University of Southern California Dissertations and Theses

Clustering and fanout optimizations of asynchronous circuits

(USC Thesis Other)

Clustering and fanout optimizations of asynchronous circuits

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content
CLUSTERING AND FANOUT OPTIMIZATIONS OF ASYNCHRONOUS
CIRCUITS

by
Georgios D. Dimou

A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ELECTRICAL ENGINEERING)

May 2009

Copyright 2009 Georgios D. Dimou
ii
Dedication
To my family
iii
Acknowledgements
I would like to thank my advisor Professor Peter A. Beerel for his valuable
guidance and support though the entire course of this work. I would also like to thank
Professors Massoud Pedram and Ken Alexander for taking the time to be part of my
defense committee.
I would also like to thank TrellisWare Technologies Inc. for their support and
flexibility they have provided me with throughout my employment with them, which
allowed me to complete this effort.
Additionally I would like to thank Fulcrum Microsystems for providing
assistance in this effort and specifically Andrew Lines for his input and the
discussions that we had during the meetings between our USC team and Fulcrum
Microsystems.
I would also like to specifically acknowledge my friend and colleague Sotirios
Zogopoulos for proof reading my work and for the exchange of ideas that we had in
the process as well as Mallika Prakash, Arash Saifhashemi and Pankaj Golani for
their help in developing parts of the flow that the testing platform for this thesis was
built on. I would also like to acknowledge Nam-Hoon Kim for the code-base that he
developed, which I was able to build upon in order to test this work.
I would also like to thank my girlfriend Kristen Grande for her emotional
support through the years.
Finally and most importantly I would like to thank my parents Dimitrios and
Evanthia-Maria Dimou for enabling me and supporting me throughout my education
iv
that lead up to this thesis and for being there for me any time I needed help or
advice. Without them this work would have never been possible.
v
Table of Contents
Dedication............................................................................................................................. ii
Acknowledgements.............................................................................................................. iii
List of Tables ...................................................................................................................... vii
List of Figures .................................................................................................................... viii
Abstract ................................................................................................................................x
Chapter 1 Introduction ...............................................................................................1
1.1 Contributions .......................................................................................................3
1.2 Outline.................................................................................................................5
Chapter 2 Background...............................................................................................7
2.1 Introduction..........................................................................................................7
2.2 Handshaking protocols ........................................................................................7
2.3 Data Encoding...................................................................................................11
2.4 Timing Assumptions ..........................................................................................13
2.5 Asynchronous Design Styles..............................................................................15
2.5.1 PCFB and PCHB............................................................................................16
2.5.2 Mousetrap......................................................................................................18
2.5.3 MLD...............................................................................................................20
2.5.4 STFB and SSTFB ..........................................................................................23
2.6 Asynchronous Circuit Synthesis.........................................................................25
2.6.1 Introduction....................................................................................................25
2.6.2 High-Level Synthesis......................................................................................25
2.6.3 HDL-Based Synthesis ....................................................................................27
2.6.3.1 De-Synchronization.................................................................................28
2.6.3.2 Phased Logic ..........................................................................................29
2.6.3.3 Null-Convention Logic .............................................................................30
2.6.4 Challenges in RTL-Based Synthesis...............................................................31
2.7 General Design Considerations .........................................................................32
2.7.1 Local and Global Cycle Time..........................................................................32
2.7.2 Handling Forks and Joins...............................................................................34
2.7.3 Fanout Optimization .......................................................................................35
2.7.4 Clustering.......................................................................................................35
Chapter 3 Clustering................................................................................................37
3.1 Definitions..........................................................................................................37
3.2 Clustering Criteria..............................................................................................45
3.2.1 Ensuring liveness ...........................................................................................46
3.2.2 Maintaining Performance................................................................................54
3.2.3 Modified Floyd-Warshall for finding distances .................................................60
3.3 Experimental Results.........................................................................................65
Chapter 4 Slack Matching and Fanout Considerations .............................................70
4.1 Introduction........................................................................................................70
4.2 Slack Matching Fanout Tree Improvements .......................................................72
4.3 Experiment Setup..............................................................................................76
4.4 Experimental Results.........................................................................................77
vi
4.5 Other considerations..........................................................................................80
Chapter 5 An Asynchronous ASIC Flow...................................................................82
5.1 Synthesis...........................................................................................................83
5.2 The Clustering Program.....................................................................................84
5.2.1 Circuit Representation....................................................................................85
5.2.2 Clustering.......................................................................................................85
5.2.3 Fanout Fixes ..................................................................................................88
5.2.4 Final Optimizations.........................................................................................88
5.2.5 Library Conversion.........................................................................................89
5.2.6 Verification.....................................................................................................90
5.3 Simulation and Back-End Design.......................................................................90
Chapter 6 Summary ................................................................................................93
6.1 Advancements...................................................................................................93
6.2 Applications.......................................................................................................95
6.3 Open Issues ......................................................................................................96
6.4 Conclusions.......................................................................................................97
References .........................................................................................................................98

vii
List of Tables

Table 1 Performance of the different circuits in terms of Local and Algorithmic
Cycle Time under different local move approval criteria. It is worth
noting that without the use of our proposed criteria more than half of the
circuits resulted in deadlock (marked DL). ...............................................66
Table 2 Area summary for different examples using different local move
approval criteria. The left columns correspond to the logic area, which is
the area of the clustered gates before the introduction of slack
matching. The columns on the right correspond to the total circuit area
after slack matching.................................................................................67
Table 3 Area increase for different examples using different local move approval
criteria On the columns on the left we see the area increase for the
clustered netlist before the addition of slack matching. On the columns
of the right we see the area increase for the final netlist after slack
matching,.................................................................................................68
Table 4 Slack matching results for the sharing and no-sharing versions of the
algorithm..................................................................................................78
Table 5 The area savings in terms of slack-matching buffer area and total circuit
area realized with the new LP formulation and sharing vs. the other
versions...................................................................................................79

viii
List of Figures

Figure 1 Summary of asynchronous handshaking protocols based on the
messages that are exchanged between sender and receiver. Based on
this classification we identify 4-phase protocols (b), 2-phase protocols
(c) and single-track 2-phase protocols......................................................9
Figure 2 An example of a 4-phase pull channel....................................................10
Figure 3 An example of an enclosed interleaving of four channels with one
sender and three receivers.....................................................................11
Figure 4 Abstract view of different channels that transmit k bits of data; (a) using
the bundled-data protocol, (b) a differential bus and (c) single track 1-
of-N communication. ..............................................................................13
Figure 5 The PCHB template is shown in (a) and the PCFB template is shown
in (b).......................................................................................................17
Figure 6 The MOUSETRAP design style for a FIFO implementation....................19
Figure 7 An abstract view of a Multi-Level Domino Pipeline. ................................22
Figure 8 An abstract view of a STFB and SSTFB communication channel is
shown in (a) and an example of two successive communications of a
1-of-4 such channel is shown in (b)........................................................24
Figure 9 Three different local move scenarios that have different impact on the
lg a
τ of the circuit. The circuit for the three scenarios before the move
((a), (c), (e) respectively) and after ((b), (d), (f) respectively) showing
how the ACT could remain unaffected ((a),(b)), increase ((c),(d)), or
even get introduced ((e),(f))....................................................................44
Figure 10 An example where 0
,
>
i j
d does not guarantee that ( )
j i
v CTFO v ∈ .
Notice that 2
9 , 2
= d , but the path includes
5
v is a TOKEN_BUFFER
thus ( )
2 9
v CTFO v ∉ . ..............................................................................53
Figure 11 Pseudo-code of the first distance-update routine. This version was
found to be very fast in practice, but its performance is not bound. ........62
Figure 12 Pseudo-code of the second distance-update routine. This version can
be shown to have performance ( )
2
BN Ο , where B is the average
branching factor at each node and N is the number of nodes in the
graph......................................................................................................64
ix
Figure 13 This example illustrates the effects of improper slack matching in a
pipeline. The circuit shown above comprises pipeline stages that can
run at a 10-transition cycle, however due to the mismatch the circuit
cannot run faster than a 12 transition cycle. ...........................................71
Figure 14 A case where the original slack-matching formulation yields sub-
optimal placement of slack buffers. The original circuit in (a) requires
on buffer stage. However the old formulation would decide to place the
buffers on the input of the gate assuming no sharing (b). The new
formulation models the sharing and places the buffer after the gate (c)..75
Figure 15 Overview of our ASIC flow. ....................................................................84
Figure 16 The clustering algorithm that ensures that LCT does not grow...............87
Figure 17 A caption of a finalized placed and routed design after a successful
pass through our asynchronous ASIC flow.............................................92
x
Abstract
As semiconductor technology advances into smaller and smaller geometries,
new challenges arise. The increased circuit integration combined with larger
variability make it harder for designers to distribute a global clock and global
interconnect signals efficiently in their designs. To combat the effects designers use
more conservative models and more complicated tools that result in longer design
times and diminishing returns from the migration to the smaller geometries. Some of
these problems can be addressed by asynchronous circuits, but there exists no well-
defined method for automated asynchronous design. Some methods have been
proposed over the years, but they leverage off existing synchronous techniques too
much, resulting in circuits that are bound by the characteristics of their synchronous
counterparts. This thesis proposes a novel approach for generating such circuits,
from any arbitrary HDL representation of a circuit by automatically clustering the
synthesized gates into pipeline stages that are then slack-matched to meet
performance goals while minimizing area. The method thus provides a form of
automatic pipelining in which the throughput of the overall design is not limited to the
clock frequency or the level of pipelining in the original RTL specification.
Consequently, the performance can generally maintain or exceed the performance of
the original circuit. The method is design-style agnostic and is thus applicable to
many asynchronous design styles.
The contributions of the thesis are two-fold. First, we define a model and
theoretical infrastructure that guides clustering to avoid the introduction of deadlocks
and achieve a target circuit performance. This provides a framework for proper
clustering that can enable the unhindered exploration of area minimization algorithms
xi
in the future and lead to optimized competitive designs. Secondly, we propose
optimizations to existing slack matching models that take advantage of fanout
optimizations of buffer trees that improve the quality of the results.

