ck);
wire
[n-1:0] a1, b1 ; // sorter’s first level outputs
wire
[n-1:0] a2, c2 ; // sorter’s second level outputs
wire
[n-1:0] b3, c3 ; // sorter’s third level outputs
assign
assign
assign
assign
assign
assign
a1
b1
a2
c2
b3
c3
=
=
=
=
=
=
(a < b) ? a : b ;
(a < b) ? b : a ;
(a1 < c) ? a1 : c
(a1 < c) ? c : a1
(b1 < c2) ? b1 : c2
(b1 < c2) ? c2 : b1
;
;
;
;
always @(ck or rst or a2 or b3 or c3)
if (rst & ck)
begin
x = {n{1’b1}}
y = {n{1’b1}}
z = {n{1’b1}}
end
else if (ck)
begin
x = a2 ;
y = b3 ;
z = c3 ;
;
;
;
11.3. SYSTOLIC SYSTEMS
329
end
endmodule
module systolicSorter #(parameter n=8, m=7)(
wire
wire
wire
assign
assign
assign
assign
output
input
input
[n-1:0] out,
[n-1:0] in1, in2,
rst, ck1, ck2);
[n-1:0] x[0:m];
[n-1:0] y[0:m-1];
[n-1:0] z[0:m-1];
y[0]
z[0]
out
x[m]
=
=
=
=
in1
in2
x[1]
{n{1’b1}}
;
;
;
;
genvar i;
generate for(i=1; i<m; i=i+1)
begin: C
systolicSorterCell systolicCell(
.a (x[i+1]),
.b (y[i-1]),
.c (z[i-1]),
.x (x[i]),
.y (y[i]),
.z (z[i]),
.rst(rst),
.ck (((i/2)*2 == i) ? ck2 : ck1));
end
endgenerate
endmodule
module systolicSorterSim #(parameter n=8);
reg
ck1, ck2, rst
;
reg
[n-1:0] in1, in2;
wire
[n-1:0] out ;
initial begin
ck1 = 0 ;
forever begin
#3 ck1 = 1
#1 ck1 = 0
;
;
#3 ck2 = 1
#1 ck2 = 0
;
;
end
end
initial begin
ck2 = 0 ;
#2 ck2 = 0 ;
forever begin
end
end
initial begin
rst = 1 ;
in2 = 0 ;
in1 = 8’b1000;
#8 rst = 0 ;
#4 in1 = 8’b0010;
#4 in1 = 8’b0100;
#4 in1 = 8’b0010;
#4 in1 = 8’b0001;
#4 in1 = 8’b11111111;
in2 = 8’b11111111;
#30 $stop;
330CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
end
systolicSorter dut( out,
in1, in2,
rst, ck1, ck2);
initial
$monitor("time = %d ck1 = %b ck2 = %b rst = %b in1 = %d in2 = %d out = %d ",
$time, ck1, ck2, rst, in1, in2, out);
endmodule
The result of simulation is:
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
time
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
0
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
0
0
0
1
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
8
8
8
8
8
8
8
8
8
8
2
2
2
2
4
4
4
4
2
2
2
2
1
1
1
1
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
255
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
out
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
x
255
255
255
255
255
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
2
2
2
2
2
2
2
2
4
4
4
11.4. INTERCONNECTION ISSUES
#
#
#
#
#
#
#
#
#
#
#
#
time
time
time
time
time
time
time
time
time
time
time
time
11.4
=
=
=
=
=
=
=
=
=
=
=
=
46
47
48
49
50
51
52
53
54
55
56
57
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
ck1
=
=
=
=
=
=
=
=
=
=
=
=
0
1
0
0
0
1
0
0
0
1
0
0
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
ck2
=
=
=
=
=
=
=
=
=
=
=
=
0
0
0
1
0
0
0
1
0
0
0
1
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
rst
=
=
=
=
=
=
=
=
=
=
=
=
0
0
0
0
0
0
0
0
0
0
0
0
331
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
in1
=
=
=
=
=
=
=
=
=
=
=
=
255
255
255
255
255
255
255
255
255
255
255
255
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
in2
=
=
=
=
=
=
=
=
=
=
=
=
255
255
255
255
255
255
255
255
255
255
255
255
out
out
out
out
out
out
out
out
out
out
out
out
=
=
=
=
=
=
=
=
=
=
=
=
4
8
8
8
8
255
255
255
255
255
255
255
Interconnection issues
Simple circuits scale up easy generating big interconnection problems.
11.4.1
Local vs. global connections
The origin of the memory wall is in the inability to avoid global connection on memory arrays,
while in the logic areas the local connections are easiest to impose.
Memory wall
11.4.2
Localizing with cellular automata
11.4.3
Many clock domains & asynchronous connections
The clock signal uses a lot of energy and area and slows down the design when the area of the
circuit became too big.
A fully synchronous design generate also power distribution issues, which come with all the
associated problems.
11.5
Neural networks
Artificial neural network (NN) is a technical construct inspired from the biological neural
networks. NN are composed of interconnected artificial neurons. An artificial neuron is a
programmed or circuit construct that mimic the property of a biological neuron. A multi-layer
NN is used as a connectionist computational model. The introductory text [Zurada ’95] is used
for a short presentation of the concept of NN.
11.5.1
The neuron
The artificial neuron (see Figure 11.8) receives the inputs x1 , . . . , xn (corresponding to n dendrites) and process them to produce an output o (synapse). The sums of each node are weighted,
using the weight vector w1 , . . . , wn and the sum, net, is passed through a non-linear function, f (net), called activation function or transfer function. The transfer functions usually have
a sigmoid shape (see Figure 11.9) or step functions.
Formally, the transfer function of a neuron:
o = f(
n
∑
i=1
wi xi ) = f (net)
332CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
w1
x1
✲
❄
×
w1 w2
w2
x2
✲
❄
wn
❄❄
×
✇
❥ ∑
✲ f (net)
net
✲
NEURON
o
✼
❄
✲
✲
x1
x2
✲
o
✲
xn
wn
xn
✲
❄
×
a.
b.
Figure 11.8: The general form of a neuron. a. The circuit structure of a n-input neuron. b.
The logic symbol.
where f , the typical activation function, is:
f (y)
✻
1
❪
♦
-3
-2
-1
1
λ=7
♦
2
✲
3
y
λ = 2 λ = 0.5
-1
Figure 11.9: The activation function.
2
−1
1 + exp(−λy)
The parameter λ determines the steepness of the continuous function f . For big value of λ the
function f becomes:
f (y) = sgn(y)
f (y) =
The neuron works as a combinational circuit performing the scalar product of the input
vector
x = [x1 x2 . . . xn ]
with the weight vector
w = [w1 w2 . . . wn ]
followed by the application of the activation function. The activation function f is simply
implemented using as a look-up table using a Read-Only Memory.
11.5. NEURAL NETWORKS
11.5.2
333
The feedforward neural network
A feedforward NN is a collection of m n-input neurons (see Figure 11.10). Each neuron receives
the same input vector
x = [x1 x2 . . . xm ]
and is characterized by its own weight vector
wi = [w1 w2 . . . wm ]
The entire NN provides the output vector
o = [o1 o2 . . . om ]t
The activation function is the same for each neuron.
x1
x2
w21 w22
xn
✲
✲
w2n
❄❄
❄
w11 w12
✲
✲
✲
N1
✲
o2
w1n
❄❄
✲
N2
✲
w1
❄
w2
❄
✲
✲
✲
✲
❄
❄
o1
N Nmn (f )
x
wm1wm2
wm
✲
o
wmn
❄❄
❄
✲
Nm
om
a.
b.
Figure 11.10: The single-layer feedforward neural network. a. The organization of a
feedforward NN having m n-input neurons. b. The logic symbol.
Each NN is characterized by the weight matrix
W=
w11
w21
..
.
w12
w22
..
.
wm1
wm2
...
...
...
...
w1n
w2n
..
.
wmn
334CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
having for each output a line, while for each input it has a column. The transition function of
the NN is:
o(t) = Γ[Wx(t)]
where:
f (·)
0
Γ[·] = ..
.
0
f (·)
..
.
...
...
0
0
..
.
...
0
0
. . . f (·)
The feedforwaed NN is of “instantaneous” type, i.e., it behaves as a combinational circuit which
provides the result in the same “cycle”. The propagation time associated do not involve storage
elements.
Example 11.6 The shaded area in Figure 11.11 must be recognized by a two-layer feedforward
NN. Four conditions must be met to define the surface:
x1 − 1 > 0 → sgn(x1 − 1) = 1
x1 − 2 < 0 → sgn(−x1 + 2) = 1
x2 > 0
→ sgn(x2 ) = 1
x2 − 3 < 0 → sgn(−x2 + 3)
For each condition a neuron from the first layer is used. The second layer determines whether
all the conditions tested by the first layer are fulfilled.
The first layer is characterized the weight matrix
W43
1
−1
=
0
0
0
0
1
−1
1
−2
0
−3
The weight vector for the second layer is
W = [1 1 1 1 3.5]
On both layers the activation function is sgn.
⋄
11.5.3
The feedback neural network
The feedback NN is a sequential system. It provides the output with a delay of a number of
clock cycles after the initialization with the input vector x. The structure of a feedback NN is
presented in Figure 11.12. The multiplexor mux is used to initialize the loop closed through
register. If init = 1 the vector x is applied to N Nmn (f ) one clock cycle, then init is switched
to 0 and the loop is closed.
In the circuit approach of this concept, after the initialization cycle the output of the network
is applied to the input through the feedback register. The transition function is:
o(t + Tclock ) = Γ[Wo(t)]
where Tclock (the clock period) is the delay on the loop. After k clock cycles the state of the
network is described by:
o(t + k × Tclock ) = Γ[WΓ[. . . Γ[Wo(t)] . . .]]
A feedback NN can be considered as an initial automaton with few final states mapping
disjoint subsets of inputs.
11.5. NEURAL NETWORKS
x1 x 2
x3 = −1
✲
✲
✲
1
0
335
1
❄ ❄ ❄
o1
N1
-1 0 -2
✲
✲
✲
❄ ❄ ❄
0
✲
✲
✲
o2
N2
1
1
1
1 3.5
❄ ❄ ❄ ❄ ❄
✲
✲
✲ Output Neuton
✲
o
✲
✲
0
❄ ❄ ❄
N3
1
o3
-1
o=1
❄ ❄ ❄
N4
✻
3
2
0 -1 -3
✲
✲
✲
x2
1
o4
✲
0
a.
1
2
3
x1
b.
Figure 11.11: A two-layer feedforward NN. a. The structure. b. The two-dimension space
mapping.
Example 11.7 Let be a feedback NN with 4 4-input neurons with one-bit inputs and outputs.
The activation function is sgn. The feedback NN can be initialized with any 4-bit binary configuration from
x = [-1 -1 -1 -1]
to
x = [1 1 1 1]
and the system has two final states:
o14 = [1 1 1 -1]
o1 = [-1 -1 -1 1]
reached in a number of clock cycles after the initialization.
The resulting discrete-time recurrent network has the following weight matrix:
W44
0
1
=
1
−1
1
0
1
−1
1
1
0
−1
−1
−1
−1
0
The resulting structure of the NN is represented in Figure 11.13, where the weight matrix is
applied on the four 4-bit inputs destined for the weight vectors.
336CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
w1
init
w2
❄
wm
❄
❄
❄
✲ mux
x
✲
✲
NNmn (f )
o
✻
register
✛ clock
✻
Figure 11.12: The single-layer feedback neural network.
[0 1 1 -1]
init
x
❄
✲ mux
❄
✲
[1 0 1 -1]
[1 1 0 -1]
❄
❄
N N44 (sgn)
[-1 -1 -1 0]
❄
✲
o
✻
register
✛ clock
✻
Figure 11.13: The feedback NN with two final states.
The sequence of transitions are computed using the form:
o(t + 1) = [sgn(net1 (t)) sgn(net2 (t)) sgn(net3 (t)) sgn(net4 (t))]
Some sequences end in o14 = [1 1 1 -1], while others in o1 = [-1 -1 -1 1].
⋄
11.5.4
The learning process
The learning process is used to determine the actual form of the matrix W. The learning process
is an iterative one. In each iteration, for each neuron the weight vector w is adjusted with ∆w,
which is proportional with the input vector x and the learning signal r. The general form of the
learning signal is:
r = r(w, x, d)
where d is the desired response (the teacher’s signal). Thus, in each step the weight vector is
adjusted as follows:
w(t + 1) = w(t) + c × r(w(t), x(t), d(t)) × x(t)
where c is the learning constant. The learning process starts from an initial form of the weight
vector (established randomly or by a simple “hand calculation”) and uses as a set of training
input vectors.
11.5. NEURAL NETWORKS
337
There are two types of learning:
unsupervised learning :
r = r(w, x)
the desired behavior is not known; the network will adapt its response by “discovering”
the appropriate values for the weight vectors by self-organization
supervised learning :
r = r(w, x, d)
the desired behavior, d, is known and can be compared with the actual behavior of the
neuron in order to find how to adjust the weight vector.
In the following both types will be exemplified using the Hebbian rule and the perceptron rule.
Unsupervised learning: Hebbian rule
The learning signal is the output of the neuron. In each step the vector w will be adjusted (see
Figure 11.14) as follows:
w(t + 1) = w(t) + c × f (w(t), x(t)) × x(t)
The learning process starts with small random values for wi .
✲×✛
❄
+✛
×✛
✻
c
w(t+1)
✲
clock
❄
wReg
w(t)
✲
x(t)
❄
NEURON
✲
o(t)
Figure 11.14: The Hebian learning rule
Example 11.8 Let be a four-input neuron with the activation function sgn. The initial weight
vector is:
w(t0 ) = [1 − 1 0 0.5]
The training inputs are:
x1 = [1 − 2 1.5 0],
x2 = [1 − 0.5 − 2 − 1.5],
x3 = [0 1 − 1 1.5]
Applying by turn the three training input vectors for c = 1 we obtain:
w(t0 + 1) = w(t0 ) + sgn(net) × x1 = w(t0 ) + sgn(3) × x1 = w(t0 ) + x1 = [2 − 3 1.5 0.5]
w(t0 +2) = w(t0 +1)+sgn(net)×x2 = w(t0 +1)+sgn(−0.25)×x2 = w(t0 +1)−x2 = [1 −2.5 3.5 2]
w(t0 +3) = w(t0 +2)+sgn(net)×x3 = w(t0 +2)+sgn(−3)×x3 = w(t0 +2)−x3 = [1 −3.5 4.5 0.5]
⋄
338CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
Supervised learning: perceptron rule
The perceptron rule performs a supervised learning. The learning is guided by the difference
between the desired output and the actual output. Thus, the learning signal for each neuron is:
r =d−o
In each step the weight vector is updated (see Figure 11.15) according to the relation:
w(t + 1) = w(t) + c × (d(t) − f (w(t), x(t))) × x(t)
The initial value for w does not matter.
✲×✛
❄
+✛
× ✛
c
✻
w(t+1)
✲
clock
❄
wReg
w(t)
✲
x(t)
❄
NEURON
- ✛
✻
d(t)
✲
o(t)
Figure 11.15: The perceptron learning rule
Example 11.9 Let be a four-input neuron with the activation function sgn. The initial weight
vector is:
w(t0 ) = [1 − 1 0 0.5]
The training inputs are:
x1 = [1 − 2 0 − 1],
x2 = [0 1.5 − 0.5 − 1],
x3 = [−1 1 0.5 − 1]
and the desired output for the three input vectors are: d1 = −1, d2 = −1, d3 = 1. The learning
constant is c = 0.1.
Applying by turn the three training input vectors for c = 1 we obtain:
step 1 : because (d − sgn(net)) ̸= 0
w(t0 +1) = w(t0 )+0.1×(−1+sgn(net))×x1 = w(t0 )+0.1×(−1−1)×x1 = [0.8 −0.6 0 0.7]
step 2 : because (d − sgn(net) ̸= 0) no correction is needed in this step
w(t0 + 2) = w(t0 + 1)
step 3 : because (d − sgn(net)) = 2
w(t0 + 3) = w(t0 + 2) + 0.1 × 2 × x3 = [0.6 − 0.4 0.1 0.5]
⋄
11.6. PROBLEMS
11.5.5
339
Neural processing
NN are currently used to model complex relationships between inputs and outputs or to find
patterns in streams of data. Although NN has the full power of a Universal Turing Machine
(some people claim that the use of irrational values for weights results in a machine with “superTuring” power), the real application of this paradigm are limited only to few functions involving
specific complex memory functions (please do not use this paradigm to implement a text editor).
They are grouped in the following categories:
• auto-association: the input (even a degraded input pattern) is associated to the closest
stored pattern
• hetero-association: the association is made between pais of patterns; distorted input patterns are accepted
• classification: divides the input patterns into a number of classes; each class is indicated
by a number (can be understood as a special case of hetero-association which returns a
number)
• recognition: is a sort of classification with input patterns which do not exactly correspond
to any of the patterns in the set
• generalization: is a sort of interpolation of new data applied to the input.
What is specific for this computational paradigm is that its “program” – the set of weight
matrices generated in the learning process – do not provide explicit information about the
functionality of the net. The content of the weight matrices can not be read and understood
as we read and understand the program performed by a conventional processor built by a
register file, an ALU, .... The representation of an actual function of a NN defies any pattern
based interpretation. Maybe this is the price we must pay for the complexity of the functions
performed by NN.
11.6
Problems
Problem 11.1 Design a stack with the first two recordings accessible.
Problem 11.2 Design a stack with the following features in reorganizing the first recordings.
Problem 11.3 Design a stack with controlled deep access.
Problem 11.4 Design an expandable stack.
Problem 11.5 The global loop on a linear cellular automata providing a pseudo-noise generator.
Problem 11.6
Problem 11.7
Problem 11.8
340CHAPTER 11. # ∗ SELF-ORGANIZING STRUCTURES:N-TH ORDER DIGITAL SYSTEMS
Chapter 12
# ∗ GLOBAL-LOOP SYSTEMS
In the previous chapter
we ended to discuss about closing local loops in digital systems introducing the n-th order
systems.
In this chapter
a new kind of loop will be introduced: the global loop. This loop has the following
characteristics:
• it is closed over an n order system
• usually carries only control information about the state of the entire system
• the feedback introduced classifies each subsystems in the context of the entire system
In the next chapter
the second part of this book starts, with a discussion about recursive functions and their
circuit implementations.
• The computation model of partial functions is presented.
• To each level (basic functions, composition, primitive recursiveness and minimalization) an digital order is associated.
• Conclusion: computation means at least 3-loop circuits.
341
342
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
A super-system is characterized by global loops closed over an n-order digital system. A global loop
receive information distributed over an array of digital modules and sends back in each digital module
information related with the whole content of the input.
12.1
Global loops in cellular automata
12.2
ConnexArrayT M based system
Connex System (see Figure 12.1) is centered on ConnexArrayT M , a first-order super-system. Its superloop is a first detection network (see 6.1.13).
flag
Atom(s)
log-depth Reduction Network
✻✻✻log-depth Global Loop
✻✻✻
✻✻
❄❄❄
❄❄
Linear Cellular Array of EUs | PUs
✻✻✻
✻✻
❄ ❄
✛
✲
P
Memory
✻
✛
✛ ✲ DMA
✛
✲
✻✻
log-depth Distribution Network
CONNEX ARRAY
✻
Instruction & Atom & Address
Figure 12.1: Connex System. It contains Connex Array with its global loop (a first detection
network), and the Reduction/Distribution network closed over the processor P used to control the entire
system. The direct memory access memory unit (DMA) allow data exchange between the array and
Memory.
Connex System is a parallel processing unit containing:
P : is a processor which fetches in each cycle from the program memory a composite instruction of form:
12.2. CONNEXARRAYT M BASED SYSTEM
343
instruction = {atomOpcCode, sequenceOpCode, const}
where:
atomOpcCode : is the code executed by the processor P, which runs the stream of instructions
stored in Memory in order to process atoms of data or to generate the stream of operations
to be performed in array on sequence of data
sequenceOpCode : is the cod sent by the processor P to be executed in each cell of the array where
a sequence of atoms is distributed
const : is the constant (used as signed integer, as address or as symbol) associated to
scalarOpcCode or vectorOpCode
Connex Array :
Distribution Network : distributes toward the cells of the array the code sequenceOpCode
associated with the value of atom[n-1:0] and of the local memory address addr[m-1:0] in
conjunction with the read/write command and address generated by the DMA module. It
contains:
preDecoder : used to
• pre-decode the sequenceOpCode operation code
• identify the cycle when the read/write command requested by DMA can be inserted
alongside the local memory address
• generate
elemCom[m+n+10:0] = {memCom[2:0], addr[m-1:0], data[n-1:0], instr[7:0]}
to be distributed in array, where:
memCom[2:0] : encodes read and write commands for the local memory; if
sequenceOpCode encode a memory operation it has priority over a similar command provided by the DMA unit
addr[m-1:0] : the address for the local memory in each cell; is, according to the
code sequenceOpCode, const or the content provided by a processor’s register; if
DMA request an accepted access to the local memory it is provided by DMA
data[n-1:0] : is, according to the code sequenceOpCode, const or the content
provided by a processor’s register
instr[7:0] : is sequenceOpCode
distrNet :
Linear Cellular Array : is composed by p identical elements with the simplest interconnection
network: one-dimension array. Each cell is connected with its left and right cell only. Thus,
the resulting system is an n-th order digital system. Each element of the array contains three
sections:
transferSection : used to insert or extract a sequence of atoms in/from array; the inserted
sequence is stored in the local memorySection, while the the extracted sequence is loaded
from the same local memory
memorySection : is the local memory used to store 2m atoms
computingSection : is an execution unit with registers (organized as a register file or as a
stack) and ALU; the code sequenceOpCode is executed according to the internal states
of each element
The array is characterized by two special sequences:
index sequence : ix = {0,1,...,p-1}
distributed alongside the cells used to identify each cell. The left most cell has the index
0.
activation sequence : is the Boolean sequence of flags
flags = {cellFlag(0),cellFlag(1)...,cellFlag(p-1)}
used to select the active cells in each cycle: if cellFlag(i) = 1, then the cell i execute the currently received sequenceOpCode, else it receives the “no operation” code
(sequenceOpCode = nop).
344
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
Global Loop : receives the sequence flags and returns two Boolean sequences:
first = {first(0),first(1),...,first(p-1)}
with first(i) = 1 only on the position of the first occurrence of 1 in flags;
back = {back(0),back(1),...,back(p-1)}
with back(i) = 1 only on the positions following the first 1 in flags.
Reduction Network : performs functions defined on sequences takeing values in the set of atoms;
it is used to send back atoms to the processor P.
DMA : is a direct memory access module used to transfer sequences between the array and Memory in
parallel with the computation process developed in array. The sum of transferSection modules
distributed in Linear Cellular Array form a p-stage shift register used to shift-in data form Memory
or the shift-out data to Memory under the control of DMA. The transfer between this big shift
register and the locally distributed memory (memorySection in each element) is “negotiated” at
the level of preDecode module between the command generated by DMA and the instruction issued
by the processor.
CONNEX SYSTEM
✻
✻
✻
❄
❄
❄
Program memory
Scalar memory
Vector memory
Figure 12.2: Connex Architecture.
connexSystem
pm2p
sm2p
clock
reset
vm2a
✲
✲
✲
✲
✲
✲
✲
✲
✲
✲ p2pm
✲ p2sm
scalarUnit
p2a
a2p
vectorUnit
✲ a2vm
Figure 12.3: connexSystem.
The psudo-Verilog description of the system is:
module connexSystem #(‘include "0_parameters.v")
(
// PROGRAM INTERFACE
output [p2pmDim-1:0]
p2pm
, // processor to program memory
12.2. CONNEXARRAYT M BASED SYSTEM
input
[pm2pDim-1:0]
pm2p
output
input
[p2smDim-1:0]
[sm2pDim-1:0]
p2sm
sm2p
output
input
[a2vmDim-1:0]
[vm2aDim-1:0]
a2vm
vm2a
, //
//
, //
, //
//
, //
, //
reset
clock
,
);
input
input
345
program memory to processor
SCALAR INTERFACE
processor to scalar memory
scalar memory to processor
VECTOR INTERFACE
array to vector memory
vector memory to array
wire ...
scalarUnit scalarUnit(...); // processor used as controller
vectorUnit vectorUnit(...); // array of execution/processing cells
endmodule
scalarUnit
✲
pm2p
a2p
✲
✲
[n+1:n] ✲
✲
✲
✲
scalarOp
✲
p2pm
✲
p2sm
✲
p2a
value
✲
controlUnit
✲
✲
✲
✲
✲
[n-1:0]
scalarOp
value
reset
tos
executionUnit
✲
✲
✲
value
vectorOp
reg
clock
Figure 12.4: scalarUnit.
module scalarUnit #(‘include "0_parameters.v")
(
output
[p2pmDim-1:0]
p2pm
, // processor to program memory
input
[pm2pDim-1:0]
pm2p
, // program memory to processor
output
[p2smDim-1:0]
p2sm
, // processor to scalar memory
input
[sm2pDim-1:0]
sm2p
, // scalar memory to processor
output reg [55:0]
p2a
,
input
[n+1:0]
a2p
,
input
reset
,
input
clock
);
wire
...
always @(posedge clock) p2a <= {tos, value, vectorOp}
assign {vectorOp, scalarOp, value} = pm2p
;
;
346
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
controlUnit controlUnit(...)
;
executionUnit executionUnit(...);
endmodule
vectorUnit
✲
✲
✲
clock
p2a
✲
✲
✲
✲
✲
vm2a
reset
connexArray
✲ ✲
✲
aFlag
rOut
a2d
d2a
dma
d2p
a2p
✲
✲
a2vm
Figure 12.5: vectorUnit.
module vectorUnit #(‘include "0_parameters.v")
(
output [a2vmDim-1:0]
a2vm
, // array to vector memroy
input
[vm2aDim-1:0]
vm2a
, // vector memory to array
input
[55:0]
p2a
, // processor2array = {tos, vectorOp}
output [n+1:0]
a2p
, // array2processor = {a2pFlag, d2pFlag, a2pScalar}
input
reset
,
input
clock
);
wire ...
assign a2p = {rOut, d2p, aFlag} ;
dma dma(...)
;
connexArray connexArray(...);
endmodule
module connexArray #(‘include "0_parameters.v")
(
input
[55:0]
p2a
, // processor2array = {tos, vectorOp}
output [31:0]
rOut
, // reduction network output
output
aFlag
, // array flag: at least one cell is active
input
[w+m+2:0]
d2a
,
output [w-1:0]
a2d
,
input
clock
);
wire ...
linearCellularArray
distributionNetwork
globalLoop
reductionNetwork
endmodule
linearCellularArray(...);
distributionNetwork(...);
globalLoop(...)
;
reductionNetwork(...)
