Introduction

to
Reconfigurable Computing

Architectures Algorithms and Application

INTRODUCTION
TO
RECONFIGURABLE COMPUTING

Architectures, Algorithms, and Application

CHRISTOPHE BOBDA
University of Kaiserslautern, Germany

Kluwer Academic Publishers

Boston/Dordrecht/London
iv RECONFIGURABLE COMPUTING
 i

To Lewin, Jan and Huguette for being so patient

Foreword

Will be written

by

Peter Cheung

and

Reiner Hartenstein
Contents

Foreword iii
Preface ix
List of Figures xi
List of Tables xix
1. INTRODUCTION 1
1 General Purpose Computing 2
2 Domain Specific Processors 5
3 Application Specific Processors 6
4 Reconfigurable Computing 8
5 Fields of Application 9
6 Organization of the Book 11
2. RECONFIGURABLE ARCHITECTURES 15
1 Early Work 15
2 Simple Programmable Logic Devices 26
3 Complex Programmable Logic Device 28
4 Field Programmable Gate Arrays 28
5 Coarse-Grained Reconfigurable Devices 49
6 Conclusion 65
3. IMPLEMENTATION 67
1 Integration 67
2 FPGA Design Flow 72
3 Logic Synthesis 74
4 Conclusion 97
vi RECONFIGURABLE COMPUTING

4. HIGH-LEVEL SYNTHESIS FOR RECONFIGURABLE DEVICES 99

1 Modelling 100
2 Temporal Partitioning Algorithms 120
3 Conclusion 148
5. TEMPORAL PLACEMENT 149
1 Off-Line Temporal Placement 151
2 On-Line Temporal Placement 160
3 Managing the Device’s Free Space with Empty Rectangles 161
4 Managing the Device’s Occupied Space 165
5 Conclusion 179
6. ON-LINE COMMUNICATION 181
1 Direct Communication 181
2 Communication Over Third Party 182
3 Bus-based Communication 183
4 Circuit Switching 183
5 Network on Chip 188
6 The Dynamic Network on Chip (DyNoC) 199
7 Conclusion 212
7. PARTIAL RECONFIGURATION DESIGN 213
1 Partial Reconfiguration on Virtex Devices 214
2 Bitstream Manipulation with JBits 215
3 The Modular Design Flow 217
4 The Early Access Design Flow 225
5 Creating Partially Reconfigurable Designs 235
6 Partial Reconfiguration using Handel-C Designs 244
7 Platform design 246
8 Enhancement in the Platform Design 256
9 Conclusion 258
8. SYSTEM ON A PROGRAMMABLE CHIP 259
1 Introduction to SoPC 259
2 Adaptive Multiprocessing on Chip 268
3 Conclusion 284
Contents vii

9. APPLICATIONS 285
1 Pattern Matching 286
2 Video Streaming 294
3 Distributed Arithmetic 298
4 Adaptive Controller 306
5 Adaptive Cryptographic Systems 310
6 Software Defined Radio 313
7 High Performance Computing 315
8 Conclusion 317
References 319
Appendices 336
A Hints to Labs 337
1 Prerequisites 338
2 Reorganization of the Project Video8 non pr 338
B Party 345
C Quick Part-Y Tutorial 349
Preface

One indicator of the the growing importance of Reconfigurable Computing

is the large number of events (conferences, workshops, meetings) organized
and devoted to this topic in the last couple of years. Also, the growth observed
in the market share of programmable logic devices, particulary the FPGAs, is
an indicator of the strong interest in reconfigurable logic.
Following this development, teaching reconfigurable computing, which was
initiated in many Universities a couple of years before, has gained more impor-
tance. The curricula in reconfigurable computing varies from simple seminars
to more heavy syllabus including lectures, exercises and labs.
Many people among whom Reiner Hartenstein have been advocating years
ago in favor of a normalized reconfigurable computing syllabus. Aware of
the importance of a teaching book in this normalization, during a bi-annual
meeting on reconfigurable computing held in Dagstuhl in 2003[11], Harten-
stein coined the importance of a text book in reconfigurable computing and
proposed the attendees to write one. He suggested to have several people writ-
ing in order to minimize the work and have the book published as faster as
possible. Unfortunately, this initiative was not pursued.
A couple of months after this initiative, in the summer term 2004, I started
teaching a course in reconfigurable computing at the university of Erlangen-
Nuremberg. With the difficulties of acquiring teaching materials and labs, I
started writing a script to ease the learning process of students. The positive
feedback gained from the student encouraged me to continue writing. Further
repetitions of the course in winter term 2004 and in winter term 2006 were used
to improve the course contents. It should be mentioned that a couple of books
[153] [220] [134] reconfigurable computing were published in between. How-
ever none of them were found to cover the complete aspects of reconfigurable
computing as I use to teach in my course.
My goal in writing this book was to provided a strong theoretical and prac-
tical background, as a contribution for a syllabus in reconfigurable computing.
x RECONFIGURABLE COMPUTING

A short overview on the content of each chapter is provided in section 6 of

chapter 1.
This book targets graduate student and lecturers in computer engineering,
computer science and electrical engineering. Also professional in the afore
mentioned field can use the book as well. We supply the book with teaching
materials (slides and labs) in order to ease the course preparation for those
willing to introduce a reconfigurable computing curricula. The teaching ma-
terial as well as the labs can be downloaded from the course web page at:
www.bobda.net/rc-book.
Finally, despite all the effort place in the review, we cannot be sure that all
the mistakes were filtered out. We will therefore be grateful to receive your
comments and feedback on possible errors.

Kaiserslautern Christophe Bobda

About the Author

Dr. Bobda received the Licence degree in mathematics from the univer-
sity of Yaounde, Cameroon, in 1992, the diploma of computer science and the
Ph.D. degree (with honors) in computer science from the university of Pader-
born in Germany in 1999 and 2003 respectively. In June 2003 he joined the
department of computer science at the University of Erlangen-Nuremberg in
Germany as post doc. In October 2005, he moved to the University of Kaiser-
slautern as Junior Professor, where he leads the working group Self-Organizing
Embedded Systems in the department of computer science. Hes research inter-
est include reconfigurable computing, self-organization in embedded systems,
multiprocessor on chip, and adaptive image processing.
Dr. Bobda received the best dissertation award 2003 from the university of
Paderborn for his work on synthesis of reconfigurable systems using temporal
partitioning and temporal placement.
Dr. Bobda is member of The IEEE Computer Society, the ACM and the
GI. He has also served in the program committee of several conferences (FPL,
FPT, RAW, RSP, ERSA, DRS) and in the DATE executive committee as pro-
ceedings chair (2004, 2005, 2006, 2007). He served as reviewer of several
journals (IEEE TC, IEEE TVLSI, Elsevier Journal of Microprocessor and Mi-
crosystems, Integration the VLSI Journal) and conferences (DAC, DATE, FPL,
FPT, SBCCI, RAW, RSP, ERSA).
List of Figures

1.1 The Von Neumann Computer architecture 2

1.2 Sequential and pipelined execution of instructions on a
Von Neumann Computer 4
1.3 ASIP implementation of Algorithm 1. 7
1.4 Flexibility vs performance of processor classes 8
2.1 Structure of the Estrin Fix-Plus Machine 17
2.2 The basic building blocks of Fix-Plus Machine 17
2.3 The wiring harness of the Estrin-Machine 18
2.4 The motherboard of the Estrin’s Fix-Plus 18
2.5 Estrin at work: Hand-Controled Reconfiguration 19
2.6 Structure of the Rammig machine 20
2.7 META-46 GOLDLAC 20
2.8 General architecture of the XPuter as implemented in
the Map oriented machine (MOM-3) prototype. 21
2.9 Structure of the XPuter’s reconfigurable ALU 22
2.10 Programmable Active Memory (PAM) architecture as
array of PAB 23
2.11 Architecture of the SPLASH II array board 25
2.12 PAL and PLA implementations of the functions F 1 =
A · B + A · B and F 2 = A · B + B · C 27
2.13 Structure of a CPLD device 28
2.14 Structure of an FPGA 29
2.15 Antifuse FPGA Technology 30
2.16 A Xillinx SRAM cell 30
2.17 Use of SRAM in FPGA-Configuration 31
xii RECONFIGURABLE COMPUTING

2.18 EEPROM Technology 32

2.19 Implementation of f=ab in a 2-input MUX 33
2.20 Implementation of a Full adder using two 4-input one
output MUX 35
2.21 The Actel basic computing blocks uses multiplexers as
function generators 36
2.22 2-input LUT 37
2.23 Implementation of a full adder in two 3-input LUTs 37
2.24 Basic block of the Xilinx FPGAs 38
2.25 CLB in the newer Xilinx FPGAs (Spartan 3, Virtex 4
and Virtex 5) 39
2.26 Logic Element in the Cyclone II 39
2.27 Stratix II Adaptive Logic Modul 40
2.28 The four basic FPGA structures 41
2.29 Symetrical array arrangement in a) the Xilinx and b)
the Atmel AT40K FPGAs 42
2.30 Virtex routing resource 42
2.31 Local connection of Atmel Cell 42
2.32 Row based arrangement on the Actel ACT3 FPGA Family 43
2.33 Actel’s ACT3 FPGA horizontal and vertical routing resources 44
2.34 Actel ProASIC local routing resources 45
2.35 Hierarchical arrangement on the Altera Stratix II FPGA 46
2.36 LAB connection on the Altera Stratix devices 46
2.37 General structure of an I/O component 48
2.38 Structure of a Xilinx Virtex II Pro FPGA with two Pow-
erPC 405 Processor blocks 49
2.39 Structure of the PACT XPP device 51
2.40 The XPP ALU Processing Array Element. The struc-
ture of the RAM ALU is similiar. 52
2.41 Structure of the NEC Dynamically Reconfigurable Processor 53
2.42 The DRP Processing Element 54
2.43 Structure of the Picochip device 55
2.44 The Quicksilver ACM hierarchical structure with 64 nodes 56
2.45 ACM Node and Routing resource 57
2.46 IPflex DAP/DNA reconfigurable processor 59
2.47 The Stretch 5530 configurable processor 60
List of Figures xiii