1
Chapter 1 Introduction
Synchronous design has dominated the VLSI and ASIC design market for many
years. Recently the design challenges of designing circuits with a global clock has been
rising, as the industry moves to smaller geometries that allow for very large circuits to be
placed on a single die. In particular, as wiring delays become more and more dominant,
the task of distributing a clock across a large design and dealing with clock delay and
skew across the entire die is becoming a growing problem. The usual trend of clock
frequencies in commercial chips increasing every year has stopped and even though
manufacturers are moving to smaller geometries clock frequencies have remained
constant for some time.
At the same time, higher chip integration is also introducing problems in terms of
power consumption, much of which is coming from the clock distribution circuits.
Processor manufacturers are resorting to more complex implementations of clocking
strategies involving clock gating, software-controlled clock frequencies and voltages
across the chip, such as PowerNow! of AMD [2] and Intel’s Enhanced SpeedStep
Technology [1]. Both technologies are used to reduce power consumption when the
processors are moderately used or idle.
Such problems are inherently solved in asynchronous circuits. The elimination of
the global clock removes all the complex timing requirements associated with its
distribution. It also allows the designer to remove timing margins built into the logic to
accommodate clock variations, and use this time for processing. Power in asynchronous
circuits is also only consumed during processing as the control circuits are data-driven.
On top of these benefits additional area becomes available on the die due to the
2
elimination of the clock generation and distribution network. Several asynchronous
design styles have been developed that span a variety of application domains, from low
power to high-performance.
Although asynchronous logic is incredibly appealing, it has yet to be adapted
widely by the semiconductor industry. Until recently asynchronous design was
synonymous to custom design, since no automated flow existed for the generation of
asynchronous circuits. For the majority of the commercial world, where ASIC flows are
the typical approach for chip design, this is a major drawback, since most companies
lack the infrastructure and the expertise to support such efforts. The cost and the time-
to-market for such efforts are usually prohibitive for companies that do not specialize in
the area of custom VLSI design. On top of that, critical issues such as verification and
testing, which is normally done by commercial tools is not supported for the majority of
design styles that are proposed in the asynchronous logic research community.
The advantages of asynchronous design, combined with the lack of a formal
automated flow for the generation of such circuits are the key drivers behind this work.
The goal is to examine existing asynchronous design styles and identify key problems
that are unique to asynchronous design. We want to characterize those problems and
formulate their solutions, so that a tool can be generated to address those
asynchronous-specific issues. Having performed that step, we can then leverage off
commercial tools and solutions used in current synchronous design flows, and complete
a full set of tools and procedures that would allow a designer to generate an
asynchronous design from behavioral RTL code. It is our belief that the existence of
such a flow would open the way for a wider adaptation of asynchronous logic and a
realization of its benefits in the commercial world.
3
1.1 Contributions
The main focus of this thesis is in the area of automated pipelining of
asynchronous circuits. The method proposed here involves starting from a gate-level
representation of a circuit, in which each gate represents a pipeline stage, and merging
two pipeline stages at a time to generate larger pipeline stages. This method allows us
to generate a pipelined circuit from an arbitrary netlist that can take advantage of
performance and design features of various asynchronous design templates. In order for
this process to be meaningful these merging transformations have to be designed based
on two important criteria. First the transformations have to be guaranteed not to cause a
circuit to become non-functional. Secondly the clustering transformation ensures that
after slack matching the target circuit performance can be achieved.
We model the circuit as a directed graph to abstract the implementation details of
the circuit. Clustering is then the iterative merging of nodes of the graph. Based on this
model this thesis includes the following contributions:
• Demonstrates criteria that help detect and prevent merges that could
generate a transformed graph that cause deadlock. The deadlock-free
maintenance criterion is expressed in two different ways for practical
reasons. The first expression uses a breadth-first search in the graph
while the other depends on the existence of all pair-wise distances in the
graph, which is a computationally intensive process. Both have the same
worst-case complexity, but the search-based one has better average-
case performance and in practice is proven to be much faster. However
4
the distance-based algorithm has the advantage that it can be seamlessly
combined with criteria that help ensure target performance.
• Proposes additional criteria that help detect and prevent graph
transformations that could introduce logic structures that would reduce
the performance of the circuit beyond the target performance that is
required for a particular design. This means that while clustering these
criteria will prevent the artificial lengthening of the critical paths of the
design beyond the point where it is impossible to satisfy the cycle time or
the end-to-end latency constraints for the design.
In the process of developing this main framework, other contributions were made
that will have an impact on the general asynchronous design process and in particular in
the area of RTL-based (ASIC) flow.
• Developed local update for the all pair-wise distance array that is
generated by a modified Floyd-Warshall algorithm so that the array can
be updated with our clustering transformations with minimal local updates
reducing complexity significantly.
• Developed practical improvement to the Slack Matching formulation
originally presented in [5] to incorporate additional information for
potential buffer sharing and thus improved the end result.
In order to test and verify all the ideas that were developed as part of this thesis a
software package was developed that enables an asynchronous ASIC flow. The
5
software implements an end-to-end tool that reads a Verilog netlist and performs the
operations described in this thesis.
• Currently reads in Verilog netlists that are gate-level synthesized netlists
from any commercial software package. It recognizes the MLD and PCHB
design styles currently and can translate to fully function netlists for either
design styles.
• Performs clustering on a graph which is design-style-agnostic and can
use all different variants of the clustering as described in Chapter 3
and.Chapter 5.
• It also performs slack matching by utilizing external linear programming
solvers and fanout checking (and fixing if necessary) on the netlist to
guarantee that all design rules are followed.
• Emits Verilog netlists that can be used as input to the back-end tools
(applicable to any software package available) as well as testbenches for
both the original and final netlists and random vector generation code that
allows pre- and post-software behavioral verification.
1.2 Outline
In order to better describe the nature of the problems that will be discussed as
part of this work, Chapter 2 has been devoted to summarize the various existing
asynchronous design styles, as well as existing synthesis flows, their features and
weaknesses. Some key challenges and definitions are also defined here to present the
6
general context for this research. Chapter 3 presents a thorough analysis of the
clustering theory for asynchronous circuits that was developed and how it can be used to
guarantee functionality and performance. Some interesting contributions to other flow
issues that are also involved in the end-to-end design process, namely slack matching
and fanout optimizations, are presented in Chapter 4. Chapter 5 presents the existing
version of our tool and flow, its features, capabilities and inefficiencies. Chapter 6
summarizes the conclusions of this work and identifies interesting problems for future
work in the area.
7
Chapter 2 Background
2.1 Introduction
In synchronous circuits all the data is synchronized by the global clock. In
between combinational logic blocks latches or flip-flops are inserted that latch the data
once per period of the clock hence achieving the synchronization of data and control
signals among the different circuit elements. In asynchronous circuits this
synchronization is achieved through handshaking protocols that are implemented to
assist the various circuit elements with the exchange of data. There are many styles of
asynchronous design libraries and flows, and almost each one has a different
handshaking mechanism associated with it. Before attempting to go into detail
describing different asynchronous design styles it is useful to go through the various
classifications of asynchronous protocols and their properties so that it is easier to place
the different options as part of a bigger picture. The classification can be done in many
dimensions, namely based on the handshaking interface, data interface and the timing
requirements that have to be satisfied for correct operation of the circuit.
2.2 Handshaking protocols
The handshaking between two asynchronous units exchanging data often starts
with the unit where the data is originating from sending a request to the receiver.
Typically the request is sent when the data is ready and depending on the protocol this
could be part of the data or a separate control signal. The receiver has to then
acknowledge the receipt of the data. Then the transmitting module knows that the data
has been consumed and can reset its value, in order to be ready to process the next set
8
of data. This Request-Acknowledgement exchange can be performed in several different
ways and handshaking protocols can be classified according to the nature of this
exchange.
First, we distinguish between protocols according to the number of phases
(transitions) that exist in the handshake. There are two distinct kinds of protocols, the 2-
phase and the 4-phase protocol. In the 4-phase protocol case the sender asserts its
request (REQ) to inform the receiving element that it holds valid data on its output. The
receiving element will then receive the data when it is ready to consume it and raise the
acknowledgment (ACK) signal when it has actually done so. The sender will then reset
its REQ signal and after that the receiver will lower its ACK signal. The second pair of
transitions could also be used to explicitly identify a data reset phase. The 2-phase
protocol only uses two active transitions to complete the communication handshake.
Therefore, all transitions of the REQ/ACK signals are used in the same way, whether
falling or rising. That means that during the first cycle the sender raises REQ and then
the receiver raises ACK to finish the handshake. Instead of resetting the signals before
the second communication, the protocol is implemented so that the sender lowers REQ
to start the next transfer, and then the receiver lowers ACK to acknowledge the data.
The request and acknowledgment signals could be individual signals or they could be
implemented across the same wire. The later is also known as single-track
communication, and could be implemented by having the sender assert the signal (REQ)
and the receiver de-asserting it. These three cases are summarized in Figure 1.
9
Figure 1 Summary of asynchronous handshaking protocols based on the
messages that are exchanged between sender and receiver. Based
on this classification we identify 4-phase protocols (b), 2-phase
protocols (c) and single-track 2-phase protocols.
Another interesting point is that the basic protocols described above could be
modified by interleaving different edges to serve different purposes. The basic forms
described above are for point-to-point communications between two adjacent units and
the communication cycle is always initiated by the sender. When the sender initiates the
protocol it is considered a push channel, and they are common in pipelined circuits.
However, in other non-pipelined circuits the receiver signals that it is ready first before
the sender produces any data. This is known as a pull channel and the initial request is
Sender Receiver
REQ
ACK
(a)
REQ
ACK
Data
(b)
REQ
ACK
(c)
1
st
Data 2
nd
Data
REQ ACK
(d)
1
st
Data 2
nd
Data
REQ ACK
10
sent by the receiver and in the reverse direction of the data flow. For example an
adaptation of the 4-phase protocol described previously for push channels can be used
for pull channel communications. The receiver asserts the REQ signal to indicate that it
is ready to accept data. When the sender has computed the data and put it on the
channel it asserts its ACK signal. The receiver then lowers its REQ signal as soon as it
has consumed the data. Finally the sender lowers its ACK signal after it has reset the
data and the channel is now ready for the next transmition. This is depicted in Figure 2.
Figure 2 An example of a 4-phase pull channel.
All the examples stated up to this point are examples of point-to-point
communications. This means that the sender sends a signal to indicate the presence of
data and releases the data when that gets acknowledged. Another quite interesting case
is called enclosed communication. It is defined as the case where the REQ signal is
asserted and then followed by an entire handshake from the receiver side (meaning the
ACK is both asserted and de-asserted), before the REQ signal gets de-asserted. This
type of behavior might not make a difference in a typical push pipelined channel,
however its usefulness becomes apparent when considering cases where we want to
perform sequential actions instead of concurrent actions. Assume that the sender wants
Sender Receiver
REQ
ACK
(a)
REQ
ACK
Data 0
(b)
11
to generate data and then there are multiple receivers that are going to operate
sequential actions based on this data. The REQ signal can then be asserted to validate
the data on the sender side. Then multiple receivers can take turns operating on the
data and the REQ signals stays high validating its presence. When the last one of the
receivers is done processing the sender can lower the REQ signal and reset the data.
The signal transitions for such a scenario are shown in Figure 3. Additionally it can also
be the case that some or all of these processes operate on the data with some level of
concurrency as well.
Figure 3 An example of an enclosed interleaving of four channels with one
sender and three receivers.
2.3 Data Encoding
Another way of classifying asynchronous channels is based on the way that the
data is encoded on the channel. The way that is closest to typical synchronous designs
is called bundled data. In bundled data the data is presented in the form of a bus of
single rail wires from the sender to the receiver. This has the benefit that only one wire
per signal is only required and that the signals could be generated by single-rail
combinational blocks just like those used for synchronous design. However there is no
way to identify that the data is valid on the receiver end by just observing the data rails,
REQ
ACK0
Data 0
ACK1
ACK2
12
hence the designer has to make sure that the data is all valid before the REQ signal
becomes visible to the receiver. For this reason the REQ path has to be delay matched
with the slowest combinational path between sender and receiver and this task is not
trivial. Post layout simulation is typically required to ensure the functionality of the circuit.
Another way to encode data on a channel is by making it dual-rail. If the dual-rail
signals are reset between transitions it is now easy to verify the presence of the data by
the data itself by making sure that at least one of the two wires representing the data has
been asserted. In this case an explicit REQ line is not necessary for the data, as a
simple OR of the two signals verifies that the data is present. Dual-rail signals can also
be grouped together in busses as in bundled data. If there is no explicit REQ like in the
bundled-data rails all the individual OR results from each signal has to be combined to
generate the global REQ signal for the bus. When one bit is transferred a single gate
delay is added to the critical path, but in the later case the impact of such a circuit to the
performance of the circuit could be significant since it could amount to several gate
delays.
Finally a more generalized for of dual-rail signaling is 1-of-N signaling. Here for
every n wires that are used one can transmit log(n) bits. Out of the n wires only one is
asserted at a time. This encoding has several benefits. Just like dual rail signaling there
is no need for an explicit REQ signal since the presence of data can be extracted from
the data itself (again assuming that the data is reset between transmitions). It has also
been shown in [17][18][20] that, with this type of data encoding, one could also omit the
acknowledgement signal as well. Moreover, there is less activity on the bus than in the
dual-rail case, since only one wire is allowed to switch during a transmition, which could
help reduce power consumption and crosstalk noise. However the width of the bus
13
grows exponentially with the amount of bits to be encoded and therefore, this approach
could quickly become impractical for the implementation of wide buses. For such wide
data paths the signals have to be broken up into smaller groups. Popular solutions are 1-
of-2 and 1-of-4 encoding since they use only 2 wires per signal, but 1-of-8 is also
common due to better power efficiency.
Figure 4 Abstract view of different channels that transmit k bits of data; (a)
using the bundled-data protocol, (b) a differential bus and (c) single
track 1-of-N communication.
2.4 Timing Assumptions
One last classifying characteristic of asynchronous communication channels is
the type of timing assumptions that are required to hold for a particular protocol to
operate correctly. In terms of the actual design process, the fewer timing assumptions
Sender Receiver
REQ
ACK
Sender Receiver
ACK
Data
2k
Sender Receiver
k
Data
2
k
(a)
(b)
(c)
14
that exist in a design the better, since timing assumptions usually have to verified
through simulation that have to be performed both pre- and post-layout.
The first timing model is one that all delays both gate and wire are allowed to
assume any value, and the circuit is guaranteed to function properly. This model is
called delay insensitive (DI), and it is the most robust model for asynchronous circuits.
However, Martin [26] showed that no practical single-output gate implementation exists
that satisfies this constraint. Realistically this means that in order to build a DI circuit one
would have to use only multi-output gates, which is incredibly restrictive for typical circuit
designs. Alternatively one would have to use gates that cannot be designed according to
the DI timing assumptions and integrate them in the circuit using DI rules at the top level.
Martin in [27] also proposed an alternative to this strict constraint called Quasi
Delay Insensitive. This has the same delay assumptions as the DI except it requires that
every fork in the design is isochronic. The isochronic fork is a fork, for which the delays
to all destinations out of that fork are equal. This realistically is very hard to achieve due
to varying wire lengths and load at the destination during design, and due to varying
operating conditions (such as voltage and crosstalk noise) during actual operation of the
circuit. More realistically this assumption can be approximated by constraining the short
path departing from a fork to be faster than slowest reconvergent path to the gate driven
by the short path. This is a constraint that is much easier to meet and verify in practice.
In [3] the QDI concept is extended from short wire fork delays to delays through a small
number of gates that depart from a fork and then re-converge further down the data
path.
15
Another category of circuits are Speed-Independent circuits (SI). In speed
independent circuits gates could have arbitrary delays, but wire delays are considered
negligible. This makes all forks isochronic, hence the QDI protocol requirement stands
by default. With process geometries constantly shrinking though, wire delays become
more and more dominant part of a path delay, and this assumption and the real delays
need to be determined post-layout and the functionality of the circuit has to be verified
again through simulation.
Scalable Delay Insensitive (SDI) is an approach that partitions the design in
smaller parts and attempts to bridge the gap between DI and SI through this approach.
Within each sub-module the design is performed by bounding the ratio of delays
between paths by a constant. It also defines a ratio related to the estimated and
observed data on the delays that is also lower and upper bound. The same constant is
used as a bound for both expressions. After each individual module is designed, the
interconnections at the top level are designed based on DI assumptions.
2.5 Asynchronous Design Styles
As one can see there are many design choices available, both in terms of the
handshaking protocol and data encoding. Each has different advantages that could be
exploited depending on the requirements of a project. Research teams have over time
defined many design styles that can be used for the implementation of asynchronous
circuits, each one having its own advantages and disadvantages. These styles are
collections of design libraries, protocol definitions and constraints that have been
designed and verified to produce functional designs. These basic cells in these libraries
can be used to produce the desired circuits, and this flow, at least theoretically, can be
16
either manual though a custom VLSI approach or automated just like a synchronous
ASIC flow. In the process of defining the context of the problems that this work is
attempting to solve, it is important to go through a few popular and proven design styles
that have been developed over the years and proven their value for certain types of
design requirements.
2.5.1 PCFB and PCHB
The Pre-Charge Half Buffer (PCHB) and Pre-Charge Full Buffer (PCFB) were
presented in [7] and are two example of a QDI template. Both templates are similar, but
PCFB uses an extra internal state variable so that it is able to store one token per stage,
and that is why it is called a Full Buffer. On the other hand a PCHB is a half buffer
meaning that one token can exist in two adjacent pipeline stages. The templates are
designed for fine-grain-pipelining, which implies that each pipeline stage is one gate
deep. The data is encoded using 1-of-N encoding and thus there is no explicit request
line associated with the data. Each gate has an input completion detection unit and the
output also has an output completion detection unit. The two templates are shown in
Figure 5.
17
Figure 5 The PCHB template is shown in (a) and the PCFB template is shown
in (b).
The function blocks are designed using dynamic logic (Domino logic [32] more
specifically) in order to reduce the size of the circuit. Another interesting property is that
the function block can actually evaluate even if not all inputs are present yet. If the
function allows it the function block can generate an output with a subset of the inputs
and data can propagate forward along the pipeline. However the C-element will not send
an acknowledgement to the left environment until all inputs arrive and the output has
been generated. That prevents premature acknowledgments from propagating
backwards to units that have not even produced data yet. The RCD is used to detect
that data has indeed been generated from the function block. In the PCHB when both
the LCD and RCD have detected valid data on both input and output the function block
gets disabled. When the next stage in the pipeline acknowledges the outputs of the
current stage then the function block will be pre-charged to be ready to receive the next
set of data.
F
C
LCD RCD
Lack
Rack
en
pc
pc
en
F
C
LCD RCD
Lack
Rack
en
pc
pc
en
C
(a) (b)
18
The LCD and RCD operate on 1-of-n encoded channels. Their operation is
performed simply by performing an OR on the two wires. The data is reset to zero during
pre-charge, therefore, the presence of data is detected when one of the two wires
produces a logic 1. If multiple channels exist the results of the OR from each channel
have to be combined together through C-elements to produce the output of the
LCD/RCD. Even though this is a simple operation one has to remember that this a fine-
grain-pipeline design style. For multi-input gates the control logic quickly becomes a
large overhead and as a result these templates are not area efficient. Also even though
the cells use dynamic logic for smaller size and better performance, there are several
levels of control involved in the critical path. With PCHB being a half-buffer the cycle
time involves multiple levels of logic as well as a completion detection unit and a C-
element. Its cycle time varies depending on the functional block, but is generally
between 14 & 18 transitions. The PCFB is a full buffer version of PCHB. It has the same
cycle time as PCHB, so its only benefit would be slack capacity. For this reason the
PCFB is not as widely used as the PCHB design style. Even though this yields good
overall performance, there are design styles available that have much smaller cycle
times.
2.5.2 Mousetrap
MOUSETRAP is a recently proposed design style [38] that seems very
interesting due to its versatility and speed. It is a bundled-data protocol, with 2-phase
control and could be used for both very fine-grain and coarser pipeline design. It has a
very small cycle time of 5 transitions for a FIFO design and although the cycle time
would increase with merges, fork and logic added to it, it still has the potential for very
19
high throughput implementations. A basic diagram of a FIFO pipeline designed with this
design style is shown in Figure 6.
Figure 6 The MOUSETRAP design style for a FIFO implementation.
The most interesting feature of this type of circuit is the simplicity of its control.
The same feature is to a large extent responsible also for its small cycle time. It
essentially consists of a single XNOR gate per pipeline stage. The data is accompanied
by a request that has alternating values for successive tokens. The data and the request
are both latched at every stage with latches that are controlled by the local control
(XNOR). The latches are initially transparent waiting for data to arrive. When data goes
through, since the requests of successive token have alternating values, when the
request goes through the latch it will change the value of the XNOR output and the latch
is made opaque. When the next stage has also fired and the data has been latched
there the XNOR will change value again and the latch is made transparent again and
new data may arrive. If the design is not just a FIFO, then logic is added on the data
path between the latches. In that case since the data is single-rail bundled data, the
REQ REQ
Data
REQ REQ REQ
Data Data Data Data
20
request has to be delay-matched with the slowest combinational path in order to avoid
errors in the interface. When the request goes through is will go through the XNOR and
close the latch, so the timing has to be designed so that the data will definitely be
through the latch before that happens.
One big advantage of this design style is that the data path can be designed from
standard gates that can be found in any synchronous library. That means that no custom
asynchronous cell library design is necessary to support the protocol and also that it
might be easier to use existing tools for an automated synthesis flow. However, this style
requires the use of delay-matched request lines and has one timing assumption that
needs to be verified for every pipeline stage. These generally require more cumbersome
verification both pre- and post-layout. The idea has been presented for FIFOs, however
it could be extended to more complex pipelines with forks, joins and cycles. In those
cases the fact that successive tokens require requests with alternating values becomes
a restriction that needs to be designed with care. Also in the case of merges and even
more specifically conditional merges it is our assessment that more complex cells with
memory of the previous state are required to handle such a pipeline.
2.5.3 MLD
Multi-Level Domino is another design style that also used bundles of wires,
however here the data is encoded using differential encoding. The data path is
constructed out of domino-logic gates in order to be more area efficient as well as faster.
This also allows the circuit to generate a request to the next stage based on the data
itself. A completion detection unit exists for each output and all the validity signals are
then combined through an AND gate tree to generate the valid flag for the entire pipeline
stage. The style is targeted more towards medium-grain pipelining and several layers of
21
logic and many data paths in parallel are typically used in a single pipeline stage. This
yields a small overhead from the addition of the pipeline stage control units and hence
an area efficient design. Several variations of this design style have been proposed over
the years in like PS0 [39] [42], and LDA [7], others using 2-phase and others using 4-
phase handshaking. Even though there are differences between the variants in terms of
the handshaking mechanism of the controllers and the generation of control signals,
abstractly the general form of these styles can be illustrated in Figure 7.
The cycle of a pipeline stage starts with the dynamic logic gates receiving data
from the previous stage and evaluating their outputs. When the data propagates to the
last stage of gates in the pipeline stage the outputs for the stage are generated and the
dual-rail signals are used to validate that all outputs are present. The valid signal is
generated for the entire stage and is used as a request to the next stage. It could also be
used internally in the stage for isolating the outputs and initiating an early pre-charge of
the logic before the final stage. When the next stage acknowledges the data, the stage
resets its outputs to all zero so that the valid signal is forced low. The data path is
connected normally just as in the case of a synchronous netlist. Any forking or merging
between stages is handled by the controller circuits. That can be accomplished by
inserting C-elements for the requests of signals reaching a merge and the
acknowledgment signals departing a fork. The introduction of such elements might
impact the cycle time of a stage, but since the data path is several stage long, this extra
delay can be offset by reducing the amount of logic levels in a particular stage.
22
Controller
Comp
Detect
Controller
Comp
Detect
en go en go
Pipeline Stage Domino
Logic
Pipeline Stage Domino
Logic
Bit 0
Bit N
Bit 0
Bit K
Bit 0
Bit L
valid K valid 0 valid L valid 0