;
12.2. CONNEXARRAYT M BASED SYSTEM
347
connexArray
✲
clock
globalLoop
loopOut
p2a
d2a
✲
✲ distribution
Network
✲ memCom4a
✲ dAddr4a
✲
✲
transIn
✲
✲
transShift
✲
bVect
linear
Cellular
Array
sVect
✲
✲
✲
vectorOp
reduction
Network
Figure 12.6: connexArray.
module linearCellularArray #(‘include "0_parameters.v")
(
output
[p*n-1:0]
sVect
,
output
[p-1:0]
bVect
,
input
[2*p-1:0]
loopOut
,
input
[p*(n+11)-1:0] distrOut
,
input
[w-1:0]
transIn
,
input
transShift ,
output reg [w-1:0]
a2d
,
input
clock
);
wire
[n:0]
interCellVector[0:p-1] ;
wire
[n-1:0] transOut[0:p-1]
;
wire
[w-1:0] data4p
;
;
genvar j;
generate for(j=0; j<l; j=j+1) begin: tr
assign data4p[(j+1)*n-1:j*n] = transOut[j]
end
endgenerate
;
genvar i;
generate for(i=0; i<p; i=i+1) begin: el
element el(.scalar
(sVect[(i+1)*n-1:i*n]
),
.boolean
(bVect[i]
),
.transOut
(transOut[i]
),
.interCellOut
(interCellVector[i]
),
.formLeftCell
(interCellVector[(p+i-1)%p]
),
.formRightCell (interCellVector[(i+1)%p]
),
.transIn
((i<(p-l)) ? transOut[i+l] :
transIn[(((i%l)+1)*n)-1:(i%l)*n]),
.transShift
(transShift
),
.loop
({loopOut[p+i], loopOut[i]}
),
.formDistr
(distrOut[(i+1)*(n+11)-1:i*(n+11)] ),
.index
(i
),
.clock
(clock
));
end
endgenerate
endmodule
module element #(‘include "0_parameters.v")
aFlag
✲
a2d
✲
rOut
✛
distrOut
always @(posedge clock) a2d <= data4p
✲
348
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
✲
fromDistr✲
loop
✲
transIn[n-1:0]
✲
✲
✲
fromRightCell
✲
fromLeftCell
distrOut
loopOut
transIn
clock
transShift
✲
transOut
scalar
✲
✲
sVect
✲
bVect
✲
a2d
element
boolean
✲
❄
interCellOut
el[0]
✲
✲
loop
✲
transIn[n-1:0]
✲
✲
✲
fromRightCell
✲
fromLeftCell
✲✲
transOut
fromDistr
scalar
✲
element
boolean
✲
interCellOut
el[1]
✲
fromDistr✲
loop
✲
transIn[n-1:0]
✲
✲
✲
fromRightCell
✲
fromLeftCell
✲
transOut
scalar
✲
element
boolean
✲
interCellOut
el[2]
✲
✲
loop
✲
transIn[n-1:0]
✲
✲
✲
fromRightCell
✲
fromLeftCell
✲
transOut
fromDistr
scalar
✲
element
boolean
✲
interCellOut
el[3]
Figure 12.7: linearCellularArray.
a2d
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
✲
✲
fromDistr
✲
✲
✲
index✲
✲
✲
✲
✲
vectorOp
elOp
c4a
clock
fromLeftCell
from RightCell
loop
✲
✲
memCom
✲
memAddr
✲
✲
349
✲
✲
✲
✲
✲
executionCell
✲ scalar
✲ boolean
✲interCelOut
✲
memCom
memOut
localMemory
transOut
transModule
transIn
transShift
Figure 12.8: element.
(
wire
reg
output [n-1:0]
output
output [n-1:0]
output [n:0]
input
[n:0]
input
[n:0]
input
[n-1:0]
input
input
[1:0]
input
[n+10:0]
input
[c-1:0]
input
[n-1:0]
memOut
[n+10:0]
elOp
scalar
boolean
transOut
interCellOut
formLeftCell
formRightCell
transIn
transShift
loop
formDistr
index
clock
;
;
always @(posedge clock) elOp <= formDistr
localMemory
transModule
executionCell
endmodule
12.3
,
,
,
,
,
,
,
,
,
,
,
);
;
localMemory(...))
;
transModule(...)
;
executionCell(...)) ;
Low level functional description
Connex System is simulated in SCHEME1 providing an easy to use tool for architectural design and
algorithm development. The Connex System (see Figure 12.9) is defined by the following set of global
variables:
1
The DrRacket implementation downloadable form http://racket-lang.org/download/.
350
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
(define
(define
(define
(define
(define
(define
(define
(define
MEMORY ()); list of scalars
ARRAY ()); 2-level list of lists
M 0); memory size
P 0); vector size
N 0); number of vectors
INDEX ()); constant vector
ACTIVE ()); vector of up/down counters
(SYSTEM) (list ARRAY ACTIVE INDEX MEMORY))
ARRAY
Vect 0
s00
s01
Vect 1
s10
s11
MEMORY
s0P −1
s0
s1
s2
Vect N-1
ACTIVE
a0
a1
aP −1
INDEX
0
1
P-1
sM −2
sM −1
SYSTEM
Figure 12.9: SYSTEM.
ARRAY represents the internal distributed memory of N P-word vectors. MEMORY is the external M-word
RAM of the system. The vector INDEX is a constant vector used to individualize the P cells of the hardware
array. Each cell has a its own local N-word RAM. The vector ACTIVE is used to select the active cell in each cycle
of computation: if ai = 0 then the i-th cell is active. Every time a cell is omitted from selection the associated
counter in ACTIVE is incremented. The selections are restored by an unsigned integer decrement of the vector
ACTIVE.
The
code
of
the
functions
used
in
simulation
is
posted
at
http://arh.pub.ro/gstefan/BCsystem/neoConnex.rkt.
12.3.1
System initialization
The following functions are used to initialize the whole system:
ResetSystem : is the function used to initialize the shape of a SYSTEM having n p-word vectors in ARRAY
and an m-word MEMORY.
(define (ResetSystem n p m) ;
(set! ARRAY (make-list n ( make-list p ’^)));
(set! MEMORY (make-list m ’^));
(set! ACTIVE (make-list p 0));
(set! N n) ;number of vectors in ARRAY
(set! P p) ;length of a vector
max P=1024
(set! M m) ; number of scalars in MEMORY
(set! INDEX (Index P))
)
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
351
The vector ACTIVE is a vector of counters used to keep track of the embedded selections. For ACTIVE[i] = 0
the cell i is active. The vector INDEX is used to index the cells of the system. INDEX[i] = i in the cell i.
Two auxiliary functions are defined:
pp : is the pretty printing SYSTEM
rs : defined as:
(define (rs) (ResetSystem 4 8 32))
to be used to make easier the initialization.
Example 12.1 A system initialization runs the function rs.
> (rs)
> (pp)
0:(^ ^
1:(^ ^
2:(^ ^
3:(^ ^
A:(0 0
M:(^ ^
^
^
^
^
0
^
^
^
^
^
0
^
^
^
^
^
0
^
^
^
^
^
0
^
^
^
^
^
0
^
^)
^)
^)
^)
0)
^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^)
provides a system with 4 8-word vectors in ARRAY, a 32-word memory and all cells selected. The initial
content of ARRAY and MEMORY is filled up with dummy symbols.
⋄
test : is defined for initializing the content of the system:
(define (test)
(ResetSystem 4 8 32)(newline)
(SetVector 0 ’(2 1 4 3 4 6 7 0))
(SetVector 1 (makeAll 5))
(SetVector 2 ’(0 1 2 3 4 5 6 7 ))
(SetVector 3 ’(7 6 5 4 3 2 1 0))
(SetActive ’(1 0 0 0 0 0 0 1))
(SetStream 0 (Index 32))
(pp)
)
where:
Index : (Index n) returns the list of n successive numbers starting with 0.
> (Index 12)
(0 1 2 3 4 5 6 7 8 9 10 11)
The function is used as (Index P) to initialize the constant vector INDEX used to name each cell in the array of
P cells.
MakeAll : (MakeAll val) returns a P-length vector with val on each position. (TBD: change name in Fill)
> (MakeAll ’13)
(13 13 13 13 13 13 13 13)
Example 12.2 Let’s apply the function test.
> (test)
0:(2 1 4
1:(5 5 5
2:(0 1 2
3:(7 6 5
A:(1 0 0
M:(0 1 2
3
5
3
4
0
3
4
5
4
3
0
4
6
5
5
2
0
5
7
5
6
1
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
Now, the content of the four vectors is initialized, the first and the last cell are inactivated and the memory
is loaded with data.
⋄
352
12.3.2
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
Addressing functions
Vec : (Vec aAddr) address the vector aAddr in ARRAY
>(Vec 3)
(7 6 5 4 3 2 1 0)
Mem : (Mem mAddr) address the location mAddr in MEMORY
>(Mem 7)
7
12.3.3
Setting functions
Copy : (Copy vSource vDest) returns vector vSource copied over the vector vDest in all places where ACTIVE
= 0.
> (Copy (Vec
(5 1 2 3 4 5
> (pp)
0:(2 1 4 3 4
1:(5 5 5 5 5
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 0 0
M:(0 1 2 3 4
> (Copy (Vec
(a 1 2 3 4 5
> (Copy ’(a
(0 a a a a a
2) (Vec 1))
6 5)
6 7 0)
5 5 5)
5 6 7)
2 1 0)
0 0 1)
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
2) ’(a a a a a a a a))
6 a)
a a a a a a a) (Vec 2))
a 7)
SetVector : (SetVector aAddr vector) initialize the vector positioned in ARRAY at aAddr with the value
vector in any active positions (only where ACTIVE[i] = 0).
> (SetVector
(5 4 3 3 6 6
> (pp)
0:(2 1 4 3 4
1:(5 4 3 3 6
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 0 0
M:(0 1 2 3 4
1 ’(4 4 3 3 6 6 7 7))
7 5)
6
6
5
2
0
5
7
7
6
1
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
The first and the last 5 in the vector 1 are untouched, because the first and the last cell in the system are inactive
(A[0] = A[7] = 0).
SetAll : (SetAll aAddr vector) initialize the vector positioned in ARRAY at aAddr with the value vector.
> (SetAll 1 ’(9 9 9 9 9 9 9 9))
(9 9 9 9 9 9 9 9)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(9 9 9 9 9 9 9 9)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
All cells in vector 1 are set on 9, whatever the content of ACTIVE vector is.
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
353
SetStream : (SetStream mAddr stream) sets the content of MEMORY at stream starting with the address
mAddr
> (SetStream
(a b c d e)
> (pp)
0:(2 1 4 3 4
1:(9 9 9 9 9
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 0 0
M:(0 1 2 3 4
23 ’(a b c d e))
6
9
5
2
0
5
7
9
6
1
0
6
0)
9)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 a b c d e 28 29 30 31)
Starting from the location 23 in MEMORY the stream a b c d e is initialized.
12.3.4
Transfer functions
Basic transfer functions
StoreVector
: (StoreVector aAddr mAddr) store in MEMORY, starting with mAddr, the vector located in
ARRAY at aAddr.
> (StoreVector
(0 1 2 3 4 5 6
> (pp)
0:(2 1 4 3 4 6
1:(9 9 9 9 9 9
2:(0 1 2 3 4 5
3:(7 6 5 4 3 2
A:(1 0 0 0 0 0
M:(0 1 2 3 4 5
2 24)
7)
7
9
6
1
0
6
0)
9)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 a 0 1 2 3 4 5 6 7)
LoadVector : (LoadVector aAddr mAddr) load in ARRAY at the location aAddr the P-length stream of
data starting in MEMORY at mAddr.
> (LoadVector 1 16)
(16 17 18 19 20 21 22 a)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(16 17 18 19 20 21 22 a)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 a 0 1 2 3 4 5 6 7)
StoreVectorStrided : (StoreVectorStrided aAddr mAddr burst stride) store in MEMORY, starting
with mAddr, the vector located in ARRAY at a Addr in bursts of burst words with address having the stride
stride.
>(test)
> (StoreVectorStrided 2 0 2 5)
((6 7) (4 5) (2 3) (0 1))
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 5 5 5 5 5 5 5)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 2 3 7 8 9 4 5 12 13 14 6 7 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
354
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
LoadVectorStrided : (LoadVectorStrided aAddr mAddr burst stride) load in ARRAY at the location
aAddr the stream of data from MEMORY starting at mAddr in bursts of burst words gathered with the stride
stride.
> (test)
> (LoadVectorStrided 1 2 2 6)
(2 3 8 9 14 15 20 21)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(2 3 8 9 14 15 20 21)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 2 3 7 8 9 4 5 12 13 14 6 7 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
12.3.5
Arithmetic functions
The arithmetic functions are defined on scalars or on vectors. Therefore, the variables x and y ar considered in
any combinations:
• (x y) : (scalar scalar) returning scalar; the operations is performed by the controller
• (x y) : (scalar vector) returning vector; the operation is performed in array with the scalar operand provided
by the controller
• (x y) : (vector scalar) returning vector; the operation is performed in array with the scalar operand provided
by the controller
• (x y) : (vector vector) returning vector; the operation is performed in array
Abs : (Abs x) computes the absolute value of x. If x is scalar, then returns a scalar. If x is a vector, the
returns a vector of absolute values.
(test)
> (SetVector 1 ’(-1 -2 -3 -4 5 6 7
(5 -2 -3 -4 5 6 7 5)
> (SetVector 1 (Abs (Vec 1)))
(5 2 3 4 5 6 7 5)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 2 3 4 5 6 7 5)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13
> (SetStream 24 ’(-7 -8 -9))
(-7 -8 -9)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 2 3 4 5 6 7 5)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13
> (SetStream 24 (Abs (Mem 24)))
(7)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 2 3 4 5 6 7 5)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13
8))
14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
14 15 16 17 18 19 20 21 22 23 -7 -8 -9 27 28 29 30 31)
14 15 16 17 18 19 20 21 22 23 7 -8 -9 27 28 29 30 31)
Inc : (Inc x) adds 1 to all components of x.
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
355
Dec : (Dec x) subtract 1 form each components of
Add : (Add x y) defines the addition in four forms, depending of how the operands are, as follows:
• if ((x y) = (scalar scalar)), then add the two scalars
• if ((x y) = (scalar vector)), then add x to each component of y
• if ((x y) = (vector scalar)), then add y to each component of x
• if ((x y) = (vector vector)), then add the two vectors
> (Add 2 8)
10
> (Add ’(1 2 3) ’(2 3 4))
(3 5 7)
> (Add 3 ’(3 4 5))
(6 7 8)
> (Add ’(2 4 6) 5)
(7 9 11)
> (Add ’(1 3 5) ’(2 4 6))
(3 7 11)
> (SetVector 2 (Add (Vec 2) (Mem 5)))
(0 6 7 8 9 10 11 7)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 2 3 4 5 6 7 5)
2:(0 6 7 8 9 10 11 7)
3:(7 7 6 5 4 3 2 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 7 -8 -9 27 28 29 30 31)
Sub : (Sub x y) defines the subtract in four forms, depending of how the operands are, as follows:
• if ((x y) = (scalar scalar)), then subtract y from x
• if ((x y) = (scalar vector)), then subtract from x each component of y providing the vector containing the
resulting differences
• if ((x y) = (vector scalar)), then subtract y from each component of x
• if ((x y) = (vector vector)), then subtract the vector y from the vector x
Mult : (Mult x y) multiply x by y.
> (Mult 2 -6)
-12
> (Mult 2 ’(1 2 3 4))
(2 4 6 8)
> (Mult ’(2 3 4 5) -3)
(-6 -9 -12 -15)
> (Mult ’(1 3 5 7) ’(1 3 5 7))
(1 9 25 49)
> (Mult ’(1 2) ’( 1 2 3))
. . map: all lists must have same size;
Div : (Div x y) divide x by y, providing the quotient as a real number.
> (Div 7 3.0)
2.3333333333333335
> (Div ’(1 3 5) 2.0)
(0.5 1.5 2.5)
> (Div 2 ’(1.0 3 5))
(2.0 2/3 2/5)
> (Div 2.0 ’(1 3 5))
(2.0 0.6666666666666666 0.4)
> (Div ’(1.0 3.0 5.0) ’(2 4 6))
(0.5 0.75 0.8333333333333334)
356
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
Quotient : (Quotient x y) integer division of x by y
> (Quotient
2
> (Quotient
(0 1 2)
> (Quotient
(2 0 0)
> (Quotient
(0 0 0)
7 3)
’(1 3 5) 2)
2 ’(1 3 5))
’(1 3 5) ’(2 4 6))
Rem : (Rem x y) provides the remaining of the division of x by y
> (Rem 3 2)
1
> (Rem ’(1 3 5) 2)
(1 1 1)
> (Rem 2 ’(1 3 5))
(0 2 2)
> (Rem ’(3 5 7 ) ’(2.0 3.0 4.0))
(1.0 2.0 3.0)
Not : (Not x)
And : (And x y)
Or : (Or x y)
Xor : (Xor x y)
12.3.6
Reduction functions
Reduction functions are defined on vectors returning scalars (they “reduces” a vector to a scalar).
RedAdd : (RedAdd vector) adds all the selected (where ACTIVE == 0) components of vector.
> (test)
> (SetStream
(30)
> (pp)
0:(2 1 4 3 4
1:(5 5 5 5 5
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 0 0
M:(0 1 2 3 4
16 (RedAdd (Vec 1)))
6
5
5
2
0
5
7
5
6
1
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 30 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
RedAddAll : (RedAddAll vector) adds all components of vector (the content of ACTIVE vector is ignored).
> (SetStream 16 (RedAddAll (Vec 1)))
(40)
RedMax : (RedMax vector) returns the maximum value from the selected components (where ACTIVE =
0) of vector
> (RedMax (Vec 3))
6
RedMaxAll : (RedMaxAll vector) returns the maximum value from the components of vector (the content
of ACTIVE vector is ignored).
> (RedMaxAll (Vec 3))
7
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
357
Select : (Select i vector) selects the i-th component from the active components (where ACTIVE = 0) of
vector vector.
> (Select 3 (Vec 3))
4
> (Select 0 (Vec 3))
ERROR-SELECT
SelectFromAll : (SelectFromAll i vector) selects the i-th component from the vector vector (the content of ACTIVE vector is ignored).
AnyActive : (AnyActive) return 1 only if at least one component of ACTIVE is 0.
> (AnyActive)
1
> (SetActive (Vec 1))
> (AnyActive)
0
12.3.7
Test functions
The test functions are defined on scalars or on vectors. Therefore, the variables x and y ar considered in any
combinations:
• (x y) : (scalar scalar) returning scalar; the operations is performed by the controller and the result is used
to control the program running on controller
• (x y) : (scalar vector) returning vector; the operation is performed in array with the scalar operand provided
by the controller and the resulting vector is used to make selections in array
• (x y) : (vector scalar) returning vector; the operation is performed in array with the scalar operand provided
by the controller and the resulting vector is used to make selections in array
• (x y) : (vector vector) returning vector; the operation is performed in array and the resulting vector is
used to make selections in array
Lt : (Lt x y) defines the Less than comparison in four forms, depending of how the operands are, as follows:
•
•
•
•
if
if
if
if
((x
((x
((x
((x
y)
y)
y)
y)
=
=
=
=
(scalar scalar)), then compare the two scalars returning 1 if x < y
(scalar vector)), then compare x to each component of y returning a Boolean vector
(vector scalar)), then compare y to each component of x returning a Boolean vector
(vector vector)), then add the two vectors returning a Boolean vector
(test)
> (Lt 2 4)
1
> (Lt (Vec 0) (Vec 1))
(1 1 1 1 1 0 0 1)
> (Lt (Vec 1) (Vec 0))
(0 0 0 0 0 1 1 0)
> (Lt (Vec 1) 3)
(0 0 0 0 0 0 0 0)
> (Lt (Vec 0) 3)
(1 1 0 0 0 0 0 1)
> (Lt 3 (Vec 0))
(0 0 1 0 1 1 1 0)
> (SetStream 15 (Lt 3 (Vec 0)))
(0 0 1 0 1 1 1 0)
> (SetVector 1 (Lt 3 (Vec 0)))
(5 0 1 0 1 1 1 5)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 0 1 0 1 1 1 5)
2:(0 1 2 3 4 5 6 7)
3:(7 6 5 4 3 2 1 0)
A:(1 0 0 0 0 0 0 1)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0 0 1 0 1 1 1 0 23 24 25 26 27 28 29 30 31)
358
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
Gt : (Gt x y) defines the Greater than comparison in the four forms.
Leq : (Leq x y) defines the Less than or equal comparison in the four forms.
Geq : (Geq x y) defines the Greater than or equal comparison in the four forms.
Eq : defines the Equal comparison in the four forms. (Eq x y)
Even? : (Even? x) returns 1 for even integers and 0 for odd integers
12.3.8
Control functions
ResetActive : (ResetActive) set the value of the vector ACTIVE to (MakeAll 0)
> (ResetActive)
> (pp)
0:(2 1 4 3 4 6 7 0)
1:(5 5 5 5 5 5 5 5)
2:(0 1 2 3 4 5 6 7)
3:(8 9 10 11 12 13 14 15)
A:(0 0 0 0 0 0 0 0)
M:(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
SetActive : (SetActive vector) set the value of the vector ACTIVE to vector
> (SetActive
> (pp)
0:(2 1 4 3 4
1:(5 0 1 0 1
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 1 1
M:(0 1 2 3 4
’(1 0 0 1 1 0 0 1))
6
1
5
2
0
5
7
1
6
1
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
SaveActive : (SaveActive vector) save the value of the vector ACTIVE in vector
... TBD
Where : (Where booleanVector) restrain the selection according to the Boolean vector vector. The ACTIVE
vector takes the value of the vector (Not booleanVector)
EndWhere : (EndWhere) resets ACTIVE.
ElseWhere : (ElseWhere) complements the components of the vector ACTIVE.
First : (First) all the 0s in ACTIVE, except the first, if any, are turned to 1.
> (SetActive
> (First)
> (pp)
0:(2 1 4 3 4
1:(5 5 5 5 5
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 1 1 1 0
M:(0 1 2 3 4
’(1 1 1 1 0 0 0 1))
6
5
5
2
1
5
7
5
6
1
1
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
FirstIndex : (FirstIndex) returns the index of the first cell where ACTIVE[i] == 0
> (SetActive ’(1 1 1 1 0 0 0 1))
> (FirstIndex)
4
12.3. LOW LEVEL FUNCTIONAL DESCRIPTION
12.3.9
359
Stream functions
Permute : (Permute v1 v2), permutes the vector v1, according to the indexes provided by the vector v2
> (Permute (Vec 2) ’(5 6 7 1 2 5 5 5))
(5 6 7 1 2 5 5 5)
ShiftLeft : (ShiftLeft many vector) shifts left the content of the vector vector with many positions and
insert 0s in the right end. The vector ACTIVE is ignored.
> (ShiftLeft 3 (Vec 2))
(3 4 5 6 7 0 0 0)
> (SetVector 1 (ShiftLeft 3 (Vec 2)))
(5 4 5 6 7 0 0 5)
ShiftLeftEl : (ShiftLeftEl many vector el) shifts left the content of the vector vector with many positions
and insert el symbols in the right end. The vector ACTIVE is ignored
ShiftRight : (ShiftRight many vector) shifts right the content of the vector vector with many positions
and insert 0s in the left end. The vector ACTIVE is ignored.
0:(2 1 4 3 4
1:(5 5 5 5 5
2:(0 1 2 3 4
3:(8 9 10 11
A:(1 0 0 0 0
M:(0 1 2 3 4
> (SetVector
(5 0 0 0 1 2
> (pp)
0:(2 1 4 3 4
1:(5 0 0 0 1
2:(0 1 2 3 4
3:(8 9 10 11
A:(1 0 0 0 0
M:(0 1 2 3 4
6 7 0)
5 5 5)
5 6 7)
12 13 14 15)
0 0 1)
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
1 (ShiftRight 3 (Vec 2)))
3 5)
6 7 0)
2 3 5)
5 6 7)
12 13 14 15)
0 0 1)
5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
ShiftRightEl : (ShiftRightEl many vector el) shifts right the content of the vector vector with many
positions and insert el symbols in the left end. The vector ACTIVE is ignored.
RotateLeft : (RotateLeft many vector) rotate left vector with many positions. The vector ACTIVE is
ignored.
> (SetVector
(5 4 5 6 7 0
> (pp)
0:(2 1 4 3 4
1:(5 4 5 6 7
2:(0 1 2 3 4
3:(7 6 5 4 3
A:(1 0 0 0 0
M:(0 1 2 3 4
1 (RotateLeft 3 (Vec 2)))
1 5)
6
0
5
2
0
5
7
1
6
1
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
RotateRight : (RotateRight many vector) rotate right vector with many positions. The vector ACTIVE is
ignored.
> (SetVector 1 (RotateRight 3 (Vec 2)))
(5 6 7 0 1 2 3 5)
>
Search : (Search el vector) in the active cells of vector the symbol el is searched: all ACTIVE[i], except
where ((ACTIVE[i] == 0) & (vector[i] == el)), are set to 1.
... // TBD
360
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
SearchCond : (SearchCond el vector) performs a conditional search: where ((ACTIVE[i-1] == 0) &
(vector[i] == el)) ACTIVE[i] = 0, elsewhere ACTIVE[i] = 1.
... // TBD
InsertPos : (InsertPos stream at vector) inserts the stream stream in the position at in the vector
vector. The vector ACTIVE is ignored.
> (InsertPos
(7 6 a 7 a 5
> (SetVector
(7 6 a 7 a 5
> (pp)
0:(2 1 4 3 4
1:(5 6 7 0 1
2:(0 1 2 3 4
3:(7 6 a 7 a
A:(1 0 0 0 0
M:(0 1 2 3 4
’(a 7 a) 2 (Vec 3))
4 3 2 1 0)
3 (InsertPos ’(a 7 a) 2 (Vec 3)))
4 0)
6
2
5
5
0
5
7
3
6
4
0
6
0)
5)
7)
0)
1)
7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31)
DeletePos : (DeletePos many at vector) deletes many components starting form at in vector filling the
empty positions with dummy symbols.
> (SetVector 1 (DeletePos 2 3 (Vec 2)))
(5 1 2 5 6 7 ^ 5)
12.4. INDEX
12.4
361
Index
GLOBAL VARIABLES:
(define (SYSTEM) (list ARRAY MEMORY))
(define ARRAY ()); array of vectors
(define MEMORY()); array of scalars
(define ACTIVE()); vector of u/d counters
(define INDEX ()); the index vector
(define M 0); m = memory size
(define P 0); p = vector size
(define N 0); n= number of vectors
(Not x)
(And x y)
(Or x y)
(Xor x y)
REDUCTION FUNCTIONS:
(RedAdd vector)
(RedAddAll vector)
(RedMax vector)
(RedMaxAll vector)
(Select i vector)
(SelectFromAll i vector); TBD
(AnyActive); -> 1 if at least one is active
(define (ResetSystem n p m) ;
(set! ARRAY (make-list n ( make-list p ’^)));
(set! MEMORY (make-list m ’^));
(set! ACTIVE (make-list p 1));
TEST FUNCTIONS:
(set! N n) ;number of vectors in ARRAY
(Lt x y)
(set! P p) ;length of a vector
(Gt x y)
(set! M m) ; number of scalars in MEMORY
(Leq x y)
(set! INDEX (Index P)))
(define (rs) (ResetSystem 4 8 32)) (define (test) (Geq x y)
(Eq x y)
(ResetSystem 4 8 32)(newline)
(Even? x)
(SetVector 0 ’(2 1 4 3 4 6 7 0))
(SetVector 1 (makeAll 5))
CONTROL FUNCTIONS
...