2.48 Pipeline Reconfiguration: Mapping of a 5 stage virtual

pipelline auf eine 3 stage 63
3.1 Architecture of a run-time reconfigurable system 69
3.2 A CPU-RPU configuration and computation step 70
3.3 The FPGA design flow 73
3.4 A structured digital system 75
3.5 Example of a boolean network with: y1 = x1 + x2 ,
y2 = x3 · x4 , y3 = x5 · x6 , y4 = y1 + y2 , z1 = y1 + y4 ,
and z2 = y2 ⊕ y3 76
3.6 BDD-representation of the function f = abc + bd + bcd 79
3.7 Example of a K-feasible cone Cv at a node v 82
3.8 Example of a graph covering with K-feasible cone and
the corresponding covering with LUTs 83
3.9 Chortle two-level decomposition 84
3.10 Example of multi-level decomposition 86
3.11 Exploiting reconvergent paths to reduce the amount of
LUTs used 86
3.12 Logic replication at fan-out nodes to reduce the number
of LUTs used 87
3.13 Construction of the network Nt from the cone Ct at
node t 90
3.14 Minimum height 3-feasible cut and node mapping 91
3.15 Illustration of the two cases in the proof of Lemma 3.8 92
3.16 ′
Transforming Nt into N t by node collapsing 93
3.17 Transforming the node cut constraints into the edge cut ones 94
3.18 Improvement of the FlowMap algorithm through effi-
cient predecessor packing 96
4.1 Computational Graph for Quadratic Root 102
4.2 Sequencing graph with a branching node linking to two
different sub graphs 103
4.3 Transformation of a sequential program into a FSMD 105
4.4 Transformation of the greatest common divisor pro-
gram into a FSMD 107
4.5 The datapath and the corresponding FSM for the GCD-FSMD 108
4.6 Dataflow graph of the functions: x = ((a × b) − (c ×
d)) + ((c × d) − (e − f )) and y = ((c × d) − (e − f )) −
((e − f ) + (g − h)) 109
xiv RECONFIGURABLE COMPUTING

4.7 HLS of the graph in figure 4.6 on a an architecture with

one instances of the resource types +, ∗ and − 110
4.8 HLS of the graph in figure 4.6 on a reconfigurable device 111
4.9 Partitioning of a coarse-grained node in the dataflow graph 113
4.10 Example of configuration graph 116
4.11 Wasted resources 117
4.12 Partitioning of the graph G with connectivity 0.24 with
an algorithm that produces a quality of 0.25 119
4.13 Partitioning of the graph G with connectivity 0.24 with
an algorithm that produces a quality of 0.45 119
4.14 Scheduling example with the ASAP-algorithm 122
4.15 ALAP Scheduling example 123
4.16 An example of list scheduling using the depth of a node
as priority 124
4.17 Levelizing effect on the list-scheduling on a dataflow graph 128
4.18 Partitioning with a better quality than the list-sceduling 128
4.19 Dataflow graph transformation into a network 134
4.20 Transformation and partitioning steps using the network
flow approach 135
4.21 1-D and 2-D spectral-based placement of a graph 136
4.22 Dataflow graph of f = ((a + b) ∗ c) − ((e + f ) + (g ∗ h)) 140
4.23 3-D spectral placement of the DFG of figure 4.22 141
4.24 Derived partitioning from the spectral placement of fig-
ure 4.23 142
4.25 Internal and external edges of a given nodes 143
4.26 Partitioning of a graph in two set with a common sets
of operators 146
4.27 Logical partitioning of the graph of figure 4.26 146
4.28 Implementation of configuration switching with the par-
titions of figure 4.27 147
5.1 Temporal placement as 3-D placement of blocks 150
5.2 First-fit temporal placement of a set of clusters 154
5.3 Valid two dimensional packing 157
5.4 A non valid two dimensional packing 158
5.5 3-D placement and corresponding interval graphs, com-
plement graphs and oreinted packing 159
5.6 Various types of empty rectangles 162
List of Figures xv

5.7 Increase of the number of MER through the insertion

of a new component. 163
5.8 Two different non-overlaping rectangles representations 164
5.9 Splitting alternatives after new insertion 164
5.10 IPR of a new module v relative to a placed module v ′ 167
5.11 Impossible and possible placement region of a compo-
nent v prior to its insertion 168
5.12 Nearest possible feasible point of an optimal location
that falls within the IPR 170
5.13 Moving out of consecutive overlapping IPRs 172
5.14 Expanding existing modules and shrinking chip area
and the new 173
5.15 Characterization of IPR of a given component: The set
of contours (left), the contour returned by the modified
CUR (middle), the contour returned by the CUR (right) 174
5.16 Placement of a component on the chip (left) guided by
the communication with its environement (right). 176
5.17 Computation of the union of contours. The point on
the boundary represent the potentially moves from the
median out of the IPR. 178
6.1 Direct communication between placed modules on a
reconfigurable device 182
6.2 Drawback of circuit switching in temporal placement
Placing a component using 4 PEs will not be possible,
although enough free resources are available 185
6.3 The RMBoC architecture 186
6.4 RMBoC FPGA implementation 187
6.5 Crosspoint architecture 187
6.6 A Network on Chip on a 2-D Mesh 189
6.7 Router Architecture 190
6.8 A general FIFO Implementation 191
6.9 General format of a packet 192
6.10 Arbiter to control the write access at output data lines 193
6.11 A general wrapper architecture 194
6.12 Implementation of a large reconfigurable module on a
Network on Chip 199
6.13 The communication infrastructure on a DyNoC 201
6.14 A impossible placement scenario 203
xvi RECONFIGURABLE COMPUTING

6.15 A stronly connected configuration on a DyNoC 204

6.16 Obstacle avoidance in the horizontal direction 206
6.17 Obstacle avoidance in the verticalal direction 207
6.18 Placement that cause an extreme long routing path 208
6.19 Router guiding in a DyNoC 209
6.20 DyNoC implementation of a traffic light controller on
a VirtexII-1000 212
7.1 Generation of bitstreams for partial reconfiguration 216
7.2 Routing tools can route the same signals on different paths 218
7.3 The recommended directory structure for the modular
design flow 219
7.4 Limitation of a PR area to a block (dark) and the actual
dimensions (light) 226
7.5 Scheme of a PR application with a traversing bus that
may not be interrupted 226
7.6 Improved directory structure for the Early Access De-
sign Flow 227
7.7 Usage of the new EA bus macros 229
7.8 Narrow (7.8(a)) and wide (7.8(b)) bus macro spanning
two or four CLBs 230
7.9 Three nested bus macros 231
7.10 Scheme of Animated Patterns 235
7.11 Two patterns that can be created with modules of An-
imated Patterns. Left: the middle beam is moving.
Right: the diagonal stripes are moving. 236
7.12 Scheme of Video8 237
7.13 Reconstructing example Video8 to place the partially
reconfigurable part and connected modules in the top-level. 238
7.14 Moving a partially reconfigurable module to the top-
level design on the example of Video8 239
7.15 Modules using resources (pins) not available in their
placement area (the complete column) must use feed-
through signals to access those resources 246
7.16 Illustration of th pin problematique on the RC200-Board 249
7.17 Architecture of the ESM-Baby board 251
7.18 Architecture of the ESM MotherBoard 254
7.19 Intermodule communication possibilities on the ESM 255
7.20 SRAM-based intermodule communication on the ESM 256
List of Figures xvii

7.21 Possible enhancement of the Erlangen Slot Machine on

the Xilinx Virtex 4 and Virtex 5 FPGAs 257
8.1 Integration of PCB modules into a single chip: from
system on PCB to SoC 260
8.2 Example of system ingration with CoreConnect buses 267
8.3 Implementation of the OPB for two maters and two slaves 268
8.4 General adaptive multiprocessor hardware infrastructure 271
8.5 Structure of the on chip network 273
8.6 Implementation of the transceiver 274
8.7 The communication protocoll 275
8.8 Automatic hardware generation flow 277
8.9 The Platform-independent Hardware generation tool (PinHat) 279
8.10 The PinHaT framework 280
8.11 software configuration flow 282
8.12 4-Processor infrastructure for the SVD on the ML310-Board 283
9.1 Sliding windows for the search of three words in parallel 288
9.2 FSM recognizers for the word "conte": a) sate diagram,
b) transition table, c) basis structure the hardware im-
plementation: 4 flip flops will be need to code a 5 × 6
transition table 291
9.3 a) Use of the common prefix to reduce the number
of flip flops of the common word detector for "par-
tir", "paris", "avale","avant". b) implementation with-
out use of common prefix and common comparator set 291
9.4 Basis structure of a FSM-based words recognizer that
exploits the common prefix and a common set of characters 292
9.5 Processing steps of the FSM for the word "tictic" 293
9.6 Implementation of a 5x5 sliding windows 296
9.7 A modular architecture for video streaming on the ESM 297
9.8 Architecture of a distributed arithmetic datapath 300
9.9 k-parallel distributed arithmetic datapath 301
9.10 Datapath of the distributed arithmetic computation for
floating popin numbers 304
9.11 An optical multimode waveguide is represented by a
multiport with several transfer paths. 305
9.12 Screenshot of the 6-parallel DA implementation of the
the recursive convolution equation on the Celoxica RC100-
PP platform 307
xviii RECONFIGURABLE COMPUTING