Figure 7 An abstract view of a Multi-Level Domino Pipeline.
The generation of the request from the data alleviates the need for delay lines
that are required in single-rail data paths such as MOUSETRAP. It also simplifies the
timing verification required for the designs. On the other hand since dual-rail dynamic
gates are not available commercially this style requires that libraries are generated
specifically for this application and this increases the design time and makes automated
synthesis harder since commercial synthesis tools are not currently capable of handling
dual-rail signaling. Another interesting point of the design style is that since the
handshaking mechanism involves the data path gates (though the generation of the valid
signal) it is not allowed to have route-through signals in a pipeline stage, since that
would cause problems with the handshaking protocol and timing. Therefore this design
style also requires the addition of buffers for all such signals.
23
2.5.4 STFB and SSTFB
Single-Track Full Buffer is a design style proposed in [18][7] for fine-grain
pipeline design. It uses 1-of-N encoding for the data and also 2-phase single-track
handshaking between gates that is embedded in the data. It has been shown to yield
very high throughput designs. There are several features of this design style that
contribute to its high performance capabilities. Firstly the gates use dynamic logic
internally for higher performance and reduced area. Secondly the gates have extremely
small forward latency of 2 transitions and a total cycle time of 6 transitions. That is
accomplished by embedding the control signals as part of the data path and the use of
2-phase handshaking.
In STFB the sender will receive data and evaluate its output and then
immediately tri-state its output. The receiver detects the presence of data and evaluates
only when all the data has been received. This is done by properly designing the stacks
of NMOS transistors so that all paths to ground use all inputs. When the receiver
evaluates its outputs it will actively drive the wires low and then tri-state the inputs. This
signals the sender that the data has been consumed and it can evaluate the next set of
data. The data is encoded in a 1-of-N fashion therefore for each communication only one
wire in the set will transition. This wire is therefore used simultaneously for the data,
request and acknowledgment signaling between the two cells. Figure 8 shows an
abstract view of such a communication channel as well as a timing diagram of two
successive token transfers on a 1-of-4 channel of this type.
24
Figure 8 An abstract view of a STFB and SSTFB communication channel is
shown in (a) and an example of two successive communications of
a 1-of-4 such channel is shown in (b).
A problem with this template is that the data wires are not actively driven at all
times. There are times that both transmitter and receiver will be in tri-state mode, hence
the data becomes more susceptible to noise and leakage. Statisizers could be used to
help alleviate this problem. An improvement for this protocol was recently published in
[18] called the Static STFB, where the data is actively driven at all times. Here once the
receiver detects the presence of data it actively holds the values present on the channel
until it consumes it.
The high speed capabilities of this design style come at the expense of increased
area, as expected from a fine-grain pipeline. The circuitry required on both ends of the
channel, which is used for detecting, driving and resetting the data is not large, but
substantial compared to the logic associated with the actual logic function of each gate.
Another problem with this design styles is that since both ends of the channel actively
drive the data lines at some time, the data communication has to be implemented with
point-to-point communications, meaning that a gate cannot drive more than one more
gate. Forks have to be implemented in special fork cells that are specifically designed for
Sender Receiver
Data
(a)
(b)
Data 0
Data 1
Data 2
Data 3
25
that purpose and which further increase the amount of logic required for implementation
by adding gates that one would not usually need when using a different template.
2.6 Asynchronous Circuit Synthesis
2.6.1 Introduction
The lack of a comprehensive ASIC design flow for asynchronous circuits, like the
one that exists for synchronous circuits, is probably the single most important reason for
the limited adoption of asynchronous technology by the semiconductor industry. Several
approaches have been proposed so far, but none of them has been able to address all
the issues associated with asynchronous design. There are two major categories of
design methodologies among the proposed work. One is attempting to define the
problems according to an asynchronous specification that incorporates the exact
description of all the operations and their relations. This approach is referred to as High-
Level Synthesis and has the potential of creating circuits that are most optimally
designed to match the given specifications. The other approach is to leverage off of tools
that already exist in the market for synchronous flows and use those for the majority of
the tasks associated with the design. That allows for faster turnaround times, since most
tools already exist. The concept there is to synthesize synchronous netlists and then
convert the results to an asynchronous design style.
2.6.2 High-Level Synthesis
Several approaches have been developed over the years for characterizing
asynchronous circuits and then generating synthesized netlists based on the
specification. These approaches defined languages able to handle the concepts
necessary for the interpretations of asynchronous communications and the design
26
process starts from specifications written in this language. Languages like CSP were
defined to be able to define the problems in the form that is unique to asynchronous
design. CSP defines a design in the form of concurrent hardware processes, and also
the way that the processes communicate. Different variant have been defined, such as
CHP that also have support for automated synthesis to a certain extent [28].
Phillips developed Tangram [21][7] which is a description language for
asynchronous processes. Syntax based translation is used to convert the processes into
circuits. The first step is to map the processes into some fixed handshake types of
functional blocks that are predefined and classified according to their handshaking
properties. These are called handshake circuits and are not mapped to any specific
library. Then the netlist is broken up into combinational logic, registers and
asynchronous elements and in the last step the combinational logic is taken into a
commercial synthesis tool and synthesized using synchronous libraries.
Bardsley et. al proposed another language called Balsa and subsequent
research has added onto this framework for top level synthesis [3][4][7]. This also starts
from a new programming language that is made explicitly to understand asynchronous-
specific commands and details. Balsa follows the synthesis approach of Tangram
although there some differences in terms of the format and the features offered.
Incremental synthesis is supported and the intermediate handshake circuits follow the
original specification architecture. A new description format is used to describe the
handshake circuits and the final mapping translates every part of the circuit into standard
cell components of given libraries.
27
This approach in general is very promising, however it starts from a high level of
abstraction that requires the use of a new language and also good understanding of
asynchronous processes and interfaces. At the same time it requires more complete tool
suites that will be used to bring the designs to the physical implementation level.
Simulation and development tools are needed to assist with the design process, which
are at a big disadvantage compared to their synchronous counterparts that have been in
development for more than two decades. Even if that was not the case companies and
engineers would be very reluctant to invest in new tools that have not been proven in the
field already. Existing designs would also have to be recoded to be adapted to the new
flow.
On the other hand there are several existing designs in HDL languages that are
extremely useful and interesting and also a large number of designers that are able to
code, partition and optimize designs using current synchronous flows. For asynchronous
design to gain leverage in the market it is imperative that flows based on RTL level
synthesis using conventional HDL and industry-accepted tools are available. In fact it is
my belief that for a good designer that is used to a particular flow and its capabilities the
results from the two approaches are going to be similar.
2.6.3 HDL-Based Synthesis
This type of synthesis starts from conventional HDL and uses tools that are well
established in the industry for the implementation of synchronous circuits. Some of these
approaches can use existing code that has already been coded for synchronous circuits
and others require certain coding details that assist the translation process. Overall most
of these approaches can be seen as synchronous translations, and especially for legacy
circuits there is a possibility that they cannot take advantage of all benefits and features
28
associated with asynchronous design. For example since timing information for the
different processes is not available, trades between concurrency and sequencing,
especially in terms of resource sharing are not possible unless the designer has already
taken these into consideration when coding at the HDL level.
2.6.3.1 De-Synchronization
De-synchronization [8][11][13][14] is a method for directly translating
synchronous netlists into asynchronous netlists with local handshaking between cells.
Synthesis is performed from regular HDL with a regular synchronous flow. The netlist
that is produced is then parsed and in the first step the flip-flops in the design are all
replaced by pairs of latches. After that is done the clock circuitry is removed from the
netlist and replaced by local controllers that implement local handshaking.
Consequently, there is combinational logic that is followed by an odd pair of latches that
is immediately followed by an even pair of latches with no logic in between. When the
control is connected together delay matched lines have to be inserted in parallel with the
combinational logic, namely between the even latch controllers and the odd latch
controllers. The delay lines are used to match the worst-case delay of the combinational
logic, and this is necessary since the data path is designed out of regular single-rail logic
gates that come from the synchronous synthesis library.
The benefit of this approach is that commercially-available standard cell libraries
and standard ASIC tools are used for the synthesis. Therefore no new libraries are
necessary and the only tool that is required is the parser that does the flip-flop
replacement and control instantiation. Another big benefit is that the circuit can easily be
proven to be equivalent to its synchronous counterpart given that all timing constraints
between the controllers are correctly verified. The correctness of the circuit is
29
mathematically proven without any exhaustive analysis of the circuit. However, the final
asynchronous circuit is so close to the originally synthesized netlist that no significant
performance or area benefits are realized with this flow. The power associated with the
clock distribution circuit is removed, which is a big benefit, but the circuit is constrained
in terms of its capabilities from the synchronous specification.
2.6.3.2 Phased Logic
Phased Logic is another approach in the quest of automated HDL based
synthesis, originally described in [22]. The flow was adapted and optimizations for both
ASIC and FPGAs [40][34][35][36][37]. The flow begins with a synchronous HDL design.
The design is synthesized using commercial tools and the netlist is then manipulated
with custom tools to produce the final netlist. The gates are replaced by their equivalent
counterparts that use two wires instead of one per signal. Each gate also has an internal
state that it uses to handshake with neighboring cells.
The two wires are not used to generate dual-rail data, but instead are used to
define not only the data, but also the state of the originating gate. This helps distinguish
“odd” data from “even” data and the gates are now capable to distinguish when the data
is valid and when not. When a gate is in a particular state – odd or even – it is ready to
fire when all inputs have the same state as the gate. As soon as it fires the gate changes
its state and waits for tokens with the same state value. It sends the output data with the
original state of the gate that produced it so that the gates further down the pipeline can
fire. Flip-flops are converted to buffers, but are initialized with a token upon reset. An
innovation in this flow is that the circuit is analyzed to verify that the netlist initial
conditions guarantee a netlist that is live and safe. This is guaranteed by making sure
that all cycles have at least one token, by inserting buffers for flip-flop-to-flip-flop paths
30
and making sure that each signal is part of a directed circuit that has one token. Dead
gates are removed and if there are any liveness or safeness problems are resolved by
adding additional logic. Other optimizations are available, such as slack-buffer insertion
and use of units that can evaluate early, however parts of this flow are done manually
with external files that include this information.
This is an integrated flow that has been used to generate circuits and proven its
value. It has the benefit of an FPGA prototyping flow and it is continuously appended
with tools that make it close to commercial standards (such as the PLFire schematic
viewer [15]). However it requires custom gate design and also some inputs from the user
for a successful conversion. For example slack-matching is done by manually specifying
files that include the position and the number of buffers to be inserted.
2.6.3.3 Null-Convention Logic
Theseus Logic proposed Null Convention Logic (NCL) [15] for the
implementation of asynchronous circuits. Later efforts have enhanced the flow and also
proposed programmable solutions for implementing this kind of circuits [22][28][30].
Reconfigurable solutions have also been proposed for the particular design style and
flow. NCL starts from conventional HDL, but the code has to be written strictly in RTL
form (no behavioral register inference) and the register acknowledgment and request
signals have to be specified. It then gets synthesized into an intermediate library called
3NCL. This library is still a single-rail library but with the addition of an extra possible
value (the NULL value) for all wires. This preserves single-rail simulation and design
capabilities, while emulating the final dual-rail gates. The final library is a full dual-rail
library, but since the control is written around the fact that data will assume the value of
NULL in a handshake cycle this is necessary for simulation. After this stage of
31
verification a second run of synthesis is performed to translate the 3NCL gates into
2NCL gates that are the true dual-rail gates that will be used for the physical design
process. In order to assure DI behavior only a limited variety of gates are used (2-input
NAND, NOR, XOR)
Generally this is a complete flow, but still requires some input from the designer
(register handshakes and placement). It uses existing synchronous tools for the majority
of the design flow and achieves results that are close to manual designs. It also closely
follows the original synchronous specification and does not optimize the design
specifically to match the asynchronous library capabilities.
2.6.4 Challenges in RTL-Based Synthesis
A lot of different approaches have been proposed to achieve automated ASIC
flows that start from HDL and result in finalized circuits using as much of the existing
synchronous ASIC toolset as possible. All these flows focus on getting a design through
the flow, but none address all the optimization issues that are specific to asynchronous
design. Although useful, they fail to address issues that prevent the generation of circuits
that can be superior to their synchronous counterparts.
The flows presented above depend on the synchronous netlist for the definition
of the pipeline stages. This is not taking advantage of key features of asynchronous
design styles and thus cannot yield more than incremental improvements. Moreover
none of these methods have any good way of guaranteeing performance constraints,
which is a basic requirement for any commercial design. Slack is not considered except
for the case of Phased Logic, and even there this is done manually from externally
defined input files that are user-specified.
32
A finer and more design-style aware pipelining mechanism can yield better
results and this is the driving force behind our clustering approach. Tailoring the circuit
pipeline to the performance requirements and the chosen design template improves the
quality of results. With a framework for correctness and performance maintenance this
leads to a more powerful method that has potential to yield results that are more
competitive than the ones provided by existing flows.
2.7 General Design Considerations
2.7.1 Local and Global Cycle Time
In the absence of a global clock, asynchronous circuit performance is
characterized using different metrics. When characterizing an asynchronous pipeline
stage (could be as small as a single cell/gate for micro-pipelines) there are two important
metrics to characterize performance. The first one is forward latency (FL) and is
measured as the time between the arrival of a new token, when the pipeline stage is
idle, and the production of valid outputs for the next stage. This is a metric that is only
dependant on the internal design of the pipeline stage. The second metric is called the
local cycle time (LCT), and it is defined as the time between the arrival of a token and
the time that the unit has reset itself back to the idle state and is ready to receive the
next token. This number is generally affected by the next pipeline stages as well since
the handshaking on the right side of the stage defines the time at which the stage can
reset its output and proceed to get ready to accept new data. Both metrics are calculated
during the design phase in terms of transitions, meaning the number of signal transitions
that have to take place for the pipeline stage to move from one state to the next. Even
33
though this is not directly translated into actual time, it is a useful first tool for tradeoff
studies, design style comparison and performance estimation.
Once the local cycle time and forward latency is known there are several
methods to do a more thorough analysis and find the performance of the entire circuit,
and potentially identify the bottlenecks in the system. This is generally a very labor-
intensive process that cannot be performed without a tool designed for this purpose, but
the basic ideas can be intuitively described using the defined metrics of forward latency
and local cycle time. The performance of a circuit is defined as the global cycle time
(GCT) of the circuit and it is essentially the metric that defines how many transitions it
takes the circuit to process a token on average. Ideally the global cycle time is equal to
the maximum of the local cycle time and the algorithmic cycle time (ACT). The
algorithmic cycle time is the maximum for all cycles of the sum of the forward latencies
of all the pipeline stages in the cycle divided by the number of tokens (data) that are in
the cycle at any time. This is the maximum performance target for a design and the
global cycle time cannot be improved beyond this point. However, the design might have
a cycle time that is higher than this value, depending on the topology and the number of
tokens in the design.
The reason that this might happen is that the performance is defined not only by
how fast data can propagate down the pipeline, but how fast the pipeline resets to
accept new tokens. The backward latency (BL) of a pipeline stage is defined as the
difference between the local cycle time and the forward latency and it can be perceived
as the time it takes for a bubble – or empty position in the pipeline – to propagate
backwards in the pipeline. Alternatively, the backward latency can also be defined as the
34
time it takes a node to complete the handshaking with its neighboring cells and reset
itself so that the next token can go through.
The forward and backward latency combined define the performance of a local
pipeline stage. However the alignment of the data in the forward direction as well as the
alignment of the bubbles in the backward direction is important to guarantee that a given
global cycle time is achievable even if both the ACT and LCTs are all smaller than the
requested global cycle time. This concept of alignment between the handshakes of the
various stages is called Slack Matching and due to its importance it will be discussed in
further detail in Chapter 4.
2.7.2 Handling Forks and Joins
Now that there is a notion of performance defined for the circuit and the pipeline
stages individually, it is easier to show what the problems are when defining an
asynchronous pipeline. The first issue that every designer is faced with when designing
an asynchronous circuit is dealing with forks and merges in the data path. Due to the
fact that the handshaking signals required for synchronization propagate along with the
data, extra gates usually have to be added to make sure that the request and/or
acknowledgement signals get combined so that the correctness of the protocol is
maintained. If care is not taken it is easy to have the system fail because of a deadlock
at a join, due to improper acknowledgement (or request) along one of the two merging
paths. These additional gates usually are on the critical path and have to be taken into
consideration in order to avoid impeding performance. Also paths going through forks
and joins are also likely to be parts of cycles or re-convergent paths and the cycle time
analysis has to be performed to verify that the design performance is not reduced.
35
2.7.3 Fanout Optimization
Another aspect of forks is the handling of fanout. In synchronous design buffers
can be added to high-fanout nets to improve performance without altering the basic
functionality of the circuit. In asynchronous circuits, however there are cases where a
high-fanout node has to be buffered, and the buffer that will be added alters the timing
and structure of the circuit in such a way that affects the global cycle time. Even worse,
in cases like SSTFB dedicated cells have to be inserted to handle nodes with fanout
grater than one, and all these cells have to be included in the design in a way that does
not cause performance degradation. This might imply modifying the shape of the fanout
tree, or adding buffers in paths parallel to the one being altered. Generally this is a labor
intensive process that currently needs to be undertaken manually during the design
process.
2.7.4 Clustering
In design styles that are targeted towards coarse-pipelines such as MLD or even
finer-grain pipelines such as MOUSETRAP, another problem arises that related to the
placement of the gates within the different pipeline stages. The more logic one places in
a stage the smaller the control overhead for the circuit. In the case of MLD for example,
though, the wider the pipeline the slower the completion detection and therefore the
circuit itself. The same is true for the number of logic levels within each stage. For
MOUSETRAP the larger the grouping the slower the design, since the requests (and
hence the cycle time of a pipeline stage) are dictated by the longest combinational path
in the pipeline stage. Generally one can trade the clustering – or the distribution of gates
within pipeline stages – on both dimensions, both in terms of depth and width, in order to
achieve better performance. It is also obvious that the grouping can further affect
36
performance, since one grouping could require less forks and joins to be inserted in the
circuit than another. Generally this is another set of tradeoffs that are currently
performed manually and it is up to a good designer to achieve a good distribution that
boosts performance and reduces area.
37
Chapter 3 Clustering
Before going through the entire design flow and the tasks performed by both the
commercial and customized pieces of software that comprise it, it is interesting to focus
on the problem of clustering. This problem is common for all the different design styles,
and it is equally important for all of them. The clustering of the gates inside larger
pipeline stages allows the circuit to reduce the control overhead and make the different
design styles competitive to not only each other, but also their synchronous
counterparts.
The goal of this chapter is to formally define the problem and its solution and to
lay the foundation for a successful application of clustering on any circuit. In particular,
this chapter develops criteria that guarantees that clustering preserves the functionality
of the circuit and does not introduce structures that make the circuit unable to meet its
performance requirements.
3.1 Definitions
Circuits usually are designed subject to performance constraints that are derived
from system requirements. Even though it is interesting to find the “fastest” a circuit can
run or the “smallest” it can be made, practically it is not very useful, since the circuit
requirements are always defined by system parameters that are not dictated by the
circuit capabilities, but by the overall system function. Therefore we plan to define our
problem in such a way that it can address a variety of design requirements.
The first step is to abstract the circuit into a more generic structure so that we
can formulate our problem mathematically. This structure is a weighted directed
38
graph ( ) m h E V G , , , = , where V is the set of nodes in the netlist.
TB CL PO PI V ∪ ∪ ∪ = , wherePI is the set of primary inputs, PO is the set of primary
outputs, CL is the set of combinational gates and TB is the set of flip-flops or
TOKEN_BUFFERS. All four set TB CL PO PI , , , are mutually disjoint sets. E is the set of
directed edges ) (VxV E⊆ . We will use the notation ) , (
, j i j i
v v e = for an edge in E to
simplify our notation for a directed edge that starts from node
i
v and ends in node
j
v .
We also require that E does not contain any self-loops
i i
e
,
. We also will define a
function
+
ℜ → E h : that is used to map an edge onto a positive real number that
represents the forward latency of the edge. We also define function { } 1 , 0 : → E m , such
that
( )


 ∈
=
otherwise
TB v
e m
i
j i
, 0
, 1
,
(1).
We define a path
j i
p
,
as a sequence of edges in E , the first edge in the
sequence starting from node
i
v , and the last edge in the sequence ending in node
j
v
and such that for all other edges in the sequence, their starting point is the ending point
of the previous edge in the path and their ending point is the starting point of the next
edge in the path. We also assume in this document that a path goes through each node
once (simple path). We will also define a cycle as a path
i i
p
,
that starts and terminates
at the same node
i
v . We also define
G
P as the set of all paths that exist in the G .
Another important input will be a target performance metric, which is defined in
terms of the target cycle time (TCT) of the circuit and will be defined as
goal
τ . We also
39
define the algorithmic cycle time (ACT or
lg a
τ ) of the circuit, which is the lower bound of
goal
τ beyond which
goal
τ is no longer achievable. Thus
lg a goal
τ τ ≥ . Having defined a
path the algorithmic cycle time is defined as:
goal
p e
k j
p e
k j
P p V v
a
i i k j
i i k j
G i i i e m
e h
τ τ ≤










=
∑
∑
∈
∈
∈ ∃ ∈
, ,
, ,
, ) (
) (
max
,
,
:
lg
(2).
We also define the weight of an edge ( ) ( ) ( )
goal j i j i j i j i
e m e h e w w τ *
, , , ,
− = = . We
have the following convention:
( ) ( ) ( )
( )
( )



∈ < −
∪ ∈ >
= − =
TB v e h
CL PI v e h
e m e h e w
i goal j i
i j i
goal j i j i j i
, 0
, 0
*
,
,
, , ,
τ
τ (3).
We also need to define the weight of a path – as an extension of the edge weight
– that is the sum of the weights of all edges in the path sequence, so
( ) ( )
∑
∈
=
j i
p e
j i j i
e w p w
,
, ,
.We also define the length of a path as the number of edges in the
sequence of the path. So ( )
j i j i
p p L
, ,
= .
We also define a distance between two nodes
i
v and
j
v as the maximum weight
of all valid paths from
i
v to
j
v , or as ∞ − if no paths exist from
i
v to
j
v , which we
denote as
( ) { }