(ResetActive); ACTIVE <= (0 0 ... 0)
(SetStream 0 (Index 32))
(SetActive vector); ACTIVE <= vector
(pp))
(SaveActive booleanVector); TBD
(Where vector)
AUXILIARY FUNCTIONS
(EndWhere)
(pp); pretty printing
(ElseWhere)
(Index n)
(First); -> leaves just first active
(MakeAll val)
(FirstIndex); -> index of the first active
ADDRESSING FUNCTIONS:
STREAM FUNCTIONS:
(Vec aAddr); ARRAY[aAddr]
(Permute v1 v2)
(Mem mAddr); MEMORY[mAddr]
(ShiftLeft many vector)
(ShiftLeftEl many vector el); TBD
SETING FUNCTIONS:
(ShiftRight many vector)
(Copy vSource vDest)
(ShiftRightEl many vector el); TBD
(SetVector aAddr vector); ARRAY[aAddr]<=vector
(RotateLeft many vector)
(SetAll aAddr vector); !!!!!!!!!!!!!!!!!! TBR
(RotateRight many vector)
(SetStream mAddr stream)
(Search el vector)
(SearchCond el vector)
TRANSFER FUNCTIONS:
(InsertPos stream at vector)
(StoreVector aAddr mAddr)
(DeletePos many at vector)
;MEMORY[mAddr:mAddr+P] <= (Vec aAddr)
(LoadVector aAddr mAddr)
NON-CONNEX FUNCTIONS:
;(Vec aAddr) <= MEMORY[mAddr:mAddr+P]
(make-list n x); list of n x
(StoreVectorStrided aAddr mAddr burst stride)
(cut v pos n); cut from pos n elem in vect v
(LoadVectorStrided aAddr mAddr burst stride)
(CheckZero x); -> #t if all elem. of v are 0
(put v pos w); set w on pos in vector v
ARITHMETIC & LOGIC FUNCTIONS:
(Abs x)
(Inc x)
(Dec x)
(Add x y)
(Sub x y)
(Mult x y)
(Div x y)
(Quotient x y)
(Rem x y)
362
12.5
CHAPTER 12. # ∗ GLOBAL-LOOP SYSTEMS
Problems
Problem 12.1
Problem 12.2
Problem 12.3
Chapter 13
DESIGNING A COMPLEX
DIGITAL SYSTEM
In the previous chapter
In this chapter
•
•
•
In the next chapter
•
•
•
363
364
CHAPTER 13. DESIGNING A COMPLEX DIGITAL SYSTEM
13.1
The Architectural Description
13.2
Complex Interconnections
13.2.1
Pipelined complex systems
13.2.2
Buffered complex systems
13.3
Complex Applications
13.4
Problems
Problem 13.1
Problem 13.2
Problem 13.3
Part III
ANNEXES
365
Appendix A
The Verilog HDL
367
368
APPENDIX A. THE VERILOG HDL
Appendix B
Binary Arithmetic
B.1
Binary numbers
B.1.1
Positive integers
B.1.2
Signed integers
B.1.3
Real numbers approximated as floating point numbers
B.2
Addition & subtract
B.2.1
Positive integer operations
B.2.2
Signed integer operations
B.2.3
Floating point operations
B.3
Multiply
B.3.1
Positive integer operations
B.3.2
Signed integer operations
B.3.3
Floating point operations
B.4
Division
B.4.1
Positive integer operations
B.4.2
Signed integer operations
B.4.3
Floating point operations
369
370
APPENDIX B. BINARY ARITHMETIC
Appendix C
Boolean functions
Searching the truth, dealing with numbers and behaving automatically are all based on logic. Starting from the
very elementary level we will see that logic can be “interpreted” arithmetically. We intend to offer a physical
support for both the numerical functions and logical mechanisms. The logic circuit is the fundamental brick used
to build the physical computational structures.
C.1
Short History
There are some significant historical steps on the way from logic to numerical circuits. In the following
some of them are pointed.
Aristotle of Stagira (382-322) a Greek philosopher considered as founder for many scientific domains. Among them logics. All his writings in logic are grouped under the name Organon, that means
instrument of scientific investigation. He worked with two logic values: true and false.
George Boole (1815-1864) is an English mathematician who formalized the Aristotelian logic like an
algebra. The algebraic logic he proposed in 1854, now called Boolean logic, deals with the truth and the
false of complex expressions of binary variables.
Claude Elwood Shannon (1916-2001) obtained a master degree in electrical engineering and PhD
in mathematics at MIT. His Master’s thesis, A Symbolic Analysis of Relay and Switching Circuits [Shannon ’38], used Boolean logic to establish a theoretical background of digital circuits.
C.2
Elementary circuits: gates
Definition C.1 A binary variable takes values in the set {0, 1}. We call it bit.
The set of numbers {0, 1} is interpreted in logic using the correspondences: 0 → f alse, 1 → true in what is
called positive logic, or 1 → f alse, 0 → true in what is called negative logic. In the following we use positive logic.
Definition C.2 We call n-bit binary variable an element of the set {0, 1}n .
Definition C.3 A logic function is a function having the form f : {0, 1}n → {0, 1}m with n ≥ 0 and m > 0.
In the following we will deal with m = 1. The parallel composition will provide the possibility to build
systems with m > 1.
C.2.1
Zero-input logic circuits
Definition C.4 The 0-bit logic function are f00 = 0 (the false-function) which generates the one bit coded 0,
and f10 = 1 (the true-function) which generate the one bit coded 1.
They are useful for generating initial values in computation (see the zero function as basic function in partial
recursivity).
371
372
APPENDIX C. BOOLEAN FUNCTIONS
C.2.2
One input logic circuits
Definition C.5 The 1-bit logic functions, represented by true-tables in Figure C.1, are:
• f01 (x) = 0 – the false function
• f11 (x) = x′ – the invert (not) function
• f21 (x) = x – the driver or identity function
• f31 (x) = 1 – the true function
f01
0
0
x
0
1
f11
1
0
f21
0
1
f31
1
1
a.
“1” = VDD
✲
x
x
x’
x
e.
d.
c.
b.
✲
Figure C.1: One-bit logic functions. a. The truth table for 1-variable logic functions. b. The
circuit for “0” (false) by connecting to the ground potential. c. The logic symbol for the inverter circuit.
d. The logic symbol for driver function. e. The circuit for “1” (true) by connecting to the high potential.
Numerical interpretation of the NOT circuit: one-bit incrementer. Indeed, the output represents the
modulo 2 increment of the inputs.
C.2.3
Two inputs logic circuits
Definition C.6 The 2-bit logic functions are represented by true-tables in Figure C.2.
x
0
0
1
1
y
0
1
0
1
f02
0
0
0
0
f12
1
0
0
0
f22
0
1
0
0
f32
1
1
0
0
f42
0
0
1
0
f52
1
0
1
0
f62
0
1
1
0
f72
1
1
1
0
f82
0
0
0
1
f92
1
0
0
1
2
fA
0
1
0
1
2
fB
1
1
0
1
2
fC
0
0
1
1
2
fD
1
0
1
1
2
fE
0
1
1
1
2
fF
1
1
1
1
a.
x
f82 = xy
y
x
f72 = (xy)′
y
c.
b.
x
2
fE
=x+y
y
x
f12 = (x + y)′
y
e.
d.
x
f62 = x ⊕ y
y
x
f92 = (x ⊕ y)′
y
f.
g.
Figure C.2: Two-bit logic functions. a. The table of all two-bit logic functions. b. AND gate
– the original gate. c. NAND gate – the most used gate. d. OR gate. e. NOR gate. f. XOR gate –
modulo2 adder. g. NXOR gate – coincidence circuit.
Interpretations for some of 2-input logic circuits:
• f82 : AND function is:
– a multiplier for 1-bit numbers
– a gate, because x opens the gate for y:
if (x = 1) output = y; else output = 0;
• f62 : XOR (exclusiv OR) is:
– the 2-modulo adder
C.3. HOW TO DEAL WITH LOGIC FUNCTIONS
373
– NEQ (not-equal) circuit, a comparator pointing out when the two 1-bit numbers on the input are
inequal
– an enabled inverter:
if x = 1 output is y ′ ; else output is y;
– a modulo 2 incrementer.
2
• fB
: the logic implication is also used to compare 1-bit numbers because the output is 1 for y < x
• f12 : NOR function detects when 2-bit numbers have the value zero.
All logic circuits are gates, even if a true gate is only the AND gate.
C.2.4
Many input logic circuits
For enumerating the 3-input function a table with 8 line is needed. On the left side there are 3 columns and on
the right side 256 columns (one for each 8-bit binary configuration defining a logic function).
n
Theorem C.1 The number of n-input one output logic (Boolean) functions is N = 22 . ⋄
Enumerating is not a solution starting with n = 3. Maybe the 3-input function can be defined using the
2-input functions.
C.3
How to Deal with Logic Functions
The systematic and formal development of the theory of logical functions means: (1) a set of elementary functions,
(2) a minimal set of axioms (of formulas considered true), and (3) some rule of deduction.
Because our approach is a pragmatic one: (1) we use an extended (non-minimal) set of elementary functions
containing: NOT, AND, OR, XOR (a minimal one contains only NAND, or only NOR), and (2) we will list a set
of useful principles, i.e., a set of equivalences.
Identity principle Even if the natural tendency of existence is becoming, we stone the value a to be
identical with itself: a = a. Here is one of the fundamental limits of digital systems and of computation based on
them.
Double negation principle The negation is a “reversible” function, i.e., if we know the output we can
deduce the input (it is a very rare, somehow unique, feature in the world of logical function): (a′ )′ ) = a. Actually,
we can not found the reversibility in existence. There are logics that don’t accept this principle (see the intuitionist
logic of Heyting & Brower).
Associativity Having 2-input gates, how can be built gates with much more inputs? For some functions
the associativity helps us.
a + (b + c) = (a + b) + c = a + b + c
a(bc) = (ab)c = abc
a ⊕ (b ⊕ c) = (a ⊕ b) ⊕ c = a ⊕ b ⊕ c.
Commutativity Commutativity allows us to connect to the inputs of some gates the variable in any order.
a+b=b+a
ab = ba
a⊕b=b⊕a
Distributivity Distributivity offers the possibility to define all logical functions as sum of products or as
product of sums.
a(b + c) = ab + ac
a + bc = (a + b)(a + c)
a(b ⊕ c) = ab ⊕ ac.
Not all distributions are possible. For example:
a ⊕ bc ̸= (a ⊕ b)(b ⊕ c).
The table in Figure C.3 can be used to prove the previous inequality.
374
APPENDIX C. BOOLEAN FUNCTIONS
a
0
0
0
0
1
1
1
1
b
0
0
1
1
0
0
1
1
c
0
1
0
1
0
1
0
1
bc
0
0
0
1
0
0
0
1
a ⊕ bc
0
0
0
1
1
1
1
0
a⊕b
0
0
1
1
1
1
0
0
a⊕c
0
1
0
1
1
0
1
0
(a⊕b)(a⊕c)
0
0
0
1
1
0
0
0
Figure C.3: Proving by tables. Proof of inequality a ⊕ bc ̸= (a ⊕ b)(b ⊕ c).
Absorbtion Absorbtion simplify the logic expression.
a + a′ = 1
a+a=a
aa′ = 0
aa = a
a + ab = a
a(a + b) = a
Tertium non datur: a + a′ = 1.
Half-absorbtion The half-absorbtion allows only a smaller, but non-neglecting, simplification.
a + a′ b = a + b
a(a′ + b) = ab.
Substitution The substitution principles say us what happen when a variable is substituted with a value.
a+0=a
a+1=1
a0 = 0
a1 = a
a⊕0=a
a ⊕ 1 = a′ .
Exclusion The most powerful simplification occurs when the exclusion principle is applicable.
ab + a′ b = b
(a + b)(a′ + b) = b.
Proof. For the first form:
ab + a′ b = b
applying successively distribution, absorbtion and substitution results:
ab + a′ b = b(a + a′ ) = b1 = b.
For the second form we have the following sequence:
(a + b)(a′ + b) = (a + b)a′ + (a + b)b = aa′ + a′ b + ab + bb =
0 + (a′ b + ab + b) = a′ b + ab + b = a′ b + b = b.
De Morgan laws Some times we are interested to use inverting gates instead of non-inverting gates, or
conversely. De Morgan laws will help us.
a + b = (a′ b′ )′
ab = (a′ + b′ )′
a′ + b′ = (ab)′
a′ b′ = (a + b)′
C.4. MINIMIZING BOOLEAN FUNCTIONS
C.4
375
Minimizing Boolean functions
Minimizing logic functions is the first operation to be done after defining a logical function. Minimizing a logical
function means to express it in the simplest form (with minimal symbols). To a simple form a small associated
circuit is expected. The minimization process starts from canonical forms.
C.4.1
Canonical forms
The initial definition of a logic function is usually expressed in a canonical form. The canonical form is given by
a truth table or by the rough expression extracted from it.
Definition C.7 A minterm associated to an n-input logic function is a logic product (AND logic function)
depending by all n binary variable. ⋄
Definition C.8 A maxterm associated to an n-input logic function is a logic sum (OR logic function) depending
by all n binary variable. ⋄
Definition C.9 The disjunctive normal form, DNF, of an n-input logic function is a logic sum of minterms.
⋄
Definition C.10 The conjunctive normal form, CNF, of an n-input logic function is a logic product of
maxterms. ⋄
Example C.1 Let be the combinational multiplier for 2 2-bit numbers described in Figure C.4. One number is
the 2-bit number {a, b} and the other is {c, d}. The result is the 4-bit number {p3, p2, p1, p0}. The logic equations
result direct as 4 DNFs, one for each output bit:
p3 = abcd
p2 = ab’cd’ + ab’cd + abcd’
p1 = a’bcd’ + a’bcd + ab’c’d + ab’cd + abc’d + abcd’
p0 = a′ bc′ d + a′ bcd + abc′ d + abcd.
Indeed, the p3 bit takes the value 1 only if a = 1 and b = 1 and c = 1 and d = 1. The bit p2 is 1 only one of the
following three 4-input ADNs takes the value 1: ab′ cd′ , ab′ cd, abcd′ . And so on for the other bits.
Applying the De Morgan rule the equations become: p3 = ((abcd)′ )′
p2 = ((ab′ cd′ )′ (ab′ cd)′ (abcd′ )′ )′
p1 = ((a′ bcd′ )′ (a′ bcd′ (ab′ c′ d)′ (ab′ cd)′ (abc′ d)′ (abcd′ )′ )′
p0 = ((a′ bc′ d)′ (a′ bcd)′ (abc′ d)′ (abcd)′ )′ .
These forms are more efficient in implementation because involve the same type of circuits (NANDs), and
because the inverting circuits are usually faster.
ab
cd
p3
p2
p1
p0
00
00
00
00
01
01
01
01
10
10
10
10
11
11
11
11
00
01
10
11
00
01
10
11
00
01
10
11
00
01
10
11
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
1
0
0
1
0
0
0
0
0
0
0
1
1
0
1
0
1
0
1
1
0
0
0
0
0
0
1
0
1
0
0
0
0
0
1
0
1
Figure C.4: Combinatinal circuit represented a a truth table. The truth table of the
combinational circuit performing 2-bit multiplication.
The resulting circuit is represented in Figure C.5. It consists in two layers of ADNs. The first layer computes
only minterms and the second “adds” the minterms thus computing the 4 outputs.
The logic depth of the circuit is 2. But in real implementation it can be bigger because of the fact that big
input gates are composed from smaller ones. Maybe a real implementation has the depth 3. The propagation time
is also influenced by the number of inputs and by the fan-out of the circuits.
The size of the resulting circuit is very big also: Smult2 = 54. ⋄
376
APPENDIX C. BOOLEAN FUNCTIONS
a
b
c
d
p3
p2
p1
p0
Figure C.5: Direct implementation of a combinational circuit. The direct implementation
starting from DNF of the 2-bit multiplier.
C.4.2
Algebraic minimization
Minimal depth minimization
Example C.2 Let’s revisit the previous example for minimizing independently each function. The least significant
output has the following form:
p0 = a′ bc′ d + a′ bcd + abc′ d + abcd.
We will apply the following steps:
p0 = (a′ bd)c′ + (a′ bd)c + (abd)c′ + (abd)c
to emphasize the possibility of applying twice the exclusion principle, resulting
p0 = a′ bd + abd.
Applying again the same principle results:
p0 = bd(a′ + a) = bd1 = bd.
The exclusion principle allowed us to reduce the size of the circuit from 22 to 2.
We continue with the next output:
p1 = a′ bcd′ + a′ bcd + ab′ c′ d + ab′ cd + abc′ d + abcd′ =
= a′ bc(d′ + d) + ab′ d(c′ + c) + abc′ d + abcd′ =
= a′ bc + ab′ d + abc′ d + abcd′ =
= bc(a′ + ad′ ) + ad(b′ + bc′ ) =
= bc(a′ + d′ ) + ad(b′ + c′ ) =
= a′ bc + bcd′ + ab′ d + ac′ d.
Now we used also the half-absorbtion principle reducing the size from 28 to 16.
Follows the minimization of p2:
p2 = ab′ cd′ + ab′ cd + abcd′ =
= ab′ c + abcd′ =
= ab′ c + acd′
The p3 output can not be minimized. De Morgan law is used to transform the expressions to be implemented with
NANDs.
p3 = ((abcd)′ )′
p2 = ((ab′ c)′ (acd′ )′ )′
p1 = ((a′ bc)′ (bcd′ )′ (ab′ d)′ (ac′ d)′ )′
p1 = ((abcd)′ )′ .
Results the circuit from Figure C.6. ⋄
C.4. MINIMIZING BOOLEAN FUNCTIONS
377
a
b
c
d
p3
p2
p1
p0
Figure C.6: Minimal depth minimiztion The first, minimal depth minimization of the 2-bit
multiplier.
Multi-level minimization
Example C.3 The same circuit for multiplying 2-bit numbers is used to exemplify the multilevel minimization.
Results:
p3 = abcd
p2 = ab′ c + acd′ = ac(b′ + d′ ) = ac(bd)′
p1 = a′ bc + bcd′ + ab′ d + ac′ d = bc(a′ + d′ ) + ad(b′ + c′ ) = bc(ad)′ + ad(bc)′ = (bc) ⊕ (ad)
p0 = bd.
Using for XOR the following form:
x ⊕ y = ((x ⊕ y)′ )′ = (xy + x′ y ′ )′ = (xy)′ (x′ y ′ )′ = (xy)′ (x + y)
results the circuit from Figure C.7 with size 22. ⋄
a
b
c
d
p3
p1
p2
p0
Figure C.7: Multi-level minimization. The second, multi-level minimization of the 2-bit multiplier.
Many output circuit minimization
Example C.4 Inspecting carefully the schematics from Figure C.7 results: (1) the output p3 can be obtained
inverting the NAND’s output from the circuit of p2, (2) the output p0 is computed by a part of the circuit used
for p2. Thus, we are encouraged to rewrite same of the functions in order to maximize the common circuits used
in implementation. Results:
x ⊕ y = (xy)′ (x + y) = ((xy) + (x + y)′ )′ .
p2 = ac(bd)′ = ((ac)′ + bd)′
allowing the simplified circuit from Figure C.8. The size is 16 and the depth is 3. But, more important: (1) the
circuits contains only 2-input gates and (2) the maximum fan-out is 2. Both last characteristics led to small area
and high speed. ⋄
378
APPENDIX C. BOOLEAN FUNCTIONS
a
b
c
d
p3
p1
p2
p0
Figure C.8: Multiple-output minimization. The third, multiple-output minimization of the 2-bit
multiplier.
C.4.3
Veitch-Karnaugh diagrams
In order to apply efficiently the exclusion principle we need to group carefully the minterms. Two dimension
diagrams allow to emphasize the best grouping. Formally, the two minterms are adjacent if the Hamming distance
in minimal.
Definition C.11 The Hamming distance between two minterms is given by the total numbers of binary variable
which occur distinct in the two minterms. ⋄
Example C.5 The Hamming distance between m9 = ab′ c′ d and m4 = a′ bc′ d′ is 3, because only the variable b
occurs in the same form in both minterms.
The Hamming distance between m9 = ab′ c′ d and m1 = a′ b′ c′ d is 1, because only the variable which occurs
distinct in the two minterms is a. ⋄
Two n-variable terms having the Hamming distance 1 are minimized, using the exclusion principle, to one
(n − 1)-variable term. The size of the associated circuit is reduced from 2(n + 1) to n − 1.
A n-input Veitch diagram is a two dimensioned surface containing 2n squares, one for each n-value minterm.
The adjacent minterms (minterms having the Hamming distance equal with 1) are placed in adjacent squares. In
Figure C.9 are presented the Veitch diagrams for 2, 3 and 4-variable logic functions. For example, the 4-input
diagram contains in the left half all minterms true for a = 1, in the upper half all minterms true for b = 1, in the
two middle columns all the minterms true for c = 1, and in the two middle lines all the minterms true for d = 1.
Results the lateral columns are adjacent and the lateral line are also adjacent. Actually the surface can be seen
as a toroid.
a
a
b
a
m3
m1
m2
m0
b
m6
m7
m3
m2
m4
m5
m1
m0
c.
b.
m4
m13 m15 m7
m5
m9 m11 m3
m1
m8 m10 m2
m0
d
c
a.
m12 m14 m6
b
c
Figure C.9: Veitch diagrams. The Veitch diagrams for 2, 3, and 4 variables.
Example C.6 Let be the function p1 and p2, two outputs of the 2-bit multiplier. Rewriting them using minterms
results::
p1 = m6 + m7 + m9 + m11 + m13 + m14
p2 = m10 + m11 + m14 .
In Figure C.10 p1 and p2 are represented.
⋄
C.4. MINIMIZING BOOLEAN FUNCTIONS
379
a
a
1
1
1
b
b
1
1
d
1
d
1
1
1
c
p1
c
p2
a.
b.
Figure C.10: Using Veitch diagrams. The Veitch diagrams for the functions p1 and p2.
The Karnaugh diagrams have the same property. The only difference is the way in which the minterms are
assigned to squares. For example, in a 4-input Karnaugh diagram each column is associated to a pair of input
variable and each line is associated with a pair containing the other variables. The columns are numbered in
Gray sequence (successive binary configurations are adjacent). The first column contains all minterms true for
ab = 00, the second column contains all minterms true for ab = 01, the third column contains all minterms true
for ab = 11, the last column contains all minterms true for ab = 10. A similar association is made for lines. The
Gray numbering provides a similar adjacency as in Veitch diagrams.
ab
cd
00
01
11
10
ab
00
01
11
10
00
m0
m1
m3
m2
0
m0
m1
m3
m2
01
m4
m5
m7
m6
1
m4
m5
m7
m6
11
m12 m13 m15 m14
10
m8
c
m9 m11 m10
Figure C.11: Karnaugh diagrams. The Karnaugh diagrams for 3 and 4 variables.
In Figure C.12 the same functions, p1 and p2, are represented. The distribution of the surface is different
but the degree of adjacency is identical.
ab
ab
00
cd
01
11
10
00
00
cd
01
11
10
00
01
1
11
1
10
1
1
p1
1
01
1
11
10
1
1
1
p2
Figure C.12: Using Karnaugh diagrams. The Karnaugh diagrams for the functions p1 and p2.
In the following we will use Veitch diagrams, but we will name the them V-K diagrams to be fair with both
Veitch and Karnaugh.
Minimizing with V-K diagrams
The rule to extract the minimized form of a function from a V-K diagram supposes:
• to define:
380
APPENDIX C. BOOLEAN FUNCTIONS
–
–
–
–
–
the smallest number
of rectangular surfaces containing only 1’s
including all the 1’s
each surface having a maximal area
and containing a power of two number of 1’s
• to extract the logic terms (logic product of Boolean variables) associated with each previously emphasized
surface
• to provide de minimized function adding logically (logical OR function) the terms associated with the
surfaces.
bcd′
acd′
a′ bc
a
a
◆
1
b
ac′ d
1
✲1
✮
1
❂
b
1
d
1
✶
1
d
✲1
ab′ c
1
ab′ d
p1
c
a.
p2
✴c
b.
Figure C.13: Minimizing with V-K diagrams. Minimizing the functions p1 and p2.
Example C.7 Let’s take the V-K diagrams from Figure C.10. In the V-K diagram for p1 there are four 2-square
surfaces. The upper horizontal surface is included in the upper half of V-K diagram where b = 1, it is also included
in the two middle columns where c = 1 and it is included in the surface formed by the two horizontal edges of the
diagram where d = 0. Therefore, the associated term is bcd′ which is true for: (b = 1)AN D(c = 1)AN D(d = 0).
Because the horizontal edges are considered adjacent, in the V-K diagram for p2 m14 and m10 are adjacent
forming a surface having acd′ as associated term.
The previously known form of p1 and p2 result if the terms resulting from the two diagrams are logically
added. ⋄
Minimizing incomplete defined functions
There are logic functions incompletely defined, which means for some binary input configurations the output value
does not matter. For example, the designer knows that some inputs do not occur anytime. This lack in definition
can be used to make an advanced minimization. In the V-K diagrams the corresponding minterms are marked as
“don’t care”s with “-”. When the surfaces are maximized the “don’t care”s can be used to increase the area of
1’s. Thus, some “don’t care”s will take the value 1 (those which are included in the surfaces of 1’s) and some of
“don’t care”s will take the value 0 (those which are not included in the surfaces of 1’s).
Example C.8 Let be the 4-input circuit receiving the binary codded decimals (from 0000 to 1001) indicating on
its output if the received number is contained in the interval [2, 7]. It is supposed the binary configurations from
1010 to 1111 are not applied on the input of the circuit. If by hazard the circuit receives a meaningless input we
do not care about the value generated by the circuit on its output.
In Figure C.14a the V-K diagram is presented for the version ignoring the “don’t care”s. Results the function:
y = a′ b + a′ c = a′ (b + c).
If “don’t care”s are considered results the V-K diagram from Figure C.14b. Now each of the two surfaces are
doubled resulting a more simplified form: y = b + c. ⋄
V-K diagrams with included functions
For various reasons in a V-K diagram we need to include instead of a logic value, 0 or 1, a logic function of variables
which are different from the variables associated with the V-K diagram. For example, a minterm depending on
a, b, c, d can be defined as taking a value which is depending on another logic 2-variable function by s, t.
A simplified rule to extract the minimized form of a function from a V-K diagram containing included functions
is the following:
C.5. PROBLEMS
381
a′ b
b
a
a
1
✌1
b
❫
-
1
1
-
-
1
1
-
✿
1
-
1
b
1
1
d
d
✐
1
1
c
′
a c
c
c
a.
b.
Figure C.14: Minimizing incomplete defined functions. a. The minimization of y (Example
1.8) ignoring the “don’t care” terms. b. The minimization of y (Example 1.8) considering the “don’t
care” terms.