9.13 Adaptive controller architecture 308

9.14 Adaptive controller architecture. Left the one slot im-
plementation, and right the two slot implemenation 310
9.15 Architecture of an adaptive cryptographic system 312
9.16 Architecture of a software defined radio system 315
C.1 Tree View after Top Assembly 351
List of Tables

2.1 Truth table of the Full adder 35

3.1 Language and tools overview for coarse-grained RPUs 72
3.2 Overview of FPGA Manufacturers and Tool Providers 74
4.1 Laplacian matrix of the graph of figure 4.22 140
6.1 Router Statistics 211
6.2 TLC and CG Statistics 211
7.1 New address ranges to be set in EDK 241
9.1 Results of the recursive convolution equation on differ-
ent platforms 306
xx RECONFIGURABLE COMPUTING
Chapter 1

INTRODUCTION

Research in architecture of computer systems has always been a central pre-

occupation of the computer science and computer engineering communities.
The investigation goals vary according to the target applications, the price of
the final equipment, the programmability of the system, the environment in
which processors will be deployed and many others.
For processors to be used in parallel machines for high performance com-
puting as it is the case in weather simulation, the focus is placed on high clock
rates, parallelism and high communication bandwidth at the expense of power.
In many embedded systems, the price of the final equipment is the governing
factor during the development. A small microcontroller is usually used to con-
trol data acquisition from sensors and provide data to actuators at a very low
frequency. In many other embedded systems, in particular in untethered sys-
tems, power and cost optimization are the central goals. In those systems, the
growing need of more computation power that contradict with power and cost
optimization, put a lot of pressure on engineers, who must find a good balance
of all contradicting goals. For an autonomous cart used to explore a given en-
vironment, the processing unit must be able to capture images, compress the
images and send the compressed images to a base station for control. Paral-
lel to this, the system must perform other actions like obstacle detection and
avoidance. In such a system, power must be optimized to allow the system to
run as long as possible. On the other hand, the processor must process image
frames as fast as possible, to avoid important frames to be missed. Obstacle
detection and avoidance must also be done as faster as possible to avoid a pos-
sible crash of the cart. The multiplicity of goals has led to the development
of several processing architectures, each optimized according to a given goal.
Those architectures can be categorized in three main groups according to their
degree of flexibility: the general purpose computing group that is based on
2 RECONFIGURABLE COMPUTING

the Von Neumann computing paradigm; domain specific processors, tailored

for a class of applications having in common a great range of characteristics;
application specific processors tailored for only one application.

1. General Purpose Computing

In 1945, the mathematician John Von Neumann demonstrated in a study of
computation that a computer could have a simple, fixed structure, able to exe-
cute any kind of computation, given a properly programmed control, without
the need for hardware modification. The Von Neumann (VN) contribution was
universally adopted and quickly became the fundament of future generations of
high-speed digital computers. One of the reason for the acceptance of the Von
Neumann approach is its simplicity of programming that follows the sequential
way of human thinking.

Figure 1.1. The Von Neumann Computer architecture

The general structure of a Von Neumann machine as shown in figure 1.1

consists of:
A memory for storing program and data. Harvard-architectures contain two
parallel accessible memories for storing program and data separately
A control unit (also called control path) featuring a program counter that
holds the address of the next instruction to be executed.
An arithmetic and logic unit (also called data path) in which instructions
are executed.
A program is coded as a set of instructions to be executed sequentially, in-
struction after instruction. At each step of the program execution, the next
Introduction 3

instruction is fetched from the memory at the address specified in the program
counter and decoded. The required operands are then collected from the mem-
ory before the instruction can be executed. After execution, the result is written
back into the memory. In this process, the control path is in charge of setting
all signals necessary to read from and write to the memory, and to allow the
data path to perform the right computation. The data path is controlled by the
control path, which interprets the instructions and sets the data path’s signals
accordingly to execute the desired operation.
In general the execution of an instruction on a Von Neumann computer can
be done in five cycles: Instruction Read (IR) in which an instruction is fetched
from the memory; Decoding (D) in which the meaning of the instruction is
determined and the operands are localized; Read Operands (R) in which the
operands are read from the memory; Execute (EX) in which the instruction is
executed with the read operands; Write Result (W) in which the result of the
execution is stored back to the memory. In each of those five cycles only the
part of the hardware involved in the computation is activated. The rest remains
idle. For example if the IR cycle is to be performed, the program counter will
be activated to get the address of the instruction, the memory will be addressed
and the instruction register to store the instruction before decoding will be
also activated. Apart from those three units (program counter, memory and
instruction register), all the other units remain idle. Fortunately, the structure
of instructions allows several of them to occupy the idle part of the processor,
thus increasing the computation throughput.

1.1 Instruction Level Parallelism

Pipelining is a transparent way to optimize the hardware utilization as well
as the performance of programs. Because the execution of one instruction cy-
cle affects only a part of the hardware, idle parts could be activated by having
many different cycles executing together. For one instruction, it is not possible
to have many cycles being executed together. For instance, any attempt to per-
form the execute cycle (EX) together with the reading of an operand (R) for
the same instruction will not work, since the data needed for the execute cycle
should first be provided by the reading of operand cycle. Nevertheless, the two
cycles (EX) and (D) can be performed in parallel for two different instructions.
Once the data has been collected for the first instruction, its execution can start
while the data are being collected for the second instruction. This overlapping
in the execution of instructions is called pipelining or instruction level par-
allelism (ILP) and its aimed at increasing the throughput in the execution of
instructions as well as the resource utilization. It should be mentioned that ILP
does not reduce the execution latency of a single execution, but increases the
throughput of a set of instructions. The maximum throughput is dictated by
4 RECONFIGURABLE COMPUTING

the impact of hazards in the computation. Those Hazards can be reduced for
example by the use of a Harvard architecture.

Figure 1.2. Sequential and pipelined execution of instructions on a Von Neumann Computer

If tcycle is the time needed to execute one cycle, then the execution of one
instruction will require 5∗tcycle to perform. If three instructions have to be exe-
cuted, then the time needed to perform the execution of those three instructions
without pipelining is 15 ∗ tcycle . As illustrated in figure 1.2. Using pipelining,
the ideal time needed to perform those three instruction, when no hazards have
to be dealed with, is 7 ∗ tcycle . In reality we must take hazards into account.
This increase the overall computation time to 9 ∗ tcycle .
The main advantage of the Von Neumann computing paradigm is its flexi-
bility, since it can be used to program almost all existing algorithms. However,
each algorithm can be implemented on a Von Neumann computer only if it is
coded according to the Von Neumann rules. We say in this case that "The al-
gorithm must adapt itself to the hardware". Also due to the temporal use of the
same hardware for a wide variety of applications, Von Neumann computation
is often characterized as "temporal computation".
With the fact that all algorithms must be sequentially programmed in order
to run on a Von Neumann computer, many algorithms cannot be executed with
their potential best performance. Algorithms, which usually perform the same
set of inherent parallel operations on a huge set of data are not good candidates
for implementation on a Von Neumann machine.
If the class of algorithms to be executed is known in advance, then the pro-
cessor can be modified to better match the computation paradigm of that class
Introduction 5

of application. In this case, the data path will be tailored to always execute the
same set of operations, thus making the memory access for instruction fetching
as well as the instruction decoding redundant. Moreover, the memory access
for data fetching and storing can also be avoided if the sources and destinations
of data are known in advance. A bus could for instance provide sensor data to
the processor, which in turn sends back the computed data to the actuators
using another bus.

2. Domain Specific Processors

A domain specific processor is a processor tailored for a class of algorithms.
As mentioned in the previous section, the data path is tailored for an optimal
execution of a common set of operations that mostly characterizes the algo-
rithms in the given class. Also, memory access is reduced as much as possible.
DSPs (Digital Signal Processor) belong to the the most used domain specific
processors.
A DSP is a specialized processor used to speed-up computation of repetitive,
numerically intensive tasks in signal processing areas like telecommunication,
multimedia, automobile, radar, sonar, seismic, image processing, etc... The
most often cited feature of the DSPs is their ability to perform one or more
multiply accumulate (MAC) operations in single cycle. Usually, MAC oper-
ations have to be performed on a huge set of data. In a MAC operation data
are first multiplied and then added to an accumulated value. The normal Von
Neumann computer would perform a MAC in 10 steps. The first instruction
(multiply) would be fetched, then decoded, then the operand would be read and
multiply, the result would be stored back and the next instruction (accumulate)
would be read, the result stored in the previous step would be read again and
added to the accumulated value and the result would be stored back. DSPs
avoid those steps by using specialized hardware which directly performs the
addition after multiplication without having to access the memory.
Because many DSP algorithms involve performing repetitive computations,
most DSP processors provide special support for efficient looping. Often a
special loop or repeat instruction is provided, which allows a loop implementa-
tion without expending any instruction cycles for updating and testing the loop
counter or branching back to the top of the loop. DSPs are also customized for
data with a given width according to the application domain. For example if a
DSP is to be used for image processing, then pixels have to be processed. If
the pixels are represented in RGB (Red Green Blue) system where each colour
is represented by a byte, then an image processing DSP will not need more that
8 bit data path. Obviously, the image processing DSP cannot be used again for
applications requiring 32 bits computation.
This specialization of the DSPs increases the performance of the processor
and improves the device utilization. However, the flexibility is reduced, since
6 RECONFIGURABLE COMPUTING

it cannot be used anymore to implement other applications other than those for
which it was optimally designed.