∞ −
∈ ∃
=
∈ ∀
otherwise
P p if p w
d
G j i j i
G p
j i
j i
,
, max
, ,
,
,
(4).
40
We also define the transitive fanout (TFO) and combinational transitive fanout
(CTFO) as well as transitive fanin (TFI) and combinational transitive fanin (CTFI) as
follows:
( ) } : {
, G j i j i
P p v v TFO ∈ ∃ = ,
( ) } : {
, G j i i j
P p v v TFI ∈ ∃ = , (5)
( ) } : : {
, , ,
TB v p e P p v v CTFO
k j i l k G j i j i
∈ ∈ ¬∃ ∧ ∈ ∃ = , and
( ) } : : {
, , ,
TB v p e P p v v CTFI
k i j l k G i j j j
∈ ∈ ¬∃ ∧ ∈ ∃ = .
So essentially the ( )
i
v CTFO (and equivalently the ( )
i
v CTFI ) is the set of all
nodes that are reachable from
i
v (or equivalently for CTFI that can reach node
i
v )
through a path that does not go through a TOKEN_BUFFER node.
We also have to formally define the local move operation before we begin our
discussion of the clustering algorithm. A local move is a function on the graph
( ) m h E V G , , , = that produces a new modified graph ( ) m h E V G , , ' , ' '= . It essentially
takes two nodes V v v
j i
∈ , and replaces them in ' V with a unified new node ' ' V v
k
∈
that contains the contents of both nodes (in circuit terms that would be the instances and
wires internal to the pipeline stages that correspond to the original nodes
i
v and
j
v ) .
The rest of the nodes of V are preserved in ' V . Mathematically:
{ } { }
k j i
v v v V V ' , ' + − = (6)
41
If both E e e
i j j i
∈ ∃ ∧ ∃
, ,
the move is not allowed because this case would
generate a self-loop in the graph (a cycle of length 1). Otherwise, when combining nodes
i
v and
j
v into the new node
k
v' , the edges in the set ' E are generated as follows:
• E e E e
j m i m
∈ ∃ ∨ ∈ ∃
, ,
with j m i m ≠ ≠ , then '
,
E e
k m
∈ ∃ .
• E e E e
m j m i
∈ ∃ ∨ ∈ ∃
, ,
with j m i m ≠ ≠ , then '
,
E e
m k
∈ ∃ . (7)
• E e
l m
∈ ∃
,
with j m i m ≠ ≠ , and j l i l ≠ ≠ , then '
,
E e
l m
∈ ∃ .
So the new node
k
v' has the combined fanin and fanout of
i
v and
j
v , except for
any edges between the two that get absorbed in the new node and are removed from
the top-level graph. In other words, nodes V v v
j i
∈ , are replaced by a single node
' ' V v
k
∈ , and V v v
j i
∈ , are also replaced by ' ' V v
k
∈ in all directed pairs (edges) in ' E .
An important observation is that if either edge
j i
e
,
or
i j
e
,
exist in E a
corresponding edge does not exist in ' E , which prevents the generation of a self-loop.
So assuming that the initial netlist has no self-loops, no new ones can be created during
the execution of local moves.
If an edge E e
j i
∈
,
gets absorbed, then E e
i m
∈ ∀
,
, '
,
E e
k m
∈ it is true that
( ) ( ) ( )
k m j i i m
e w e w e w
, , ,
' ≥ + . This means that an absorbed edge can increase the weights
of all incoming edges to the new node ' ' V v
k
∈ that before the execution of the move we
incoming edges to
i
v , but at most by ( )
j i
e w
,
. If E e
j i
∉
,
then the local move does not
change the weights of any edges. It is also important to note that such a move is only
42
allowed when ( ) 0
,
>
j i
e w . This is due to the fact that ( ) 0
,
<
j i
e w implies from definition
(3) that TB v
i
∈ . A move that absorbs a token buffer is not allowed because of the
special functionality that token buffers serve in the circuit guaranteeing liveness around
loops.
Finally it is important to note the following relationships, since they are very
useful in understanding the effects of clustering on the connectivity of the graph model of
the circuit. They represent the relationship between the TFI, TFO, CTFI and CTFO of the
old nodes V v v
j i
∈ , and the new merged node ' ' V v∈ after the execution of a local
move. These relationships can be easily derived from the definition of ' E that was
presented previously.
( ) ( ) ( ) { } { }
j i j i
v v v TFI v TFI v TFI , ' − ∪ =
( ) ( ) ( ) { } { }
j i j i
v v v TFO v TFO v TFO , ' − ∪ = (8)
( ) ( ) ( ) { } { }
j i j i
v v v CTFI v CTFI v CTFI , ' − ∪ =
( ) ( ) ( ) { } { }
j i j i
v v v CTFO v CTFO v CTFO , ' − ∪ =
43
v
4
v
3
v
2
v
1
Algorithmic Cycle
v'
2
v
3
v
1
Algorithmic Cycle

(a) (b)
v
4
v
3
v
2
v
1
Algorithmic Cycle
v
3
v
1
v'
2
Algorithmic Cycle

(c) (d)
v
5
TOKEN_BUFFER
v
6
v
4
v'
2
v
3
v
1
(e)
44
v
5
TOKEN_BUFFER
v
4
v'
2
v
3
v'
1
(f)
Figure 9 Three different local move scenarios that have different impact on
the
lg a
τ of the circuit. The circuit for the three scenarios before the
move ((a), (c), (e) respectively) and after ((b), (d), (f) respectively)
showing how the ACT could remain unaffected ((a),(b)), increase
((c),(d)), or even get introduced ((e),(f)).
Some examples of possible moves are shown in Figure 9. It is interesting to see
the different scenarios and the effects that they could have either on the weight function
( )
j i
e w
,
and/or on the algorithmic cycle of the circuit:
• In the first scenario, which is depicted in Figure 9 (a) and (b) we can see an
example of two nodes merging that are part of parallel paths and hence the
execution of the move does not affect the ACT of the circuit, since the levels
of logic in the nodes that are part of the path that defines the ACT are
unaffected.
45
• In the second scenario, which is depicted in Figure 9 (c) and (d) there is an
edge connecting the two nodes which gets absorbed. In this case the new
logic will artificially inflate the ACT since a new level of logic is added in a
node that is part of the critical path. Even though the levels do not really
change in terms of the actual data path, this node will now have a delayed
handshaking sequence due to the new logic, which affects the critical path. It
should be noted here that additional merging with v
3
could remove this effect
assuming this move was possible.
• In the third scenario, which is depicted in Figure 9 (e) and (f) there was no
cycle in the portion of the circuit that is depicted. However after merging the
two nodes v
1
and v
6
there is now a new cycle in the design. This cycle is
again not introduced in the logic, since the circuitry is not modified by
clustering, but introduced in terms of the control handshakes between the
different nodes of the graph.
Having defined the general mathematical framework it is now time to look into a
more formal representation of the clustering. In the following section we will define our
present goals from the clustering and some key theorems that are necessary for the
clustering to provide a functional solution that meets the user performance requirements.
3.2 Clustering Criteria
Clustering is a sequence of local moves that serve the purpose of minimizing
control area. We have chosen to define local moves as the merging of two nodes, since
all merges can be broken down into this basic two-way merge, and the two-way merge is
46
easier to characterize and study. Since every pipeline cluster will ultimately need to have
its own control unit as well as left and right C-element trees for multiple fanins and
fanouts, every local move results in a drop in total area.
The ultimate goal is to find the clustering of the circuit into pipeline stages that
achieves the minimum overall area while hitting a target performance. However
practically this means that this has to take into consideration not only the clustering
process, but also the effects of slack matching and fanout optimization. In general since
this is a new area of research we found that before even considering different
optimization algorithms and approaches, there were more fundamental problems that
need to be addressed before area optimality. As we will show in Section 5.2.2 we chose
a heuristic algorithm using a steepest descent approach and local constraints for the
area optimizations. The focus of this work was maintaining correctness and performance
during clustering. Having this foundation will allow us to further explore the optimization
process in the future.
3.2.1 Ensuring liveness
The handshaking nature of asynchronous circuits requires that one constraint is
satisfied to ensure the circuit is live [14] (also referred to as liveness of a circuit).
Informally a circuit is live, if every cycle in the circuit should have at least one data
TOKEN. This is guaranteed during the design process, by ensuring in every cycle in the
design at least one TOKEN_BUFFER cell. A TOKEN_BUFFER is a special gate in the
netlist that upon reset (or startup) will get initialized with a token (data). All other gates in
the netlist are empty during initialization.
47
Based on the definitions in Section 3.1 the liveness criteria can be formalized in
the context of our proposed graph model. A graph ( ) m h E V G , , , = is live if every cycle
i i
p
,
includes at least an edge e that starts at a node TB v∈ . Based on our convention,
equivalently the graph is live if every cycle
i i
p
,
includes at least an edge 0 ) ( : < e w e .
However, with arbitrary clustering it is easy to see how a cycle can get created
that generates a new cycle that violates this principle. So our first task is to make sure
that clustering does not destroy the liveness of a circuit and that we find a criterion that
allows us to prevent all moves that could cause that from ever being executed.
Several of our proofs are based on the modified graph
{ } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = , which does not include the two special edges that have the
possibility of being absorbed during a local move. The reason for this is that those are
the only two edges in the graph that are treated differently than others and by
considering the modified graph that does not include them we can generalize our theory
without having to consider the multitude of special cases. An important first conclusion is
that in such a modified graph cannot include any paths between the two merging nodes
that is of length 1. This is used in the proofs that follow.
Lemma 1: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . In the modified graph
{ } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = if
* , G n l
P p ∈ ∃ and ( ) ( ) j n i n j l i l = ∨ = ∧ = ∨ = then
1
,
>
n l
p .
48
Proof: Let us assume that
* , G n l
P p ∈ ∃ . If ( ) ( ) j n i n j l i l = ∨ = ∧ = ∨ = and 1
,
=
n l
p .
Then we will show that this statement cannot be true through contradiction. Indeed the
possible cases are as follows:
1. ( ) i n l = = ⇒ { }
G i i i i
P e p ∈ =
, ,
which contradicts our convention that E e
i i
∉
,

2. ( ) j n l = = ⇒ { }
G j j j j
P e p ∈ =
, ,
which contradicts our convention that E e
j j
∉
,

3. ( ) j n i l = ∧ = ⇒ { }
* , , G j i j i
P e p ∈ = which contradicts our proposition that
{ }
i j j i j i
e e E e
, , ,
, − ∉
4. ( ) i n j l = ∧ = ⇒ { }
* , , G i j i j
P e p ∈ = which contradicts our proposition that
{ }
i j j i i j
e e E e
, , ,
, − ∉
Since all possible combinations contradict our original convention or proposition it is
proven 1
,
>
n l
p .
Now we would like to show that every path that exists in the modified original
graph { } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = also exists in the resulting graph after the move. This
is an important conclusion since it helps us to easily qualify and evaluate the results of a
possible move and its consequences in terms of the connectivity of the graph.
Lemma 2: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . In the modified graph
{ } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = if
* , G n l
P p ∈ ∃ then
' , G q o
P p ∈ ∃ such that
Case 1: If ( ) ( ) j n i n j l i l ≠ ∧ ≠ ∧ ≠ ∧ ≠ then l o= and n q=
Case 2: If ( ) ( ) j n i n j l i l ≠ ∧ ≠ ∧ = ∧ = then k o= and n q=
Case 3: If ( ) ( ) j n i n j l i l = ∧ = ∧ ≠ ∧ ≠ then l o= and k q=
49
Case 4: If ( ) ( ) j n i n j l i l = ∧ = ∧ = ∧ = then k o= and k q=
Proof: First we are going to prove all the cases for the special case that a path
* , G n l
P p ∈ does not go though nodes V v v
j i
∈ , , unless
j i
v v , are one of the endpoints of
the path.
Case 1: ( ) ( ) j n i n j l i l ≠ ∧ ≠ ∧ ≠ ∧ ≠ Assume that ( )
n l
p v v e
, 2 1
, ∈ = ∀ it is true that
i
v v ≠
1
,
i
v v ≠
2
,
j
v v ≠
1
,
j
v v ≠
2
, meaning the path does not include any edge
that includes any of the merging nodes. Then based on the definition of the
local move
n l
p e
, 2 , 1
∈ ∀ and '
2 , 1
E e ∈ ∃ therefore
' , G n l
P p ∈ ∃ .
Case 2: ( ) ( ) j n i n j l i l ≠ ∨ ≠ ∧ = ∨ = ( )
n l
p v v e
, 2 1
, ∈ = ∀
i
v v ≠
2
,
j
v v ≠
2
, then V v
r
∈ ∃
such that { }
n r r l n l
p e p
, , ,
, = , and from the definition of the local move if
j l i l E e
r l
= ∨ = ∈ ∃ :
,
⇒ '
,
E e
r k
∈ ∃ . So if
* , G n l
P p ∈ ∃ and
( ) ( ) j n i n j l i l ≠ ∨ ≠ ∧ = ∨ = then { }
' , , ,
,
G n r r k n k
P p e p ∈ = ∃ .
Case 3: ( ) ( ) j n i n j l i l = ∨ = ∧ ≠ ∨ ≠ and ( )
n l
p v v e
, 2 1
, ∈ = ∀
i
v v ≠
1
,
j
v v ≠
1
, then
V v
r
∈ ∃ such that { }
n r r l n l
e p p
, , ,
, = , and from the definition of the local move if
j n i n E e
n r
= ∨ = ∈ ∃ :
,
⇒ '
,
E e
k r
∈ ∃ . So if
* , G n l
P p ∈ ∃ and
( ) ( ) j n i n j l i l = ∨ = ∧ ≠ ∨ ≠ then { }
' , , ,
,
G k r r l k l
P e p p ∈ = ∃
Case 4: Now if we assume that ( ) ( ) j n i n j l i l = ∨ = ∧ = ∨ = and the path does not
include nodes
j i
v v , on any edge other than its starting and ending point. Using
Lemma 1 1
,
>
n l
p , therefore we can write that { }
n r r l n l
p p p
, , ,
, = for some node
50
r
v along the path. But if
' , ,
*
G r k r l
P p G p ∈ ∃ ⇒ ∈ ∃ , and
' , ,
*
G k r n r
P p G p ∈ ∃ ⇒ ∈ ∃ , so { }
' , , ,
,
G k r r k k k
P p p p ∈ = ∃ .
We have now proven all four cases for paths that do not go through nodes V v v
j i
∈ , .
We need to also show that all cases also stand when the path does go through nodes
V v v
j i
∈ , .
Next we cover the case where
* , G n l
P p ∈ ∃ that goes through node
( ) ( )
j i
v v v v v = ∨ =
1 1 1
: and
m l
v v v v ≠ ∧ ≠
1 1
. But then we can write that
{ }
n l n l
p p p
, 1 1 , ,
, = . Using the conclusions of Cases 1-4 above on the sub-paths
n l
p p
, 1 1 ,
,
and for any value of n l, we know that if
' , 1 , G k o G l
P p P p ∈ ∃ ⇒ ∈ ∃ and if
' , , 1 G n k G n
P p P p ∈ ∃ ⇒ ∈ ∃ . This implies { }
' , , ,
,
G n k k l n l
P p p p ∈ = . So the relationships in
Cases 1-4 stand also when the path goes through the nodes V v v
j i
∈ ,
With Lemma 2 we have shown that the new graph ( ) m h E V G , , ' , ' '= after the
local move maintains all the connectivity as that in the modified initial graph
{ } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = . However, the initial graph ( ) m h E V G , , , = is our real initial
graph and therefore, it is useful to also show that the connectivity in G is maintained in
' G even if the two special edges
i j j i
e e
, ,
, where part of some original path.
Lemma 3: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . If
G n l
P p ∈ ∃
,
and { }
j i n l
e p
, ,
≠ , { }
i j n l
e p
, ,
≠
then
' , G q o
P p ∈ ∃ such that
Case 1: If ( ) ( ) j n i n j l i l ≠ ∧ ≠ ∧ ≠ ∧ ≠ then l o= and n q=
51
Case 2: If ( ) ( ) j n i n j l i l ≠ ∧ ≠ ∧ = ∧ = then k o= and n q=
Case 3: If ( ) ( ) j n i n j l i l = ∧ = ∧ ≠ ∧ ≠ then l o= and k q=
Case 4: If ( ) ( ) j n i n j l i l = ∧ = ∧ = ∧ = then k o= and k q=
Proof: Lemma 2 proves the proposition for
n l i j n l j i G n l
p e p e P p
, , , , ,
: ∉ ∧ ∉ ∈ ∀ . So we
need to prove the proposition for
n l i j n l j i G n l
p e p e P p
, , , , ,
: ∈ ∨ ∈ ∈ ∀ . Let us assume that
{ }
n j j i i l n l G n l
p e p p P p
, , , , ,
, , : = ∈ ∃ . Even though in '
,
E e
j i
∉ it is true that if
G i l
P p ∈ ∃
,
then
' , G k l
P p ∈ ∃ and if
G n j
P p ∈ ∃
,
then
' , G n k
P p ∈ ∃ therefore { }
n k k l n l G n l
p p p P p
, , , ' ,
, : = ∈ ∃ .
The three lemmas show that except for the single-edge path that connects the
two merging nodes, the new graph ' G includes all other paths that existed in G . This is
a very useful conclusion that will be used to show the necessary conditions that need to
be satisfied to maintain both liveness and performance on the new graph.
Theorem 1: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . The graph ' G is non-live iff in the modified
graph { } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = , ( )
j i
v CTFO v ∈ or ( )
i j
v CTFO v ∈ .
Proof:
(
←
) Assume in * G it is true that ( )
i j
v CTFO v ∈ . Then from Lemma 1,
1 :
, * ,
> ∈ ∃
j i G j i
p P p . Consequently ( )
j m i m
v v v CTFO v ≠ ∈ ∃ : and
j m m i
p p
, ,
, ∃ in * G that
only traverse combinational nodes. And from Lemma 2 it is also true that
k m m k
p p
, ,
, ∃ , so
there exists a combinational cycle, therefore the graph ' G is not live. Similarly we can
show the same if we assume that ( )
j i
v CTFO v ∈ , which proves the one side of the
proposition.
52
(
→
) Now let us assume that ' G is not live (but G was) and that ' ' V v
k
∈ is part of a
combinational cycle that was generated during the local move. We will assume that
( )
j i
v CTFO v ∉ and ( )
i j
v CTFO v ∉ and show that we reach a contradiction.
From the definition of the local move guarantees the absence of self-loops on nodes
so ' V v
m
∈ ∃ for which
k m m k
p p
, ,
, ∃ in
' G
P and ( )
m k
v CTFO v ∈ ' and ( )
k m
v CTFO v ' ∈ . But if
k m m k
p p
, ,
, ∃ in
' G
P it is true due to Lemma 2 that
n m m l
p p
, ,
, ∃ in
* G
P such that either:
• i l = and j n= ⇒ ( )
i j
v CTFO v ∈ , which contradicts our assumptions or
• j l = and i n= ⇒ ( )
j i
v CTFO v ∈ , which contradicts our assumptions or
• i n l = = ⇒ * G not live ⇒ G not live, which contradicts our assumptions or
• j n l = = ⇒ * G not live ⇒ G not live, which contradicts our assumptions.
Therefore all possible cases contradict our assumption, therefore if ' G is not live, it must
be that ( )
j i
v CTFO v ∈ or ( )
i j
v CTFO v ∈ , therefore our proposition stands.
Theorem 2: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . Graph ' G will be live if in the modified graph
{ } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = , 0
,
≤
j i
d and 0
,
≤
i j
d
Proof: From Theorem 1, we know that the new graph will not be live iff ( )
i j
v CTFO v ∈
or ( )
j i
v CTFO v ∈ . But if ( )
i j
v CTFO v ∈ ⇒ ( )
j i j i
p e e w p
, ,
0 : ∈ ∀ > ∃ by the definition of
CTFO . And from the definition of
j i
d
,
it is clear that 0
,
>
j i
d . So if
( )
i j
v CTFO v ∈ ⇒ 0
,
>
j i
d and similarly if ( )
j i
v CTFO v ∈ ⇒ 0
,
>
i j
d . Therefore if
0
,
≤
j i
d and 0
,
≤
i j
d , then ( )
j i
v CTFO v ∉ and ( )
i j
v CTFO v ∉ , so the resulting graph will
be live due to Theorem 1, which completes the proof.
53
However it should be noted that the reverse argument is not always true,
meaning that 0
,
>
i j
d does not guarantee that ( )
j i
v CTFO v ∈ . For example, consider a
situation of a long path from a PI through a TOKEN_BUFFER to a PO. It might be
possible in that case to have a path for which 0
,
>
j i
d , but which after the merge
includes a TOKEN_BUFFER in the cycle. Such an example if presented in Figure 10.
v
5
TOKEN_BUFFER
v
6
v
4
v
2
v
3
v
1
v
7
v
8
v
9
v
5
TOKEN_BUFFER
v
6
v
4
v'
2
v
3
v
1
v
7
v
8
2
2 2 2
2 -10 2 2
2 2 2 2 2 2
PO
PO
PI
PI
2
2
-10
2