1. consider first only the 1s from the diagram and the rest of the diagram filed only with 0s and extract the
resulting function
2. consider the 1s as “don’t care”s for surfaces containing the same function and extract the resulting function
“multiplying” the terms with the function
3. “add” the two functions.
bc’d
a
a
acde
a
a’bc’
c’de’
❯
1
b
1
e
1
b
✌1
1
1
1
1
a.
c
-
1
d
e′
-
1
b
❯
❯
e
-
d
e′
b
1
1
1
1
✣c
b’c
d
e′
c.
-
-
-
-
e′
c
Figure C.15: An example of V-K diagram with included functions. a. The initial form.
b. The form considered in the first step. c. The form considered in the second step.
Example C.9 Let be the function defined in Figure C.15a. The first step means to define the surfaces of 1s
ignoring the squares containing functions. In Figure C.15b are defined 3 surfaces which provide the first form
depending only by the variables a, b, c, d:
bc′ d + a′ bc′ + b′ c
The second step is based on the diagram represented in Figure C.15c, where a surface (c′ d) is defined for the
function e′ and a smaller one (acd) for the function e. Results:
c′ de′ + acde
In the third step the two forms are “added” resulting:
f (a, b, c, d, e) = bc′ d + a′ bc′ + b′ c + c′ de′ + acde.
⋄
Sometimes, an additional algebraic minimization is needed. But, it deserves because including functions in
V-K diagrams is a way to expand the number of variable of the functions represented with a manageable V-K
diagram.
C.5
Problems
Problem C.1
382
APPENDIX C. BOOLEAN FUNCTIONS
Appendix D
Basic circuits
Basic CMOS circuits implementing the main logic gates are described in this appendix. They are based on simple
switching circuits realized using MOS transistors. The inverting circuit consists in a pair of two complementary
transistors (see the third section). The main gates described are the NAND gate and the NOR gate. They are
built by appropriately connecting two pairs of complementary MOS transistors (see the fourth section). Tristate
buffers generate an additional, third “state” (the Hi-Z state) to the output of a logic circuit, when the output pair
of complementary MOS transistors are driven by appropriate signals (see the sixth section). Parallel connecting
a pair of complementary MOS transistors provides the transmission gate (see the seventh section).
D.1
Actual digital signals
The ideal logic signals are 0 Volts for false, or 0, and VDD for true, or 1. Real signals are more complex. The
first step in defining real parameters is represented in Figure D.1, where is defined the boundary between the
values interpreted as 0 and the values interpreted as 1.
VDD
VDD
v
✻
✻
Valid 1
”1”
VDD /2
❄
✻
”0”
❄
VHmin = VLmax
VDD /2
Valid 0
✲
0
time
Figure D.1: Defining 0-logic and 1-logic. The circuit is supposed to interpret any value under
VDD /2 as 0, and any value bigger than VDD /2 are interpreted as 1.
This first definition is impossible to be applied because supposes:
VHmin = VLmax .
There is no engineering method to apply the previous relation. A practical solution supposes:
VHmin > VLmax
generating a “forbidden region” for any actual logic signal. Results a more refined definition of the logic signals
represented in Figure D.2, where VL < VLmax and VH > VHmin .
In real applications we‘are faced with nasty realities. A signal generated to the output of a gate is sometimes
received to the input of the receiving gate distorted by parasitic signals. In Figure D.3 the noise generator simulate
the parasitic effects of the circuits switching in a small neighborhood.
Because of the noise captured from the “environment” a noise margin must be added to expand the forbidden
region with two noise margin regions, one for 0 level, N M0 , and another for 1 level, N M1 . They are defined as
follows:
383
384
APPENDIX D. BASIC CIRCUITS
VDD
v
✻
Valid 1
Forbidden
region
VHmin
VHmin
VLmax
VLmax
Valid 0
✲
time
0
Figure D.2: Defining the “forbidden region” for logic values. A robust design asks a net
distinction between the electrical values interpreted as 0 and the electrical values interpreted as 1.
-
+
Receiver
Sender
gate
gate
noise generator
VDD
v
✻
VOH
Valid 1
VOH
1 noise
margin
VIH
Forbidden
region
VIL
VOL
✲
VOL
time
0
0 noise
margin
VIH
VIL
Valid 0
Figure D.3: The noise margin. The output signal must be generated with more restrictions to allow
the receivers to “understand” correct input signals loaded with noise.
N M0 = VIL − VOL
N M1 = VOH − VIH
making the necessary distinctions between the VOH , the 1 at the output of the sender gate, and VIH , the 1 at the
input of the receiver gate.
D.2
CMOS switches
A logic gates consists in a network of interconnected switches implemented using the two type of MOS transistors:
p-MOS and n-MOS. How behaves the two type of transistors in specific configurations is presented in Figure D.4.
A switch connected to VDD is implemented using a p-MOS transistor. It is represented in Figure D.4a off
(generating z, which means Hi-Z: no signal, neither 0, nor 1) and in Figure D.4b it is represented on (generating
1 logic, or truth).
A switch connected to ground is implemented using a n-MOS transistor. It is represented in Figure D.4c off
(generating z, which means Hi-Z: no signal, neither 0, nor 1) and in Figure D.4e it is represented on (generating
0 logic, or false).
A MOS transistor works very well as an on-off switch connecting its drain to a certain potential. A p-MOS
transistor can be used to connect its drain to a high potential when its gates is connected to ground, and an
n-MOS transistor can connect its drain to ground if its gates is connected to a high potential. This complementary
behavior is used to build the elementary logic circuits.
D.3. THE INVERTER
385
VDD
VDD
z
z
0
VDD
VDD
VDD
z
✛
VDD
0
✛
✲
z
✲
VDD
a.
c.
b.
d.
Figure D.4: Basic switches. a. Open switch connected to VDD . b. Closed switch connected to VDD .
c. Open switch connected to ground. d. Closed switch connected to ground.
In Figure D.5 is presented the switch-resistor-capacitor model (SRC). If VGS < VT then the transistor is off,
if VGS ≥ VT then the transistor is on. In both cases the input of the transistor behaves like a capacitor, the
gate-source capacitor CGS .
When the transistor is on its drain-source resistance is:
RON = Rn
L
W
where: L is the channel length, W is the channel width, and Rn is the resistance per square. The length L is a
constant characterizing a certain technology. For example, if L = 0.13µm this means it is about a 0.13µm process.
The input capacitor has the value:
ϵOX LW
.
CGS =
d
The value:
ϵOX
COX =
d
where: ϵOX ≈ 3.9ϵ0 is the permittivity of the silicon dioxide, is the gate-to-channel capacitance per unit area of
the MOSFET gate.
In this conditions the gate input current is:
iG = CGS
dvGS
dt
Example D.1 For an AND gate with low strength, with W = 1.8µm, in 0.13µm technology, supposing COX =
4f F/µm2 , results the input capacitance:
CGS = 4 × 0.13 × 1.8f F = 0.936f F
Assuming Rn = 5KΩ, results for the same gate:
RON = 5 ×
0.13
KΩ = 361Ω
1.8
⋄
D.3
D.3.1
The Inverter
The static behavior
The smallest and simplest logic circuit – the invertor – can be built using a pair of complementary transistors,
connecting together the two gates as input and the two drains as output, while the n-MOS source is connected
386
APPENDIX D. BASIC CIRCUITS
D
D
D
G
RON
✲
G
G
CGS
CGS
S
S
VGS < VT
S
VGS ≥ VT
Figure D.5: The MOSFET switch. The switch-resistor-capacitor model consists in the two states:
OF (VGS < VT ), and ON (VGS ≥ VT ). In both states the input is defined by the capacitor CGS .
to ground (interpreted as logic 0) and the p-MOS source to VDD (interpreted as logic 1). Results the circuit
represented in Figure D.6.
The behavior of the invertor consist in combining the behaviors of the two switches previously defined. For
in = 0 pMOS is on and nMOS is of f the output generating VDD which means 1. For in = 1 pMOS is of f and
nMOS is on the output generating 0.
The static behavior of the inverter (or NOT) circuit can be easy explained starting from the switches described
in Figure D.4. Connecting together a switch generating z with a switch generating 1 or 0, the connection point
will generate 0 or 1.
D.3.2
Dynamic behavior
The propagation time of an inverter can be analyzed using the two serially connected invertors represented in
Figure D.7. The delay of the first invertor is generated by its capacitive load, CL , composed by:
• its parasitic drain/bulk capacitance, CDB , is the intrinsic output capacitance of the first invertor
• wiring capacitance, Cwire , which depends on the length of the wire (of width Ww and of length Lw )
connected between the two invertors:
Cwire = Cthickox Ww Lw
• next stage input capacitance, CG , approximated by summing the gate capacitance for pMOC and nMOS
transistors:
CG = CGp + CGn = Cox (Wp Lp + Wn Ln )
The total load capacitance
CL = CDB + Cwire + CG
is sometimes dominated by Cwire . For short connections CG dominates, while for big fan-out both, Cwire and
CG must be considered.
The signal VA is used to measure the propagation time of the first NOT in Figure D.7a. It is generated by
an ideal pulse generator with output impedance 0. Thus, the rising time and the falling time of this signal are
considered 0 (the input capacitance of the NOT circuit is charged or discharged in no time).
The two delay times (see Figure D.7c) associated to an invertor (to a gate in the general case) are defined as
follows:
• tpLH : the time interval between the moment the input switches in 0 and the output reaches VOH /2 coming
from 0
• tpHL : the time interval between the moment the input switches in 1 and the output reaches VOH /2 coming
from VOH
Let us consider the transition of VA from 0 to VOH at tr (rise edge). Before transition, at t−
r , CL is fully
charged and VB = VOH . In Figure D.7b is represented the equivalent circuit at t+
r , when pMOS is off and nMOS
is on. In this moment starts the process of discharging the capacitance CL at the constant current
IDn(sat) =
Wn
1
µn Cox
(VOH − VT n )2
2
Ln
D.3. THE INVERTER
pMOS
lin
lin
sat
cut
nMOS
cut
sat
lin
lin
B
C
D
Vout
VDD
VDD
✛
X
387
✻A
X
X’
X’
✲
✲
VT n
VDD
VDD /2
VDD − |VT p |
a.
Vin
c.
b.
Figure D.6: Building an invertor. a. The invertor circuit. b. The logic symbol for the invertor
circuit.
A
VA
B
✻
VOH
✲
tr
a.
i(t)
✛
VB
v(t)
+
VDD
-
VOH /2
CL
✲ ✛
b.
time
✻
VOH
RON n
tf
CGSn + CGSp
c.
tpHL
✲
✲ ✛ time
tpLH
Figure D.7: The propagation time.
+
In Figure D.8, at t−
r the transistor is cut, IDn = 0. At tr the nMOS transistor switch in saturation and becomes
an ideal constant current generator which starts to discharge CL linearly at the constant current IDn(sat) . The
process continue until VOU T = VOH , according to the definition of tpHL .
In order to compute tpHL we take into consideration the constant value of the discharging current which
provide a linear variation of vOU T .
−IDn(sat)
d qL
dvout
= (
)=
dt
dt CL
CL
dvout
=
dt
VOH
2
− VOH
tpHL
We solve the equations for tpHL :
tpHL = CL
1
VOH
n
V
− VT n
µn Cox W
(V
−
V
)
OH
OH
Tn
Ln
Because:
RON n =
1
n
µn Cox W
(VOH
Ln
− VT n )
388
APPENDIX D. BASIC CIRCUITS
tr + tpHL
t−
r
ID
IDn
✻
s
✢
VIN = VOH
t−
r
✢✲V
VIN = 0
VOH
VOH /2
OU T
Figure D.8: The output characteristic of the nMOS transistor.
results:
tpHL = CL RON n
where:
1
1−
VT n
VOH
= kn RON n CL = kn τnL
• τnL is the constant time associated to the H-L transition
• kn is a constant associated to the technology we use; it goes down when VOH increases or VT decreases
The speed of a gate depends by its dimension and by the capacitive load it drives. For a big W the value of
RON is small charging or discharging CL faster.
For tpLH the approach is similar. Results: tpLH = kp τpL .
By definition the propagation time associated to a circuit is:
tp = (tpLH + tpHL )/2
its value being dominated by the value of CL and the size (width) of the two transistors, Wn and Wp .
D.3.3
Buffering
It is usual to be confronted, in designing a big systems, with the buffering problem: a logic signal generated by a
small, “weak” driver must be used to drive a big, “strong” circuit (see Figure D.9a) maintaining in the same time
a high clock frequency. The driver is an invertor with a small Wn = Wp = Wdrive (to make the model simple),
unable to provide an enough small RON to move fast the charge from the load capacitance of a circuit with a big
Wn = Wp = Wload . Therefore the delay introduced between A and B is very big. For our simplified model,
tp = tp0
Wload
Wdriver
where: tp0 is the propagation time when the driver circuit and the load circuit are of the same size.
The solution is to interpose, between the small driver and the big load, additional drivers with progressively
increased area as in Figure D.9b. The logic is preserved, because two NOTs are serially connected. While the
no-buffer solution provides, between A and B, the propagation time:
tp(no−buf f er) = tp0
Wload
Wdriver
the buffered solution provide the propagation time:
tp(buf f ered) = tp0 (
W2
Wload
W1
+
+
)
Wdriver
W1
W2
How are related the area of the circuits in order to obtain a minimal delay, i.e., how are related Wdriver , W1 and
W2 ? The relation is given by the minimizing the delay introduced by the two intermediary circuits. Then, the
first derivative of
Wload
W2
+
W1
W2
D.3. THE INVERTER
389
must be 0. Results:
W2 =
√
W1 Wload
Wload
W2
=
=
W1
W2
√
Wload
W1
We conclude: in order to add a minimal delay, the size ratio of successive drivers in a chain must be the same.
Wload
Wdriver
Wdriver
A
B
Wload
W2
W1
A
B
a.
b.
Figure D.9: Buffered connection. a. An invertor with small W is not able to handle at high
frequency a circuit with big W . b. The buffered connection with two intermediary buffers.
In order to design the size of the circuits in Figure D.9b, let us consider
W1
Wdriver
=
= n. Then,
√
Wload
W2
=
= 3n
W1
W2
The acceleration is
α=
Wload
Wdriver
tp(no−buf f er)
=
tp(buf f ered)
√
3
n2
3
For example, for n = 1000 the acceleration is α = 33.3.
The hand calculation, just presented, is approximative, but has the advantages to provide an intuitive understanding about the propagation phenomenon, with emphasis on the buffering mechanism.
The price for the acceleration obtained by buffering is the area and energy consumed by the two additional
circuits.
D.3.4
Power dissipation
There are three major physical processes involved in the energy requested by a digital circuit to work:
• switching energy: due to charging and discharging of load capacitances, CL
• short-circuit energy: due to non-zero rise/fall times of the signals
• leakage current energy: which becomes more and more important with the decreasing of device sizes
From the power supply, which provide VDD with enough current, the circuit absorbs as much as needed current.
Switching power
The average switching power dissipated is the energy dissipated in a clock cycle divided by the clock cycle time,
T . Suppose the clock is applied to the input of an invertor. When clock = 0 the load capacitor is loaded from
the power supply with the charge:
QL = CL VDD
We assume in T /2 the capacitor is charged (else the frequency is too big for the investigated circuit). During
the next half-period, when clock = 1, the same charge is transferred from the capacitor to ground. Therefore
the charge QL is transferred from VDD to ground in the time T . The amount of energy used for this transfer is
VDD QL , and the switching power results:
pswitch =
VDD CL VDD
2
fclock
= CL VDD
T
While a big VOH = VDD helped us in reducing tp , now we have difficulties due to the square dependency of
switching power by the same VDD .
390
APPENDIX D. BASIC CIRCUITS
VDD
RON p
vout
i
vout
i
❄
RON n
CL
❄
CL
Figure D.10: The main power consuming process. For Vin = 0 CL is loaded by the current
provided by RON p . The charge from CL is transferred to the ground through RON n for Vin = VOH .
VIN
✻
✲
time
IDD
✻
IDD(mean)
✲
time
Figure D.11: Direct flow of current from VDD to ground. This current due to the non-zero
edge to the circuit input can be neglected.
Short-circuit power
When the output of the invertor switches between the two logic levels, for a very short time interval around
the moment when VOU T = VDD /2, both transistors have IDD ̸= 0 (see Figure D.11). Thus is consumed the
short-circuit power.
The amount of power wasted by these temporary short-cuts is:
psc = IDD(mean) VDD
where IDD(mean) is the mean value of the current spikes. If the edge of the signal is short and the mean frequency
of switchings is low, then the resulting value is low.
Leakage power
The last source of energy waste is generated by the leakage current. It will start to be very important in sub 65nm
technologies (for 65nm the leakage power is 40% of the total power consumption). The leakage current and the
associated power is increasing exponentially with each new technology generation and is expected to become the
dominant part of total power. Device threshold voltage scaling, shrinking device dimensions, and larger circuit
sizes are causing this dramatic increase in leakage. Thus, increasing the amount of leakage is critical for power
constraint integrated circuits.
pleakage = Ileakage VDD
where Ileakage is the sum of subthreshold and gate oxide leakage current. In Figure D.12 the two components of
the leakage current are presented for a NOT circuit with Vin = 0.
D.4
Gates
The 2-input AND circuit, a · b, works like a “gate” opened by the signal a for the signal b. Indeed, the gate is
“open” for b only if a = 1. This is the reason for which the AND circuit was baptised gate. Then, the use imposed
D.4. GATES
391
VDD
✛
✻
Vout = VDD
✲
✸✲
⑥
❄
Diode leakage
❄
✻
Gate leakage
Sub-threshold leakage
Figure D.12: The two main components of the leakage current. .
this alias as the generic name for any logic circuit. Thus, AND, OR, XOR, NAND, ... are all called gates.
D.4.1
NAND & NOR gates
The static behavior of gates
For 2-input NAND and 2-input NOR gates the same principle will be applied, interconnecting 2 pairs of complementary transistors to obtain the needed behaviors.
There are two kind of interconnecting rules for the same type of transistors, p-MOS or n-MOS. They can be
interconnected serially or parallel.
A serial connection will establish an on configuration only if both transistors of the same type are on, and
the connection is off if at least one transistor is off.
A parallel connection will establish an on configuration if at least one is on, and the connection is off only if
both are off.
Applying the previous rules result the circuits presented in Figures D.13 and D.14.
VDD
✛
✲
✲
A
(AB)’
A
(AB)’
B
✲
✛
B
a.
b.
Figure D.13: The NAND gate. a. The internal structure of a NAND gate: the output is 1 when
at least one input is 0. b. The logic symbol for NAND.
For the NAND gate the output is 0 if both n-MOS transistors are on, and the output is one when at least
on p-MOS transistor is on. Indeed, if A = B = 1 both n transistors are on and both p transistors are off. The
output corresponds with the definition, it is 0. If A = 0 or B = 0 the output is 1, because at least one p transistor
is on and at least one n transistor is off.
A similar explanation works for the NOR gate. The main idea is to design a gate so as to avoid the simultaneous connection of VDD and ground potential to the output of the gate.
392
APPENDIX D. BASIC CIRCUITS
VDD
✛
✲
✲
A
(A+B)’
A
✲
✛
(A+B)’
B
B
a.
b.
Figure D.14: The NOR gate. a. The internal structure of a NOR gate: the output is 1 only when
both inputs are 0. b. The logic symbol for NOR.
For designing an AND or an OR gate we will use an additional NOT connected to the output of an AND or
an OR gate. The area will be a little bigger (maybe!), but the strength of the circuit will be increased because
the NOT circuit works as a buffer improving the time performance of the non-inverting gate.
The propagation time for the 2-input main gates is computed in a similar way as the propagation for NOT
circuit is computed. The only differences are due to the fact that sometimes RON must be substituted with
2 × RON .
Propagation time
Propagation time for NAND gate becomes, in the worst case when only one input switches:
tHL = kn (2RON n )CL
tLH = kp (RON p )CL
because the capacitor CL is charged through one pMOS transistor and is discharged through two, serially connected, nMOS transistors.
Propagation time for NOR gate becomes, in the worst case when only one input switches:
tHL = kn (RON n )CL
tLH = kp (2RON p )CL
because the capacitor CL is charged through two, serially connected, pMOS transistors and is discharged through
one nMOS transistor.
It is obvious that we must prefer, when is is possible, the use of NAND gates instead of NOR gates, because,
for the same area, RON p > RON n .
Power consumption & switching activity
The power consumption is determined by the 0 to 1 transitions of the output of a logic gates. The problem is
meaningless for a NOT circuit because the transitions of the output has the same probability as of the transition
of the input. But, for a n-input gate the probability of an output transition depends on the function performed
by the gate.
For a certain gate, with unbiased 0 and 1 applied on the inputs, the output probability of switching from 0
to 1, P0−1 , is given by the logic function. We define switching activity, σ, this probability of switching from 0
to 1.
D.4. GATES
393
Switching activity for 2-input AND with the inputs A and B is:
σ = P0−1 = POU T =0 POU T =1 = PA PB (1 − PA PB )
where: PA is the probability of having 1 on the input A, PB is the probability of having 1 on the input B, and
POU T =0 is the probability of having 0 on output, while POU T =1 = PAB is the probability of having 1 on output
(see Figure D.15a).
σ = 3/16
PA = 1/2
A
σ = 7/64
PAB = 1/4
PABC = 1/8
A
B
PB = 1/2
B
C
PC = 1/2
a.
σ = 15/256
PABCD = 1/16
A
B
C
D
PD = 1/2
c.
b.
Figure D.15: Switching activity σ and the output probability of 1. a. For 2-input AND.
b. For 3-input AND. c. For 4-input AND.
If the input are not conditioned, PA = PB = 0.5, then the switching activity for a 2-input NAND is σN AN D2 =
3/16 (see Figure D.15a).
Switching activity for 3-input AND with the inputs A, B, and C is σN AN D3 = 7/64 (see Figure
D.158). The probability of 1 to the output of a 3-input AND is only 1/8 leading to a smaller σ.
Switching activity for n-input AND is:
σN AN Dn =
1
2n − 1
≃ n
22n
2
The switching activity decreases exponentially with the number of inputs in AND, OR, NAND, NOR gates.
This is a very good news.
Now, we must reconsider the computation of the power substituting CL with σCL :
2
pswitch = σCL VDD
fclock
In big systems, a conservative assumption is that the mean value of the inputs of the logic gates is 3, and,
therefore a global value for switching activity could be σglobal ≃ 1/8. Actual measurements provide frequently
σglobal ≃ 1/10.
Power consumption & glitching
In the previous paragraph we learned that the output of a circuit switch due to the change on the inputs. This
is an ideal situation. Depending on the circuit configuration and on the various delays introduced by gates,
unexpected “activity” manifests sometimes in our network of gates. See the simple example form Figure D.16.
From the logical point of view, when the inputs switch form ABC = 010 to ABC = 111 the output must maintain
its value on 1. Unfortunately, because the effect of the inputs A and B are affected by the extra delay introduced
by the first gate, the unexpected glitch manifests to the output. Follow the wave forms form Figure D.16 to
understand why.
The glitch is undesired for various reasons. The most important are two:
• the signal can be latched by a memory circuit (such an elementary latch), thus triggering the switch of a
memory circuit; a careful design can avoid this effect
• the temporary, useless transition discharge and charge back the load capacitor increasing the energy consumed by the circuit.
Let us go back to the Zero circuit represented in two versions in Figure 2.1c and Figure 2.1d. We have now
an additional reason to prefer the second version. The balanced delays to the inputs of the intermediary circuits
allow us to avoid almost totaly the glitching contribution to the power consumption.
394
APPENDIX D. BASIC CIRCUITS
A=C
✻
✲
O1
O1
A
B=1
O2
✲
✛
t
tpHLO1
✻
C
O2
✻
✲
✲
✛
✛
✲
t
tpLHO2
tpHLO2
✲
t
Figure D.16: Glitching effect. When the input value switch from ABC = 010 to ABC = 111
the output of the circuit must remain on 1. But, a short glitch occurs because of the delay, tpHLO1 ,
introduced by the first NAND.
D.4.2
Many-Input Gates
How can be built 3-input NAND or a 3-input NOR applying the same rule? For a 3-input NAND 3 n-MOS
transistors will be connected serially and 3 p-MOS transistors will be connected parallel. Similar for the 3-input
NOR gate.
How “much” this rule can be applied to built n-input gates? Not too much because of the propagation time
which is increased when too many serially connected RON resistors will be used to transfer the electrical charge
in or out from the load capacitor CL . A 4-input NAND, for example, discharge CL trough 4 serially connected
RON n , while a 4-input NOR loads CL with a constant time 4RON p CL . The mean worst case (when only one
input switches) time constants used to compute tp become:
(4RON n + RON p )/2
for NAND, and
(4RON p + RON n )/2
for NOR.
Fortunately, there is another way to increase the number of inputs of a certain gate. It is by composing the
function using an appropriate number of 2-input gates organized as a balanced binary tree.
For example, an 8-input NAND gate, see Figure D.17a, is recommended to be designed as a binary tree of
two input gates, see Figure D.17b, as follows:
(a · b · c · d · e · f · g · h)′ = (((a · b)′ + (c · d)′ )′ · ((e · f )′ + (g · h)′ )′ )′
The form results as the application of the De Morgan law.
In the first case, represented in Figure D.17a, an 8-input NAND uses a similar arrangement as in Figure
D.13a, where instead of two parallel connected pMOS transistors and two serially connected nMOS transistors
are used 8 pMOSs and 8 nMOSs. Generally speaking, for each new input an additional pair, nMOS & pMOS, is
added.
Increasing in this way the number of inputs the propagation time is increased linearly because of the serially
connected channels of the nMOS transistors. The load capacitor is discharged to the ground through m × RON ,
where m represents the number of inputs.
The second solution, see Figure D.17b, is to build a balanced tree of gates. In the first case the propagation
time is in O(n), while in the second it is in O(log n) for implementations using transistors having the same size.
For an m-input gate results a log2 m depth network of 2-input gates. For example, see Figure D.17, where
an 8-input NAND is implemented using a 3-level network of gates (first to the 8-input gate the divide & impera
principle is applied, and then the De Morgan rule transformed the first level of four ANDs in four NANDs and
the second level of two ANDs in two NORs). While the maximum propagation time for the 8-input NAND is
tpHL(one−level) = kn × 8 × RON n × (n × Cin )
where Cin is the value of the input capacitor in a typical gate and n is the fan-out of the circuit, the maximum
propagation time for the equivalent log-depth net of gates is
tpHL(log−levels) = kn ((2 × 2 × RON n × Cin ) + 2 × RON n × (n × Cin ))
D.4. GATES
395
..
.. Cin
.
..
..
.
..
.. Cin
.
..
..
.
..
.. n × Cin
.
..
..
.
..
.. Cin
.
..
..
.
..
.. Cin
.
..
..
.
a.
..
.. Cin
.
..
..
.
..
.. n × Cin
.
..
..
.
..
.. Cin
.
..
..
.
b.
Figure D.17: How to manage a many-input gate. a An N AN D8 gate with fan-out n. b. The
log-depth equivalent circuit.
For n = 3 results a 2.4 times faster circuit if the log-depth version is adopted, while for n = 4 the acceleration is
2.67.