3. Application Specific Processors

Although DSPs incorporate a degree of application specific features like
MAC and data width optimization, they still incorporate the Von Neumann
approach and, therefore, remain sequential machines. Their performance is
limited. If a processor has to be used for only one application, which is known
and fixed in advance, then the processing unit could be designed and optimized
for that particular application. In this case, we say that "the hardware adapts
itself to the application".
In multimedia processing, processors are usually designed to perform the
compression of video frames according to a video compression standard. Such
processors cannot be used for something else than compression. Even in com-
pression, the standard must exactly match the one implemented in the proces-
sors. A processor designed for only one application is called an Application
Specific Processor (ASIP). In an ASIP, the instruction cycles (IR, D, EX, W)
are eliminated. The instruction set of the application is directly implemented in
hardware. Input data stream in the processor through its inputs, the processor
performs the required computation and the results can be collected at the out-
puts of the processor. ASIPs are usually implemented as a single chips called
Application Specific Integrated Circuit (ASIC)

Example 1.1 If algorithm 1 has to execute on a Von Neumann computer,

then at least 3 instructions are required.

Algorithm 1
if a < b then
d=a+b
c=a·b
else
d=b+1
c=a−1
end if

With tcycle being the instruction cycle, the program will be executed in 3 ∗
5 ∗ tcycles = 15 ∗ tcycle without pipelining.
Let us now consider the implementation of the same algorithm in an ASIP.
We can implement the instructions d = a + b and c = a ∗ b in parallel. The
same is also true for d = b + 1, c = a − 1 as illustrated in figure 1.3
The four instruction a + b, a ∗ b, b + 1, a − 1 as well as the comparison
a < b will be executed in parallel in a first stage. Depending on the value of
Introduction 7

Figure 1.3. ASIP implementation of Algorithm 1.

the comparison a < b, the correct values of the previous stage computations
will be assigned to c and d as defined in the program. Let tmax be the longest
signal needed by a signal to move from one point to another in the physical
implementation of the processor (this will happen on the path Input-multiply-
multiplex). tmax is also called the cycle time of the ASIP processor. For two
inputs a and b, the results c and d can be computed in time tmax . The Von
Neumann processor can compete with this ASIP only if 15 ∗ tcycle < tmax , i.e.
tcycle < tmax /15. The Von Neumann must be at least 15 times faster than the
ASIP to be competitive. Obviously we have assumed a Von Neumann without
pipeline. The case, where a Von Neumann computer with a pipeline is used,
can be treated in the same way.

ASIPs use a spatial approach to implement only one application. The func-
tional units needed for the computation of all parts of the application must be
available on the surface of the final processor. This kind of computation is
called "Spatial Computing".
Once again, an ASIP that is built to perform a given computation cannot be
used for other tasks other than those for which it has been originally designed.
8 RECONFIGURABLE COMPUTING

4. Reconfigurable Computing
From the discussion in the previous sections, where we studied three differ-
ent kinds of processing units, we can identify two main means to characterize
processors: flexibility and performance.

The Von Neumann computers are very flexible because they are able to
compute any kind of task. This is the reason why the terminology GPP
(General Purpose Processor) is used for the Von Neumann machine. They
don’t bring so much performance, because they cannot compute in parallel.
Moreover the five steps (IR, D, R, EX, W) needed to perform one instruc-
tion becomes a major drawback, in particular if the same instruction has
to be executed on huge sets of data. Flexibility is possible because "the
application must always adapt to the hardware" in order to be executed.

ASIPs bring much performance because they are optimized for a particular
application. The instruction set required for that application can then be
built in a chip. Performance is possible because "the hardware is always
adapted to the application".

If we consider two scales, one for the performance and the other for the flex-
ibility, then the Von Neumann computers can be placed at one end and the
ASIPs at the other end as illustrated in figure 1.4

Figure 1.4. Flexibility vs performance of processor classes

Introduction 9

Between the GPPs and the ASIPs are a large numbers of processors. De-
pending on their performance and their flexibility, they can be placed near or
far from the GPPs on the two scales.
Given this, how can we choose a processor adapted to our computation
needs? If the range of applications for which the processor will be used is large
or if it is not even defined at all, then the GPP should be chosen. However if
the processor is to be used for one application like it is the case in embedded
systems, then the best approach will be to design a new ASIP optimized for
that application.
Ideally, we would like to have the flexibility of the GPP and the performance
of the ASIP in the same device. We would like to have a device able "to adapt
to the application" on the fly. We call such a hardware device a reconfigurable
hardware or reconfigurable device or reconfigurable processing unit (RPU)
in analogy the Central Processing Unit (CPU). Following this, we provide a
definition of the term reconfigurable computing. More on the taxonomy in
reconfigurable computing can be found in [111] [112].

Definition 1.2 (Reconfigurable Computing) Reconfigurable com-

puting is defined as the study of computation using reconfigurable devices.

For a given application, at a given time, the spatial structure of the device
will be modified such as to use the best computing approach to speed up that
application. If a new application has to be computed, the device structure
will be modified again to match the new application. Contrary to the Von
Neumann computers, which are programmed by a set of instructions to be
executed sequentially, the structure of reconfigurable devices are changed by
modifying all or part of the hardware at compile-time or at run-time, usually
by downloading a so called bitstream into the device.

Definition 1.3 (Configuration, Reconfiguration) Configuration

respectively reconfiguration is the process of changing the structure of a recon-
figurable device at star-up-time respectively at run-time

Progress in reconfiguration has been amazing in the last two decades. This
is mostly due to the wide acceptance of the FPGAs (Field Programmable Gate
Array) that are now established as the most widely used reconfigurable devices.
The number of workshops, conferences and meetings dealing with this topics
has also grown following the FPGA evolution. Reconfigurable devices can be
used in a wide number of fields, from which we list some in the next section.

5. Fields of Application
In this section, we would like to present a non exhaustive list of fields, where
the use of reconfiguration can be of great interest. Because the field is still
10 RECONFIGURABLE COMPUTING

growing, several new fields of application are likely to be developed in the

future.

5.1 Rapid Prototyping

Rapid prototyping is certainly one of the most important fields of applica-
tion of reconfiguration. The development of an ASIC (Application Specific
Integrated Circuit) that can be seen as physical implementation of ASIPs, is a
cumbersome process consisting of several steps, from the specification down to
the layout of the chip and the final production. Due to the different optimiza-
tion goals in development, several teams of engineers are usually involved.
This increases the NRE (Non Recurring Engineering) cost that can only be
amortized if the final product is produced in a very large quantity. Contrary to
software development, errors discovered after the production mean enormous
loss because the produced pieces become either unusable or a great adaptation
effort must be spent for the deployment. Rapid prototyping allows a device
to be tested in real hardware before the final production. In this sense, errors
can be corrected without affecting the pieces already produced. A reconfig-
urable device is useful here, since it can be used several times to implement
different versions of the final product until an error free state. One of the con-
cepts related to rapid prototyping is hardware emulation, in analogy to soft-
ware simulation. With hardware emulation, a hardware module is tested under
real operating conditions in the environment where it will be deployed later.

5.2 In System Customization

Time to market has become one of the main challenges that electronic man-
ufacturers face today. In order to secure market segments, manufacturers must
release their products as quickly as possible. In many cases, a well working
product can be released with less functionalities. The manufacturer can then
upgrade the product on the field to incorporate new functionalities. While this
approach works with normal Von Neumann processors, it is not the case with
ASIPs. A reconfigurable device provides such capabilities of being upgraded
in the field, by changing the configuration.
In system customization can also be used to upgrade systems that are de-
ployed into non accessible or very difficult to access locations. One example
is the Mars rover vehicle in which some FPGAs, that can be modified from the
earth, are used.

5.3 Multi modal Computation

The number of electronic devices we interact with is permanently increas-
ing. Many people hold besides a mobile phone, other devices like handhelds,
portable mp3 player, portable video player, etc... Besides those mobile devices,
Introduction 11

fixed devices like navigation systems, music and video players as well as TV
devices are available in cars and at home. All those devices are equipped with
electronic control units that run the desire application on the desired device.
Furthermore, many of the devices are use in a time multiplexed fashion. It is
difficult to imagine someone playing mp3 songs while watching a video clip
and given a phone call. For a group of devices used exclusively in a time mul-
tiplexed way, only one electronic control unit can be used. Whenever a service
is needed, the control unit is connect to the corresponding device at the correct
location and reconfigure with the adequate configuration. For instance a do-
mestic mp3, a domestic DVD player, a car mp3, a car DVD player as well as a
mobile mp3 player and a mobile video player can all share the same electronic
unit, if they are always used by the same person. However, if several persons
have access to the same devices (this can happen in a household with several
people), then sharing of the electronic control unit will be rather difficult. In
the first case, the user just needs to remove the control unit from the domestic
devices and connect them to one car device when going to work. The control
unit can be removed from the car and connected to a mobile device if the user
decides to go for a walk. Coming back home, the electronic control unit is
removed from the mobile device and used for watching video.

5.4 Adaptive Computing Systems

Advances in computation and communication are helping the development
of ubiquitous and pervasive computing. Computer anytime and everywhere is
now becoming reality.
The design of ubiquitous computing system is cumbersome task that cannot
be deal with only at compile time. Due to uncertainty and unpredictability of
such systems, it is impossible, at compile-time, to address all scenarios that
can happen at run-time, due to unpredictable changes in the environment. We
need computing systems that are able to adapt their behavior and structure to
changing operating and environmental conditions, to time-varying optimizing
objectives, and to physical constraints like changing protocols and new stan-
dards. We call those computing systems Adaptive Computing System.
Reconfiguration can provide a good fundament for the realization of adap-
tive systems, because it allows system to quickly reacts to changes by adopting
the optimal behavior for a given run-time scenario.