Figure 10 An example where 0
,
>
i j
d does not guarantee that ( )
j i
v CTFO v ∈ .
Notice that 2
9 , 2
= d , but the path includes
5
v is a TOKEN_BUFFER
thus ( )
2 9
v CTFO v ∉ .
Therefore this criterion is weaker than the one presented in Theorem 1, in the
sense that it could exclude local moves that would not necessarily cause a non-live
graph. The value of Theorem 2 is that once we have obtained the distances for the
entire graph, the check can be implemented with a single lookup, while the operation of
finding the CTFO is much more computationally intensive. It is also much more effective
when combined with the performance criteria, which are discussed in the next section
and are also based on distances. For example if in we like to avoid creating an
algorithmic cycle that is longer than our target cycle time then this move should be
54
avoided anyway. This will become more apparent in the following section, where the
performance criteria are described.
3.2.2 Maintaining Performance
The goal of our analysis here is to define some criteria so that our local move
operations can maintain the performance of the original circuit. There are two
performance measures that are interesting here and one could potentially choose to
even enforce them separately. The first one has to do with the TCT or
goal
τ . In essence
what that means is that the local move should not introduce any new cycles that could
make the ACT any larger than the TCT thus making the circuit slower than requested by
the user. The other one has to do with end to end latency. It is sometimes important that
the PI-to-PO latency in a circuit does not increase. In that case we should be preventing
any moves from being executed that could increase the latency from any PI to any PO in
the circuit. Another limiting factor for the performance is the LCT. This is taken care of by
local criteria that prevent moves from being executed that would slow down a channel to
the point that it hurts performance. These are not discussed in this section but rather in
the implementation section.
In order to prove our criteria we are first going to prove a Lemma that will help
make our further discussion simpler. In particular we want to show that there are certain
distance criteria that can be used to easily determine, whether the graph is satisfying the
performance target for the TCT or not. This allows us to evaluate the performance of the
graph (circuit) by marely evaluating the distances between the different nodes of the
graph.
Lemma 4: In a live graph G the
lg a
τ is satisfied iff V v
i
∈ ∀ it is true that 0
,
≤
i i
d .
55
Proof: For all nodes that are not part of any cycle it is trivially proven since in that case
−∞ =
i i
d
,
. So we need to prove this for all nodes that are part of a cycle. So let’s assume
that the graph G is live and also meets the
lg a
τ . Then










= ≥
∑
∑
∈
∈
∈
i i k j
i i k j
i
p e
k j
p e
k j
V v
a goal
e m
e h
, ,
, ,
) (
) (
max
,
,
lg
τ τ ,or 0
) (
) (
max
, ,
, ,
,
,
≤ −










∑
∑
∈
∈
∈
goal
p e
k j
p e
k j
V v
i i k j
i i k j
i e m
e h
τ , and since
goal
τ is a
positive constant this can also be transformed as:
0
) (
) ( * ) (
max
, ,
, , , ,
,
, ,
≤









 −
∑
∑ ∑
∈
∈ ∈
∈
i i k j
i i k j i i k j
i
p e
k j
p e
k j
p e
goal k j
V v
e m
e m e h τ

We can further write that:
( ) { } [ ]
[ ]










− =










− = =
∑ ∑
∑
∈ ∈
∈
∈
∈ ∈
i i k j i i k j
i
i i k j
i i i
p e
k j goal
p e
k j
V v
p e
goal k j k j
V v
i i
G p
i i
e m e h
e m e h p w d
, , , ,
, ,
,
) ( * ) ( max
* ) ( ) ( max max
, ,
, , , ,
τ
τ

And since
∑
∈
i i k j
p e
k j
e m
, ,
) (
,
is the number of TOKEN_BUFFERS on that particular path and
therefore and the graph is live 1 ) (
, ,
,
≥
∑
∈
i i k j
p e
k j
e m , so:
0
) (
) ( * ) (
max
, ,
, , , ,
,
, ,
,
≤









 −
≤
∑
∑ ∑
∈
∈ ∈
∈
i i k j
i i k j i i k j
i
p e
k j
p e
k j
p e
goal k j
V v
i i
e m
e m e h
d
τ

We can also trivially prove the reverse side of that relationship by contradiction. Assume
that the graph is live, V v
i
∈ ∀ it is true that 0
,
≤
i i
d , but the
lg a
τ is violated, therefore
56
lg a goal
τ τ ≤ . Let us assume that the ACT is determined by a path
i i
p
,
for node V v
i
∈ .
Then similarly to before we can write that:
( ) [ ] [ ]
∑ ∑ ∑
∈ ∈ ∈
− = − = =
i i k j i i k j i i k j
p e
k j goal
p e
k j
p e
goal k j k j i i i i
e m e h e m e h p w d
, , , , , ,
) ( * ) ( * ) ( ) (
, , , , , ,
τ τ
But 0
,
≤
i i
d so
[ ] 0 ) ( * ) (
, , , ,
, ,
≤ −
∑ ∑
∈ ∈
i i k j i i k j
p e
k j goal
p e
k j
e m e h τ [ ]
∑ ∑
∈ ∈
≤ ⇔
i i k j i i k j
p e
k j goal
p e
k j
e m e h
, , , ,
) ( * ) (
, ,
τ
⇔
[ ]
goal
p e
k j
p e
k j
i i k j
i i k j
e m
e h
τ ≤
∑
∑
∈
∈
, ,
, ,
) (
) (
,
,
, since 1 ) (
, ,
,
≥
∑
∈
i i k j
p e
k j
e m
⇔
goal a
τ τ ≤
lg

Which contradicts the original assumption therefore the proposition stands.
Lemma 5: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . If G is live and the
lg a
τ and/or PI-to-PO
latency constraints are satisfied and a local move is executed, any violating path in
( ) m h E V G , , ' , ' '= will go through the newly formed node
k
v' .
Proof: We can prove this through contradiction. Assume that we have a graph G where
the
lg a
τ and/or PI-to-PO latency constraints are satisfied. Assume that the new node is
' V v
k
∈ that is the merge of V v v
j i
∈ , . Assume that there exists a path that violates
either the
lg a
τ or PI-to-PO latency constraints in * G that does not go through the new
node. However by the definition of the local move the edges into
k
v
are{ } { } { } { }
i j j i i n i l k m
e e E e E e E e
, , , , ,
, ' − ∈ ∪ ∈ = ∈ which means that if the path does not go
through ' V v
k
∈ it cannot go through either V v v
j i
∈ , . And since G and * G are
57
identical for all nodes and edges it means that if this violating path exists in * G it existed
in G as well. Therefore G must have a violating path, which contradicts our assumption.

Theorem 3: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . If G is live and the
lg a
τ constraints are
satisfied, the local move will not create a path violating the
lg a
τ constraint iff in the
modified graph { } ( ) m h e e E V G
i j j i
, , , , *
, ,
− = the following distance relationships are true
0
,
≤ +a d
j i
, 0
,
≤ +a d
j j
, 0
,
≤ +b d
i j
and 0
,
≤ +b d
i i
where
( )



∈ ∃
∈ ¬∃
=
E e if e w
E e if
a
i j i j
i j
, ,
,
,
, 0

and
( )



∈ ∃
∈ ¬∃
=
E e if e w
E e if
b
j i j i
j i
, ,
,
,
, 0

Proof:
Case 1: Let us first assume that E e e
i j j i
∉
, ,
. . Then based on the definition of the local
move the weights of the edges of E are the same as those in ' E . From Lemmas 3, 4
and 5 we can write the following relationship for the new distances ' d in ' G and the old
ones d in G :







≤
≤
≤
≤
⇔ ≤
0
0
0
0
0 '
,
,
,
,
,
i j
j i
j j
i i
k k
d
and d
and d
and d
d
In fact the first two relationships are true since
i i
d
,
and
j j
d
,
must be smaller or equal to
0 since G satisfies the
lg a
τ .
58
Case 2: Let us first assume that E e E e
i j j i
∈ ∧ ∉
, ,
. Then
k k k l
p e
, ,
' ∈ ∃ for which we know
by definition that ( ) ( ) ( )
k l i j j l
e w e w e w
, , ,
' ≥ + . This implies that the path that defines
k k
d
,
'
could be greater by ( )
i j
e w
,
than the distance of the path it originated from if
i j
e
,
was not
part of the original path of iit would otherwise be smaller or equal to it. This means that
we can write:
( )
( )







≤
≤ +
≤ +
≤
⇔ ≤
0
0
0
0
0 '
,
, ,
, ,
,
,
i j
i j j i
i j j j
i i
k k
d
and e w d
and e w d
and d
d
The reason why the ( )
i j
e w
,
is not added to the other two paths is that since they end in
node
i
v it is impossible that they could be followed by
i j
e
,
.
Case 3: Let us first assume that E e E e
i j j i
∉ ∧ ∈
, ,
. Then
k k k l
p e
, ,
' ∈ ∃ for which we know
by definition that ( ) ( ) ( )
k l j i i l
e w e w e w
, , ,
' ≥ + . This implies that the path that defines
k k
d
,
'
could be greater by ( )
j i
e w
,
than the distance of the path it originated from if
j i
e
,
was not
part of the original path or it would otherwise be smaller or equal to it. This means that
we can write:
( )
( ) 






≤ +
≤
≤
≤ +
⇔ ≤
0
0
0
0
0 '
, ,
,
,
, ,
,
j i i j
j i
j j
j i i i
k k
e w d
d
d
e w d
d
The reason why the ( )
j i
e w
,
is not added to the other two paths is that since they end in
node
j
v it is impossible that they could be followed by
j i
e
,
.

59
Theorem 4: Let a local move merge two nodes V v v
j i
∈ , in graph ( ) m h E V G , , , = into
node ' ' V v
k
∈ in the graph ( ) m h E V G , , ' , ' '= . If G and ' G are live and satisfy the
goal
τ
constraint, the move will not increase the latency of any path from PI to PO iff PI v
m
∈ ∀
and PO v
l
∈ ∀ it is true that:
{ } { }
l i l j j m i m l m
d d d d d
, , , , ,
, max , max + ≥
Proof: The latency between PI v
m
∈ ∀ and PO v
l
∈ ∀ can be described by the distance
between
m
v and
l
v that is
l m
d
,
. The distance according to Lemma 5 will not increase if
the length of all paths through the new node
k
v' have length smaller or equal to
l m
d
,
.
We denote with d the distances of paths in
G
P and with ' d the paths in
' G
P . The length
of the longest path through
k
v' can be written as
l k k m
d d
, ,
' ' + therefore we can write that
the statement will be true as long as
l k k m l m
d d d
, , ,
' ' + ≥ . For simplicity we are going to
break down the proof into three cases:
Case 1: If E e E e
i j j i
∉ ∧ ∉
, ,
. Assume that the local move generates a path from node
m
v to node
l
v . But since E e E e
i j j i
∉ ∧ ∉
, ,
based on the definition of the local move no
edge will change weight so } , max{ '
, , , j m i m k m
d d d = and similarly } , max{ '
, , , j l i l l k
d d d = . In
other words the longest path from the input
m
v to the new node
k
v' is equal to the
longest path to either of the two nodes that the move merges. Similarly the longest path
from the new node
k
v' to the output
l
v is the longest path from either of the two nodes
that the move merges. So the longest path from input
m
v to output
l
v that goes through
60
the new node
k
v' will have a length of { } { }
l i l j j m i m l k k m
d d d d d d
, , , , , ,
, max , max ' ' + = + . In
order for the latency not to increase from input
m
v to output
l
v we can therefore write:
{ } { }
l i l j j m i m l k k m l m l m
d d d d d d d d
, , , , , , , ,
, max , max ' ' ' + = + ≥ =
Case 2: If E e E e
i j j i
∉ ∧ ∈
, ,
. In this case we know by definition that
( ) ( ) ( )
k m j i i m
e w e w e w
, , ,
' ≥ + and ( ) 0
,
>
j i
e w , so a path in the original graph from input
m
v
to node
i
v could get longer by ( )
j i
e w
,
. However it is true that
( )
j i i m j i i m j m
e w d d d d
, , , , ,
+ ≥ + ≥ . So the equation proven in Case 1 still holds, in fact we
can simplify it and write { }
l i l j j m l k k m l m l m
d d d d d d d
, , , , , , ,
, max ' ' ' + = + ≥ =
Case 3: If E e E e
i j j i
∈ ∧ ∉
, ,
. Similarly due to symmetry with Case 2 the equation that
was proven in Case 1 holds and can be simplified and written as
{ }
l i l j i m l k k m l m l m
d d d d d d d
, , , , , , ,
, max ' ' ' + = + ≥ = .
3.2.3 Modified Floyd-Warshall for finding distances
The Floyd-Warshall algorithm is used to find all pair-wise distances in the graph
so that the constraints set above can be checked quickly by a simple look-up. The
algorithm is originally designed to find all the minimum pair-wise distances in a graph
and cannot be used if there are any negative-weight cycles in the graph. This works well
for our case, because we need to find the maximum distances between all pairs of
nodes. And assuming that the
lg a
τ is met in the graph originally according to Lemma 4
V v
i
∈ ∀ it is true that 0
,
≤
i i
d . Therefore there are no positive weight cycles in the
graph, which allows to replace all min operations in the original algorithm with max
operations, and still achieve convergence.
61
The complexity of the algorithm is ( )
3
V Θ , so it is very expensive
computationally. Moreover, the complexity does not change no matter how many nodes
really need to be updated. After each local move is executed, the pairwise distances in
the graph change and an update is required. Running the entire Floyd-Warshall
algorithm was attempted, but it was quickly realized that this was impractical and so slow
that made the use of the distance-based algorithms impractical.
A local update has been implemented that updates the array with simple
operations only around the neighborhood of the new node after each move so that the
Floyd-Warshall algorithm needs to run only once. However, the complexity of even one
execution of the algorithm is really prohibitive for circuits that include tens of thousands
of nodes. This is one of the reasons why when it comes to liveness two theorems were
developed. Theorem 1 only requires a local search and in very large netlists where
calculating all the distances is hard to do, one can choose to ignore performance in
order to obtain results quickly and still maintain a functional circuit in the end of the
operation, which however may or may not meet the performance requirements.
The local update that was developed to speed up the processing between move
executions is extremely fast, reducing the overhead of the Floyd-Warshall to practically
just the initial run that finds the initial distances. It takes advantage of the knowledge of
the graph interconnect as well as the nature of the move so that it can speed up the
processing and avoid updating any unnecessary values. The algorithm is shown in
Figure 11.
The update function for the distances (update_distance_from_node) has
complexity ( ) N B Ο , where B is the average branching factor at each node and N is
62
the number of nodes in the graph. It is based on the realization that each path from a
node has to go through one of its fanout nodes, therefore all the distances from a given
node can be calculated using just the distances to its fanout nodes and the distance
vectors stored at each of the fanout nodes. This version of the algorithm was used
extensively in our implementation of the distance algorithm presented in 3.2.2 and in
practice it performs several times better than the Floyd-Warshall algorithm, however
theoretically its performance could not be proven to be any better than the ( )
3
N Ο that
the Floyd-Warshall algorithm achieves, since this algorithm will not prevent a node from
being visited several times.

Figure 11 Pseudo-code of the first distance-update routine. This version was
found to be very fast in practice, but its performance is not bound.
63
A slightly modified version was then generated that has a performance that can
be bound by ( )
2
N B Ο , where B is the average branching factor at each node and N is
the number of nodes in the graph. The modified version is shown in Figure 12. This
algorithm takes advantage of the fact that the pair wise distances in the rest of the graph
did not change. So first the distances of the new node to all other nodes are reset and
an update is executed to calculate the distances from the new node to all other nodes
using just information from its fanout. Any path from the new node to any other node has
to go through its fanout so this operation is enough to give us the new distances. Then
the update is executed on the fanins of the new node, so that any new paths generated
by the new node are updated on its fanin. Since the distance from all other nodes to the
fanins of the new node did not change a final update on all nodes of the graph using
these three nodes is enough to update the entire graph. This last loop is the longest
operation in this update algorithm and is executed ( ) N B Ο times each containing
N updates so the total complexity of the update proposed is ( )
2
N B Ο . For the typical
circuits that we have studied it is true that the branching factor B is negligible in size
compared to the number of nodes N in the graph and thus for such graphs we can say
that the complexity of the update is of complexity approximately equal to ( )
2
V Ο .
This algorithm was designed, but not implemented. The previous version was
proven to be very fast in practice and the results were satisfactory, so this new algorithm
was not deemed necessary. Thus no run-time comparisons were generated to compare
the two.

64

Figure 12 Pseudo-code of the second distance-update routine.
This version can be shown to have performance ( )
2
BN Ο , where B is
the average branching factor at each node and N is the number of
nodes in the graph.