Generally, for fan-in equal with m and fan-out equal with n result the acceleration for the log-depth solutions,
α, expressed by the formula:
m×n
α=
2 × (n − 1 + log m)
Example: n = 4, m = 32, α = 8.
The log-depth circuit has two advantages:
• the intermediary (−1 + log m) stages are loaded with a constant and minimal capacitor – Cin – given by
only one input
• only the final stage drives the real load of the circuit – n × Cin – but its driving capability does not depend
by fan-in.
Various other solutions can be used to speed-up a many-input gate. For example:
(a · b · c · d · e · f · g · h)′ = (((a · b · c · d)′ + (e · f · g · h)′ )′ )′
could be a better solution for an 8-input NAND, mainly because the output is generated by a NOT circuit and
the internal capacitors are minimal, making the 4-input NANDs harmless.
D.4.3
AND-NOR gates
For implementing the logic function:
(AB + CD)′
besides the solution of composing it from the previously described circuits, there is a direct solution using 4 CMOS
pairs of transistors, one associated for each input. The resulting circuit is represented in Figure D.18.
The size of the circuit according to Definition 2.2 is 4. (Implementing the function using 2 NANDs, 2 invertors,
and a NOR provides the size 8. Even if the de Morgan rule is applied results 3 NANDs and and invertor, which
means the size is 7.)
The same rule can be applied for implementing any NOR of ANDs. For example, the circuit performing the
logic function
f (A, B, C) = (A(B + C))′
has a simple implementation using a similar approach. The price will be the limited speed or the over-dimensioned
transistors.
396
APPENDIX D. BASIC CIRCUITS
VDD
✛
✲
C
D
✛
✲
A
A
B
✲(AB + CD)’
A
B
(AB + CD)’
C
✲
✛
C
✲
✛
D
D
b.
B
a.
Figure D.18: The AND-NOR gate. a. The circuit. b. The logic symbol for the AND-NOR gate.
D.5
The Tristate Buffers
A tristate circuit has the output able to generate three values: 0, 1, x (which means nothing). The output value
x is unable to impose a specific value, we say the output of the circuit is unconnected or it is off.
Two versions of this kind of circuit are presented in Figures D.19 and D.20.
The inverting version of the tristate buffer uses one additional pair of complementary transistors to disconnect
the output from any potential. If enable = 0 the CMOS transistors connected to the output are both off. Only
if enable = 1 the circuit works as an inverter.
For the non-inverting version the two additional logic gates are used to control the gates of the two output
transistors. Only if enable = 0 the two logic gates transfer the input signal inverted to the gates of the two output
transistor.
D.6
The Transmission Gate
A simple and small version of a gate is the transmission gate which works connecting directly the signal from
a source to a destination. Figure D.21a represents the CMOS version. If enable = 1 then out = in because at
least on transistors is on. If in = 0 the signal is transmitted by the n-MOS transistor, else, if in = 1 the signal is
transmitted by the p-MOS transistor.
The transmission gate is not a regenerative gate in contrast to the previously described gates which were
regenerative gates. A transmission gate performs a true two-direction electrical connection, with all its goods and
bad involved.
The main limitation introduced by the transmission gate is its RON which is serially connected to the CL
increasing the constant time associated to the delay.
The main advantage of this gate is the absence of a connection to the ground or to VDD . Thus, the energy
consumed by this gate is lowered.
One of the frequently used application of the transmission gate is the inverting multiplexor (see Figure D.21c).
The two transmission gates are enabled by in a complementary mode. Thus, only one gate is active at a time,
avoiding the “fight” of two opposite signals to impose the value to the inverter’s input.
When the propagation time is not critical the use of this gate is recommended because, both, area and power
are saved.
D.7. MEMORY CIRCUITS
397
VDD
✛
enable’
✛
enable’
out
enable’
✲
out
in
enable
out
enable
✲
in
✲
a.
enable
b.
✛ ✲
in/out
in
c.
en1’
System 1
en2’
en1
System 2
en2
d.
Figure D.19: Tristate inverting buffer. a. The circuit. b. The logic symbol for the inverting
tristate buffer. c. Two-direction connection on one wire. For enable = 1, in/out = out’, while for
enable = 0, in = in/out’. d. Interconnecting two systems. For en1 = 1, en2 = 0, System
1 sends and System 2 receives; for en1 = 0, en2 = 1, System 2 sends and System 1 receives;
en1 = en2 = 0 booth systems are receivers, while en1 = en2 = 1 is not allowed.
D.7
D.7.1
Memory Circuits
Flip-flops
Data latches and their transparency
Master-slave DF-F
Resetable DF-F
D.7.2
# Static memory cell
D.7.3
# Array of cells
D.7.4
# Dynamic memory cell
D.8
Problems
Gates
Problem D.1
Problem D.2
Problem D.3
398
APPENDIX D. BASIC CIRCUITS
VDD
✛
in
enable’
✲
out
in
out
enable’
✲
a.
b.
Figure D.20: Tristate non-inverting buffer. a. The circuit. b. The logic symbol for the noninverting tristate buffer.
en’
en’
in1
en’
✻
en
out
in
out
out
in
en
✻
en
in0
en
a.
b.
en’
c.
Figure D.21: The transmission gate. a. The complementary transmission gate. b. The logic
symbol. c. An application: the elementary inverting multiplexer.
Problem D.4
Flop-flops
Problem D.5
Problem D.6
Problem D.7
Problem D.8
D.8. PROBLEMS
399
CK
D
D CK
Q
Q
Figure D.22: Data latches. a. Transparent from D to Q (D = Q) for ck = 0. For ck = 1 the loop
is closed and D input has no effect on output. b. Transparent from D to Q for ck = 1. For ck = 0 the
loop is closed and D input has no effect on output.
ck
Slave inverting MUX
TG3
ck
Master inverting MUX
TG1
Q
ck’
D
ck’
TG4
TG2
SiMUX
ck
MiMUX
ck’
ck
clock
ck
a.
TG3
TG1
TG4
Q
D
TG2
b.
TG3
TG1
TG4
Q
D
TG2
c.
Figure D.23: Master-slave delay flip-flop (DF-F) with the clock signal active on the
negative transition. a. Implemented with data latches based on transmission gates. b. The
equivalent schematic for ck = 1. c. The equivalent schematic for ck = 0.
400
APPENDIX D. BASIC CIRCUITS
RST’
D
SiMUX
Q
MiMUX
RST
Figure D.24: Master-slave delay flip-flop with asynchronous reset.
Appendix E
Standard cell libraries
Modern design procedures suppose standard cell libraries containing gates, latches, flip-flops, and many other
frequently used typical modules.
E.1
Main parameters
Each component of a cell library is characterized for area, propagation time, and power1 .
Each factory has three type of processes: slow, typical, and fast. A derating factor is associated with each
process:
Kprocess (typical) = 1
Kprocess (slow) = 1.3
Kprocess (f ast) = 0.8
The power supply variation introduces two derating factors, one for the negative variations and another for
the positive variation:
Kvolt (−10%) = −2.1/V
Kvolt (+10%) = −1.4/V
The temperature variation introduce the last two derating factors:
Ktemp (f rom 250 C to − 400 C) = 0.00090 C
Ktemp (f rom 250 C to 1250 C) = 0.00070 C
Figure E.1 represents how the propagation time, tp , is measured in combinational circuits.
input
1
✻
0.5
✲
time
output
1
✻
✛ pLH ✲
✛ pHL ✲
t
t
0.5
✲
time
inverted output
1
✻
✛ pHL✲
✛
t
0.5
tpLH
✲
✲
time
Figure E.1: Propagation time in combinational circuits.
1
The values used in the following are typical values for a 13µm process in a good factory.
401
402
APPENDIX E. STANDARD CELL LIBRARIES
Figure E.2 represent temporal relations in a clocked circuit, where the active edge of clock is the positive
edge. The frequency of clock is:
1
fck =
tckw1 + tckw0
The following parameters are defined:
set-up time : tSU because a clocked circuit must receive the input data before the active edge of clock (the
circuit must “learn” before what to do)
hold time : tH because the input data can not disappears immediately after the active edge of clock (the circuit
must “understand” firmly what to do)
recovery time : tR is necessary to change the control of the circuit from the imperium of set/reset devices to
the imperium of clock (the set/reset operation is very “stressful” and the circuits must recover)
propagation time : tpHL or tpLH is the response time of the output of the circuit.
clock
1
✻
✛
tckw1
✲✛tckw0✲
0.5
✲
time
tSU
✛
✲✛tH✲
data input
1
✻
0.5
✲
time
✛
set or reset
1
✻
tR
✲
0.5
✲
time
output
1
✻
✛pHL✲
t
✛
✲
tpLH
0.5
✲
time
Figure E.2: Time relations in a clocked system.
The propagation time for a circuit can be estimated using the following forms:
tp = Kprocess × (1 + Kvolt × ∆VDD ) × (1 + Ktemp × ∆T ) × ttypical
ttypical = tintrinsic + Kload × Cload
where:
• tp – propagation time in ns taking into account all derating factors
• ttypical – propagation time in ns at nominal VDD ,, 250 C, for typical process
• tintrinsic – circuit delay in ns with no loaded output
• Kload – load factor expressed in ns/pF
• Cload – output load expressed in pF .
E.2
Basic cells
A standard library contains cells having the same height. An usual value is around 3.7µm. The differences
between inputs are ignored. The power is expressed for each input pin in µW/M Hz. The tables contain mean
typical values.
E.2. BASIC CELLS
E.2.1
403
Gates
Two input AND gate: AND2
Implementation : y = ((a · b)′ )′
Strength
1
2
4
Width(µm)
1.8
1.8
2.8
tintrinsic (LH/HL)
0.053 / 0.073
0.047 / 0.069
0.044 / 0.057
Zload (LH/HL)
4.28 / 2.2
1.9 / 1.1
0.96 / 0.56
Cpin
0.0014
0.002
0.004
Power
0.0044
0.006
0.012
tintrinsic (LH/HL)
0.045 / 0.095
0.04 / 0.085
0.04 / 0.075
Zload (LH/HL)
4.2 / 2.3
1.9 / 1.25
1.1 / 0.6
Cpin
0.0015
0.0025
0.004
Power
0.004
0.006
0.01
Zload (LH/HL)
4.3 / 2.2
1.9 / 1.2
0.97 / 0.56
Cpin
0.0016
0.0023
0.0042
Power
0.0047
0.0075
0.013
Zload (LH/HL)
4.3 / 2.5
1.9 / 1.2
1 / 0.6
Cpin
0.0017
0.0028
0.0054
Power
0.005
0.008
0.015
Zload (LH/HL)
4.3 / 2.2
1.9 / 1.1
0.97 / 0.56
Cpin
0.0017
0.0027
0.0053
Power
0.0051
0.009
0.016
Zload (LH/HL)
4.3 / 2.8
2 / 1.3
1 / 0.65
Cpin
0.002
0.003
0.006
Power
0.0055
0.009
0.016
Two input OR gate: OR2
Implementation : y = ((a + b)′ )′
Strength
1
2
4
Width(µm)
1.8
2.3
2.8
Three input AND gate: AND3
Implementation : y = ((a · b · c)′ )′
Strength
1
2
4
Width(µm)
2.3
2.3
3.2
tintrinsic (LH/HL)
0.073 / 0.090
0.061 / 0.086
0.065 / 0.067
Three input OR gate: OR3
Implementation : y = ((a + b + c)′ )′
Strength
1
2
4
Width(µm)
2.75
2.75
3.68
tintrinsic (LH/HL)
0.05 / 0.14
0.05 / 0.1
0.045 / 0.1
Four input AND gate: AND4
Implementation : y = ((a · b · c · d)′ )′
Strength
1
2
4
Width(µm)
2.8
3.2
4.6
tintrinsic (LH/HL)
0.082 / 0.097
0.075 / 0.092
0.073 / 0.082
Four input OR gate: OR4
Implementation : y = ((a + b + c + d)′ )′
Strength
1
2
4
Width(µm)
2.8
3.2
4.6
tintrinsic (LH/HL)
0.06/ 0.17
0.05 / 0.13
0.05 / 0.12
404
APPENDIX E. STANDARD CELL LIBRARIES
Four input AND-OR gate: ANDOR22
Implementation : y = ((a · b + c · d)′ )′
Strength
1
2
4
Width(µm)
2.8
3.2
4.2
tintrinsic (LH/HL)
0.07 / 0.14
0.065 / 0.13
0.06 / 0.1
Zload (LH/HL)
4.3 / 2.4
1.9 / 1.3
1 / 0.6
Cpin
0.0017
0.0027
0.0053
Power
0.0056
0.0094
0.016
tintrinsic (LH/HL)
0.015 / 0.014
0.014 / 0.0135
0.013 / 0.014
0.013 / 0.013
0.0125 / 0.013
Zload (LH/HL)
2.9 / 2.7
1.9 / 1.9
0.95 / 1
0.48 / 0.48
0.12 / 0.13
Cpin
0.0017
0.0027
0.0053
0.0053
0.0053
Power
0.0026
0.0037
0.0067
0.0137
0.053
tintrinsic (LH/HL)
0.027 / 0.016
0.013 / 0.025
0.013 / 0.025
Zload (LH/HL)
4.2 / 2.4
1.9 / 1.2
1 / 0.7
Cpin
0.003
0.0054
0.01
Power
0.0028
0.0053
0.0095
tintrinsic (LH/HL)
0.026 / 0.014
0.025 / 0.0135
0.023 / 0.0134
Zload (LH/HL)
6.1 / 2
3 / 1
1.5 / 0.5
Cpin
0.003
0.0054
0.01
Power
0.003
0.0055
0.01
Zload (LH/HL)
4.25 / 2.9
2 / 1.5
1.1 / 0.85
Cpin
0.003
0.0061
0.0105
Power
0.0048
0.0086
0.015
Zload (LH/HL)
7.7 / 2.3
3.8 / 1.2
1.7 / 0.6
Cpin
0.0032
0.0063
0.0133
Power
0.0058
0.0095
0.0195
Invertor: NOT
Implementation : y = a′
Strength
1
2
4
8
16
Width(µm)
0.92
1.38
1.84
2.76
7.36
Two input NAND gate: NAND2
Implementation : y = (a · b)′
Strength
1
2
4
Width(µm)
1.38
2.3
3.2
Two input NOR gate: NOR2
Implementation : y = (a + b)′
Strength
1
2
4
Width(µm)
1.38
2.3
3.2
Three input NAND gate: NAND3
Implementation : y = (a · b · c)′
Strength
1
2
4
Width(µm)
1.84
3.22
4.6
tintrinsic (LH/HL)
0.035 / 0.023
0.03 / 0.021
0.031 / 0.022
Three input NOR gate: NOR3
Implementation : y = (a + b + c)′
Strength
1
2
4
Width(µm)
1.84
2.76
2.52
tintrinsic (LH/HL)
0.056 / 0.022
0.046 / 0.021
0.042 / 0.021
E.2. BASIC CELLS
405
Four input NAND gate: NAND4
Implementation : y = (a · b · c · d)′
Strength
1
2
4
Width(µm)
2.3
3.7
7.35
tintrinsic (LH/HL)
0.04 / 0.03
0.035 / 0.025
0.034 / 0.024
Zload (LH/HL)
4.26 / 3.4
2.2 / 2
1 / 0.85
Cpin
0.0033
0.0056
0.0127
Power
0.0055
0.009
0.02
Zload (LH/HL)
8 / 2
4.5 / 1.1
2.4 / 0.6
Cpin
0.0036
0.007
0.013
Power
0.0061
0.012
0.02
Zload (LH/HL)
4.25 / 2.3
1.9 / 1.2
1.1 / 0.6
Cpin
0.0038
0.006
0.0072
Power
0.0072
0.012
0.018
Zload (LH/HL)
5.1 / 2.7
2.6 / 1.5
1.2 / 0.7
Cpin
0.0043
0.0073
0.0124
Power
0.006
0.0115
0.02
Cpin
0.003
0.005
0.01
Power
0.007
0.012
0.02
Four input NOR gate: NOR4
Implementation : y = (a + b + c + d)′
Strength
1
2
4
Width(µm)
2.3
4.14
7.82
tintrinsic (LH/HL)
0.06 / 0.022
0.063 / 0.023
0.061 / 0.023
Two input multiplexer: MUX2
Implementation : y = s′ · a + s · b
Strength
1
2
4
Width(µm)
3.7
4.15
5.5
tintrinsic (LH/HL)
0.09 / 0.075
0.085 / 0.07
0.09 / 0.075
Two input inverting multiplexer: NMUX2
Implementation : y = s′ · a + s · b)′
Strength
1
2
4
Width(µm)
3.22
4.6
6.9
tintrinsic (LH/HL)
0.045 / 0.06
0.04 / 0.06
0.035 / 0.05
Two input XOR: XOR2
Implementation : y = a · b′ + a′ · b
Strength
1
2
4
E.2.2
Width(µm)
3.22
5.5
8.3
# Flip-flops
tintrinsic (LH/HL)
0.035 / 0.055
0.035 / 0.05
0.03 / 0.04
Zload (LH/HL)
4 / 2.2
2 / 1.1
1 / 0.6
406
APPENDIX E. STANDARD CELL LIBRARIES
Appendix F
Memory compilers
407
408
APPENDIX F. MEMORY COMPILERS
Appendix G
Finite Automata
The finite automaton is a very elaborated concept. The digital synthesis tools take a big part of the “responsibilities” in deciding how to implement a HDL definition of a finite automaton. For this reason only few basic design
aspects are discussed in this appendix.
G.1
Basic definitions in automata theory
Definition G.1 An automaton, A, is defined by the following 5-uple:
A = (X, Y, Q, f, g)
where:
X : the finite set of input variables
Y : the finite set of output variables
Q : the set of state variables
f : the state transition function, described by f : X × Q → Q
g : the output transition function, with one of the following two definitions:
• g : X × Q → Y for Mealy type automaton
• g : Q → Y for Moore type automaton.
At each clock cycle the state of the automaton switches and the output takes the value according to the new state
(and the current input, in Mealy’s approach). ⋄
Definition G.2 A half-automaton, A1/2 , is defined by the following triplet:
A1/2 = (X, Q, f )
where X, y and f have the same meaning as in the previous definition. ⋄
An half-automaton is an automaton “without output”. It is only a concept theoretically useful.
Definition G.3 A finite automaton, FA, is an automaton with Q a finite set. ⋄
FA is a complex circuit because the size of its definition depends by |Q|.
Definition G.4 A recursively defined n-state automaton, nSA, is an automaton with |Q| ∈ O(f (n)). ⋄
An n SA has a finite (usually short) definition depending by one or many parameters. Its size will depend by
parameters. Therefore, it is a simple circuit.
Definition G.5 An initial state is a state having no predecessor state. ⋄
Definition G.6 An initial automaton is an automaton having a set of initial states, Q′ , which is a subset of
Q, Q′ ⊂ Q. ⋄
409
410
APPENDIX G. FINITE AUTOMATA
Definition G.7 A strict initial automaton is an automaton having only one initial state, Q′ = {q0 }. ⋄
A strict initial automaton is defined by:
A = (X, Y, Q, f, g; q0 )
and has a special input, called reset, used to led the automaton in the initial state q0 . If the automaton is initial
only, the input reset switches the automaton in one, specially selected, initial state.
Definition G.8 The delayed (Mealy or Moore) automaton is an automaton with the output values generated
through a (delay) register, thus the current output value corresponds to the previous internal state of the automaton,
instead of the current value of the state, as in non-delayed variant. ⋄
Definition G.9 The delayed (Mealy or Moore) automaton is an automaton with the output values generated
through a (delay) register, thus the current output value corresponds to the previous internal state of the automaton,
instead of the current value of the state, as in non-delayed variant. ⋄
Theorem G.1 The time relation between the input value and the output value is the following for the four types
of automata:
1. for Mealy automaton the output to the moment t, y(t) ∈ Y depends on the current input value, x(t) ∈ X,
and by the current state, q(t) ∈ Q, i.e., y(t) = g(x(t), q(t))
2. for delayed Mealy automaton and Moore automaton the output corresponds with the input value from the
previous clock cycle:
• y(t) = g(x(t − 1), q(t − 1)) for Mealy delayed automaton
• y(t) = g(q(t)) = g(f (x(t − 1), q(t − 1)) for Moore automaton
3. for delayed Moore automaton the input transition acts on the output transition delayed with two clock
cycles:
y(t) = g(q(t − 1)) = g(f (x(t − 2), q(t − 2)).⋄
Proof The proof is evident starting from the previous two definitions. ⋄
The possibility emphasized by this theorem is that we dispose of automata with different time reaction to
the input variations. The Mealy automaton follows immediate the input transitions, delayed Mealy and Moore
automata react with one clock cycle delay to the input transitions and delayed Moore automaton delays with two
cycles the response to the input.
The symbols from the sets X, Y , and Q are binary coded using bits specified by X0 , X1 , . . . for X, Y0 , Y1 , . . .
for Y , Q0 , Q1 , . . . for Q.
G.2
How behaves a finite automata
In the real world of our mind we have the filling of a continuous flowing time. But, the time in the real word
of the automata evolve sequentially, like a tick-tack. In a world evolving sequentially all the events manifest in
distinct and periodic moments of time. Time’s evolution is marked by a periodic signal called clock (see Figure
G.1). All events in a theoretical sequential environment are triggered by the active edge of clock.
In real life a lot of things happen between the two successive active edges of the clock, but they are managed only
by engineers.
amplitude
✻
✰
✻
tick
active edge
✴✠
✻
✻
✛
tack
✻
✻
✻
✲
✻
✻
✲
time
T
Figure G.1: The sequential clock. The clock signal is cyclic one. It has the period T: the time
interval between two successive active edges (between two ticks).
G.2. HOW BEHAVES A FINITE AUTOMATA
411
While our mind perceives globally the string of letters
aaabb
and analyze it in continuous time, a sequential machine receives the same string letter by letter. At each active
edge of the clock another letter is applied to the input of a sequential machine. In other words, each active edge
of the clock triggers the switch of value applied to the input.
The internal state of the automaton switches synchronized by the same clock. Accordingly, the output of the
sequential machine changes synchronized by the same event.
The clock is the big boss allowing all the changes that must be performed in a sequential synchronous system.
reset
❄
in
✲
out
clock
a.
clock
✲
DIVIDER
✻
✻
✻
✻
✻
✻
✻
✻
✻
✲
time
reset
✻
1
0
0
0
0
0
0
0
✲
time
in
✻
0
1
1
0
1
1
1
0
✲
time
out
✻
X
0
0
1
1
1
0
0
T0
T1
T2
T3
T4
T5
T6
T7
✲
time
b.
Figure G.2: Divider: a simple sequential machine. a. The block diagram of the system. b.
The wave forms describing its behavior.
Example G.1 Let be the sequential machine, called Divide, represented in Figure G.2a. It behaves as follows:
its output takes the value 0 if the signal reset is activated ignoring the value applied on in, then, when reset =
0, after each two 1s received on its input the output switch in the complementary value (from 1 to 0, or from 0 to
1). The wave forms describing the behavior of the system are represented in Figure G.2b.
Before receiving reset = 1 the state of the output is unknown, out = x (x stands here for unknown).
The stream of 1s and 0s is meaningful for the machine only after it is reset. Therefore, the stream of bits
received in the example illustrated in Figure G.2b is: 1101110. The 0 applied to the input in T0 is ignored by the
machine because the input reset has priority. ⋄
412
APPENDIX G. FINITE AUTOMATA
The value generated on the output of the machine DIVIDER can not be explained investigating only the value
applied in the same time to its input. For in=0 we find sometimes out = 0, and other times out = 1. Similar,
for in=1. It is obvious that an internal hidden variable evolves generating a more complex behavior than one can
be explained starting only from the current value on input.
G.3
Representing finite automata
A finite automaton is represented by defining its transition functions f , the state transition function, and g, the
output transition function. For a half-automaton only the function f defined.
G.3.1
Flow-charts
A flow-chart contains for each state a circle and for each type of transition an arrow. In each clock cycle the
automaton “runs” on an arrow going from the current state to the next state. In our simple model the “race” on
arrow is done in the moment of the active edge of the clock.
The flow-chart for a half-automaton
The first version is a pure symbolic representation, where the flow chart is marked on each circle with the name of
the state, and on each arrow with the transition condition, if any. The initial states can be additionally marked
with the minus sign (-), and the final states can be additionally marked with the plus sign (+).
reset
a
q0
b
❘
❘
q0 , -
0
q1
■
1
X0
a
q2 , +
q1
✛
b
1
X0
0
✒
q2
a.
b.
Figure G.3: Example of flow-chart for a half-automaton. The machine is a “double b
detector”. It stops when the first bb occurs.
The second version is used when the input are considered in the binary form. Instead of arches are used
rhombuses containing the symbol denoting a binary variable.
Example G.2 Let be a finite half-automaton that receives on its input strings containing symbols from the alphabet X = {a, b}. The machine stops in the final state when the first sequence bb is received. The first version
of the associated flow-chart is in Figure G.3a. Here is how the machine works:
• the initial state is q0 ; if a is received the machine remains in the same state, else, if b is received, then the
machine switch in the state q1
• in the state q1 the machine “knows” that one b was just received; if a is received the half-automaton switch
back in q0 , else, if b is received, then the machine switch in q2
• q2 is the final state; the next state is unconditionally q2 .
G.3. REPRESENTING FINITE AUTOMATA
413
The second version uses tests represented by a rhombus containing the tested binary input variable (see (Figure
G.3b). The input I takes the binary value 0 for the the symbol a and the binary value 1 for the symbol b. ⋄
The second version is used mainly when a circuit implementation is envisaged.
The flow-chart for a Moore automaton
When an automaton is represented the output behavior must be also included.
The first, pure symbolic version contains in each circle besides, the name of the sate, the value of the output in
that sates. The output of the automaton shows something which is meaningful for the user. Each state generates
an output value that can be different from the state’s name. The output set of value are used to classify the state
set. The input events are mapped into the state set, and the state set is mapped into the output set.
reset
a
q0
0
b
❘
0
❘
q0 /0, -
q1 /0
■
1
X0
q1
0
a
1
q2 /1, +
✛
b
1
X0
0
q2
✒
a.
b.
Figure G.4: Example of flow-chart for a Moore automaton. The output of this automaton
tells us: “bb was already detected”.
The second uses for each pair state/output one rectangle. Inside of the rectangle is the value of the output
and near to it is marked the state (by its name, by its binary code,, or both).
Example G.3 The problem solved in the previous example is revisited using an automaton. The output set is
Y = {0, 1}. If the output takes the value 1, then we learn that a double b was already received. The state set
Q = {q0 , q1 , q2 } is divided in two classes: Q0 = {q0 , q1 } and Q1 = {q2 }. If the automaton stays in Q0 with out
= 1, then it is looking for bb. If the automaton stays in Q1 with out = 1, then it stopped investigating the input
because a double b was already received.
The associated flow-chart is in, in the first version represented by Figure G.4a. The states q0 and q1 belong
to Q0 because in the corresponding circles we have q0 /0 and q1 /0. The state q2 belongs to Q1 because in the
corresponding circle we have q2 /1. Because the evolution from q2 does not depend by input, the arrow emerging
from the corresponding circle is not labelled.