6. Organization of the Book

The rest of the book is organized in 9 chapters, ranging from the architec-
ture to the applications. In those chapters, we try to provide a mixture of the
theoretical as well as the practical background needed to understand, develop
and deploy reconfigurable systems.
12 RECONFIGURABLE COMPUTING

Architecture of reconfigurable systems: After a brief tour in the earlier

systems, this chapter considers the technology as well as the coupling pos-
sibilities of reconfigurable systems, from the fine grained look up table
(LUT) based reconfigurable systems like the field programmable gate ar-
rays (FPGA) to the new coarse-grained technology.
Design and implementation: This chapter deals with the implementation
on reconfigurable system. It covers the steps needed (design entry, func-
tional simulation, logic synthesis, technology mapping, place and route and
bit stream generation) to implement today’s FPGAs. We focus deeply on
the logic synthesis for FPGAs, in particular LUT technology mapping.
High-Level Synthesis for Reconfigurable Devices: This chapter consid-
ers the high-level synthesis for reconfigurable systems, also known as tem-
poral partitioning. It covers the implementation of large functions, which
cannot fit into one RPU. Several temporal partitioning techniques are pre-
sented and explained.
Temporal placement: In this chapter, stand alone reconfigurable systems
are considered. We assume that a kind of operating systems for reconfig-
urable systems is in charge of managing the resources of a given system and
allocate space on a device for the computation of incoming tasks. We there-
fore present several temporal placement approaches for off-line as well as
on-line placement.
On-line Communication: Modules dynamically placed at run-time on a
given device need to communicate with each other in order to exchange
data. Therefore, they dynamically create a need of communication chanels
on the chip. This chapter reviews and explains the different approaches to
solve this dynamic intercommunication need.
Designing for Partial Reconfiguration on Xilinx Virtex FPGA: This
chapter considers the implementation for partial reconfiguration on the Xil-
inx FPGAs that are one of the few one on the market with this feature. We
present the different possibilities to produces the partial bitstream needed
at run-time to reconfigure only part of the device. Also, based on a case
study, some hints are provided on the design of viable platforms for partial
reconfiguration.
System on Programmable Chip: System on programmable chip is a hot
topic in reconfigurable computing. This is mainly the integration of a sys-
tem made upon some peripheral (UART, Ethernet, VGA, etc.), but also
computational (coding, filter, etc.) hardware modules on one programmable
chip. We present the current usable solutions: Furthermore, we focus on
the development of adaptive multiprocessors on chip. Those are systems
Introduction 13

consisting of a set of Processors and exchangeable hardware accelerators

connected in a network for a parallel implementation of applications on the
chip.
Applications: This section presents applications of reconfigurable systems.
It covers the use of reconfigurable system in computer architecture (rapid
prototyping, reconfigurable supercomputer, reconfigurable massively par-
allel computers) and algorithm better adapted for reconfigurable systems
(distributed arithmetic, pattern matching, control, software defined radio,
cryptography, etc...)
Chapter 2

RECONFIGURABLE ARCHITECTURES

In the previous chapter, we presented the reconfigurable devices as piece of

hardware able to dynamically adapt to algorithms. How can it be possible that
a hardware device, whose structure is normally fixed at fabrication time and
cannot be changed anymore during the lifetime, can be readapted at run-time
to dynamically match the application requirements? In this chapter we bring
some light into this issue by presenting the architecture of the commonly used
reconfigurable devices.
After taking a brief tour in the evolution of reconfigurable computing archi-
tectures, we will cover in the first part of the chapter the so called fine-grained
reconfigurable devices. Those are hardware devices, whose functionality can
be modified on a very low level of granularity. For instance, the device can
be modified such as to add, or remove a single inverter or a single two input
NAND-gate. Fine-grained reconfigurable devices are mostly represented by
Programmable Logic Devices (PLD). This consists of Programmable Logic
Arrays (PLA), Programmable Array Logics (PAL), Complex Programmable
Logic Devices (CPLD) and the Field Programmable Gate Arrays (FPGA).
In the second part of the chapter, we cover the pseudo reconfigurable device
also called coarse-grained reconfigurable devices. Those devices usually con-
sist of a set of coarse-grained elements like ALUs that allow for the efficient
implementation of a dataflow function.

1. Early Work
Attempts to have a flexible hardware structure that can be dynamically mod-
ified at run-time to compute a desired function are almost as older as the devel-
opment of other computing paradigms. In order to overcome the non flexibility
of the first computer, the ENIAC (Electronical Numerical Integrator and Com-
puter) that could be programmed only by hardwiring an algorithm, Jon Von
16 RECONFIGURABLE COMPUTING

Neumann proposed a first universal architecture made upon three main blocks
(memory, datapath and control path), able to run any given and well coded
program. The Von Neumann approach was not intended to provide the most
efficient hardwired structure for each application, but a platform to run all type
of programs without spending too much effort in the rearrangement of the un-
derlying hardware. Thee previous effort of Von Neumann was then pursued by
other researchers with goal of always have the best computation structure of a
given application. We present some of those investigations in this section.

1.1 The Estrin Fix-plus machine

In 1959, Gerald Estrin, a computer scientist of the university of California
at Los Angeles introduced the concept of reconfigurable computing. The fol-
lowing fragment of an Estrin publication in 1960 [81] the fix-plus machine,
defines the concept of reconfigurable computing paradigm.
“Pragmatic problem studies predicts gains in computation speeds in a va-
riety of computational tasks when executed on appropriate problem-oriented
configurations of the variable structure computer. The economic feasibility of
the system is based on utilization of essentially the same hardware in a variety
of special purpose structures. This capability is achieved by programmed or
physical restructuring of a part of the hardware.”
To implement its vision, Estrin designed a computing system, the fix-plus
machine [81]. Like many reconfigurable computing systems available today,
the fix-plus machine consists of 3 main elements as shown in figure 2.1.
A high speed general purpose computer, the fixed part (F) that can be im-
plemented on any general purpose processor. In its fix-plus machine, the
IBM 7090 were used.
A variable part (V) consisting of various size high speed digital substruc-
tures that can be reorganized in a problem oriented special purpose config-
urations.
The supervisory control (SC) coordinates operations between the fix mod-
ule and the variable module.
The Variable Part (V) was made upon a set of problem specific optimized
functional units in the basic configuration (trigonometric functions, logarithm,
exponentials, n-th power, roots, complex arithmetic, hyperbolic, matrix opera-
tion).
The basic block modules (figure 2.2) could be inserted into any of 36 po-
sitions on a motherboard (figure 2.12(a)) that provides the functionality for a
given application. The functionality of the motherboard could be manually
modified by replacing some basic blocks by new ones. Two types of basic
building were available in the fix-plus machine: The first basic element that
Reconfigurable Architectures 17

Figure 2.1. Structure of the Estrin Fix-Plus Machine

Figure 2.2. The basic building blocks of Fix-Plus Machine

entailed four amplifiers and associated input logic for signal inversion, ampli-
fication, or high-speed storage. The second basic block used as combinatoric
that was made upon ten diodes and four output drivers.
The connection between the modules is done by wiring harness (figure 2.3).
With this architecture, the reconfiguration was done manually by replacing
18 RECONFIGURABLE COMPUTING

Figure 2.3. The wiring harness of the Estrin-Machine

Figure 2.4. The motherboard of the Estrin’s Fix-Plus

some modules on the mother board (figure 2.4) or by changing a wiring harness
for a new connection among the existing modules.
The fix-plus machine was intended to be used for accelerating Eigenvalues
computation of matrices and has shown a speed gain of 2.5 to 1000 over the
IBM7090 [80] [79]. The available technology at that time however made the
use of the fix-plus machine difficult. Reconfiguration had to be done manually,
and substantial software efforts were required to implement applications.
Reconfigurable Architectures 19

Figure 2.5. Estrin at work: Hand-Controled Reconfiguration

1.2 The Rammig Machine

In the year 1977, Franz J. Rammig, a researcher at the university of Dort-
mund proposed a concept for editing hardware [182]. The goal was the “in-
vestigation of a system, which, with no manual or mechanical interference,
permits the building, changing, processing and destruction of real (not simu-
lated) digital hardware”.
Rammig materialized he’s concept by developing a hardware editor similar
to the today’s Field Programmable Gate Array architecture. The editor was
made upon a set of modules, a set of pins and a one-to-one mapping function on
the set of pins. The circuitry of a given function was then defined as a “string”
on an alphabet of two letters (w = “wired” and u = “unwired”). In order
to build the hardware editor, selectors were provided with the modules outputs
connected to the input of the selectors and the output of the selectors connected
to the input of the modules. The overall system architecture is shown in figure
2.6.
The implementation of the {wired, unwired} property was done through
a programmable crossbar switch, made upon an array of selectors. The bit
strings were provided by storing the selector control in registers, and by mak-
ing these registers accessible from a host computer, the PDP11 in those days.
The modules were provided on a library board similar to that of Estrin’s Fix-
Plus. Each board could be selected under software control. The mapping from
module I/Os to pins was realized manually, by a wiring of the provided library
boards, i.e. fixed per library board.
20 RECONFIGURABLE COMPUTING

Figure 2.6. Structure of the Rammig machine

Figure 2.7. META-46 GOLDLAC

The Rammig machine was heavily used as emulation platform, which is also
one of the largest application fields of today’s FPGAs. It was therefore possible
to control the complete behaviour of a circuit under observation from the soft-
ware side. This was done by buffering module outputs in registers and transfer
the contents of these registers to the host before clocking the system. The sys-
tem was implemented with a 128x192 crossbar and got the name META-46
GOLDLAC (figure 2.7).

1.3 Hartenstein’s XPuter

The Xputer’s [113] [216] concept was presented in early 80’s by Reiner
Hartenstein, a researcher at the University of Kaiserslautern in Germany.
Reconfigurable Architectures 21

Figure 2.8. General architecture of the XPuter as implemented in the Map oriented machine
(MOM-3) prototype.