65
3.3 Experimental Results
Both aspects of the clustering theory were tested using our software platform and
ASIC flow that will be discussed in Chapter 5. In all cases we use the same clustering
core algorithm, which is a greedy, steepest descent algorithm that picks the local moves
that reduce the total area of the design the most. It stops when the algorithm determines
that there are no possible moves left that do not violate the rules defined in each case.
Since our slack matching and clustering rules also depend on the assumption that all
LCTs are smaller all equal to the
lg a
τ and hence not critical, another local check was
added that ensures that no moves are permitted that would grow an LCT beyond the
allowed TCT target. Depending on the design style selected, other minor local
constraints are also enforced to guarantee legal circuit implementations for the particular
design style, but these details are minor and are not going to be discussed here.
We use for testing various examples from the ISCAS benchmark set that
includes examples that are purely combinational (hence include no cycles) or mixed
sequential and combinational (include state elements and cycles). We also added a few
examples that are generally common in commercial circuit designs. We use four variants
of clustering for each of the netlists. The first one applies no rules to the greedy
clustering. The second one only applies the liveness rule described in Theorem 1. The
third one adds the TCT constraints of Theorem 3 and also uses the expressions in
Theorem 2 for liveness to reduce the computational complexity. Finally the last one adds
the criteria of Theorem 4 to control the PI-to-PO overall latency of the circuit. For the
examples we report the logic area after clustering, the total circuit area after slack
matching and the GCT as measured by our software after clustering and slack matching
is complete.
66
Local & Global Cycle Time (transitions)

Area
Guided
Liveness
Liveness
& TCT
Liveness,
TCT &
Latency
Design LCT ACT LCT ACT LCT ACT LCT ACT
c3540 28 DL 32 32 34 34 30 30
MAC16 28 68 32 32 34 34 28 28
MAC32 32 72 32 32 32 32 32 32
s1196 34 34 34 34 32 32 30 30
s1238 28 DL 32 32 34 34 30 30
s13207 28 DL 32 32 30 30 30 30
s1423 28 DL 32 50 32 32 32 32
s1488 30 30 30 30 32 32 28 28
s15850 30 DL 38 64 34 34 34 34
s27 26 26 26 26 24 26 24 26
s298 26 26 26 26 26 26 26 26
s344 26 26 26 26 28 28 26 26
s349 26 26 26 26 26 26 26 26
s382 26 DL 26 26 26 26 26 26
s386 28 28 28 28 26 26 26 26
s400 26 DL 26 26 26 26 26 26
s420 26 26 26 26 26 26 26 26
s444 26 DL 30 30 26 26 26 26
s510 28 28 28 28 28 28 28 28
s526 26 DL 32 32 28 28 30 30
s5378 32 32 32 32 32 32 30 30
s641 26 DL 32 32 28 28 26 26
s713 26 DL 32 32 32 32 26 26
s820 28 DL 32 32 30 30 28 28
s832 32 32 32 32 32 32 28 28
s838 26 DL 32 32 30 30 26 26
s9234 32 32 32 32 32 32 26 26
s953 28 DL 32 32 32 32 26 26
SISO 34 DL 36 76 46 46 30 34
Table 1 Performance of the different circuits in terms of Local and Algorithmic
Cycle Time under different local move approval criteria. It is worth
noting that without the use of our proposed criteria more than half of
the circuits resulted in deadlock (marked DL).
Table 1 shows that 15 of 29 designs (51.7%) resulted in deadlock and were not
functional after clustering. So it is clear that it is very important to be able to maintain the
liveness of the circuit throughout these transformations, since the likelihood of destroying
its functionality is very high. It is also important to note that a non-live circuit will also fail
67
slack matching due to infeasible constraints, therefore in most cases the area guided
results are skewed since there were no slack buffers added. Since such results are not
meaningful area comparisons for the area-guided approach are not further analyzed.
Logic Area (μm
2
) Total Area (μm
2
)
Design Liveness
Liveness
& TCT
Liveness,
TCT &
Latency
Liveness
Liveness
& TCT
Liveness,
TCT &
Latency
c3540 53906.7 55238.4 67044.1 95369.5 92305.2 89650.9
MAC16 73659 79578 123170 218770 220988 209570
MAC32 61990.7 70978.2 84036.9 111006 140453 118325
s1196 22883.9 24017.5 27561 41813.6 45172.8 42873.4
s1238 24158.6 24916.1 28344.4 51539.3 54246 44670.5
s13207 132051 134391 144565 265536 273949 254701
s1423 26546.64 28230.9 31879.8 33606.1 59998.5 58762.9
s1488 29227.4 30055.7 31940.4 80592.8 80444.2 71767.3
s15850 117041.6 182966 182966 150210 297931 297931
s27 387.072 728.066 705.026 1013.76 1041.41 1018.37
s298 5170.18 5276.16 5234.69 8787.46 8865.79 8603.14
s344 5852.18 6165.54 6690.8 11796.5 12192.8 11699.7
s349 5824.5 6128.6 6667.8 12215.8 11828.7 11621.4
s382 6981.09 7326.7 7787.53 12575.2 13261.8 12722.7
s386 6420.66 6494.43 6570.98 11858.1 11259.1 12630.5
s400 7068.65 7455.76 7847.39 11994.6 12432.4 11713.5
s420 7225.34 7414.25 8262.14 15091.2 15390.7 14114.3
s444 6469.61 6907.43 7828.96 11649 12073 11593.7
s510 11920.9 12354 12570.7 31085.6 30034.9 27795.5
s526 8068.6 8193 9202.22 16132.6 16372.2 15699.5
s5378 61099.2 63975.6 68370.9 127436 129294 121197
s641 7326.73 7294.5 9400.3 15777.8 14667.3 14008.3
s713 7308.3 7345.12 9321.96 17459.7 14418.4 14008.3
s820 14742.8 15249.6 15931.6 33202.4 33225.4 29654.2
s832 13916.2 14303.3 15727.1 28781.6 30795.3 26993.7
s838 15694.9 15819.3 16985.1 34260.5 36159 29205.5
s9234 46361.1 48940.4 51255.4 79276 95743.9 91948.6
s953 18312.2 18266.1 20312.09 33039.4 34956.3 29707.8
SISO 22939.8 43118 100449.6 22939.8 167119 114762
Table 2 Area summary for different examples using different local move
approval criteria. The left columns correspond to the logic area, which
is the area of the clustered gates before the introduction of slack
matching. The columns on the right correspond to the total circuit area
after slack matching.
68
% Logic Area Increase % Total Area Increase
Design Liveness
Liveness
& TCT
Liveness,
TCT &
Latency
Liveness
Liveness
& TCT
Liveness,
TCT &
Latency
c3540 0.0% 2.5% 24.4% 0.0% -3.2% -6.0%
MAC16 0.0% 8.0% 67.2% 0.0% 1.0% -4.2%
MAC32 0.0% 14.5% 35.6% 0.0% 26.5% 6.6%
s1196 0.0% 5.0% 20.4% 0.0% 8.0% 2.5%
s1238 0.0% 3.1% 17.3% 0.0% 5.3% -13.3%
s13207 0.0% 1.8% 9.5% 0.0% 3.2% -4.1%
s1423 0.0% 6.3% 20.1% 0.0% 78.5% 74.9%
s1488 0.0% 2.8% 9.3% 0.0% -0.2% -11.0%
s15850 0.0% 56.3% 56.3% 0.0% 98.3% 98.3%
s27 0.0% 88.1% 82.1% 0.0% 2.7% 0.5%
s298 0.0% 2.0% 1.2% 0.0% 0.9% -2.1%
s344 0.0% 5.4% 14.3% 0.0% 3.4% -0.8%
s349 0.0% 5.2% 14.5% 0.0% -3.2% -4.9%
s382 0.0% 5.0% 11.6% 0.0% 5.5% 1.2%
s386 0.0% 1.1% 2.3% 0.0% -5.1% 6.5%
s400 0.0% 5.5% 11.0% 0.0% 3.6% -2.3%
s420 0.0% 2.6% 14.3% 0.0% 2.0% -6.5%
s444 0.0% 6.8% 21.0% 0.0% 3.6% -0.5%
s510 0.0% 3.6% 5.5% 0.0% -3.4% -10.6%
s526 0.0% 1.5% 14.0% 0.0% 1.5% -2.7%
s5378 0.0% 4.7% 11.9% 0.0% 1.5% -4.9%
s641 0.0% -0.4% 28.3% 0.0% -7.0% -11.2%
s713 0.0% 0.5% 27.6% 0.0% -17.4% -19.8%
s820 0.0% 3.4% 8.1% 0.0% 0.1% -10.7%
s832 0.0% 2.8% 13.0% 0.0% 7.0% -6.2%
s838 0.0% 0.8% 8.2% 0.0% 5.5% -14.8%
s9234 0.0% 5.6% 10.6% 0.0% 20.8% 16.0%
s953 0.0% -0.3% 10.9% 0.0% 5.8% -10.1%
SISO 0.0% 88.0% 337.9% 0.0% 628.5% 400.3%
Average 0.0% 11.5% 31.3% 0.0% 30.1% 15.9%
Average excluding SISO,s1423,s15850 0.0% 2.6% -4.4%
Table 3 Area increase for different examples using different local move
approval criteria On the columns on the left we see the area increase for
the clustered netlist before the addition of slack matching. On the
columns of the right we see the area increase for the final netlist after
slack matching,
From the Table 2 and Table 3 we can see that the simpler CTFO criterion-based
algorithm performs the most amount of clustering, as expected, due to fewer constraints.
The logic area for this algorithm is the smallest of the three variants analyzed. The
69
addition of the ACT maintenance criteria poses extra constraints, yielding clustered
netlists that have 11.5% more area. There is a total area increase on average of 30.1%
once the slack matching is factored in. However if we exclude the three examples where
CTFO yielded extremely large GCT essentially requiring no slack matching (SISO,
s15850 and s1423) the average increase in total area is only 2.6%. Adding the latency
constraints further constrains clustering, and one can see that the clustered area
increase is now 31.3%. The slack matching cost drops though, since the latency
constraints were intended to force better alignment of paths and reduce slack matching
as well as maintaining latency. On average after slack-matching, this algorithm suffers a
15.9% area increase. In fact, if we again exclude the 3 examples that the liveness-only
algorithm yielded excessive GCT on, this third algorithm is actually 4.4% better on
average than the liveness-only one on total area. If one also factors in the fact that the
latency-aware algorithm beat the others almost uniformly on the final GCT it is clear that
it is the best one of the three.
Intuitively this is expected as this additional criterion is likely to force better
alignment of nearby pipeline stages and prevent merging of stages that are further apart
from each other. In general given that this algorithm is able to maintain performance and
liveness. However it should be pointed out that this algorithm requires all pairwise
distances in the circuit, which is a computationally expensive operation. On the other
hand the liveness-only one has small complexity and thus may be more practical for
larger circuits.

70
Chapter 4 Slack Matching and Fanout
Considerations
4.1 Introduction
This section will discuss the concept of slack matching and improvements that
have been made to existing algorithms. Although describing the entire theory behind
slack matching as well as its modeling and formulation is beyond the scope of this work,
it is useful to show the basic principle behind its use and motivation through an example.
This allows the reader to understand the general context of the optimizations described
here, without getting into some of the more intricate details of the process.
Slack matching can be thought as the process of properly aligning the timing of
the handshakes between the pipeline stages in the design, so that a circuit can
maximize its performance. If a stage generates data late, forcing another stage to wait
then the receiving stage will have to stall its other inputs as it cannot process without all
its inputs being present. This forward latency matching is straightforward and analogous
to the latency matching that is frequently performed in synchronous circuits. In
asynchronous circuits, though, it is also true that there is a backward latency, that
defines the time it takes a stage to reset itself and get ready to receive data again. This
forces an alignment constraint for the backward latencies as well, which is harder to
visualize and which is realized in re-convergent paths. In both cases the mismatches
can be handled by adding additional pipeline stages called slack buffers. The slack cells
commonly are faster than regular logic cells and that allows them to be able to address
larger mismatches than a common cell due to their excess slack which is defined that
the difference between the circuit GCT and the cell LCT.
71

Tok
Buf
l
0
= 2
τ
0
=10
l
1
= 2
τ
1
=10
l
2
= 2
τ
2
=10
l
3
= 2
τ
3
=10
l
4
= 2
τ
4
=10
l
5
= 2
τ
5
=10
Token arrives at t = 0
Token arrives at t = 10
(stalled by 4 w hile
output channel resets)
Token arrives at t = 24
Token arrives at t = 34
(stalled by 4 w hile
output channel resets)
Token arrives at t = 6
Token arrives at t =16
Token arrives at t = 30
(starved by 4)
Token arrives at t = 40
Token arrives at t = 2
(stalled by 4 w hile
w aiting for other input)
Token arrives at t = 16
Token arrives at t = 26
(stalled by 4 w hile
w aiting for other input)
Token arrives at t = 40
Token arrives at t = 2
Token arrives at t = 12
Token arrives at t = 26
(starved by 4)
Token arrives at t = 36
Token arrives at t = 4
Token arrives at t = 14
Token arrives at t = 28
(starved by 4)
Token arrives at t = 38
Token arrives at t = 0
Token arrives at t = 10
Token arrives at t = 24
(starved by 4)
Token arrives at t = 34
Token arrives at t = 8
Token arrives at t =18
(stalled by 4 w hile output
channel rests)
Token arrives at t = 32
Token arrives at t = 42
(stalled by 4 w hile output
channel rests)

Figure 13 This example illustrates the effects of improper slack matching in a
pipeline. The circuit shown above comprises pipeline stages that
can run at a 10-transition cycle, however due to the mismatch the
circuit cannot run faster than a 12 transition cycle.
The above intuitive explanation is illustrated in Figure 13. Notice that all pipeline
stages can run at a cycle time of 10- transition delays. This should enable the circuit to
also run at that speed, but due to improper slack matching it can not go faster than an
average cycle of 12 transitions. One can start tracing the causes of this slow down by
looking at the arrival times of the data at the different stages assuming that data is
launched at time t = 0. Notice that the token from the top branch arrives at stage 4 at
time t = 2. However the other branch is not ready causing it to wait for 4 transitions. This
is the first stall and is happening due to the top path being too fast. Now stage 5 can only
start resetting at time t = 6 when the data is used by stage 4, so it cannot accept data
until t = 14. This in turn causes a stall between stages 0 and 5. The stalls inevitably keep
72
propagating backwards, and on average it turns out that the circuit can never exceed 12
transitions per cycle performance on average.
The slack matching problem is commonly modeled using a Petri net model
commonly referred to as the Full-Buffer Channel Net model proposed in [5]. As the name
suggests the underlying assumption is that the stages are full-buffer stages, although in
practice it has been shown to work well even for half-buffer cells. The problem can be
presented as MILP problem, but due to complexity it has been also approximated in [5]
with an LP. This section is devoted on optimizations that can be performed on this
original formulation.
4.2 Slack Matching Fanout Tree Improvements
Slack matching adds buffers on the connections between pipeline stages to
enhance the performance of the circuit. In reality slack matching creates a small buffer
tree at the output of a cluster that ensures the alignment of the data at the leaf cells of
the tree at the desired times. The cost of slack matching is high in many asynchronous
circuit templates that could account to up to 33% of the total circuit area. Therefore it is
worth investigating improvements to the existing models and methods, which could be
used to reduce the area overhead that makes this process so costly.
Slack matching is done using the Full-Buffer-Channel-Net (FBCN) model as
presented in [5]. In the FBCN model slack matching is done on channels. Channels are
point-to-point edges in the graph and are used to describe abstractly a connection
between two pipeline clusters or nodes in the graph. So if a connection between the two
73
exists, irrespective of how many physical wires it includes a single channel is created.
The formulation is fairly straight forward:
( ) V v v f h s f e w a a
j i slack slack j i j i j i i j
∈ ∀ + ∗ + + + = , , ) (
, , ,

Goal Function is to Minimize
∑ j i
s
,

In this equation,
j i
a a , represent the arrival time at node V v v
j i
∈ , .
j i
f
,
is the free
slack of the channels and is equal to the difference between the TCT and the LCT for
the particular channel.
j i
s
,
represents the number of slack cells that need to be
introduced on the channel so that the circuit is slack-matched.
slack
h is the forward
latency of a slack cell, which is a constant that is known at design time and
slack
f is the
free slack of the slack cell, which is also constant and corresponds to the difference
between the TCT and the LCT if the slack cell is inserted and depending on the
complexity of the circuit can either be precisely calculated or estimated. This is because
depending on the design style the LCT of the slack cell could depend on the fanin and/or
fanout and/or width of the bus that is routed through it, none of which parameters are
known in advance and they actually depend on the solution of the arrival time problem.
This model is fairly accurate for simple pipelines and it has been proven to work
well in most cases. However, this formulation does not account for hardware
optimizations that could be made and could make this model more accurate. In most
design styles, it is possible that several channels leave a particular cluster for several
different destinations. However all these channels could be the same wire actually
forking to many targets in the netlist. So in this case it is possible to actually share any
slack buffers that need to be placed for channels from the same source. The original
74
formulation does not explore this possibility and could think that several buffers are
needed in a location that one buffer could be used for all the channels that run in
parallel.
In most cases – assuming that we merge the buffers after slack matching
anyway – this inaccuracy actually just results in the LP program overestimating the
number of buffers needed. However, this problem could be further exaggerated in some
cases if the LP formulation concludes that buffers are better placed in a different
location. For example if a 2-input cluster has 3 fanout channels due to a wire feeding into
3 other clusters the original formulation would make the LP solver conclude that it is
better to place buffers before the cluster than after it, and ultimately use 2 slack buffers
instead of 1. This is shown in Figure 14.
A modification was made to the goal function of the LP problem to reflect this
inaccuracy from the goal function. The formulation of the arrival times remains identical,
but an additional set of parameters is required to be stored and used for evaluation
along with some extra constraints that are however linear in terms of the problem size
(number of nodes). So again:
( ) V v v f h s f e w a a
j i slack slack j i j i j i i j
∈ ∀ + ∗ + + + = , , ) (
, , ,

But also we add the extra parameters
i
z and respective constraints such as
i j i
z s ≤
,

Goal Function is to Minimize
∑ i
z
75
And since the minimization of the sums of
i
z is the goal function it will be true
that { }
j i
j
i i j i
s z z s
, ,
max = ⇔ ≤ . With this formulation, the linear program solver will
attempt to find the best solution assuming that it can use a line of buffers that can be
shared among all the outgoing channels. This is accurate for cases that slack buffers
can support arbitrary fanout and the only cases where it is problematic is situations
where a particular design styles poses a hard limit on the fanout of every gate/node. In
those cases it might be necessary to have buffer trees rather than buffer lines for slack
and in that case our formulation is inaccurate and also optimistic.
(a)
(b)
(c)

Figure 14 A case where the original slack-matching formulation yields sub-
optimal placement of slack buffers. The original circuit in (a)
requires on buffer stage. However the old formulation would decide
to place the buffers on the input of the gate assuming no sharing (b).
The new formulation models the sharing and places the buffer after
the gate (c).
76