The second version (see Figure G.4b) uses three rectangles, one for each state. ⋄
A meaningful event on the input of a Moore automaton is shown on the output with a delay of a clock cycle.
All goes through the state set. In the previous example, if the second b from bb is applied on the input in the
period Ti of the clock cycle, then the automaton points out the event in the period Ti+1 of the clock cycle.
The flow-chart for a Mealy automaton
The first, pure symbolic version contains on each arrow besides, the name of the condition, the value of the output
generated in the state where the arrow starts with the input specified on the arrow.
The Mealy automaton reacts on its outputs more promptly to a meaningful input event. The output value
depends on the input value from the same clock cycle.
The second, implementation oriented version uses rectangles to specify the output’s behavior.
414
APPENDIX G. FINITE AUTOMATA
reset
q0
0
0
1
X0
a/0
q1
b/0
❘
❘
q0 , -
1
q1
■
1
a/0
X0
0
0
a/1, b/1 = -/1
q2 , +
✛
b/1
q2
✒
1
a.
b.
Figure G.5: Example of flow-chart for a Mealy automaton. The occurrence of the second b
from bb is detected as fast as possible.
Example G.4 Let us solve again the same problem of bb detection using a Mealy automaton. The resulting
flow-chart is in Figure G.5a. Now the output is activated (out = 1) when the automaton is in the state q1 (one b
was detected in the previous cycle) and the input takes the value b. The same condition triggers the switch in the
state q2 . In the final state q2 the output is unconditionally 1. In the notation −/1 the sign − stands for “don’t
care”.
Figure G.5b represents the second representation. ⋄
We can say the Mealy automaton is a “transparent” automaton, because a meaningful change on its inputs
goes directly to its output.
G.3.2
Transition diagrams
Flow-charts are very good to offer an intuitive image about how automata behave. The concept is very well
represented. But, automata are also actual machines. In order to help us to provide the real design we need
different representation. Transition diagrams are less intuitive, but they work better for helping us to provide the
image of the circuit performing the function of a certain automaton.
Transition diagrams uses Vetch-Karnaugh diagrams, VKD, for representing the transition functions. The
representation maps the VKD describing the state set of the automaton into the VKDs defining the function f
and the function g.
Transition diagrams are about real stuff. Therefore, the symbols like a, b, q0 , . . . must be codded binary,
because a real machine work with bits, 0 and 1, not with symbols.
The output is already codded binary. For the input symbols the code is established by “the user” of the
machine (similarly the output codes have been established by “the user”). Let say, for the input variable, X0 ,
was decided the following codification: a → X0 = 0 and b → X0 = 1.
G.3. REPRESENTING FINITE AUTOMATA
415
Because the actual value of the state is “hidden” from the user, the designer has the freedom to assign the
binary values according to its own (engineering) criteria. Because the present approach is a theoretical one, we
do not have engineering criteria. Therefore, we are completely free to assign the binary codes. Two option are
presented:
option 1: q0 = 00, q1 = 01, q2 = 10
option 2: q0 = 00, q1 = 10, q2 = 11
For both the external behavior of the automaton must be the same.
Transition diagrams for half-automata
The transition diagram maps the reference VKD into the next state VKD, thus defining the state transition
function. Results a representation ready to be used to design and to optimize the physical structure of a finite
half-automaton.
Example G.5 The flow-chart from Figure G.3 has two different correspondent representations as transition
diagrams in Figure G.6, one for the option 1 of coding (Figure G.6a), and another for the option 2 (Figure G.6b).
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
❘
S1
S0
1
1
0
1
1
0
0
0
S1
S0
-
- X0 0
1
0
0 X0
1
-
+
+
S1
, S0
S1 , S 0
a.
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
❘
S1
S0
1
1
0
1
1
0
0
0
S1 , S 0
S0
S1
1
-
X0 X0 X0 0
+
+
S1
, S0
b.
Figure G.6: Example of transition diagram for a half-automaton. a. For the option 1 of
coding. b. For the option 2 of coding.
In VKD S1 , S0 each box contains a 2-bit code. Three of them are used to code the states, and one will be
ignored. VKD S1+ , S0+ represents the transition from the corresponding states. Thus, for the first coding option:
• from the state codded 00 the automaton switch in the state 0x, that is to say:
– if X0 = 0 then the next state is 00 (q0 )
– if X0 = 1 then the next state is 01 (q1 )
• from the state codded 01 the automaton switch in the state x0, that is to say:
– if X0 = 0 then the next state is 00 (q0 )
– if X0 = 1 then the next state is 10 (q2 )
• from the state codded 10 the automaton switch in the same state, 10 that is the final state
• the transition from 11 is not defined.
If in the clock cycle Ti the state of the automaton is S1 , S0 (defined in the reference VKD), then in the next clock
cycle, Ti+1 , the automaton switches in the state S1+ , S0+ (defined in the next state VKD).
For the second coding option:
• from the state codded 00 the automaton switch in the state X0 0, that is to say:
– if X0 = 0 then the next state is 00 (q0 )
– if X0 = 1 then the next state is 10 (q1 )
416
APPENDIX G. FINITE AUTOMATA
• from the state codded 10 the automaton switch in the state X0 X0 , that is to say:
– if X0 = 0 then the next state is 00 (q0 )
– if X0 = 1 then the next state is 11 (q2 )
• from the state codded 11 the automaton switch in the same state, 11 that is the final state
⋄
• the transition from 01 is not defined.
The transition diagram can be used to extract the Boolean functions of the loop of the half-automaton.
Example G.6 The Boolean function of the half-automaton working as “double b detector” can be extracted from
the transition diagram represented in Figure G.6a (for the first coding option). Results:
S1+ = S1 + X0 S0
S0+ = X0 S1′ S0′
⋄
Transition diagrams Moore automata
The transition diagrams define the two transition functions of a finite automaton. To the VKDs describing the
associated half-automaton is added another VKD describing the output’s behavior.
Example G.7 The flow-chart from Figure G.4 have a correspondent representation in the transition diagrams
from Figure G.7a or Figure G.7b. Besides the transition diagram for the state, the output transition diagrams are
presented for the two coding options.
For the first coding option:
• for the states coded with 00 and 01 the output has the value 0
• for the state coded with 10 the output has the value 1
• we do not care about how works the function g for the state coded with 11 because this code is not used in
defining our automaton (the output value can 0 or 1 with no consequences on the automaton’s behavior).
⋄
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
out = g(S1 , S0 )
✠
S1
S0
-
0
1
0
❘
S1
S0
1
1
0
1
1
0
0
0
-
- X0 0
1
0
0 X0
1
-
+
+
S1
, S0
S 1 , S0
out
S1
S0
a.
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
out = g(S1 , S0 )
✠
S1
S0
1
-
0
0
out
❘
S1
S0
1
1
0
1
1
0
0
0
S1 , S 0
S0
S1
1
-
X0 X0 X0 0
+
+
S1
, S0
b.
Figure G.7: Example of transition diagram for a Moore automaton.
Example G.8 The resulting output function is:
out = S1 .
Now the resulting automaton circuit can be physically implemented, in the version resulting from the first coding
option, as a system containing a 2-bit register and few gates. Results the circuit in Figure G.8, where:
G.3. REPRESENTING FINITE AUTOMATA
417
• the 2-bit register is implemented using two resetable D flip-flops
• the combinational loop for state transition function consists in few simple gates
• the output transition function is so simple as no circuit are needed to implement it.
When reset = 1 the two flip-flops switch in 0. When reset = 0 the circuit starts to analyze the stream received
on input symbol by symbol. In each clock cycle a new symbol is received and the automaton switches according to
the new state computed by three gates. ⋄
loop combinational circuit
in = X0
✙
+
S1
+
S0
clock
reset
✲
D
R
DFF
Q
Q’
✲
D
R
DFF
Q
Q’
✯
✲
S0
2-bit register
S1
out
Figure G.8: The Moore version of “bb detector” automaton.
Transition diagrams Mealy automata
The transition diagrams for a Mealy automaton are a little different from those of Moore, because the output
transition function depends also by the input variable. Therefore the VKD defining g contains, besides 0s and 1s,
the input variable.
Example G.9 Revisiting the same problem result, in Figure G.9 the transition diagrams associated to the flowchart from Figure G.5.
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
out = g(X0 , S1 , S0 )
✠
S1
S0
-
X0
1
0
❘
S1
S0
1
1
0
1
1
0
0
0
-
- X0 0
1
0
0 X0
1
-
+
+
S1
, S0
S 1 , S0
out
S1
S0
a.
+
+
{S1
, S0
} = f (X0 , S1 , S0 )
out = g(X0 , S1 , S0 )
✠
S1
S0
out
1
-
X0
0
❘
S1
S0
1
1
0
1
1
0
0
0
S1 , S 0
S0
S1
1
-
X0 X0 X0 0
+
+
S1
, S0
b.
Figure G.9: Example of transition diagram for a Mealy automaton.
418
APPENDIX G. FINITE AUTOMATA
The two functions f are the same. The function g is defined for the first coding option (Figure G.9a) as
follows:
• in the state coded by 00 (q0 ) the output takes value 0
• in the state coded by 01 (q1 ) the output takes value x
• in the state coded by 10 (q2 ) the output takes value 1
• in the state coded by 11 (unused) the output takes the “don’t care” value
Extracting the function out results:
out = S1 + X0 S0
a more complex from compared with the Moore version. (But fortunately out = S1+ , and the same circuits can be
used to compute both functions. Please ignore. Engineering stuff.) ⋄
G.3.3
Procedures
The examples used to explain how the finite automata can be represented are simple because of obvious reasons.
The real life is much more complex and we need tools to face its real challenges. For real problems software tools
are used to provide actual machines. Therefore, software oriented representation must be provided for representing
automata. The so called Hardware Description Languages, HDLs, are widely used to manage complex applications.
(The Verilog HDL is used to exemplify the procedural way to specify a finite automaton.)
HDL representations for Moore automata
A HDL (Verilog, in our example) representation consists in a program module describing the connections and the
behavior of the automaton.
Example G.10 The same “bb detector” is used to exemplify the procedures used for the Moore automaton representation.
module moore_automaton(out, in, reset, clock);
// input codes
parameter
a = 1’b0,
b = 1’b1;
// state codes
parameter
init_state = 2’b00,
// the initial state
one_b_state = 2’b01,
// the state for one b received
final_state = 2’b10;
// the final state
// output codes
parameter
no = 1’b0, // no bb yet received
yes = 1’b1; // two successive b have been received
// external connections
input
in, reset, clock;
output
out;
reg
[1:0]
state; // state register
reg
out;
// output variable
// f: the state sequential transition function
always @(posedge clock)
if (reset) state <= init_state;
else case(state)
init_state : if (in == b) state <= one_b_state;
else
state <= init_state;
one_b_state : if (in == b) state <= final_state;
else
state <= init_state;
final_state :
state <= final_state;
default
state <= 2’bx;
endcase
// g: the output combinational transition function
always @(state) case(state)
init_state : out = no;
one_b_state : out = no;
final_state : out = yes;
G.4. STATE CODDING
default
endcase
419
out = 1’bx;
endmodule
⋄
HDL representations for Mealy automata
A Verilog description consists in a program module describing the connections and the behavior of the automaton.
Example G.11 The same “bb detector” is used to exemplify the procedures used for the Mealy automaton representation.
module mealy_automaton(out, in, reset, clock);
parameter
a = 1’b0,
b = 1’b1;
parameter
init_state = 2’b00,
// the initial state
one_b_state = 2’b01,
// the state for one b received
final_state = 2’b10;
// the final state
parameter
no = 1’b0, // no bb yet received
yes = 1’b1; // two successive b have been receiveda
input
in, reset, clock;
output
out;
reg
[1:0]
state;
reg
out;
always @(posedge clock)
if (reset) state <= init_state;
else case(state)
init_state : if (in == b) state <= one_b_state;
else
state <= init_state;
one_b_state : if (in == b) state <= final_state;
else
state <= init_state;
final_state :
state <= final_state;
default
state <= 2’bx;
endcase
always @(state or in) case(state)
init_state : out = no;
one_b_state : if (in == b) out = yes;
else
out = no;
final_state : out = yes;
default
out = 1’bx;
endcase
endmodule
⋄
The procedural representations are used as inputs for automatic design tools.
G.4
State codding
The function performed by an automaton does not depend by the way its states are encoded, because the value
of the state is a “hidden variable”. But, the actual structure of a finite automaton and its proper functioning are
very sensitive to the state encoding.
The designer uses the freedom to code in different way the internal state of a finite automaton for its own
purposes. A finite automaton is a concept embodied in physical structures. The transition from concept to an
actual structure is a process with many traps and corner cases. Many of them are avoided using an appropriate
codding style.
Example G.12 Let be a first example showing the structural dependency by the state encoding. The automaton
described in Figure G.10a has three state. The first codding version for this automaton is: q0 = 00, q1 = 01, q2 =
10. We compute the next state Q1 , Q+
0 , and the output Y1 , Y0 using the first two VK transition diagrams from
Figure G.10b:
420
APPENDIX G. FINITE AUTOMATA
′
Q+
1 = Q0 + X0 Q1
′
′
′
Q+
0 = Q1 Q0 X0
Y1 = Q0 + X0 Q′1
Y0 = Q′1 Q′0 .
The second codding version for the same automaton is: q0 = 00, q1 = 01, q2 = 11. Only the code for q2 is
different. Results, using the last two VK transition diagrams from Figure G.10b:
Q1
Q0
q0
0
X0
00
1
1
0
1
1
0
0
0
Q1 Q0
1
Q1
0
1
1
1
Q0
10
01
q1
q2
-
-
0
0 X0 X0′
0
0
1
0
Q0
Q+
Q+
1
0
(11)
1
Q1
Q0
0
0 X0 1
1
0
Q1
1
1
0
0
-
- X0 1
Q0
Q+
Q+
0
1
a.
-
Y1 Y0
Q1
0
-
1
0
0
0
-
- X0 1
Y1 Y0
b.
Figure G.10: A 3-state automaton with two different state encoding. a. The flow-chart
describing the behavior. b. The VK diagrams used to implement the automaton: the reference diagram
for states, two transition diagrams used for the first code assignment, and two for the second state
assignment.
CK
Q+
0
X0
D
D-FF1
Q’ Q
Q+
1
D
D-FF0
Q’ Q
Y1
Y0
Figure G.11: The resulting circuit It is done for the second state assignment of the automaton
defined in Figure G.10a.
′ ′
′
′
Q+
1 = Q1 Q0 + X0 Q1 = (Q1 + (Q0 + X0 ) )
′
Q+
0 = Q1
Y1 = Q′1 Q0 + X0 Q′1 = (Q1 + (Q0 + X0 )′ )′
G.4. STATE CODDING
421
Y0 = Q′0 .
Obviously the second codding version provides a simpler and smaller combinational circuit associated to the
same external behavior. In Figure G.11 the resulting circuit is represented. ⋄
G.4.1
Minimal variation encoding
Minimal variation state assignment (or encoding) refers to the codes assigned to successive states.
Definition G.10 Codding with minimal variation means successive state are codded with minimal Hamming
distance. ⋄
qi
000
qi
000
qj
001
X0
X0
qj
ql
001
010
qk
?
qk
101
a.
ql
011
b.
Figure G.12: Minimal variation encoding. a. An example. b. An example where the minimal
variation encoding is not possible.
Example G.13 Let be the fragment of a flow chart represented in Figure G.12a. The state qi is followed by the
state qj and the assigned codes differ only by the least significant bit. The same for qk and ql which both follow
the state qj . ⋄
Example G.14 Some times the minimal variation encoding is not possible. An example is presented in Figure
G.12b, where qk can not be codded with minimal variation. ⋄
The minimal variation codding generates a minimal difference between the reference VK diagram and the
state transition diagram. Therefore, the state transition logical function extracted form the VK diagram can be
minimal.
G.4.2
Reduced dependency encoding
Reduced dependency encoding refers to states which conditionally follow the same state. The reduced dependency
is related to the condition tested.
Definition G.11 Reduced dependency encoding means the states which conditionally follow a certain state to be
codded with binary configurations which differs minimal (have the Hamming distance minimal). ⋄
Example G.15 In Figure G.13a the states qj and qk follow, conditioned by the value of 1-bit variable X0 , the
state qi . The assigned codes for the first two differ only in the most significant bit, and they are not related with
the code of their predecessor. The most significant bit used to code the successors of qi depends by X0 , and it is
X0′ . We say: the next states of qi are X0′ 11, for X0 =0 the next state is 111, and for X0 =1 it is 011. Reduced
dependency means: only one bit of the codes associated with the successors of qi depends by X0 , the variable tested
in qi . ⋄
422
APPENDIX G. FINITE AUTOMATA
qi
000
qi
0
010
1
X0
qj
0
001
qj
qk
011
111
0
X1
1
(001)
100
(011)
a.
1
X0
b.
qk
ql
101
(101)
Figure G.13: Examples of reduced dependency encoding. a. The transition from the state
is conditioned by the value of a single 1-bit variable. b. The transition from the state is conditioned by
two 1-bit variables.
Example G.16 In Figure G.13b the transition from the state qi depends by two 1-bit variable, X0 and X1 . A
reduced dependency codding is possible by only one of them. Without parenthesis is a reduced dependency codding
by the variable X1 . With parenthesis is a reduced dependency codding by X0 . The reader is invited to provide the
proof for the following theorem. ⋄
Theorem G.2 If the transition from a certain state depends by more than one 1-bit variable, the reduced dependency encoding can not be provided for more than one of them. ⋄
The reduced dependency encoding is used to minimize the transition function because it allows to minimize
the number of included variables in the VK state transition diagrams. Also, we will learn soon that this encoding
style is very helpful in dealing with asynchronous input variables.
G.4.3
Incremental codding
The incremental encoding provides an efficient encoding when we are able to use simple circuits to compute the
value of the next state. An incrementer is the simple circuit used to design the simple automaton called counter.
The incremental encoding allows sometimes to center the implementation of a big half-automaton on a presetable
counter.
Definition G.12 Incremental encoding means to assign, whenever it is possible, for a state following qi a code
determined by incrementing the code of qi . ⋄
Incremental encoding can be useful for reducing the complexity of a big automaton, even if sometimes the
price will be to increase the size. But, as we more frequently learn, bigger size is a good price for reducing
complexity.
G.4.4
One-hot state encoding
The register is the simple part of an automaton and the combinational circuits computing the state transition
function and the output function represent the complex part of the automaton. More, the speed of the automaton
is limited mainly by the size and depth of the associated combinational circuits. Therefore, in order to increase the
simplicity and the speed of an automaton we can use a codding stile which increase the dimension of the register
reducing in the same time the size and the depth of the combinational circuits. Many times a good balance can
be established using the one-hot state encoding.
Definition G.13 The one-hot state encoding associates to each state a bit, and consequently the state register
has a number of flip-flops equal with the number of states. ⋄
All previous state encodings used a log-number of bits to encode the state. The size of the state register will
grow, using one-hot encoding, from O(log n) to O(n) for an n-state finite automaton. Deserves to pay sometimes
this price for various reasons, such as speed, signal accuracy, simplicity, . . ..
G.5. MINIMIZING FINITE AUTOMATA
G.5
423
Minimizing finite automata
There are formal procedure to minimize an automaton by minimizing the number of internal states. All these
procedures refer to the concept. When the conceptual aspects are solved remain the problems related with the
minimal physical implementation. Follow a short discussion about minimizing the size and about minimizing the
complexity.
G.5.1
Minimizing the size by an appropriate state codding
There are some simple rules to be applied in order to generate the possibility to reach a minimal implementation.
Applying all of these rules is not always possible or an easy task and the result is not always guarantee. But it is
good to try to apply them as much as possible.
A secure and simple way to optimize the state assignment process is to evaluate all possible codding versions
and to choose the one which provide a minimal implementation. But this is not an effective way to solve the
problem because the number of different versions is in O(n!). For this reason are very useful some simple rules
able to provide a good solution instead of an optimal one.
A lucky, inspired, or trained designer will discover an almost optimal solution applying the following rule in
the order they are enounced.
Rule 1 : apply the reduced dependency codding style whenever it is possible. This rule allows a minimal
occurrence of the input variable in the VK state transition diagrams. Almost all the time this minimal occurrence
has as the main effect reducing the size of the state transition combinational circuits.
Rule 2 : the states having the same successor with identical test conditions (if it is the case) will have assigned
adjacent codes (with the Hamming distance 1). It is useful because brings in adjacent locations of a VK diagrams
identical codes, thus generating the conditions to maximize the arrays defined in the minimizing process.
Rule 3 : apply minimal variation for unconditioned transitions. This rule generates the conditions in which
the VK transition diagram differs minimally from the reference diagram, thus increasing the chance to find bigger
surfaces in the minimizing process.
Rule 4 : the states with identical outputs are codded with minimal Hamming distance (1 if possible). Generates similar effects as Rule 2.
To see at work these rules let’s take an example.
Example G.17 Let be the finite automaton described by the flow-chart from Figure G.14. Are proposed two
codding versions, a good one (the first), using the codding rules previously listed, and a bad one (the second with
the codes written in parenthesis), ignoring the rules.
For the first codding version results the expressions:
′
′
Q+
2 = Q2 Q0 + Q2 Q1
′
′
′
′
Q+
1 = Q1 Q0 + Q2 Q1 Q0 + Q2 Q0 X0
′
′
′
′
Q+
0 = Q0 + Q2 Q1 X0
Y2 = Q2 + Q1 Q0
Y1 = Q2 Q1 Q′0 + Q′2 Q′1
Y0 = Q2 + Q′1 + Q′0
the resulting circuit having the size SCLCver1 = 37.
For the second codding version results the expressions:
′
′
′
′
′
Q+
2 = Q2 Q1 Q0 + Q1 Q0 + Q2 Q0 X0 + Q1 Q0 X0
′
′
′
′
′
Q+
1 = Q1 Q0 + Q2 Q1 + Q2 X 0
′
′
′
′
Q+
0 = Q1 Q0 + Q2 Q1 + Q2 X 0
Y2 = Q2 Q′0 + Q2 Q1 + Q′2 Q′1 Q0 + Q1 Q′0
Y1 = Q′2 Q0 + Q′2 Q′1
Y0 = Q2 + Q′1 + Q0
the resulting circuit having the size SCLCver2 = 50. ⋄
424
APPENDIX G. FINITE AUTOMATA
000
Q2
011
(000)
Q1
110
111
011
010
100
101
001
000
001
011
(011)
Q0
Q2 Q1 Q0
0
1
X0
011
010
100
001
(010)
(101)
0
1
X0
100
110
101
111
(110)
(001)
101
111
101
101
(100)
(111)
Q2
Q1
Q2
111
000
1X0 0
111
101
000
01X0′
001
Q1
111
101
100
001
101
101
011
011
Version 1
Q0
Q+
Q+
Q+
2
1
0
Q0
Y2 Y1 Y0
Q2
Q1
Q2
100
000
000
111
X0 X0′ X0 X0′ X0′ X0
Q1
101
101
011
100
101
001
111
011
Version 2
Q+
Q+
Q+
2
1
0
111
011
Q0
Q0
Y2 Y1 Y0
Figure G.14: Minimizing the structure of a finite automaton. Applying appropriate codding
rules the occurrence of the input variable X0 in the transition diagrams can be minimized, thus resulting
smaller Boolean forms.
G.5. MINIMIZING FINITE AUTOMATA
G.5.2
425
Minimizing the complexity by one-hot encoding
Implementing an automaton with one-hot encoded states means increasing the simple part of the structure,
the state register. It is expected at least a part of this additional structure to be compensated by a reduced
combinational circuit used to compute the transition functions. But, for sure the entire complexity is reduced
because of a simpler combinational circuit.
Example G.18 Let be the automaton described by the flow-chart from Figure G.15, for which two codding version
are proposed: a one-hot encoding using 6 bits (Q6 . . . Q1 ), and a compact binary encoding using only 3 bits
(Q2 Q1 Q0 ).
Q1 = 1
Y0=1
000
0
Q2 = 1
1
X0
Q3 = 1
Y2=1
Y3=1
011
0
Q4 = 1
Y4=1
X0
Q5 = 1
1
111
0
Y5=1
010
X0
1
Q6 = 1
Y6=1
110
100
Figure G.15: Minimizing the complexity using one-hot encoding.
The outputs are Y6 , . . . , Y1 each active in a distinct state.
+
Version 1: with ”one-hot” encoding The state transition functions, Q+
i , i = 1, . . . , Q6 , can be
written directly inspecting the definition. Results:
Q+ 1 = Q4 + Q5 + Q6
Q+ 2 = Q1 X0
Q+ 3 = Q1 X0′
Q+ 4 = Q2 X0′
Q+ 5 = Q2 X0 + Q3 X0′
Q+ 6 = Q3 X0
Because in each state only one output bit is active, results:
Yi = Qi , pentru i = 1, . . . , 6.
The combinational circuit associated with the state transition function is very simple, and for outputs no circuits
are needed. The size of the entire combinational circuit is SCLC,var1 = 18, with the big advantage that the outputs
come directly from a flip-flop without additional unbalanced delays or other parasitic effects (like different kinds
of hazards).
Version 2: compact binary codding The state transition functions for this codding version (see
Figure G.15 for the actual binary codes) are:
Q+ 2 = Q2 Q0 + Q0 X0 + Q′2 Q′1 X0
Q+ 1 = Q′2 Q0 + Q′2 Q′1 + Q0 X0′
Q+ 0 = Q′2 Q′1
For the output transition function an additional decoder, DCD3 , is needed. The resulting combinational circuit has
the size SCLC,var2 = 44, with the additional disadvantage of generating the outputs signal using a combinational
circuit, the decoder. ⋄
426
G.6
APPENDIX G. FINITE AUTOMATA
Parasitic effects in automata
A real automaton is connected to the “external world” from which it receives of where it sends signals only
partially are controlled. This happens mainly when the connection is not sequential, mediated by a synchronous
register, because sometimes this is not possible. The designer controls very well the signals on the loop. But, the
uncontrolled arriving signals can by very dangerous for the proper functioning of an automaton. Similarly, an
uncontrolled output signal can have “hazardous” behaviors.
G.6.1
Asynchronous inputs
An automaton is implemented as a synchronous circuit changing its internal states at each active (positive or
negative) edge of clock. Let us remember the main restrictions imposed by the set-up time and hold time related
to the active edge of a clock applied to a flip-flop. No input signal can change in the time interval beginning tSU
before the clock transition and ending tH after the clock transition. Call it the prohibited time interval. But, if
at least one input of a certain finite automaton determines a switch on at least one input of the state register,
then no one can guarantee a proper functioning of that automaton.
Let be a finite automaton with one input, X0 , changing unrelated with the system clock. Its transition can
determine a transition on the input of a state flip-flop in the prohibited time interval. We call this kind of variable
asynchronous input variable or simply asynchronous variable, and we use for it the notation X0∗ . If, in a certain
state the automaton test X0∗ and switches in 1X0 X0 0 (which means in 1000 if X0∗ = 0, or 1110 is X0∗ = 1), then
we are in trouble. The actual behavior of the automaton will allow also the transition in 1X0′ A0 and in 1X0 X0′ 0,
which means the actual transition of the automaton will be in fact in 1xx0, where x ∈ {0, 1}. Indeed, if X0∗
determine the transition of two state flip-flops in the prohibited time interval, any binary configuration can be
loaded in that flip-flops, not only 11 or 00.