The goal was to have a very high degree of programmable parallelism into
the hardware, at lowest possible level, in order to obtain performance not possi-
ble with the Von Neumann computers. Instead of sequencing the instructions,
the Xputer used to sequences data, thus exploiting the regularity in the data
dependencies of some class of applications like in image processing, where a
repetitive processing is performed on a large amount of data. An Xputer con-
sists of three main parts: the data sequencer, the data memory and the reconfig-
urable ALU (rALU) that permits the run-time configuration of communication
at levels below instruction set level. Within a loop, data to be processed were
accessed via a data structure called scan window. Data manipulation was done
by the rALU that had access to many scan windows. The most essential part of
the data sequencer were the Generic Address Generator (GAG) that were able
to produce address sequences corresponding to the data of up to three nested
loops. A rALU subnet that could be configured to perform all computations on
the data of a scan window is generally, was required for each level of a nested
loop.
The general XPuter architecture is presented on figure 2.8. Shown is the
realization of the XPuter as a Map Oriented Machine (MoM). The overall sys-
tem was made upon of a host processor, whose memory was accessible by the
MoM.
The rALU subnets received their data directly from local memory or from
the host main memory via the MoM bus. Communication was also possible
22 RECONFIGURABLE COMPUTING

Figure 2.9. Structure of the XPuter’s reconfigurable ALU

among the rALUs via direct serial connections. Several XPuters could also be
connected in order to provide more performance.
For the executing a program, the hardware had first to be configured. If no
reconfiguration would take place at run-time, then only the data memory was
necessary. Otherwise, a configuration memory would be required to hold all
the configurations to be used at run-time.
The basic building block of the reconfigurable ALU was the so called recon-
figurable Datapath Unit (rDPU) (figure 2.9). Several rDPUs were used within
an rALU for data manipulation. Each rDPU had two registered inputs and two
registered ouputs with a datawidth of 32 bit. Input data were provided either
from the north or from the west, while the south and east were used for the
output. Besides the interconnection lines for the rALUs, a global I/O-Bus is
available for the connection of designs to the external world. The I/O bus was
principally used for accessing the scan windows.
The global view of the reconfigurable datapath attached to a host processor
is given on figure 2.9. It consists of two main parts: the control unit and a field
of rDPUs. The register file was used for optimizing the memory access when
the GAG operated with overlapping scan windows. In this case, data in the
actual scan window position will be reused in the following positions. Those
data could therefore be temporary stored in registers and copied back into the
memory if they were no more needed.
The control implemented a program that is loaded on reconfiguration to
control different units of the rALU. Its instruction set consisted of instructions
for loading the data as well as instructions for collecting results from the field.
Reconfigurable Architectures 23

Figure 2.10. Programmable Active Memory (PAM) architecture as array of PAB

Application of the XPuters were in image processing, systolic array and sig-
nal processing.

The introduction of commercial Field Programmable Gate Arrays by Xilinx

in the mid-1980s increased interest in reconfigurable computing. Many sys-
tems in high performance computing were built with farm of FPGAs on dif-
ferent boards. The "drawbacks" of FPGAs also inspired many research groups
that propose new architectures. We briefly present some of the most popular
experiments. Included are the PAM developed at the DEC research center in
Paris, the SPLASH system of the Supercomputer Research Center, the PRISM,
the DISC and the GARP.

1.4 The PAM Machine

Introduced by Bertin et al. [25] from the Digital Equipment corporation, a
Programmable Active Memory (PAM) is defined as a uniform array of identical
cells, the Processing Array Blocks (PAB) that are connected in a given regular
fashion. The architecture is built on the FPGA model that will be presented
in a later in this chapter. The basic building block in a PAM is a cell, which
consist of a flip flop and a combinatorial element with four inputs and four out-
puts. The combinatorial element consists of five look-up tables, each of which
implements an output as a function of the five inputs. 160 bits are therefore,
necessary to configure a combinatorial element. The flip flop provides a mem-
ory for the realization of finite state machines, but can be used for local data
storage. Figure 2.10 shows a PAM consisting of a 4x4 array of PAB.
Like in the FPGAs, the PAM configuration is defined by a bitstream that set
the functions of the combinatorial elements as well as the state of the flip flops.
24 RECONFIGURABLE COMPUTING

A prototype of the PAM, the Perle was built using a 5x5 array of Logic
Array Cell (LCA) , a CMOS cell designed by Xillinx. Perle-0 features a VME
bus to interface a host CPU to which it is coupled. The host is in charge of
downloading the bitstream to the Perle. The configuration controller, as well
as the host-bus communication protocol are programmed into two extra LCAs,
statically configured at power-up time from a PROM.
Example of applications for which the Perle was used to implement, are data
compression, cryptography, image processing, energetic physics.

1.5 The SPLASH II

SPLASH II [97], [39] is a computing system made upon a set of FPGA
modules solded on different printed circuit boards, connected together to build
a massive parallel computer. The system consists of a set of boards, called ar-
ray boards (figure 2.11), that can be reconfigured according to the application
currently being implemented. The array boards are connected to a host com-
puter (in this case a SUN SPARC Station 2) via a system Bus (SBUS) and an
interface board. Each array board features 17 Xilinx XC4010 FPGAs. While
sixteen of the seventeen FPGAs are used for computation, the seventeenth is
use for data broadcasting. Each of the FPGAs has access to a separate 512
Kbytes memory. The SLASH II can be viewed as a large serial data path from
the interface board to the array board. The SIMD bus is used to transmit data
to the first FPGA (X0) on each board, which can then broadcast the data to the
other sixteen FPGAs on its array. Alternatively, the data can also be moved in
a linear way on an array board, from FPGA to FPGA. In this case, the data are
received by the first FPGA (X1) that serially transmit the data to the next FP-
GAs on the board. The data is transmitted to the next board by the last FPGA
in the chain.
Some examples of applications that have benefited from the SPLASH II
performance are image processing, text searching in genetic database and fin-
gerprint matching.

1.6 The PRISM Paradigm

PRISM is an acronym for Processor Reconfiguration through Instruction Set
Metamorphosis. Developed by Athanas et al [12], the idea behind the PRISM
is to use information extracted at compiled-time from a given application and,
to build a new set of instructions tailored to that application. The so gener-
ated instructions can then be implemented in a reconfigurable hardware unit,
tightly coupled to the processor. In order to ease the programmability of such
systems, the identification and synthesis of the operations in hardware should
be done automatically and transparent to the user. This task is performed by
a configuration compiler, which automatically performs a hardware/software
Reconfigurable Architectures 25

Figure 2.11. Architecture of the SPLASH II array board

partitioning. This results in the generation of software and a hardware im-

age. The software image is a sequential program to be executed on the pro-
cessor, while the hardware image defines the configuration of the attached re-
configurable hardware. As prof of concept, a PRISM prototype consisting of a
10MHz processor M68010 attached to a board with four Xilinx 3090 FPGAs
was constructed and tested.

1.7 The Garp Approach

The approach proposed by Hauser and Wawrzynek [114] is very close to
today’s embedded FPGA systems. The Garp was proposed as attempt to over-
come the obstacles of reconfigurable computing machines, mostly consisting
of a processor attached to a set of FPGA boards. The problems identified were
for instance, the small capacity of FPGA, their expensive reconfiguration time,
the non availability of internal memory to temporally stored data to be com-
puted and the lack of a clear defined standard for coupling FPGAs to processor.
Hauser and Wawrzynek proposed a system consisting of a microprocessor and
a reconfigurable array on the same die. The system was not different than
those already proposed. However, having the processor on the same die with
the reconfigurable device was a great benefit. The time for transferring data
to the reconfigurable array could be drastically reduced. Instructions and data
caches were also foreseen on the chip. Almost all the FPGAs available today
on the market offer a similar structure, with processors that can be synthesized
according to the user’s need. Despite the nice concept and positive simulation
results, a Garp-chip could not be produced and commercialized.
26 RECONFIGURABLE COMPUTING

1.8 DISC
The purpose of the Dynamic Instruction Set Computer (DISC) presented in
[219] is to support demand-driven instruction set modifications. In contrast to
the PRISM approach where the specific instructions are synthesized and fixed
at compiled-time, the DISC approach uses partial reconfiguration of FPGAs
to place on the FPGA, hardware modules, each of which implements a given
instruction. The relocation was also proposed as a means to reduced defrag-
mentation of the FPGA. Due to its partial reconfiguration capabilities, the Na-
tional Semiconductor Configurable Logic Array (Clay) was chosen for build-
ing a prototype consisting of a printed circuit board with two CLA31 FPGA
and some memory. While the first FPGA was used to control the configuration
process, the second FPGA was used for implementing the instruction specific
hardware blocks. Via an ISA Bus, the board was attached to a host processor
running on Linux. A simple image mean filter first implemented as applica-
tion specific module, and later as sequence of general purpose instructions was
used to show the viability of the platform.

1.9 The DPGA

With similar goals as the Garp architecture, Dehon’s Dynamically Pro-
grammable Gate Array was proposed as a mean to extend the capability of
microprocessor [68]. However, DGPAs provide a better implementation of the
reconfigurable array proposed in the Garp. The concept proposed by Dehon is
in used in many coarse-grained devices available today. In order to reduce the
reconfiguration time, one of the main bottlenecks of early FPGAs, the DGPA
should be able to quickly switched among preloaded configurations. There-
fore, redundant look up tables are used to broadcast configurations in a local
area on a cycle-by-cycle basis, thus allowing a clockwise reconfiguration of
the FPGAs.

Despite the recent development of flexible coardse-grained device, almost

all systems that require flexibility in the hardware structure rely on the pro-
grammable devices to which the PALs and PLAs, the CPLDs and the FPGA
belongs. In the next section, we provide a short view on the structure of those
devices.