4.3 Experiment Setup
In order to test the improvements that can be accomplished by the sharing of the
buffers an experiment was setup that focuses on the case of PCHB pipelines. The
experiment involved slack matching a netlist that was using the PCHB design style. The
formulation was based on the LP approximation described in [5] with the use of a ceiling
function. That is the amount of slack per channel is formulated and solved as a real
number and the result of the LP problem is rounded up to give the actual number of
buffer stages that will be instantiated. This was done because the MILP formulation that
is actually accurate, where the slack variables are actually described and solved as
integer parameters, is too slow and not practical as also observed in [5] and only very
few extremely small examples could be tested.
The objective of our tests was to show the benefits of the buffer-sharing concept
while slack matching and characterize the improvements that can be realized. The
results are subject to the LP approximation inaccuracy, but have practical value since
this approximation allows us to test on larger circuits with tens of thousands of gates,
which are closer to real life examples that this method could be applicable to. There are
a couple of additional constraints to the PCHB template that was used that slightly affect
the accuracy of the results. Those include maximum fanout constraints on gates and an
additional requirement that Primary Inputs (PIs) should only fanout to a single gate.
When either constraint is violated additional buffers need to be added to yield a
functional circuit. Our software was designed to yield functional circuits that can be
simulated in Verilog for correctness, therefore those constraints needed to be enforced
in order to be able to verify the functionality and performance of the final netlists.
77
4.4 Experimental Results
For the sake of our experiments we have used 26 examples from the standard
ISCAS benchmark circuits as well are 3 of our own examples that represent different
sized mathematical operations (2 Multiply Accumulate Units) and a SISO module (Soft-
In-Soft-Out used in Error Correcting Decoders in Telecommunication Systems). The
results are summarized in Table 4.
78
Orig. LP - No Sharing Orig. LP - Sharing New LP - Sharing
Example Insts
Buffer
Area
Total Area
Buffer
Area
Total
Area
Buffer
Area
Total
Area
c3540 731 6804.17 27225 5235.15 25474.9 3982.69 24121.5
MAC16 1604 11344 48254.1 9893.84 47153.7 8519.73 45508.6
MAC32 1092 8584.7 32071.7 6780.67 30150.1 5696.87 29015.2
s1196 387 4357.32 14964.5 3472.59 13929.1 3388.26 13763.2
s1238 409 4866.05 16109.1 3852.75 14918.9 3693.77 14450.2
s13207 1738 14034.1 52840.9 12908.9 51906.4 8720.18 47919.5
s1423 443 2134.43 12430.5 1835.83 12138.9 1581.47 11901.1
s1488 455 2842.21 14061.8 2021.07 13535.1 1679.62 13275.2
s15850 2229 12408.4 64173.8 11310.8 63170.2 8967.63 60804.9
s27 9 99.5328 299.981 99.5328 299.981 77.4144 277.862
s298 76 597.197 2144.1 503.194 2050.1 470.016 2012.77
s344 82 796.262 2641.77 707.789 2589.24 508.723 2322.43
s349 80 608.256 2430.26 594.432 2437.17 519.782 2368.05
s382 101 829.44 3035.75 707.789 2903.04 663.552 2827.01
s386 89 619.315 2683.24 505.958 2528.41 450.662 2470.35
s400 102 785.203 3002.57 693.965 2907.19 637.286 2822.86
s420 99 1396.22 3533.41 1238.63 3341.26 663.552 2786.92
s444 99 763.085 3015.01 713.318 2956.95 677.376 2894.75
s510 172 1205.45 5374.77 879.206 5213.03 666.317 5011.2
s526 125 873.677 3581.8 821.146 3532.03 837.734 3543.09
s5378 960 7633.61 28001.9 6571.93 27002.4 4813.52 25161.1
s641 135 2092.95 5059.58 1889.74 4788.63 731.29 3642.62
s713 134 2012.77 4954.52 1891.12 4776.19 641.434 3508.53
s820 219 1504.05 6947.94 1241.4 6852.56 740.966 6245.68
s832 213 1492.99 6672.84 1222.04 6498.66 803.174 6097.77
s838 208 2233.96 6903.71 2054.25 6711.55 1194.39 5847.55
s9234 632 3859.66 18323.7 3157.4 17592.4 2742.68 17403
s953 281 1957.48 9277.29 1515.11 8760.27 1338.16 8555.67
SISO 1300 1470.87 28431.8 1426.64 28391.7 1166.75 28126.3
Table 4 Slack matching results for the sharing and no-sharing versions of the
algorithm

79
Area Savings
Example
Buffer
vs. Orig.
LP - No
Sh.
Buffer
vs.
Orig.
LP -
Sharing
Total
vs.
Orig.
LP - No
Sh.
Total
vs.
Orig.
LP -
Sharing
c3540 41.47% 23.92% 11.40% 5.31%
MAC16 24.90% 13.89% 5.69% 3.49%
MAC32 33.64% 15.98% 9.53% 3.76%
s1196 22.24% 2.43% 8.03% 1.19%
s1238 24.09% 4.13% 10.30% 3.14%
s13207 37.86% 32.45% 9.31% 7.68%
s1423 25.91% 13.86% 4.26% 1.96%
s1488 40.90% 16.89% 5.59% 1.92%
s15850 27.73% 20.72% 5.25% 3.74%
s27 22.22% 22.22% 7.37% 7.37%
s298 21.30% 6.59% 6.13% 1.82%
s344 36.11% 28.13% 12.09% 10.30%
s349 14.55% 12.56% 2.56% 2.84%
s382 20.00% 6.25% 6.88% 2.62%
s386 27.23% 10.93% 7.93% 2.30%
s400 18.84% 8.17% 5.99% 2.90%
s420 52.48% 46.43% 21.13% 16.59%
s444 11.23% 5.04% 3.99% 2.10%
s510 44.72% 24.21% 6.76% 3.87%
s526 4.11% -2.02% 1.08% -0.31%
s5378 36.94% 26.76% 10.15% 6.82%
s641 65.06% 61.30% 28.01% 23.93%
s713 68.13% 66.08% 29.19% 26.54%
s820 50.74% 40.31% 10.11% 8.86%
s832 46.20% 34.28% 8.62% 6.17%
s838 46.53% 41.86% 15.30% 12.87%
s9234 28.94% 13.13% 5.02% 1.08%
s953 31.64% 11.68% 7.78% 2.34%
SISO 20.68% 18.22% 1.07% 0.93%
Average 32.63% 21.60% 9.19% 6.00%
Table 5 The area savings in terms of slack-matching buffer area and total circuit
area realized with the new LP formulation and sharing vs. the other
versions.
The results show that the proposed sharing algorithm and the new formulation
achieve significant savings over the previous formulation with no sharing as well as over
the previous formulation with sharing allowed. There is only one example that the new
algorithm did worse compared to the original formulation with sharing. In this particular
80
case the new formulation did worse by just one buffer. This can be attributed to the fact
that in both cases the formulations are approximate, in the sense that the results are
rounded up to instantiate an integer number of buffers, and in this context an difference
of one buffer can be considered a reasonable approximation error.
Overall the sharing of existing buffers seems to achieve average buffer area
savings of approximately 32.63% vs. the original LP formulation and no sharing and
21.6% vs. the original LP formulation with sharing. In terms of the total circuit area the
savings are approximately 9.19% and 6% respectively on average.
4.5 Other considerations
The new formulation as well as the sharing of buffers have area benefits that are
clear from all our experiments. However it is interesting to note a couple of other issues
that require careful consideration when choosing which variant to use. Firstly, the new
formulation requires additional variables as many as the nodes in the circuit and
additional constraints as many as the channels in the circuit, thus making the linear
formulation more complex. This results in more memory usage and larger runtime. As
the circuits grow in size the new formulation becomes more and more expensive and
ultimately impractical. The older formulation can generally tackle larger problems in the
same amount of time and thus could be preferential for larger netlists.
Another interesting observation is that the sharing assumes that arrival times of
shared buffers can satisfy all leaf nodes attached to them. However this is not always
true and as a result sometimes it results in additional buffers being added on a
subsequent pass. Without sharing, the arrival times have no such assumptions and thus
81
convergence is achieved usually faster. Both formulations are approximate due to the
rounding, so addition of buffers in subsequent paths is always possible, but without
sharing it is more rare. This is important when runtime is critical and an additional path
might require a very large amount of time that is prohibitive for a given design.
82
Chapter 5 An Asynchronous ASIC Flow
It is easier to understand the context of this work and the general motivation
behind it as part of the entire flow. In this chapter we are going to describe a new flow
that has been developed to enable an ASIC design flow for the asynchronous design
styles. It will show the commercial tools used and all the customized tools generated and
how this flow works from end to end. This is will give the reader a perspective as to what
is the value of our work and what its position is relative to the rest of the flow.
Our goal is to generate the framework to enable an asynchronous ASIC flow that
will enable engineers to design circuits using asynchronous design styles with the same
ease and efficiency as they do for synchronous designs currently. To achieve this we
have generated a tool flow that resembles the standard synchronous ASIC flow and in
fact reuses most of the existing parts and standard commercial tools. The designer
starts with standard HDL input and uses commercial synthesis engines to generate a
synthesized image of the logic. We have then designed a tool that takes the synthesized
netlist and applies all the design transformation and optimizations that are unique to
asynchronous design. After our customized tool is done it generates a netlist back into a
standard format so that it can be imported into a standard back-end flow. That enables
us to use standard Place and Route, verification and simulation tools that are well known
in the industry.
We have also focused throughout this work on the goal of guaranteeing
performance. There are three limiting factors for the performance of an asynchronous
circuit. The first one is algorithmic loops that were addressed through the distance-based
clustering criteria presented in Section 3.2.2. The second one is unbalanced paths and
83
misaligned data in the pipeline, which is a problem solved through slack matching that
was discussed extensively in Chapter 4. The third limiting factor is the LCT of a channel,
which could make a local handshake the critical path of a circuit. This is addressed in
this chapter in Section 5.2.2, where we describe the Clustering algorithm that is
implemented as part of this flow.
5.1 Synthesis
The first process in the flow is synthesis. This is done in commercial synthesis
tools to convert the design to a synchronous Verilog netlist. The initial synchronous
specification is done from a behavioral HDL or RTL HDL, which could be defined in any
language that is supported by the commercial synthesis tools. A second input to the tool
is the library information. The library information is really an “image” library that abstracts
the asynchronous gate details for the synthesis tool (such as whether they are dual-rail
or single-rail, their handshaking ports and protocols, etc.), but includes all the information
that is useful for it to do a proper design (such as timing, power, area, etc.).
The benefit of this flow instead of trying to generate a new tool is that we can
leverage off tools that have been refined through years of research to perform good
optimizations, and include predefined and optimized design components such as adders
and multipliers that can be used in the data path. So by placing appropriate constraints
one could use ripple-carry or carry look-ahead adders to trade between latency and
area. Other tasks, such as buffer insertion and gate resizing could be performed as well,
by changing the input constraints for the design. In most cases relaxed constraints are
enough, since they yield the simplest data path, but if one wanted to shorten the data
84
path due to a large algorithmic cycle time in the final design, modified constraints could
be used to force the compiler to generate a different netlist.
Figure 15 Overview of our ASIC flow.
5.2 The Clustering Program
Most asynchronous design styles allow for multiple gates to be placed within the
same pipeline stage. Generally that is done to reduce the area overhead associated with
the control logic that is used for handshaking between every pipeline stage and its
neighbors. The control logic is large and the more efficiently it gets shared among many
gates the smaller the penalty that one has to pay for the asynchronous design style
conversion. Depending on the design style the number of gates per pipeline cluster
Clustering Program
Clustering
Slack Matching
Template Conversion
HDL Input
Image Verilog Netlist
Mapped Verilog Netlist
Final Verilog Netlist GDSII
Back-End
Clustering Program
Synthesis
Image .lib
Target .lib
85
varies widely, but it is believed that the same principles apply in all cases. Also most
asynchronous design styles share other design rules, such as limits on fanin/fanouts per
pipeline cluster and others. Finally all asynchronous circuits have to obey similar rules
about Local Cycle Time, Target Cycle Time and Algorithmic Cycle Time as discussed in
Section 2.7.1. And all of these parameters depend on the size and interconnect of the
different clusters in the final netlist. So the goal is to minimize the area of the final netlist
by grouping gates together as well as possible, but at the same time obeying all the
design constraints that are specific to a particular design style and also maintain the
functionality and the performance requirements that the designer specified when
defining the HDL description of the circuit.
5.2.1 Circuit Representation
Immediately after the original Verilog input file for the circuit is read, the netlist is
converted to a generic directed graph structure, with the instances and primary IO
converted into nodes and the wire interconnect represented by edges. The original
Verilog input file is converted into a generic graph structure that can be manipulate more
efficiently during the different optimization operations that are run during processing.
Initially the tool reads the entire netlist (flat netlist with no hierarchy) and creates a single
pipeline stage out of every gate in the original netlist. The goal of this is to give the tool
the most flexibility in terms of clustering, and also start with very simple pipeline clusters
that should generally meet any give performance targets more easily.
5.2.2 Clustering
Having formed the fundamental one-gate clusters the next step is to try and
group them in larger groups without violating the performance requirements of the
design and at the same time preserve the original functionality intended by the user. Our
86
present algorithm does not try to find an optimal clustering solution. Instead it uses a
heuristic optimization based on area and performance criteria that implements a
steepest-descent-type of algorithm. The main focus of our work so far has been on
correctness of the emitted circuit as well as preservation of the performance
requirements. The runtime of the heuristic approach is also a big advantage since it
allows us to test larger circuits.
The clustering is done by merging two clusters into one (we refer to this as a
local move) and executing one such move at a time. During each iteration the software
looks at all possible moves that are available and executes the one that has the largest
performance benefit. It also estimates the area gains from the potential merging, and
uses the area benefits to break ties in the case that many moves have the same
performance improvements. The area metric is an area estimate of the control logic that
can be removed by executing the particular move. The performance metric is an error
metric associated with the Target Cycle Time of the circuit and the Local Cycle Time of
the individual channels. For each channel an error metric is calculated that is equal to
the amount that the LCT violates the GCT. For each move the performance metric is the
difference between the sum of errors for all channels before the move and after the
move. The largest the metric the more LCT improvements the move will achieve.
After each move the algorithm discards all moves that are no longer feasible and
generates new ones from the region of the circuit that was affected and then repeats the
process. This avoids costly recalculation of all move data at each step. The algorithm
does not select any move that would make the metric worse and thus in practice avoids
making LCT worse than the GCT and affecting the performance of the circuit, thus
addressing the last and final parameter that could affect performance.
87
The moves are also checked for other local rules that are associated with
particular design restrictions for the design style that is being targeted and are not
related to either performance or correctness. For example TOKEN_BUFFER cells are
cells that for all templates are initialized during reset with particular data values, unlike
regular cells that do not hold data after reset. Those can only be placed in pipeline
stages that host similar types of cells. When the clustering algorithm cannot find any
candidate local moves to execute it stops. The algorithm is summarized in Figure 16.

Figure 16 The clustering algorithm that ensures that LCT does not grow.
88
5.2.3 Fanout Fixes
The circuit is generated using synchronous tools and design libraries. There are
certain restrictions in some of the different design styles in terms of the number of fanout
gates that each gate will drive. Sometimes this is due to the particular protocol – for
example single-track 1-of-N channels (such as those in SSTFB) only support point-to-
point connections – or due to performance restrictions (fanout load of the gate or depth
of C-element trees required for merging ACK signals). When the given netlist does not
obey these restrictions the program has to intervene and correct these faults by
instantiating fanout trees for problematic signals using buffer cells, in a process that
bears a large amount of resemblance to the buffering and fanout optimization process of
the synchronous netlists, only with different criteria and decision metrics. This step is
executed if necessary before moving on to create a final netlist.
5.2.4 Final Optimizations
After clustering is done the software also performs slack matching on the netlist.
The concept of slack-matching is analyzed in detail in [5], and the tool uses a similar
formulation as described there to achieve the performance requested by the user. Some
modifications have been made to improve the results based on some practical
observations, which will be analyzed later in Chapter 4.
Other local optimizations are done in the netlist during this final stage to ensure
the best quality of results. Gate replication is used to reduce the levels of logic in certain
cases and buffers are inserted in some cases if it is deemed that the operation might
help performance. In general these are minor implementation details that help a practical
design, but are probably outside the scope of this work and will not be analyzed as part
of this document.
89
5.2.5 Library Conversion
The next step is to generate a new netlist that instantiates the actual gates that
will be used for implementation from the selected asynchronous implementation design.
The program also needs to generate and instantiate all the additional control logic that is
necessary for the handshaking between the pipeline stages that were defined during
clustering and slack matching.
For the design styles that have been imported in the tool so far the final gates
that will be used are dual rail gates, even though for other design styles that is not the
case. All the nets in the original netlist are in this case converted into pairs of nets each
representing the true and false rails of the original single rail signal. The gates are also
converted to their dual rail counterparts and the wires are interconnected appropriately.
With a dual rail library inversions of inputs and outputs could be handled by swapping
the two wires attached to a gate. During this step all gates are converted to their non-
inverting counterparts, and the inversion of inputs and outputs takes place by inverting
the dual-rail wires. Even though the single rail library uses image gates that include all
permutations of possible inversions of inputs and outputs, the final library includes only
the non-inverted versions, reducing the amount of library cells that are required for a
complete library implementation. Special signals such as power, ground and a global
asynchronous reset are also generated. When all the gates have been converted the
completion detection trees are formed and the controllers for the pipeline stages are
instantiated. The merge and fork cells for the primary inputs and outputs are also
instantiated, and the handshaking signals for the left and right environment are added to
the top level module.
90
5.2.6 Verification
The software package finally creates additional information that is required for
verification of the resulting design. A header file is written that contains some important
parameters for the design, such as the number of inputs, outputs and vectors that are
going to be used. Another file that contains a random number generator is also copied to
the design directory. Then the software will generate a testbench for the synchronous
netlist as well as a testbench for the asynchronous one. A script is also written out that
includes the commands needed for running the tests.
The script file commands will first call the random generator module and
generate an input file for both simulations that uses random data at all inputs. After this
the synchronous netlist is run given the input vectors and the output vectors are
recorded at every cycle and sent to a file. The script then executed the command to
simulate the asynchronous netlist with the same input file. The results of this run are
compared against the results of the synchronous netlist and if they match the testbench
will indicate that the test completed successfully. Otherwise it will indicate a failure. At
the same time the testbench samples the design outputs and average the number of
tokens received over time to calculate the global cycle time in the design.
5.3 Simulation and Back-End Design
The default simulation that our program performs is done in NC-Sim to verify that
the translation was successful. However since the netlist is in regular Verilog format,
further simulation in any commercial simulator is possible. The Verilog Netlist can also
be used as is by commercial back-end tools to perform place and route and verify the
performance of the circuit. Back-annotation could also be used for post-layout
91
simulations if desired. However, the flow currently does not support ECO flows and post-
layout optimizations. Therefore if the netlist is not yielding the desired results and
changes are necessary, the design might have to be processed again from the
beginning. Alternatively hand-editing might be able to alleviate the problem if it is easy to
identify and fix. Post-layout analysis and ECO-type fixes to the netlist is an interesting
and potentially necessary future step.
The flow has been proven in practice to work and we have taken several
example designs through, including place & route to produce GDSII. The netlists started
from RTL-level code (Verilog) and taken through synthesis, our clustering and slack-
matching tool and then Place and Route through Cadence’s Encounter tool suite. The
results are encouraging showing functional netlists that can successfully place and route
and produce functional circuits that would be ready for fabrication. Obviously more is
necessary for this flow to reach the levels of sophistication that synchronous flows have
reached, but it is a proof of concept that an automated ASIC flow for performance-aware
asynchronous design is possible.
92