What is the solution for this pathological behavior induced by s asynchronous variable? To use reduced
dependency codding for the transition from the state in which X0∗ is tested. If the state assignment will allow, for
example, a transition to 11X0 0, then the behavior of the automaton becomes coherent. Indeed, if X0∗ determine
a transition in the prohibited time interval on only one state flip-flop, then the next state will be only 1110 or
1100. In both cases the automaton behaves according to its definition. If the transition of X0∗ is considered is
correct, but even if the transition is not catched it will be considered at the net clock cycle.
Example G.19 In Figure G.16 is defined a 3-state automaton with the asynchronous input variable X0∗ . Two
code assignment are proposed. The first one uses the minimal variation kind of codding, and the second uses for
the transition from the state q0 a reduced dependency codding.
The first codding is:
q0 = 01, q1 = 00, q2 = 10, q3 = 11.
Results the following circuits for state transition:
′
′
Q+
1 = Q0 + Q1 X0
Q+
0 = Q1 + Q0 X 0 .
The transition from the state Q1 Q0 = 01 is dangerous for the proper functioning of the finite automaton. Indeed,
from q0 the transition is defined by:
+
Q+
1 = X0 , Q 0 = X 0
and the transition of X0 can generate changing signals on the state flip-flops in the prohibited time interval.
Therefore, the state q0 can be followed by any state.
The second codding, with reduced (minimal) dependency, is:
q0 = 01, q1 = 00, q2 = 11, q3 = 10
Results the following equations describing the loop circuits:
′
′
Q+
1 = Q1 Q0 + Q1 Q0 + Q0 X 0
′
Q+
0 = Q0 .
The transition from the critical state, q0 , is
+
′
Q+
1 = A, Q0 = Q0 .
Only Q+
1 depends by the asynchronous input.
The size, the depth and the complexity of the resulting circuit is similar, but the behavior is correct only for
the second version. The correctness is achieved only by a proper encoding. ⋄
Obviously, transition determined by more than one asynchronous variable must be avoided, because, as we
already know, the reduced dependency codding can be done only for one asynchronous variable in each state.
But, what is the solution for more than one asynchronous input variable? Introducing new states in the definition
of the automaton, so as in each state no more than one asynchronous variable will be tested.
G.6. PARASITIC EFFECTS IN AUTOMATA
427
q0
01 [01]
0
00
[00]
q1
10
[11]
q2
X0∗
1
q3
11 [10]
CK
Version 1
D
X0
D
1
Q′
Q1
0
Q
Q′
Q
Q0
Q+
0
01
X0 X0
11
10
Q+
, Q+
1
0
Q+
1
CK
Version 2
X0
D
D
1
Q′
0
Q
Q′
Q1
Q
Q+
0
Q0
Q+
1
10
X0 0
01
11
Q+
, Q+
1
0
Figure G.16: Implementing a finite automaton with an asynchronous input.
G.6.2
The Hazard
Some additional problems must be solved to provide accurate signals to the outputs of the immediate finite
automata. The output combinational circuit introduces, besides a delay due to the propagation time through the
gate used to build it, some parasitic effects due to a kind of “indecision” in establishing the output value. Each
bit on the output is computed using a different network o gates and the effect of an input switch reaches the
output going trough different logic path. The propagation trough these various circuits can provide hazardous
transient behaviors on certain outputs.
Hazard generated by asynchronous inputs
A first form of hazardous behavior, or simply hazard, is generated by the impossibility to have synchronous
transitions to the input of a combinational circuit.
Let be circuit from Figure G.17a representing the typical gates receiving the signal A and B, ideally represented in Figure G.17b. Ideally means the two signals switches synchronously. They are considered ideal because
no synchronous signal can be actually generated. In Figure G.17c and Figure G.17d two actual relations between
the signals A and B are represented (other two are possible, but our purpose this two cases will allow to emphasize
the main effects of the actual asynchronicity).
Ideally, the AND gate must have the output continuously on 0, and the OR and XOR gates on 1. Because
of the inherent asynchronnicity between the input signals some parasitic transitions occur to the outputs of the
three gates (see Figure G.17c and Figure G.17d). Ideally, to the inputs of the three gates are applied only two
binary configurations: AB = 10 and AB = 01. But, because of the asynchronicity between the two inputs, all
428
APPENDIX G. FINITE AUTOMATA
possible binary configurations are applied, two of them for long time (AB = 10 and AB = 01) and the other
two (AB = 00 and AB = 11) only for short (transitory) time. Consequently, transitory effects are generated, by
hazard, on the outputs of the three circuits.
Some times the transitory unexpected effects can be ignored including them into the transition time of the
circuit. But, there are applications where they can generate big disfunctionalities. For example, when one of the
hazardous output is applied on a set or reset input of a latch.
A
✻
A
✲
B
The ideal
situation
a.
C
D
B
✲
A
✻
✻
situations
B
✲
t
0
t
0
B
✻
✻
✲
C
✲
t
0
t
0
C
✻
✻
✲
D
✲
t
0
t
0
D
✻
✻
✲
E
✲
t
0
0
t
0
E
✻
✲
c.
t
0
✲
The actual
✻
E
b.
A
t
0
✻
✲
t
d.
0
t
Figure G.17: How the asynchronous inputs generate hazard.
In order to offer an additional explanation for this kind of hazard VK diagrams are used in Figure G.18, where
in the first column of diagrams the ideal case is presented (the input switches directly to the desired value). In
the next two column the input reach the final value through an intermediary value. Some times the intermediary
value is associated with a parasitic transition of the output.
When between two subsystems multi-bit binary configurations are transferred, parasitic configuration must
be considered because of the asynchronicity. The hazardous effects can be “healed” being “patient” waiting for
the hazardous transition to disappear. But, we can wait only if we know when the transition occurred, i.e., the
hazard is easy to be avoided in synchronous systems.
Simply, when more than one input of a combinational is changing we must expect hazardous transitions at
least on some outputs.
Propagation hazard
Besides the hazard generated by the two or many switching inputs there exists hazard due to the transition of
only one input. In this case the internal propagations inside of the combinational circuit generate the hazard. It
G.6. PARASITIC EFFECTS IN AUTOMATA
A
B
429
A
1
B
A
1
✒
B
1
✲✻
✲
✻
IDEAL
”A < B”
”B < A”
A
A
A
AND
B
1
1
B
1
✒
1
B
1
✲✻
1
1
✲
✻
1
1
IDEAL
”A < B”
”B < A”
A
A
A
OR
1
B
✛
❄
B
✒
1
IDEAL
1
1
”A < B”
1
B
✛❄
1
”B < A”
XOR
Figure G.18: VK diagrams explaining the hazard due to the asynchronous inputs. ”A < B”
means the input A switch before the input B, and ”A > B” means the input B switch before
the input A.
could by a sort of asynchronicity generated by the different propagation paths inside the circuit.
Let be a simple example of the circuit represented in Figure G.19a, where two input are stable (A = C = 1)
and only one input switches. The problem of asynchronoous inputs is not an issue because only one input is in
transition. In Figure G.19b the detailed wave forms allow us to emphasize a parasitic transition on the output
D. For A = C = 1 the output must stay on 1 independent by the value applied on B. The actual behavior of
the circuit introduces a parasitic (hazardous) transition in 0 due to the switch of B from 1 to 0. An ideal circuit
with zero propagation times should maintain its output on 1.
A simple way to explain this kind of hazard is to say that in the VK diagram of the circuit (see Figure G.19c)
when the input “flies” from one surface of 1s to another it goes through the 0 surface generating a temporary
transition to 0. In order to avoid this transitory journey through the 0 surface an additional surface (see Figure
G.19d) is added to transform the VK diagram in a surface containing two contiguous surfaces, one for 0s and
one for 1s. The resulting equation of the circuit has an additional term: AC. The circuit with the same logic
behavior, but without hazardous transitions is represented in Figure G.19e.
Example G.20 Let be the function presented in VK diagram from Figure G.20a. An immediate solution is shown
in Figure G.20b, where a square surface is added in the middle. But this solution is partial because ignores the fact
that the VK diagram is defined as a thor, with a three-dimensional adjacency. Consequently the surfaces A′ BCD ′
and A′ B ′ CD′ are adjacent, and the same for AB ′ C ′ D and A′ B ′ C ′ D. Therefore, the solution to completely avoid
the hazard is presented in Figure G.20c, where two additional surfaces are added. ⋄
Theorem G.3 If the expression of the Boolean function
f (xn−1 , . . . x0 )
takes the form
xi + x′i
for at least one combination of the other variables than xi , then the actual associated circuit generates hazard
when xi switches. (The theorem of hazard) ⋄
430
APPENDIX G. FINITE AUTOMATA
B
B
✻
✲
A
t
0
C
X
X
A
✻
1
B
Y
✲
Z
t
0
Y
a.
1
C
d.
✲
Z
A
1
1
B
t
✻
1
A
C
✲
t
0
1
D
c.
1
✻
D
0
B
1
C
✻
✲
b.
0
t
e.
D
Figure G.19: The typical example of propagation hazard. a. The circuit. b. The wave forms.
c. The VK diagram of the function executed by the circuit. d. The added surface allowing the
behavior of the circuit to have a continuous 1 surface. e. The equivalent circuit without hazard.
Example G.21 The function f (A, B, C) = AB ′ + BC, is hazardous because: f (1, B, 1) = B ′ + B.
The function g(A, B, C, D) = AD+BC +A′ B ′ is hazardous because: g(A, 0, −, 1) = A+A′ , and g(0, B, 1, −) =
B + B ′ . Therefore, there are 4 input binary configuration generating hazardous conditions. ⋄
Dynamic hazard
The hazard generated by asynchronous inputs occurs in circuits after a first level of gates. The propagation hazard
needs a logic sum of products (2 or 3 levels of gates). The dynamic hazard is generated by similar causes but
manifests in circuits having more than 3 layers of gates. In Figure G.21 few simple dynamic hazards are shown.
There are complex and not very efficient techniques to avoid dynamic hazard. Usually it is preferred to
transform the logic in sums of products (enlarging the circuit) and to apply procedures used to remove propagation
hazard (enlarging again the circuit).
G.7
Fundamental limits in implementing automata
Because of the problems generated in the real world by the hazardous behaviors some fundamental limitations
are applied when an actual automaton works.
The asynchronous input bits can be interpreted only independently in distinct
states. In each clock cycle the automaton interprets the bits used to determine the transition form the current
state. If more than one of these bits are asynchronous the reduced dependency coding style must be applied
for all of them. But, as we know, this is impossible, only one bit can be considered with reduced dependency.
Therefore, in each state no more than one tested bit can be asynchronous. If more than one is asynchronous,
then the definition of the automaton must be modified introducing additional states.
G.7. FUNDAMENTAL LIMITS IN IMPLEMENTING AUTOMATA
A
A
1
1
A
1
B
1
B
1
1
1
a.
1
1
1
1
1
1
C
1
1
1
1
1
1
1
1
1
1
1
B
1
1
D
1
431
D
1
1
1
1
1
1
D
c.
b.
Figure G.20: a. A hazardous combinational circuit. b. A partial solution to avoid the hazard.
c. A full protection with two additional surfaces.
✻
✻
✲
✲
t
0
t
0
✻
✻
✲
0
✲
t
0
t
Figure G.21: Examples of dynamic hazards.
Immediate Mealy automaton with asynchronous inputs has no actual implementation The outputs of an immediate Mealy automaton are combinational conditioned by inputs. Therefore, an
asynchronous input will determine untolerable asynchronous transitions on some or on all of the outputs.
Delayed Mealy automaton can not be implemented with asynchronous input variables Even if all the asynchronous inputs are took into consideration properly when the state code are assigned,
the assemble formed by the state register plus the output register works wrong. Indeed, if at least one state bit
and one output bit change triggered by an asynchronous input there is the risk that the output register to be
loaded with a value unrelated with the value loaded into the state register.
Hazard free Moore automaton with asynchronous inputs has no actual implementation Asynchronous inputs involve coding with reduced dependency encoding. Hazard free outputs ask coding
with minimal variation. But, these two codding styles are incompatible.
432
APPENDIX G. FINITE AUTOMATA
Appendix H
FPGA
433
434
APPENDIX H. FPGA
Appendix I
How to make a project
I.1
How to organize a project
Project specification
Project architecture
Behavioral description
Simulation & test
Synthesis
Testing the design
Evaluation
I.2
An example: FIFO memory
I.3
About team work
There are few principle governing the team work:
1. each member of the team must know all the aspects of the project but must work only on a specific part
of the project
2. even if the final contribution of each member will be unbalanced, the effort of each member must be
maximized
3. a project is a creative work and creation has multiple facets to be covered by the diversity of skills and
talents of the members of the team
4. a leader is good but not mandatory for a small team
5. a good chemistry between the members of the team is a good starting point
6. the professional value of the team is not the “sum” of the professional values of its components (a team of
“geniuses” sometimes will be unable to finalize a project)
7. the team must work having in mind that the current project is not the last the team faces.
The previous list is open.
435
436
APPENDIX I. HOW TO MAKE A PROJECT
Appendix J
Designing a simple CISC processor
J.1
The project
cisc_processor.v
cisc_alu.v
control_automaton.v
mux2.v
mux4.v
mux2.v
register_file.v
J.2
RTL code
The module cisc processor.v
module cisc_processor(input
clock
,
input
reset
,
output
reg [31:0] addr_reg,
// memory address
output
reg [1:0]
com_reg ,
// memory command
output
reg [31:0] out_reg ,
// data output
input
[31:0] in
); // data/inst input
// INTERNAL CONNECTIONS
wire
[25:0] command;
wire
flag;
wire
[31:0] alu_out, left, right, left_out, right_out;
// INPUT & OUTPUT BUFFER REGISTERS
reg [31:0] data_reg, inst_reg;
always @(posedge clock) begin
if (command[25]) inst_reg <= in;
data_reg <= in;
addr_reg <= left_out
;
out_reg <= right_out
;
com_reg <= command[1:0];
end
// CONTROL AUTOMATON
control_automaton control_automaton(.clock (clock
),
.reset (reset
),
.inst
(inst_reg[31:11]),
.command(command
),
.flag
(alu_out[0]
));
// REGISTER FILE
register_file register_file(.left_out
(left_out
),
.right_out
(right_out
),
.result
(alu_out
),
.left_addr
(command[18:14] ),
.right_addr
(command[9:5]
),
.dest_addr
(command[23:19] ),
437
438
APPENDIX J. DESIGNING A SIMPLE CISC PROCESSOR
.write_enable
.clock
(command[24]
(clock
),
));
// MULTIPLEXERS
mux2 left_mux(
.out(left
),
.in0(left_out
),
.in1(data_reg
),
.sel(command[4]));
mux4 right_mux( .out(right
.in0(right_out
.in1({{21{inst_reg[10]}}, inst_reg[10:0]}
.in2({16’b0, inst_reg[15:0]}
.in3({inst_reg[15:0], 16’b0}
.sel(command[3:2]
// ARITHMETIC & LOGIC UNIT
cisc_alu alu(
.alu_out(alu_out
),
.left
(left
),
.right (right
),
.alu_com(command[13:10] ));
endmodule
The module cisc alu.v
module cisc_alu(output reg [31:0]
input
[31:0]
input
[31:0]
input
[3:0]
wire
[32:0] add, sub;
assign
alu_out
left
right
alu_com
,
,
,
);
add = left + right,
sub = left - right;
always @(alu_com or left
case(alu_com)
4’b0000: alu_out
4’b0001: alu_out
4’b0010: alu_out
4’b0011: alu_out
4’b0100: alu_out
4’b0101: alu_out
4’b0110: alu_out
4’b0111: alu_out
4’b1000: alu_out
4’b1001: alu_out
4’b1010: alu_out
4’b1011: alu_out
4’b1100: alu_out
4’b1101: alu_out
4’b1110: alu_out
4’b1111: alu_out
endcase
endmodule
or right or add or sub)
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
left
right
left + 1
left - 1
add[31:0]
sub[31:0]
{1’b0, left[31:1]}
{left[31], left[31:1]}
{31’b0, (left == 0)}
{31’b0, (left == right)}
{31’b0, (left < right)}
{31’b0, add[32]}
{31’b0, sub[32]}
left & right
left | right
left ^ right
The module control automaton.v
module control_automaton(
input
input
input
output
input
// THE STRUCTURE OF ’inst’
wire
[5:0]
opcode ;
clock
reset
[31:11] inst
[25:0] command
flag
// operation code
,
,
,
,
);
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
),
),
),
),
),
));
J.2. RTL CODE
wire
[4:0]
dest
,
//
left_op ,
//
right_op;
//
assign {opcode, dest, left_op,
439
selects destination register
selects left operand register
selects right operand register
right_op} = inst;
// THE STRUCTURE OF ’command’
reg
en_inst
; // enable load a new instruction in inst_reg
reg
write_enable; // writes the output of alu at dest_addr
reg [4:0] dest_addr
; // selects the destination register
reg [4:0] left_addr
; // selects the left operand in file register
reg [3:0] alu_com
; // selects the operation performed by the alu
reg [4:0] right_addr ; // selects the right operand in file register
reg
left_sel
; // selects the source of the left operand
reg [1:0] right_sel
; // selects the source of the right operand
reg [1:0] mem_com
; // generates the command for memory
assign command = {en_inst, write_enable, dest_addr, left_addr,
alu_com, right_addr, left_sel, right_sel, mem_com};
// MICRO-ARCHITECTURE
// en_inst
parameter
no_load
= 1’b0, // disable instruction register
load_inst
= 1’b1; // enable instruction register
// write_enable
parameter
no_write
= 1’b0,
write_back = 1’b1; // write back the current ALU output
// alu_func
parameter
alu_left
= 4’b0000, // alu_out = left
alu_right
= 4’b0001, // alu_out = right
alu_inc
= 4’b0010, // alu_out = left + 1
alu_dec
= 4’b0011, // alu_out = left - 1
alu_add
= 4’b0100, // alu_out = left + right
alu_sub
= 4’b0101, // alu_out = left - right
alu_shl
= 4’b0110, // alu_out = {1’b0, left[31:1]}
alu_half
= 4’b0111, // alu_out = {left[31], left[31:1]}
alu_zero
= 4’b1000, // alu_out = {31’b0, (left == 0)}
alu_equal
= 4’b1001, // alu_out = {31’b0, (left == right)}
alu_less
= 4’b1010, // alu_out = {31’b0, (left < right)}
alu_carry
= 4’b1011, // alu_out = {31’b0, add[32]}
alu_borrow = 4’b1100, // alu_out = {31’b0, sub[32]}
alu_and
= 4’b1101, // alu_out = left & right
alu_or
= 4’b1110, // alu_out = left | right
alu_xor
= 4’b1111; // alu_out = left ^ right
// left_sel
parameter
left_out
= 1’b0, // left out of the reg file as left op
from_mem
= 1’b1; // data from memory as left op
// right_sel
parameter
right_out
= 2’b00, // right out of the reg file as right op
jmp_addr
= 2’b01, // right op = {{22{inst[10]}}, inst[10:0]}
low_value
= 2’b10, // right op = {{16{inst[15]}}, inst[15:0]}
high_value = 2’b11; // right op = {inst[15:0], 16’b0}
// mem_com
parameter
mem_nop
= 2’b00,
mem_read
= 2’b10, // read from memory
mem_write
= 2’b11; // write to memory
// INSTRUCTION SET ARCHITECTURE (only samples)
440
APPENDIX J. DESIGNING A SIMPLE CISC PROCESSOR
// arithmetic & logic instructions & pc = pc + 1
parameter
move
= 6’b10_0000, // dest_reg = left_out
inc
= 6’b10_0010, // dest_reg = left_out + 1
dec
= 6’b10_0011, // dest_reg = left_out - 1
add
= 6’b10_0100, // dest_reg = left_out + right_out
sub
= 6’b10_0101, // dest_reg = left_out - right_out
bwxor
= 6’b10_1111; // dest_reg = left_out ^ right_out
// ...
// data move instructions & pc = pc + 1
parameter
read
= 6’b01_0000, // dest_reg = mem(left_out)
rdinc
= 6’b01_0001, // dest_reg = mem(left_out + value)
write
= 6’b01_1000, // mem(left_out) = right_out
wrinc
= 6’b01_1001; // mem(left_out + value) = right_out
// ...
// control instructions
parameter
nop
= 6’b11_0000, // pc = pc + 1
jmp
= 6’b11_0001, // pc = pc + value
call
= 6’b11_0010, // pc = value, ra = pc + 1
ret
= 6’b11_0011, // pc = ra
jzero
= 6’b11_0100, // if (left_out = 0) pc = pc + value;
// else pc = pc + 1
jnzero = 6’b11_0101; // if (left_out != 0) pc = pc + value;
// else pc = pc + 1
// ...
// THE STATE REGISTER
reg
[5:0]
state_reg
; // the state register
reg
[5:0]
next_state ; // a "register" used as variable
always @(posedge clock) if (reset) state_reg <= 0
;
else
state_reg <= next_state ;
// THE CONTROL AUTOMATON’S LOOP
always @(state_reg or opcode or dest or left_op or right_op or flag)
begin
en_inst
= 1’bx;//no_load
;
write_enable = 1’bx;//no_write ;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= 2’bxx;//mem_nop
;
next_state
= 6’bxxxxxx;//state_reg + 1;
// INITIALIZE THE PROCESSOR
if (state_reg == 6’b00_0000)
// pc = 0
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_xor
;
right_addr
= 5’b11111
;
left_sel
= left_out
;
right_sel
= right_out
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
// INSTRUCTION FETCH
J.2. RTL CODE
if (state_reg == 6’b00_0001)
// rquest for a new instruction & increment pc
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_inc
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_read
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0010)
// wait for memory to read doing nothing
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0011)
// load the new instruction in instr_reg
begin
en_inst
= load_inst
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b00_0100)
// initialize the control automaton
begin
en_inst
= load_inst
;
write_enable = write_back
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= opcode[5:0] ;
end
// EXECUTE THE ONE CYCLE FUNCTIONAL INSTRUCTIONS
if (state_reg[5:4] == 2’b10)
// dest = left_op OPERATION right_op
begin
en_inst
= no_load
;
write_enable = write_back
;
441
442
APPENDIX J. DESIGNING A SIMPLE CISC PROCESSOR
dest_addr
left_addr
alu_com
right_addr
left_sel
right_sel
mem_com
next_state
=
=
=
=
=
=
=
=
dest
left_op
opcode[3:0]
right_op
left_out
right_out
mem_nop
6’b00_0001
;
;
;
;
;
;
;
;
end
// EXECUTE MEMORY READ INSTRUCTIONS
if (state_reg == 6’b01_0000)
// read from left_reg in dest_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_read
;
next_state
= 6’b01_0010
;
end
if (state_reg == 6’b01_0001)
// read from left_reg + <value> in dest_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_read
;
next_state
= 6’b01_0010
;
end
if (state_reg == 6’b01_0010)
// wait for memory to read doing nothing
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= state_reg + 1;
end
if (state_reg == 6’b01_0011)
// the data from memory is loaded in data_reg
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
J.2. RTL CODE
443
right_sel
mem_com
next_state
= 2’bxx
;
= mem_nop
;
= state_reg + 1;
end
if (state_reg == 6’b01_0100)
// data_reg is loaded in dest_reg & go to fetch
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= dest
;
left_addr
= 5’bxxxxx
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= from_mem
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
// EXECUTE MEMORY WRITE INSTRUCTIONS
if (state_reg == 6’b01_1000)
// write right_op to left_op & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_left
;
right_addr
= right_op
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_write
;
next_state
= 6’b00_0001
;
end
if (state_reg == 6’b01_1000)
// write right_op to left_op + <value> & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= left_op
;
alu_com
= alu_add
;
right_addr
= right_op
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_write
;
next_state
= 6’b00_0001
;
end
// CONTROL INSTRUCTIONS
if (state_reg == 6’b11_0000)
// no operation & go to fetch
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
444
APPENDIX J. DESIGNING A SIMPLE CISC PROCESSOR
if (state_reg == 6’b11_0001)
// jump to (pc + <value>) & go to fetch
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if (state_reg == 6’b11_0010)
// call: first step: ra = pc + 1
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11110
;
left_addr
= 5’b11111
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b11_0110;
end
if (state_reg == 8’b0011_0110)
// call: second step: pc = value
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’bxxxxx
;
alu_com
= alu_right
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= jmp_addr
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if (state_reg == 6’b11_0011)
// ret: pc = ra
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11110
;
alu_com
= alu_left
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && flag)
// jzero: if (left_out = 0) pc = pc + value;
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
J.2. RTL CODE
445
left_addr
alu_com
right_addr
left_sel
right_sel
mem_com
next_state
=
=
=
=
=
=
=
5’b11111
alu_add
5’bxxxxx
left_out
low_value
mem_nop
6’b00_0001
;
;
;
;
;
;
;
end
if ((state_reg == 6’b11_0100) && ~flag)
// jzero: if (left_out = 1) pc = pc + 1;
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && ~flag)
// jnzero: if (left_out = 1) pc = pc + value;
begin
en_inst
= no_load
;
write_enable = write_back
;
dest_addr
= 5’b11111
;
left_addr
= 5’b11111
;
alu_com
= alu_add
;
right_addr
= 5’bxxxxx
;
left_sel
= left_out
;
right_sel
= low_value
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
if ((state_reg == 6’b11_0100) && flag)
// jnzero: if (left_out = 0) pc = pc + 1;
begin
en_inst
= no_load
;
write_enable = no_write
;
dest_addr
= 5’bxxxxx
;
left_addr
= 5’bxxxxx
;
alu_com
= 4’bxxxx
;
right_addr
= 5’bxxxxx
;
left_sel
= 1’bx
;
right_sel
= 2’bxx
;
mem_com
= mem_nop
;
next_state
= 6’b00_0001
;
end
end
endmodule
The module register file.v
module register_file(
output
output
input
input
input
input
input
[31:0]
[31:0]
[31:0]
[4:0]
[4:0]
[4:0]
left_out
,
right_out
,
result
,
left_addr
,
right_addr ,
dest_addr
,
write_enable,
446
APPENDIX J. DESIGNING A SIMPLE CISC PROCESSOR
input
reg [31:0]
assign
clock
);
file[0:31];
left_out
right_out
= file[left_addr]
= file[right_addr]
,
;
always @(posedge clock) if (write_enable) file[dest_addr] <= result;
endmodule
The module mux4.v
module mux4(output [31:0] out,
input
[31:0] in0,
input
[31:0] in1,
input
[31:0] in2,
input
[31:0] in3,
input
[1:0]
sel);
wire[31:0]
out1, out0;
mux2
mux(out, out0, out1, sel[1]);
mux2
mux1(out1, in2, in3, sel[0]),
mux0(out0, in0, in1, sel[0]);
endmodule
The module mux2.v
module mux2(output [31:0]
input
[31:0]
input
[31:0]
input
assign out = sel ? in1
endmodule
J.3
out,
in0,
in1,
sel);
: in0;
Testing cisc processor.v
Appendix K
# Meta-stability
Any asynchronous signal applied the the input of a clocked circuit is a source of meta-stability [webRef 1]
[Alfke ’05]. There is a dangerous timing window “centered” on the clock transition edge specified by the
sum of set-up time, edge transition time and hold time. If the data input of a D-FF switches in this window, then
there are three possible behaviors for its output:
• the output does not change according to the change on the flip-flop’s input (the flip-flop does not catch
the input variation)
• the output change according to the change on the flip-flop’s input (the flip-flop catches the input variation)
• the output goes meta-stable for tM S , then goes unpredictable in 1 or 0 (see the wave forms [webRef 2]).