2. Simple Programmable Logic Devices

Programmable logic arrays (PLA) and programmable array logic (PAL)
consist of a plane of AND-gates connected to a plane of OR-gates. The inputs
signals as well as their negations are connected to the inputs of the AND-Gates
in the AND-plane. The outputs of the AND-gates are use as input for the OR-
Reconfigurable Architectures 27

(a) PAL: The OR-connections are fixed (b) PLA: Programmable OR-connections

Figure 2.12. PAL and PLA implementations of the functions F 1 = A · B + A · B and

F2 = A · B + B · C

gate in the OR-plane whose outputs correspond to those of the PAL/PLA. The
connections in the two planes are programmable by the user.
Because every Boolean function can be written as a sum of products, the
PLAs and PALs can be used to program any Boolean function after fabrica-
tion. The products are implemented in the AND-plane by connecting the wires
accordingly and the sums of the product are realized in the OR-plane. While in
the PLAs both fields can be programmed by the user, this is not the case in the
PALs where only the AND-plane is programmable. The OR-plane is fixed by
the manufacturer. Therefore, the PALs can be seen as a subclass of the PLAs.
Example 2.1 Figure 2.12 shows the PLA and PAL implementations of the
functions F 1 = A·B+A·B and F 2 = A·B+B·C. While the OR-plane of the
PAL is no more programmable, by modifiying the connection in the OR-Plane
of the PLA, diferent sums of the products can be generated.
Further enhancements (feeding the OR-output to a flip flop or feeding back a
OR-output to a AND-input) are made on PLAs to allow more complex circuits
to be implemented.
PALs and PLAs are well suited to implement two-level circuits, those are
circuits made upon the sum of product as described earlier. At the first level, all
the products are implemented and all the sum are implemented on the second
level.
The main limitation of PLAs and PALs is their low capacity, which is due
to the nature of the AND-OR-plane. The size of the plane grows too quickly
as the number of inputs increases. Due to their low complexities, PALs and
PLAs belong to the class of devices called simple programmable logic devices
(SPLD).
28 RECONFIGURABLE COMPUTING

Figure 2.13. Structure of a CPLD device

3. Complex Programmable Logic Device

As stated in the previous section, PALs and PLAs are only available in small
sizes, equivalent to a few hundred logic gates. For large logic circuits, Complex
Programmable Logic Devices (CPLD) can be used.
A CPLD consist of a set of macro cells, Input/Output blocks and an in-
terconnection network. The connection between the Input/Output blocks and
the macro cells and those between macro cells and macro cells can be made
through the programmable interconnection network (figure 2.13). A macro
cell typically contains several PLAs and flip flops. Despite their relative large
capacity (few hundreds thousands of logic gates), compared to those of PLAs,
CPLDs are still too small for the use in reconfigurable computing devices.
They are usually used as glue logic, or to implement small functions. Due to
their non volatility, CPLDs are used in many systems for configuration of the
main reconfigurable device at star up.

4. Field Programmable Gate Arrays

Introduced in 1985 by the company Xilinx, a Field Programmable Gate Ar-
ray (FPGA) is a programmable device consisting, like the CPLDs, of three
main parts. a set of programmable logic cells also called logic blocks or con-
figurable logic blocks, a programmable interconnection network and a set of
input and output cells around the device (figure 2.14). A function to be imple-
mented in FPGA is partitioned in modules, each of which can be implemented
in a logic block. The logic blocks are then connected together using the pro-
grammable interconnection. All three basic components of an FPGA (logic
Reconfigurable Architectures 29

Figure 2.14. Structure of an FPGA

block, interconnection and input output) can be programmed by the user in

the field. FPGAs can be programmed once or several times depending on the
technology used.

4.1 Technology
The technology defines how the different blocks (logic blocks, intercon-
nect, input/output) are physically realized. Basically, two major technologies
exist: Antifuse and memory-based. While the antifuse paradigm is limited to
the realization of interconnexion, the memory-based paradigm is used for the
computation as well as the interconnection. In the memory-based category, we
can list the SRAM the EEPROM and the Flash based FPGAs.

4.1.1 Antifuse
Contrary to a fuse an antifuse is normally an open circuit. An antifuse based
FPGAs uses special antifuses included at each connection customization point.
The two-terminal elements are connected to the upper and lower layer of the
antifuse, in the middle of which a dielectric is placed (figure 2.15). In its initial
state, the high resistance of the dielectric does not allow any current to flow
between the two layers. Applying a high voltage causes large power dissi-
pation in a small area, which melts the dielectric. This operation drastically
reduces the resistance and a link can be be built, which permanently connects
the two layers. The two types of antifuses actually commercialized are: The
PLICE (Programmable Low-Impedance Circuit Element) (2.15(a)), which is
30 RECONFIGURABLE COMPUTING

(a) Q-Logic Vialink antifuse (b) Actel PLICE antifuse

Figure 2.15. Antifuse FPGA Technology

Figure 2.16. A Xillinx SRAM cell

manufactured by the by the company Actel and the Metal Antifuse also called
ViaLink made by the company QuickLogic.
The PLICE antifuse consists of an Oxygen-Nitrogen-Oxygen (ONO) dielec-
tric layer sandwiched between a polysilicon and an n+ diffusion layer that
serves as conductor. The ViaLink antifuse (2.15(b)) is composed of a sandwich
of very high resistance layer of programmable amorphous silicon between two
metal layers. When a programming voltage is applied, a metal-to-metal link is
formed by permanently converting the silicon to a low resistance state.
The main advantage of the antifuse chips is their small area and their sig-
nificantly lower resistance and parasitic capacitance compared to transistors.
This help to reduce the RC delays in the routing. However, anti-fuse based
FPGAs are not suitable for devices which must be frequently reprogrammed,
as it is the case in reconfigurable computing. Antifuse FPGAs are normally
programmed once by the user and will not change anymore. For this reason,
they are also known as one-time programmable FPGAs.

4.1.2 SRAM
Unlike the antifuse that is used mostly used to configure the connection, a
Static RAM (SRAM) is use to configure the logic blocks and the connection
as well. SRAM-based FPGAs are the most widely used.
In an SRAM-based FPGA, the states of the logic blocks, i.e. their function-
ality bits as well as that of the interconnections are controlled by the output of
SRAM Cells (figure 2.16).
Reconfigurable Architectures 31

(a) Wire connection (b) MUX programming (c) Function generation

Figure 2.17. Use of SRAM in FPGA-Configuration

As shown in figure 2.17(a), a connection between two wires can be done by a

pass transistor connected to the wires (one wire on a terminal). The state of the
transistor can then be controlled by the output of a SRAM cell which allows
the current to flow or not to flow between the two wires. Copying different
values into the SRAM allows for changing the behaviour of the connection.
The same principle applies for the programming of the logic modules, the so
called function generators. The selector input of a multiplexer (figure 2.17(b))
as well as the content of a look-up table representing the values of a function
for different input combinations (figure 2.17(c)) can be reprogrammed by just
copying new values into the corresponding SRAM cells.
The basic SRAM configuration cell (figure 2.16) is constructed from two
cross-coupled inverters and uses a standard CMOS-process. The Q-output of
the SRAM cell is connected to the module (pass transistor, multiplexer or Look
up table) that is controlled by configuration. The read and write word line is
use either to read or write data from or to the SRAM through the data line.
The major advantage of this technology is that FPGAs can be programmed
(configured) indefinitely. We just need to change the value into the SRAM-
cells to realize a new connection or a new function. Moreover, the device can
be done in-circuit very quickly and allow the reconfiguration to be done on the
fly.
The disadvantages of SRAM-Based FPGAs are the chip area required by the
SRAM approach is relatively large. The total size of an SRAM-configuration
cell plus the transistor switch that the SRAM-cell drives is also larger than the
programming devices used in the antifuse technology. Furthermore, the device
is volatile, i.e. the configuration of the device stored in the SRAM-cells is lost
if the power is cut off. Therefore external storage or non-volatile devices like
CPLDs, EPROM, or Flash devices, are required to store the configuration and
load it into the FPGA-device at power-on.
32 RECONFIGURABLE COMPUTING

(a) EPROM Cell (b) Actel Flash-EEPROM

Figure 2.18. EEPROM Technology

4.1.3 EPROM
Erasable Programmable Read Only Memory (EPROM) devices are based on
a floating gate (figure 2.18(a)). The device can be permanently programmed
by applying a high voltage (10-21 V) between the control gate and the drain
of the transistor (12 V). This causes the floating gate to be permanently and
negatively charged. This negative potential on the floating gate compensates
the voltage on the control gate and keeps the transistor closed.
In an UV (Ultra Violet) PROM, the programming process can be reversed
by exposing the floating gate to UV-light. This process reduces the threshold
voltage and makes the transistor function normally. For this purpose, the device
must be remove from the system in which it operates and plug into a special
device.
In Electrically Erasable and Programmable ROM (EEPROMs) as well as
in Flash-EPROM, the the erase operation is accomplished electrically, rather
than by exposure to ultraviolet light. A high negative voltage must therefore
be applied at the control gate. This process is faster than using a UV lamp and
the chip does not have to be removed from the system. In EEPROM-based
devices, two or more transistors are typically used in a ROM cell: one access
and one programmed transistor. The programmed transistor performs the same
function as the floating gate in an EPROM, with both charge and discharge
being done electrically.
In the Flash-EEPROMs that are used as logic tile cell in the Actel ProASIC
chips (figure 2.18(b)), two transistors share the floating gate, which stores the
programming information. The sensing transistor is only used for writing and
verification of the floating gate voltage while the other is used as switch. This
can be used to connect or disconnect routing nets to or from the configured
logic. The switch is also used to erase the floating gate.
Reconfigurable Architectures 33

Figure 2.19. Implementation of f=ab in a 2-input MUX

4.2 Function generators

A reconfigurable hardware device should provide the users with the possibil-
ity to dynamically implement and reimplement new functions. This is usually
done by means of function generators that can be seen as the basic computing
unit in the device. Two types of function generators are in use in commercial
FPGAs: The multiplexers and the look-up tables.