Figure 17 A caption of a finalized placed and routed design after a successful
pass through our asynchronous ASIC flow.
93
Chapter 6 Summary
6.1 Advancements
Asynchronous circuit design has been proposed many times, but has only been
adopted in very few isolated cases by the design community. The main reason for this,
in our view, is the lack of an automated set of tools that would allow a designer to
generate a circuit quickly from a behavioral Hardware Description Language (HDL), just
like the ASIC flow that has existed for years for synchronous circuits. As part of this
effort we created an automated flow that can automatically generate asynchronous
circuits from any HDL using a mixture of custom and existing industry tools.
Due to the fact that asynchronous circuits require a handshaking controller for
every pipeline stage, which is used to interface to adjacent pipeline stages, the logic
overhead of such circuits is large. By grouping the circuits appropriately one can reduce
this overhead and yield circuits that have competitive or even superior characteristics
that their synchronous counterparts. Without this grouping asynchronous circuits are
most likely going to be far less efficient in terms of throughput per area, even though
they might have a substantial absolute performance advantage over their synchronous
equivalents.
It is also important to note that asynchronous circuits are data driven in nature.
This allows clustering to redefine the pipeline stages irrespective of the pipeline stages
defined at the RTL without altering the behavior of the circuit. Thus clustering allows
automatic re-pipelining based on performance-driven criteria that can create a circuit that
meets the desired performance without changes to the RTL. This makes the design
94
easier and saves designers time and effort that would otherwise be required in adjusting
their pipeline stages to meet the required performance.
As we were investigating the tradeoffs early on we discovered that arbitrary
clustering with only local constraints had a very high likelihood of yielding non-functional
circuits that would deadlock, and very frequently circuits that had performance that was
far worse than that of the initial circuit or the desired performance target. Consequently
this work was focused on guaranteeing that these two important criteria were maintained
by exploring the proper design criteria that would ensure functionality and performance
maintenance.
In Chapter 3 we proposed and proved several different criteria that allow us to
ensure that the functionality of a circuit is retained through clustering as well as other
additional ones that ensure that both algorithmic cycle time and end-to-end latency do
not deteriorate. This along with the directed graph model that is proposed is the most
important contribution of this thesis and it is an enabling factor for automated pipelining
of asynchronous circuits. Using the theory of Chapter 3 we can ensure that the circuits
will always work and meet performance goals desired by the designer.
Additionally, our algorithm relies heavily on maintaining a a record of all
distances from all the nodes to all the nodes of a graph. We have used the Floyd-
Warshall algorithm for deriving the initial distances, but as we modified the graph it was
quickly obvious that the algorithm was too complex ( ( )
3
V Θ ) to be executed multiple
times for updating the distances after every change in clustering. Thus we proposed
different practical update algorithms that require far less computation and can be used in
practice. One of them was proven to have ( )
2
V Ο for circuits with limited fanout. This is
95
an update algorithm that could be useful to a variety of applications that require the
maintenance of such an all-pair wise distance array.
Finally Chapter 4 focuses on implementation improvements to the slack matching
formulation that was earlier proposed in [5]. The improvements are based on a different
formulation for the Linear Program that can be used to solve the Slack Matching problem
for a circuit. Even though both methods discussed are approximations of the optimal
solution and the model proposed is only a small improvement over the original one the
section offers an approach that has practical value. The method discussed along with
the new model for sharing slack matching buffers between channels results in overall
saving of 13% in terms of total buffering area and approximately 3% in overall circuit
area.
6.2 Applications
In this thesis we have created a modeling framework that allows a designer to
perform this clustering of the logic gates into pipeline stages by modeling the logic onto
an implementation-agnostic graph. This allows us to cluster any arbitrary circuit from its
gate-level logic representation into pipeline stages of an asynchronous circuit of any
desired design style. Depending on the implementation style the clustering constraints
have to be modified to yield a functional circuit that abides by all design rules for the
particular style. Area models might also vary quite a bit. But the basic principles of our
proposed flow are applicable to all of them.
With the right formulation of the local constraints and appropriate performance
and area models for a particular design style one can use this work to design any type of
96
asynchronous pipeline that one desires starting from a gate-level representation of the
desired circuit. That means that one could start from any regular HDL representation of a
circuit and use a conventional synthesis engine to convert it to gates. Then using this
model one could use it to create a pipelined circuit for any design style of their choice. It
is believed that this method, since it maintain the functionality and performance of the
original circuit, while optimizing its area by clustering, could be used as part of a much
broader ASIC flow that would be applicable to all types of asynchronous circuits.
6.3 Open Issues
The focus of this thesis was on maintaining correctness and performance
throughout our processing. For all our experiments we used a greedy method for
selecting our clustering moves one at a time. This is essentially a steepest-descent
method that is most likely far from optimal. Clustering is a hard problem and several
heuristic methods have been proposed to address it in different contexts. It is believed
that many such methods are also applicable here and it remains an open issue to find
practical and good methods for this particular application. Optimization techniques from
dynamic programming, simulated annealing, and other optimization techniques are all
thought to be applicable for this problem. It is our belief that there deserves to be follow-
up work that builds on this framework and will focus now on improving the area results
for such circuits, by attempting to find suitable area models and optimization techniques
that solve this problem in a good and practical fashion.
97
6.4 Conclusions
This thesis proposed a method for modeling an arbitrary gate-level circuit using a
graph model and methods that allow us to perform area-reducing transformations to it,
which are modeled as clustering operations on the graph. The goal is to provide a
theoretical foundation for a method that could lead to a fully automated flow for
designing asynchronous circuits from an arbitrary gate-level representation that is
implementation-agnostic. The long term goal is to provide a framework for further
optimizations that could be done to generate asynchronous circuits in an ASIC-like flow
that can ultimately compete in performance, area and ease of design with their
synchronous counterparts, thus enabling designers to take advantage of asynchronous
design techniques in commercial applications. It is our belief that this work is a
significant step along a path that can ultimately lead to significant wider adoption of
asynchronous design technology.
98
References
[1]. Enchanced Intel® SpeedStep® Technology.
http://download.intel.com/design/processor/datashts/31327801.pdf
[2]. AMD PowerNow!™ Technology.
http://www.amd.com/us-en/assets/content_type/DownloadableAssets/Power_Now2.pdf
[3]. A. Bardsley and D. A. Edwards, “The Balsa Asynchronous Circuit Synthesis System,” in
Proc. Forum on Design Languages, Sept. 2000.
[4]. A. Bardsley and D. A. Edwards, “Synthesizing an asynchronous DMA controller with
Balsa,” Journal of Systems Architecture, vol. 46, pp. 1309–1319, 2000.
[5]. P.A. Beerel, A. Lines,. M. Davies, N. H. Kim, “Slack matching asynchronous designs”,
IEEE International Symposium on Asynchronous Circuits and Systems (ASYNC’06),
2006,
[6]. K. van Berkel, F. Huberts, A. Peeters. “Stretching Quasi Delay Insensitivity by Means of
Extended Isochronic Forks,” in Proc. Second working conference on Asynchronous
Design Methodologies, pp. 99–106, May, 1995.
[7]. K. van Berkel, J. Kessels, M. Roncken, R. Saeijs, and F. Schalij, “The VLSI-Programming
Language Tangram and Its Translation into Handshake Circuits,” in Proc. European
Conference on Design Automation (EDAC), 1991, pp. 384–389.
[8]. I. Blunno, J. Cortadella, A. Kondratyev, L. Lavagno, K. Lwin, and C. Sotiriou, “Handshake
Protocols for De-Synchronization,” in Proc. International Symposium on Advanced
Research in Asynchronous Circuits and Systems, Apr. 2004, pp. 149–158.
[9]. T. Chelcea , S. M. Nowick, “Resynthesis and peephole transformations for the
optimization of large-scale asynchronous systems,” Proceedings of the 39th conference
on Design automation, 2002,
[10]. T. Chelcea, A. Bardsley, D. Edwards, and S. M. Nowick, “A Burst-Mode Oriented Back-
End for the Balsa Synthesis System,” in Proc. Design, Automation and Test in Europe
(DATE), Mar. 2002, pp. 330–337.
[11]. C.S. Choy, J. Butas, J. Povazanic, C. F. Chan. “A new control circuit for asynchronous
micropipelines”, IEEE Transactions on Computers Volume 50, Issue 9, Sept. 2001
pp:992 – 997
[12]. J. Cortadella, A. Kondratyev, L. Lavagno and C. Sotiriou. “A Concurrent Model for De-
synchronization.” In Handouts of the International Workshop on Logic Synthesis, pp. 294-
301, 2003.
[13]. J. Cortadella, A. Kondratyev, L. Lavagno, K. Lwin, and C. Sotiriou, “From Synchronous to
Asynchronous: an Automatic Approach,” in Proc. of International, Design and Test in
Europe Conference and Exhibition, 2004.
[14]. J. Cortadella, A. Kondratyev, L. Lavagno, and C. Sotiriou, “De-synchronization: Synthesis
of Asynchronous Circuits from Synchronous Specification.” in IEEE Transactions on
Computer-Aided Design of Integrated Circuits and Systems. Oct. 2006.
99
[15]. K.M. Fant, S.A. Brandt, “NULL Convention Logic
TM
: a complete and consistent logic for
asynchronous digital circuit synthesis” Proceedings of International Conference on
Application Specific Systems, Architectures and Processors, Aug. 1996.
[16]. K. Fazel, R.B. Reese; M.A. Thornton, “PLFire: A Visualization Tool for Asynchronous
Phased Logic Designs”, Proc. of International, Design and Test in Europe Conference
and Exhibition, 2003.
[17]. M. Ferretti, “Single-Track Asynchronous Pipeline Template, PhD Thesis”, University of
Southern California, 2004.
[18]. M. Ferretti, P. A. Beerel, “Single-Track Asynchronous Pipeline Templates Using 1-of-N
Encoding”, IEEE Proceedings of the 2002 Design, Automation and Test in Europe
Conference and Exhibition (DATE.02)
[19]. S. B. Furber and J. Liu. “Dynamic logic in four-phase Micropipelines”. In Proc.
International Symposium on Advanced Research in Asynchronous Circuits and Systems.
IEEE Computer Society Press, March 1996.
[20]. P. Golani, P. A. Beerel, “High-Performance Noise-Robust Asynchronous Circuits,” IEEE
Computer Society Annual Symposium on Emerging VLSI Technologies and Architectures
(ISVLSI'06), 2006, pp. 173-178,.
[21]. J. Kessels and A. Peeters, “The Tangram Framework: Asynchronous Circuits for Low
Power,” in Proc. of Asia and South Pacific Design Automation Conference, Feb. 2001,
pp. 255–260.
[22]. A. Kondratyev and K. Lwin. “Design of asynchronous circuits by synchronous CAD tools.”
In IEEE Design and Test of Computers, vol. 19, no. 4, pp. 107-117, July/August 2002.
[23]. M. Ligthart, K. Fant, R. Smith, A. Taubin, and A. Kondratyev, “Asynchronous Design
Using Commercial HDL Synthesis Tools,” in Proc. International Symposium on Advanced
Research in Asynchronous Circuits and Systems, Apr. 2000, pp. 114–125.
[24]. D. H. Linder and J. C. Harden, “Phased Logic: Supporting the Synchronous Design
Paradigm with Delay-Insensitive Circuitry,” IEEE Transactions on Computers, vol. 45, no.
9, pp. 1031–1044, Sept. 1996.
[25]. A. M. Lines, “Asynchronous Pipelined Circuits”, Master’s Thesis, California Institute of
Technology 1995, Revised 1998.
[26]. A. J. Martin, “The Limitations to Delay-Insensitivity in Asynchronous Circuits.” Proc of
Sixth MIT Conference on Advanced Research in VLSI, ed. W. J. Dally, 263-278, MIT
Press, 1990.
[27]. A. J. Martin, “Compiling Communicating Process into Delay-Insensitive VLSI Circuits.”
Distributed Computing 1(4):226-234, 1986.
[28]. A. J. Martin. Programming in VLSI: From communicating processes to delay-insensitive
circuits. In C. A. R. Hoare, editor, Developments in Concurrency and Communication, UT
Year of Programming Series pages 1-64. Addison-Wesley, 1990.
[29]. S. Masteller, L. Sorenson, “Cycle decomposition in NCL”, IEEE Design & Test of
Computers, Volume 20, Issue 6, Nov.-Dec. 2003
100
[30]. K. Meekins, D. Ferguson, M. Basta, “Delay insensitive NCL reconfigurable logic”, IEEE
Aerospace Conference Proceedings, Volume 4, 2002
[31]. T. Nanya, A. Takamura, M. Kuwako, M. Imai, T. Fujii, M. Ozawa, I. Fukasaku, Y. Ueno,
F. Okamoto, H. Fujimoto, O. Fujita, M. Yamashina, and M. Fukuma. “TITAC-2: A 32-bit
scalable-delay-insensitive microprocessor.” In Symposium Record of HOT Chips IX, pp.
19–32, August 1997.
[32]. J. M. Rabaey, “Digital Integrated Circuits A Design Perspective”, Prentice Hall Electronics
and VLSI Series, 1996.
[33]. R.B. Reese; M.A. Thornton; Traver, C., “Arithmetic logic circuits using self-timed bit level
dataflow and early evaluation”, Proceedings of the International Conference on Computer
Design, 2001.
[34]. R.B. Reese; M.A. Thornton; Traver, C., “A fine-grain Phased Logic CPU”, Proceedings of
the IEEE Computer Society Annual Symposium on VLSI, Feb. 2003.
[35]. R.B. Reese; M.A. Thornton; Traver, C., “A coarse-grain phased logic CPU”, Proceedings
of the Ninth International Symposium on Asynchronous Circuits and Systems, May 2003
[36]. R.B. Reese; M.A. Thornton; Traver, C.; Hemmendinger, D., “Early evaluation for
performance enhancement in phased logic”, IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems, April 2005
[37]. R.B. Reese; M.A. Thornton; C. Traver, “A coarse-grain phased logic CPU” IEEE
Transactions on Computers, July 2005.
[38]. M. Singh and S.M. Nowick. “MOUSETRAP: Ultra-High-Speed Transition-Signaling
Asynchronous Pipelines.” Proc. of IEEE Intl. Conf. on Computer Design (ICCD-01), Sept.
2001
[39]. M. Singh, and S.M. Nowick. “High-throughput asynchronous pipelines for fine grain
dynamic datapaths”. In Proc. of ASYNC, 2000, pp. 198–209.
[40]. C. Traver.; R.B. Reese.; M.A. Thornton, “Cell designs for self-timed FPGAs”,
Proceedings of the 14th Annual IEEE International ASIC/SOC Conference, 2001.
[41]. T.E. Williams. “Latency and Throughput Tradeoffs in Self-Timed Speed-Independent
Pipelines and Rings”., PhD Thesis Stanford University, June 1991 Technical Report CSL-
TR-90-431
[42]. T.E. Williams, and M.A. Horowitz. “A Zero-overhead self-timed 160ns 54b CMOS
divider”. ISSCC Digest of Technical Papers, 1991, pp. 98-296.

Abstract (if available)

Abstract As semiconductor technology advances into smaller and smaller geometries, new challenges arise. The increased circuit integration combined with larger variability make it harder for designers to distribute a global clock and global interconnect signals efficiently in their designs. To combat the effects designers use more conservative models and more complicated tools that result in longer design times and diminishing returns from the migration to the smaller geometries. Some of these problems can be addressed by asynchronous circuits, but there exists no welldefined method for automated asynchronous design. Some methods have been proposed over the years, but they leverage off existing synchronous techniques too much, resulting in circuits that are bound by the characteristics of their synchronous counterparts. This thesis proposes a novel approach for generating such circuits, from any arbitrary HDL representation of a circuit by automatically clustering the synthesized gates into pipeline stages that are then slack-matched to meet performance goals while minimizing area. The method thus provides a form of automatic pipelining in which the throughput of the overall design is not limited to the clock frequency or the level of pipelining in the original RTL specification. Consequently, the performance can generally maintain or exceed the performance of the original circuit. The method is design-style agnostic and is thus applicable to many asynchronous design styles.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Average-case performance analysis and optimization of conditional asynchronous circuits

PDF

Gated Multi-Level Domino: a high-speed, low power asynchronous circuit template

PDF

Radiation hardened by design asynchronous framework

PDF

Power optimization of asynchronous pipelines using conditioning and reconditioning based on a three-valued logic model

PDF

An asynchronous resilient circuit template and automated design flow

PDF

Formal equivalence checking and logic re-synthesis for asynchronous VLSI designs

PDF

Power efficient design of SRAM arrays and optimal design of signal and power distribution networks in VLSI circuits

PDF

Library characterization and static timing analysis of asynchornous circuits

PDF

Theory, implementations and applications of single-track designs

PDF

Electronic design automation algorithms for physical design and optimization of single flux quantum logic circuits

PDF

Variation-aware circuit and chip level power optimization in digital VLSI systems

PDF

Advanced cell design and reconfigurable circuits for single flux quantum technology

PDF

Thermal analysis and multiobjective optimization for three dimensional integrated circuits

PDF

Clocking solutions for SFQ circuits

PDF

Static timing analysis of GasP

PDF

Optimizing power delivery networks in VLSI platforms

PDF

Verification and testing of rapid single-flux-quantum (RSFQ) circuit for certifying logical correctness and performance

PDF

Trustworthiness of integrated circuits: a new testing framework for hardware Trojans

PDF

Energy-efficient shutdown of circuit components and computing systems

PDF

Production-level test issues in delay line based asynchronous designs

Asset Metadata

Creator Dimou, Georgios D. (author)

Core Title Clustering and fanout optimizations of asynchronous circuits

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Electrical Engineering (VLSI Design)

Publication Date 03/03/2010

Defense Date 12/03/2008

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag area-driven,ASIC,asynchronous,clustering,fanout,OAI-PMH Harvest,performance-driven,slack-matching

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Pedram, Massoud (committee chair), Alexander, Kenneth S. (committee member), Beerel, Peter A. (committee member)

Creator Email dimou@usc.edu,jorjinio4u@yahoo.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-m1996

Unique identifier UC1434487

Identifier etd-Dimou-2579 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-146892 (legacy record id),usctheses-m1996 (legacy record id)

Legacy Identifier etd-Dimou-2579.pdf

Dmrecord 146892

Document Type Dissertation

Rights Dimou, Georgios D.

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email cisadmin@lib.usc.edu