447
448
APPENDIX K. # META-STABILITY
Bibliography
[Alfke ’05] Peter Alfke: “Metastable Recovery in Virtex-II Pro FPGAs”, Application Note: Virtex-II Pro
Family, http://www.xilinx.com/support/documentation/application notes/xapp094.pdf, XILINX, 2005.
[Andonie ’95] Rǎzvan Andonie, Ilie Gârbacea: Algoritmi fundamentali. O perspectivǎ C ++ , Ed. Libris,
Cluj-Napoca, 1995. (in Roumanian)
[Ajtai ’83] M. Ajtai, et al.: “An O(n log n) sorting network”, Proc. 15th Ann. ACM Symp. on Theory
of Computing, Boston, Mass., 1983.
[Batcher ’68] K. E. Batcher: “Sorting networks and their applications”, in Proc. AFIPS Spring Joint
Computer Conference, vol. 32, 1968.
[Benes ’68] Václav E. Beneš: Mathematical Theory of Connecting Networks and Telephone Traffic. New
York: Academic, 1968.
[Blakeslee ’79] T. R. Blakeslee: Digital Design with Standard MSI and LSI, John Wiley & Sons, 1979.
[Booth ’67] T. L. Booth: Sequential Machines and Automata Theory, John Wiley & Sons, Inc., 1967.
[Bremermann ’62] H. G. Bremermann: “Optimization through Evolution and Recombination”, in SelfOrganizing Systems, ed.: M. C. Yovits, S. Cameron, Washington DC, Spartan, 1962.
[Calude ’82] Cristian Calude: Complexitatea calculului. Aspecte calitative (The Complexity of Computation. Qualitative Aspects), Ed. Stiintifica si Enciclopedica, Bucuresti, 1982.
[Calude ’94] Cristian Calude: Information and Randomness, Springer-Verlag, 1994.
[Casti ’92] John L. Casti: Reality Rules: II. Picturing the World in Mathematics - The Frontier, John
Wiley & Sons, Inc., 1992.
[Cavanagh ’07] Joseph Cavanagh: Sequential Logic. Analysis and Synthesis, CRC Taylor & Francis, 2007.
[Chaitin ’66] Gregory Chaitin: “On the Length of Programs for Computing Binary Sequences”, J. of the
ACM, Oct., 1966.
[Chaitin ’70] Gregory Chaitin: “On the Difficulty of Computation”, in IEEE Transactions of Information
Theory, ian. 1970.
[Chaitin ’77] Gregory Chaitin: “Algorithmic Information Theory”, in IBM J. Res. Develop., Iulie, 1977.
[Chaitin ’87] Gregory Chaitin: Algorithmic Information Theory, Cambridge University Press, 1987.
[Chaitin ’90] Gregory Chaitin: Information, Randomness and Incompletness, World Scientific,1990.
[Chaitin ’94] Gregory Chaitin: The Limits of Mathematics IV, IBM Research Report RC 19671, e-print
chaodyn/9407009, July 1994.
[Chaitin ’06] Gregory Chaitin: “The Limit of Rason”, in Scientific American, Martie, 2006.
[Chomsky ’56] Noam Chomsky, “Three Models for the Description of Languages”, IEEE Trans. on Information Theory, 2:3 , 1956.
[Chomsky ’59] Noam Chomsky, “On Certain Formal Properties of Grammars”, Information and Control,
2:2, 1959.
449
450
BIBLIOGRAPHY
[Chomsky ’63] Noam Chomsky, “Formal Properties of Grammars”, Handbook of Mathematical Psychology, Wiley, New-York, 1963.
[Church ’36] Alonzo Church: “An Unsolvable Problem of Elementary Number Theory”, in American
Journal of Mathematics, vol. 58, pag. 345-363, 1936.
[Clare ’72] C. Clare: Designing Logic Systems Using State Machines, Mc Graw-Hill, Inc., 1972.
[Cormen ’90] Thomas H. Cormen, Charles E. Leiserson, Donsld R. Rivest: Introduction to Algorithms,
MIT Press, 1990.
[Dascǎlu ’98] Monica Dascǎlu, Eduard Franţi, Gheorghe Ştefan: “Modeling Production with Artificial
Societies: the Emergence of Social Structure”, in S. Bandini, R. Serra, F. Suggi Liverani (Eds.):
Cellular Automata: Research Towars Industry. ACRI ’98 - Proceedings of the Third Conference on
Cellular Automata for Research and Industry, Trieste, 7 - 9 October 1998, Springer Verlag, 1998. p
218 - 229.
[Dascǎlu ’98a] Monica Dascǎlu, Eduard Franţi, Gheorghe Ştefan: “Artificial Societies: a New Paradigm
for Complex Systems’ Modeling”, in IFAC Conference on Supplemental Ways for Improving International Stability - SWIIIS ’98, May 14-16, Sinaia, 1998. p.62-67.
[Drǎgǎnescu ’84] Mihai Drǎgǎnescu: “Information, Heuristics, Creation”, in Plauder, I. (ed): Artificial
Inteligence and Information Control System of Robots, Elsevier Publishers B. V. (North-Holland),
1984.
[Drǎgǎnescu ’91] Mihai Drǎgǎnescu, Gheorghe Ştefan, Cornel Burileanu: Electronica functionalǎ, Ed.
Tehnicǎ, Bucureşti, 1991 (in Roumanian).
[Einspruch ’86] N. G. Einspruch ed.: VLSI Electronics. Microstructure Science. vol. 14 : VLSI Design,
Academic Press, Inc., 1986.
[Einspruch ’91] N. G. Einspruch, J. L. Hilbert: Application Specific Integrated Circuits (ASIC) Technology, Academic Press, Inc., 1991.
[Ercegovac ’04] Miloš D. Ercegovac, Tomás Lang: Digital Arithmetic, Morgan Kaufman, 2004.
[Flynn ’72] Flynn, M.J.: “Some computer organization and their affectiveness”, IEEE Trans. Comp.
C21:9 (Sept. 1972), pp. 948-960.
[Glushkov ’66] V. M. Glushkov: Introduction to Cybernetics, Academic Press, 1966.
[Gödels ’31] Kurt Gödel: “On Formally Decidable Propositions of Principia Mathematica and Related
Systems I”, reprinted in S. Fefermann et all.: Collected Works I: Publications 1929 - 1936, Oxford
Univ. Press, New York, 1986.
[Hartley ’95] Richard I. Hartley: Digit-Serial Computation, Kulwer Academic Pub., 1995.
[Hascsi ’95] Zoltan Hascsi, Gheorghe Ştefan: “The Connex Content Addressable Memory (C 2 AM )”, Proceedings of the Twenty-first European Solid-State Circuits Conference, Lille -France, 19-21 September
1995, pp. 422-425.
[Hascsi ’96] Zoltan Hascsi, Bogdan Mı̂ţu, Mariana Petre, Gheorghe Ştefan, “High-Level Synthesis of an
Enchanced Connex memory”, in Proceedings of the International Semiconductor Conference, Sinaia,
October 1996, p. 163-166.
[Head ’87] T. Head: “Formal Language Theory and DNA: an Analysis of the Generative Capacity of
Specific Recombinant Behaviours”, in Bull. Math. Biology, 49, p. 737-759, 1987.
[Helbing 89’] Walter A. Helbing, Veljko M. Milutinovic: “Architecture and Design of a 32-bit GaAs
Microprocessor”, in [Milutinovic 89’].
[Hennessy ’07] John L. Hennessy, David A. Patterson: Computer Architecture: A Quantitative Approach,
Fourth Edition, Morgan Kaufmann, 2007.
[Hennie ’68] F. C. Hennie: Finite-State Models for Logical Machine, John Wiley & Sons, Inc., 1968.
[Hillis ’85] W. D. Hillis: The Connection Machine, The MIT Press, Cambridge, Mass., 1985.
BIBLIOGRAPHY
451
[Kaeslin ’01] Hubert Kaeslin: Digital Integrated Circuit Design, Cambridge Univ. Press, 2008.
[Keeth ’01] Brent Keeth, R. jacob Baker: DRAM Circuit Design. A Tutorial, IEEE Press, 2001.
[Kleene ’36] Stephen C. Kleene: “General Recursive Functions of Natural Numbers”, in Math. Ann., 112,
1936.
[Karim ’08] Mohammad A. Karim, Xinghao Chen: Digital Design, CRC Press, 2008.
[Knuth ’73] D. E. Knuth: The Art of Programming. Sorting and Searching, Addison-Wesley, 1973.
[Kolmogorov ’65] A.A. Kolmogorov: “Three Approaches to the Definition of the Concept “Quantity of
Information” “, in Probl. Peredachi Inform., vol. 1, pag. 3-11, 1965.
[Kung ’79] H. T. Kung, C. E. Leiserson: “Algorithms for VLSI processor arrays”, in [Mead ’79].
[Ladner ’80] R. E. ladner, M. J. Fischer: “Parallel prefix computation”, J. ACM, Oct. 1980.
[Lindenmayer ’68] Lindenmayer, A.: ”Mathematical Models of Cellular Interactions in Development I,
II”, Journal of Theor. Biology, 18, 1968.
[Maliţa ’06] Mihaela Maliţa, Gheorghe Ştefan, Marius Stoian: “Complex vs. Intensive in Parallel Computation”, in International Multi-Conference on Computing in the Global Information Technology Challenges for the Next Generation of IT&C - ICCGI, 2006 Bucharest, Romania, August 1-3, 2006
[Maliţa ’07] Mihaela Maliţa, Gheorghe Ştefan, Dominique Thiébaut: “Not Multi-, but Many-Core: Designing Integral Parallel Architectures for Embedded Computation” in International Workshop on
Advanced Low Power Systems held in conjunction with 21st International Conference on Supercomputing June 17, 2007 Seattle, WA, USA.
[Markov ’54] Markov, A. A.: ”The Theory of Algorithms”, Trudy Matem. Instituta im V. A. Steklova,
vol. 42, 1954. (Translated from Russian by J. J. Schorr-kon, U. S. Dept. of Commerce, Office of
Technical Services, no. OTS 60-51085, 1954)
[Mead ’79] Carver Mead, Lynn Convay: Introduction to VLSI Systems, Addison-Wesley Pub, 1979.
[MicroBlaze] *** MicroBlaze Processor. Reference Guide. posted at:
http://www.xilinx.com/support/documentation/sw manuals/xilinx14 1/mb ref guide.pdf
[Milutinovic 89’] Veljko M. Milutinovic (ed.): High-Level Language Computer Architecture, Computer
Science Press, 1989.
[Mindell ’00] Arnold Mindell: Quantum Mind. The Edge Between Physics and Psychology, Lao Tse Press,
2000.
[Minsky ’67] M. L. Minsky: Computation: Finite and Infinite Machine, Prentice - Hall, Inc., 1967.
[Mı̂ţu ’00] Bogdan Mı̂ţu, Gheorghe Ştefan, “Low-Power Oriented Microcontroller Architecture”, in CAS
2000 Proceedings, Oct. 2000, Sinaia, Romania
[Moto-Oka ’82] T. Moto-Oka (ed.): Fifth Generation Computer Systems, North-HollandPub. Comp.,
1982.
[Omondi ’94] Amos R. Omondi: Computer Arithmetic. Algorithm, Architecture and Implementation,
Prentice Hall, 1994.
[Palnitkar ’96] Samir Palnitkar: Verilog HDL. AGuide to Digital Design and Synthesis, SunSoft Press,
1996.
[Parberry 87] Ian Parberry: Parallel Complexity Theory. Research Notes in Theoretical Computer science. Pitman Publishing, London, 1987.
[Parberry 94] Ian Parberry: Circuit Complexity and Neural Networks, The MIT Presss, 1994.
[Patterson ’05] David A. Patterson, John L.Hennessy: Computer Organization & Design. The Hardware
/ Software Interface, Third Edition, Morgan Kaufmann, 2005.
[Pǎun ’95a] Pǎun, G. (ed.): Artificial Life. Grammatical Models, Black Sea University Press, 1995.
452
BIBLIOGRAPHY
[Pǎun ’85] A. Pǎun, Gh. Ştefan, A. Birnbaum, V. Bistriceanu, “DIALISP - experiment de structurare neconventionala a unei masini LISP”, in Calculatoarele electronice ale generatiei a cincea, Ed.
Academiei RSR, Bucuresti 1985. p. 160 - 165.
[Post ’36] Emil Post: “Finite Combinatory Processes. Formulation I”, inThe Journal of Symbolic Logic,
vol. 1, p. 103 -105, 1936.
[Prince ’99] Betty Prince: High Performance Memories. New architecture DRAMs and SRAMs evolution
anad function, John Wiley & Sons, 1999.
[Rafiquzzaman ’05] Mohamed Rafiquzzaman: Fundamentals of Digital Logic and Microcomputer Design,
Fifth Edition, Wiley – Interscience, 2005.
[Salomaa ’69] Arto Salomaa: Theory of Automata, Pergamon Press, 1969.
[Salomaa ’73] Arto Salomaa: Formal Languages, Academic Press, Inc., 1973.
[Salomaa ’81] Arto Salomaa: Jewels of Formal Language Theory, Computer Science Press, Inc., 1981.
[Savage ’87] John Savage: The Complexity of Computing, Robert E. Krieger Pub. Comp., 1987.
[Shankar ’89] R. Shankar, E. B. Fernandez: VLSI Computer Architecture, Academic Press, Inc., 1989.
[Shannon ’38] C. E. Shannon: “A Symbolic Annalysis of Relay and Switching Circuits”, Trans. AIEE,
vol. 57, p.713-723, 1938.
[Shannon ’48] C. E. Shannon: “A Mathematical Theory of Communication”, Bell System Tech. J., Vol.
27, 1948.
[Shannon ’56] C. E. Shannon: “A Universal Turing Machine with Two Internal States”, in Annals of
Mathematics Studies, No. 34: Automata Studies, Princeton Univ. Press, pp 157-165, 1956.
[Sharma ’97] Ashok K. Sharma: Semiconductor Memories. Techology, Testing, and Reliability, Wiley –
Interscience, 1997.
[Sharma ’03] Ashok K. Sharma: Advanced Smiconductor Memories. Architectures, Designs, and Applications, Whiley-Interscience, 2003.
[Solomonoff ’64] R. J. Solomonoff: “A Formal Theory of Inductive Inference”, in Information and Control, vol. 7, pag. 1- 22 , pag. 224-254, 1964.
[Spira ’71] P. M. Spira: “On time-Hardware Complexity Tradeoff for Boolean Functions”, in Preceedings
of Fourth Hawaii International Symposium on System Sciences, pp. 525-527, 1971.
[Stoian ’07] Marius Stoian, Gheorghe Ştefan: “Stacks or File-Registers in Cellular Computing?”, in CAS,
Sinaia 2007.
[Streinu ’85] Ileana Streinu: “Explicit Computation of an Independent Gödel Sentence”, in Recursive
Functions Theory Newsletter, June 1985.
[Ştefan ’97] Denisa Ştefan, Gheorghe Ştefan, “Bi-thread Microcontroller as Digital Signal Processor”,
in CAS ’97 Proceedings, 1997 International Semiconductor Conference, October 7 -11, 1997, Sinaia,
Romania.
[Ştefan ’99] Denisa Ştefan, Gheorghe Ştefan: “A Procesor Network without Interconnectio Path”, in
CAS 99 Proceedings, Oct., 1999, Sinaia, Romania. p. 305-308.
[Ştefan ’80] Gheorghe Ştefan: LSI Circuits for Processors, Ph.D. Thesis (in Roumanian), Politechnical
Institute of Bucharest, 1980.
[Ştefan ’83] Gheorghe Ştefan: “Structurari neechilibrate in sisteme de prelucrare a informatiei”, in Inteligenta artificiala si robotica, Ed. Academiei RSR, Bucuresti, 1983. p. 129 - 140.
[Ştefan ’83] Gheorghe Ştefan, et al.: Circuite integrate digitale, Ed. Did. si Ped., Bucuresti, 1983.
[Ştefan ’84] Gheorghe Ştefan, et al.: “DIALISP - a LISP Machine”, in Proceedings of the ACM Symposium on LISP and Functional Programming, Austin, Texas, Aug. 1984. p. 123 - 128.
BIBLIOGRAPHY
453
[Ştefan ’85] Gheorghe Ştefan, A. Pǎun, “Compatibilitatea functie - structura ca mecanism al evolutiei
arhitecturale”, in Calculatoarele electronice ale generatiei a cincea, Ed. Academiei RSR, Bucuresti,
1985. p. 113 - 135.
[Ştefan ’85a] Gheorghe Ştefan, V. Bistriceanu, A. Pǎun, “Catre un mod natural de implementare a LISPului”, in Sisteme cu inteligenta artificiala, Ed. Academiei Romane, Bucuresti, 1991 (paper at Al doilea
simpozion national de inteligenta artificiala, Sept. 1985). p. 218 - 224.
[Ştefan ’86] Gheorghe Stefan, M. Bodea, “Note de lectura la volumul lui T. Blakeslee: Proiectarea cu
circuite MSI si LSI”, in T. Blakeslee: Prioectarea cu circuite integrate MSI si LSI, Ed. Tehnica,
Bucuresti, 1986 (translated from English by M. Bodea, M. Hancu, Gh. Stefan). p. 338 - 364.
[Ştefan ’86a] Gheorghe Stefan, “Memorie conexa” in CNETAC 1986 Vol. 2, IPB, Bucuresti, 1986, p. 79
- 81.
[Ştefan ’91] Gheorghe Ştefan: Functie si structura in sistemele digitale, Ed. Academiei Romane, 1991.
[Ştefan ’91] Gheorghe Ştefan, Drǎghici, F.: “Memory Management Unit - a New Principle for LRU
Implementation”, Proceedings of 6th Mediterranean Electrotechnical Conference, Ljubljana, Yugoslavia,
May 1991, pp. 281-284.
[Ştefan ’93] Gheorghe Ştefan: Circuite integrate digitale. Ed. Denix, 1993.
[Ştefan ’95] Gheorghe Ştefan, Maliţa, M.: “The Eco-Chip: A Physical Support for Artificial Life Systems”, Artificial Life. Grammatical Models, ed. by Gh. Pǎun, Black Sea University Press, Bucharest,
1995, pp. 260-275.
[Ştefan ’96] Gheorghe Ştefan, Mihaela Maliţa: “Chaitin’s Toy-Lisp on Connex Memory Machine”, Journal of Universal Computer Science, vol. 2, no. 5, 1996, pp. 410-426.
[Ştefan ’97] Gheorghe Ştefan, Mihaela Maliţa: “DNA Computing with the Connex Memory”, in RECOMB 97 First International Conference on Computational Molecular Biology. January 20 - 23, Santa
Fe, New Mexico, 1997. p. 97-98.
[Ştefan ’97a] Gheorghe Ştefan, Mihaela Maliţa: “ The Splicing Mechanism and the Connex Memory”,
Proceedings of the 1997 IEEE International Conference on Evolutionary Computation, Indianapolis,
April 13 - 16, 1997. p. 225-229.
[Ştefan ’98] Gheorghe Ştefan, “Silicon or Molecules? What’s the Best for Splicing”, in Gheorghe Pǎun
(ed.): Computing with Bio-Molecules. Theory and Experiments. Springer, 1998. p. 158-181
[Ştefan ’98a] Gheorghe Ştefan, “ “Looking for the Lost Noise” ”, in CAS ’98 Proceedings, Oct. 6 - 10,
1998, Sinaia, Romania. p.579 - 582.
[Ştefan ’98b] Gheorghe Ştefan, “The Connex Memory: A Physical Support for Tree / List Processing”
in The Roumanian Journal of Information Science and Technology, Vol.1, Number 1, 1998, p. 85 - 104.
[Ştefan ’98] Gheorghe Ştefan, Robrt Benea: “Connex Memories & Rewrieting Systems”, in MELECON
’98, Tel-Aviv, May 18 -20, 1998.
[Ştefan ’99] Gheorghe Ştefan, Robert Benea: “Experimente in info cu acizi nucleici”, in M. Drǎgǎnescu,
Ştefan Trǎusan-Matu (eds): Natura realitatii fizice si a informatiei, Editura Tempus, 1999.
[Ştefan ’99a] Gheorghe Ştefan: “A Multi-Thread Approach in Order to Avoid Pipeline Penalties”, in
Proceedings of 12th International Conference on Control Systems and Computer Science, Vol. II, May
26-29, 1999, Bucharest, Romania. p. 157-162.
[Ştefan ’00] Gheorghe Ştefan: “Parallel Architecturing starting from Natural Computational Models”,
in Proceedings of the Romanian Academy, Series A: Mathematics, Physics, Technical Sciences, Information Science, vol. 1, no. 3 Sept-Dec 2000.
[Ştefan ’01] Gheorghe Ştefan, Dominique Thiébaut, “Hardware-Assisted String-Matching Algorithms”,
in WABI 2001, 1st Workshop on Algorithms in BioInformatics, BRICS, University of Aarhaus, Danemark, August 28-31, 2001.
454
BIBLIOGRAPHY
[Ştefan ’04] Gheorghe Ştefan, Mihaela Maliţa: “Granularity and Complexity in Parallel Systems”, in
Proceedings of the 15 IASTED International Conf, 2004, Marina Del Rey, CA, ISBN 0-88986-391-1,
pp.442-447.
[Ştefan ’06] Gheorghe Ştefan: “Integral Parallel Computation”, in Proceedings of the Romanian
Academy, Series A: Mathematics, Physics, Technical Sciences, Information Science, vol. 7, no. 3
Sept-Dec 2006, p.233-240.
[Ştefan ’06a] Gheorghe Ştefan: “A Universal Turing Machine with Zero Internal States”, in Romanian
Journal of Information Science and Technology, Vol. 9, no. 3, 2006, p. 227-243
[Ştefan ’06b] Gheorghe Ştefan: “The CA1024: SoC with Integral Parallel Architecture for HDTV Processing”, invited paper at 4th International System-on-Chip (SoC) Conference & Exhibit, November 1
& 2, 2006, Radisson Hotel Newport Beach, CA
[Ştefan ’06c] Gheorghe Ştefan, Anand Sheel, Bogdan Mı̂ţu, Tom Thomson, Dan Tomescu: “The CA1024:
A Fully Programmable System-On-Chip for Cost-Effective HDTV Media Processing”, in Hot Chips:
A Symposium on High Performance Chips, Memorial Auditorium, Stanford University, August 20 to
22, 2006.
[Ştefan ’06d] Gheorghe Ştefan: “The CA1024: A Massively Parallel Processor for Cost-Effective HDTV”,
in SPRING PROCESSOR FORUM: Power-Efficient Design, May 15-17, 2006, Doubletree Hotel, San
Jose, CA.
[Ştefan ’06e] Gheorghe Ştefan: “The CA1024: A Massively Parallel Processor for Cost-Effective HDTV”,
in SPRING PROCESSOR FORUM JAPAN, June 8-9, 2006, Tokyo.
[Ştefan ’07] Gheorghe Ştefan: “Membrane Computing in Connex Environment”, invited paper at 8th
Workshop on Membrane Computing (WMC8) June 25-28, 2007 Thessaloniki, Greece
[Ştefan ’07a] Gheorghe Ştefan, Marius Stoian: “The efficiency of the register file based architectures in
OOP languages era”, in SINTES13 Craiova, 2007.
[Ştefan ’07b] Gheorghe Ştefan: “Chomsky’s Hierarchy & a Loop-Based Taxonomy for Digital Systems”,
in Romanian Journal of Information Science and Technology vol. 10, no. 2, 2007.
[Sutherland ’02] Stuart Sutherland: Verilog 2001. A Guide to the New Features of the Verilog Hardware
Description Language, Kluwer Academic Publishers, 2002.
[Tabak ’91] D. Tabak: Advanced Microprocessors, McGrow- Hill, Inc., 1991.
[Tanenbaum ’90] A. S. Tanenbaum: Structured Computer Organisation third edition, Prentice-Hall, 1990.
[Thiébaut ’06] Dominique Thiébaut, Gheorghe Ştefan, Mihaela Maliţa: “DNA search and the Connex
technology” in International Multi-Conference on Computing in the Global Information Technology Challenges for the Next Generation of IT&C - ICCGI, 2006 Bucharest, Romania, August 1-3, 2006
[Tokheim ’94] Roger L. Tokheim: Digital Principles, Third Edition, McGraw-Hill, 1994.
[Turing ’36] Alan M. Turing: “On computable Numbers with an Application to the Eintscheidungsproblem”, in Proc. London Mathematical Society, 42 (1936), 43 (1937).
[Vahid ’06] Frank Vahid: Digital Design, Wiley, 2006.
[von Neumann ’45] John von Neumann: “First Draft of a Report on the EDVAC”, reprinted in IEEE
Annals of the History of Computing, Vol. 5, No. 4, 1993.
[Uyemura ’02] John P. Uyemura: CMOS Logic Circuit Design, Kluver Academic Publishers, 2002.
[Ward ’90] S. A. Ward, R. H. Halstead: Computation Structures, The MIT Press, McGraw-Hill Book
Company, 1990.
[Wedig ’89] Robert G. Wedig: “Direct Correspondence Architectures: Principles, Architecture, and Design” in [Milutinovic ’89].
[Waksman ’68] Abraham Waksman, ”A permutation network,” in J. Ass. Comput. Mach., vol. 15, pp.
159-163, Jan. 1968.
BIBLIOGRAPHY
455
[webRef 1] http://www.fpga-faq.com/FAQ_Pages/0017_Tell_me_about_metastables.htm
[webRef 2] http://www.fpga-faq.com/Images/meta_pic_1.jpg
[Weste ’94] Neil H. E. Weste, Kamran Eshraghian: Principle of CMOS VLSI Design. ASystem Perspective, Second Edition, Addisson Wesley, 1994.
[Wolfram ’02] Stephen Wolfram: A New Kind of Science, Wolfram Media, Inc., 2002.
[Zurada ’95] Jacek M. Zurada: Introductin to Artificial Neural network, PWS Pub. Company, 1995.
[Yanushkevich ’08] Svetlana N. Yanushkevich, Vlad P. Shmerko: Introduction to Logic Design, CRC
Press, 2008.
Index
D flip-flop: delay flip-flop, is the master-slave structure
used for the embodiment of registers, 37
D-latch: is the elementary clocked latch with R=S=D,
33
fan-out: the number on input gates connected to the
output of a digital circuit, 386
flip-flop: the edge clocked basic storage element, 35
gate, 390
master-slave principle: the mechanism used to remove
the transparency of a storage element, 35
456