4.2.1 Multiplexer
A 2:1 (2n -input-1) multiplexer (MUX) is a selector circuit with 2n inputs and
one output. Its function is to allow only one input line to be fed at the output.
The line to be fed at the output can be selected using some selector inputs. In
order to select one of the 2n possible inputs, n selectors lines are required. A
MUX can be used to implement a given function. The straihthforward way is
to place the possible results of the function at the 2n inputs of the MUX and to
place the function arguments at the selector inputs. In this case, the MUX will
work like a look-up table that will be explained in the next section. Several
possibilities exist to implement a function in a MUX, for instance by using
some arguments as inputs and other as selectors. Figure 2.19 illustrates this
case for the function f = a×b. The argument a is used as input in combination
with a second input 0 and the second argument b is used as selector.
The Shannon expansion theorem can be used to decompose a function and
implement it into a MUX. This theorem states that a given Boolean logic func-
tion F (x1 , .xn ) of n variables can be written as shown in the following equa-
tion.

F (x1 , · · · , xn ) = F (x1 , · · · , xi = 1, · · · , xn ) × xi +
F (x1 , · · · , xi = 0, · · · , xn ) × xi

Where F (x1 , · · · , xi = 1, · · · , xn ) is the function obtained by replacing xi

with one and F (x1 , · · · , xi = 0, · · · , xn ) the function obtained by replacing
xi by zero in F . The functions F1 = F (x1 , · · · , xi = 1, · · · , xn ) and F2 =
F (x1 , · · · , xi = 0, · · · , xn ) are called cofactors.
34 RECONFIGURABLE COMPUTING

A 2:1 multiplexer with Inputs I0 , I1 , output O and selector S is defined by

the following equation: O(S, I0 , I1 ) = S(I0 ) + S(I1 ). The function F can be
implemented on a 2:1 MUX with inputs I0 = F0 and I1 = F1 , and selector
S = xi .
A complex function can be implemented in many multiplexers connected
together. The function is first broken down into small pieces. Each piece is
then implemented on a multiplexer. The multiplexers will then be connected
to build the given function. This process is called technology mapping and is
part of the logic synthesis that we consider in the next chapter.
In order to implement a function on a 2n :1 MUX, the Shannon expansion
theorem should be refined such as to have n fixed variables and 2n cofactors.
For n = 2 the function of the multiplexer is defined as follow:

O(S0 , S1 , I0 , I1 , I2 , I3 ) = S 0 S 1 (I0 )+S0 S1 (I1 )+S0 S 1 (I2 )+S0 S1 (I3 ) (4.1)

This is equivalent to the Shannon refinement of a function with two fixed

variables as described below:

F (x1 ,··· ,xn )=F (x1 ,··· ,xi =1,··· ,xn )·xi +F (x1 ,··· ,xi =0,··· ,xn )·xi
=[F (x1 ,··· ,xi =1,··· ,xj =1,..xn )·xi +F (x1 ,··· ,xi =0,··· ,xj =1,··· ,xn )·xi ]xj
+[F (x1 ,··· ,xi =1,··· ,xj =0,··· ,xn )·xi +F (x1 ,··· ,xi =0,··· ,xj =0,··· ,xn )·xi ]xj
=F (x1 ,··· ,xi =1,··· ,xj =1,··· ,xn )·xi xj +F (x1 ,··· ,xi =0,··· ,xj =1,··· ,xn )·xi xj

+F (x1 ,··· ,xi =1,··· ,xj =0,··· ,xn )·xi xj +F (x1 ,··· ,xi =0,··· ,xj =0,··· ,xn )·xi xj

I0 =F0 =F (x1 ,··· ,xi =0,··· ,xj =0,··· ,xn ),

I1 =F1 =F (x1 ,··· ,xi =0,··· ,xj =1,··· ,xn ),
I2 =F2 =F (x1 ,··· ,xi =1,··· ,xj =0,··· ,xn ),
I3 =F3 =F (x1 ,··· ,xi =1,··· ,xj =1,··· ,xn )

Example 2.2 As example, we consider the implementation of a full adder

using 4:1 MUX. Let ai , bi be the operand and ci−1 be the carry on a previous
level, then si is the sum and ci the carry for the next level. According to the
truth table of the full adder (table 2.1), we can write si and ci as:
si = ai bi ci−1 + ai bi ci−1 + ai bi ci−1 + ai bi ci−1
ci = ai bi ∗ 0 + ai bi ci−1 + ai bi ci−1 + ai bi ∗ 1
According to those two equations and to the function of a 4:1 MUX given
above, two multiplexers can be configured as shown in figure 2.20.
Reconfigurable Architectures 35

ai bi ci−1 si ci
0 0 0 0 0
0 0 1 1 0
0 1 0 1 0
0 1 1 0 1
1 0 0 1 0
1 0 1 0 1
1 1 0 0 1
1 1 1 1 1
Table 2.1. Truth table of the Full adder

Figure 2.20. Implementation of a Full adder using two 4-input one output MUX

As stated earlier, a MUX provides much more possibilities than look-up

tables, because the inputs of the MUX, which are equivalent of the SRAM
output to a look-up table, are not reserved to the function result like as it is the
case in look-up tables. Function arguments can be connected to the MUX-input
as well as to he selector inputs.

The Actel ACT X Logic module. Multiplexer are used as function genera-
tors in the Actel FPGA devices (figure 2.21). In the Actel ACT1 device, the
basic computing element is the logic module. It is an 8-input 1-output logic
circuit that can implement a wide range of functions. Besides combinatorial
functions, the logic module can also implement a variety of D-latches. The
C-modules present in the second generation of Actel devices, the ACT2, are
similar to the logic module. The S-modules, which are found in the second and
third generation of actel devices, contain an additional dedicated fipl-flop. This
avoid the building of flip flop from the combinatorial logic as it is the case in
the logic module.
36 RECONFIGURABLE COMPUTING

Figure 2.21. The Actel basic computing blocks uses multiplexers as function generators

4.2.2 Look-Up Tables

A Look-Up Tables (LUT) is a group of memory cells, which contain all the
possible results of a given function for a given set of input values. Usually,
the set of possible function values corresponds to the possible combinations of
the inputs. The values of the function are stored in such a way that they can
be retrieved by the corresponding input values. A n-input LUT can be use to
n
implement up to 22 different functions, each of which can take 2n possible
values. Therefore, an n-input LUT must provide 2n cells for storing the possi-
ble values of an n-input function. In FPGAs, a LUT physically consists of a set
of SRAM-cells to store the values and a decoder that is used to access the cor-
rect SRAM location and retrieve the result of the function, which corresponds
to the input combination (figure 2.22).
In order to implement a complex function in a LUT-based FPGA, the func-
tion must be divided into small pieces, each of which can be implemented in
a single LUT. The interconnections are used to connect small pieces together
and form the complete function

Example 2.3 The full adder of the previous section can be implemented us-
ing two 3-inputs, 1 output (3:1) LUT as shown in figure 2.23. The sum is
implemented in the first LUT and the carry-out in the second LUT.
The first three columns of table 2.1 represent the input values. They build
the address used to retrieve the function value (corresponding to the value
in the third column) from the corresponding LUT location. The content of
the fourth and fifth columns of the truth table must therefore be copied in the
Reconfigurable Architectures 37

Figure 2.22. 2-input LUT

Figure 2.23. Implementation of a full adder in two 3-input LUTs

corresponding LUTs as shown in figure 2.23. The sum values are copied in the
upper LUT, while the carry values are compiled in the lower LUT.

SRAM-based LUT are used in the most commercial FPGAs as function gen-
erators. Several LUTs are usually grouped in a large module in which other
functional elements like flip flops and multiplexers are available. The connec-
tion between the LUTs inside such modules is faster than connections via the
routing network, because dedicated wires are used. We consider examples of
devices using LUT as function generator, the Xilinx FPGA and those of Altera,
and we next explain how the LUT are used in those devices.

The Xilinx Configurable Logic Block. The basic computing block in the
Xilinx FPGAs consists of a LUT with variable number of inputs, a set of mul-
tiplexers, arithmetic logic and a storage element (figure 2.24).
The LUT is used to store the configuration while the multiplexers selects the
right inputs for the LUT and the storage element as well as the right output of
the block.
38 RECONFIGURABLE COMPUTING

Figure 2.24. Basic block of the Xilinx FPGAs

The arithmetic logic provides some facilities like XOR-gate and faster carry
chain to build faster adder without wasting too much LUT-resources.
Several basic computing blocks are grouped in a coarse-grained element
called the Configurable Logic Block (CLB)(figure 2.25). The number of basic
blocks in a CLB varies from device to device. In the older devices like the
4000 series, the Virtex and Virtex E and the Spartan devices, two basic blocks
were available in a CLB. In the newer devices like the Spartan 3, the Virtex
II, the Virtex II-Pro and the Virtex 4, the CLBs are divided in four slices each
of which contains two basic blocks. The CLBs in the Virtex 5 devices contain
only two slices, each of which contains 4 basic blocks.
In the newer devices, the left part slices of a CLB, also called SLICEM can
be configure either as combinatorial logic, or can be use as 16-bit SRAM or
as shift register while righ-hand slices, the SLICEL, can only be configured as
combinatorial logic.
Except for the Virtex 5, all LUTs in Xilinx devices have four inputs and one
output. In the Virtex 5 each LUT has six inputs and two outputs. The LUT can
be configured either as a 6-input LUT, in which case only one output can be
used, or as two 5-input LUTs, in which case each of the two outputs is used as
output of a 5-input LUT.

The Altera Logic Array Block. Like the Xilinx devices, Altera’s FPGAs
(Cyclone, FLEX and Stratix) are also LUT based. In the Cyclone II as well
as in the FLEX architecture, the basic unit of logic is the logic element (LE)
that typically contains a LUT, a filp-flop, a multiplexer and additional logic
for carry chain and register chain. Figure 2.26 shows the structure of the logic
element in the Cyclone FPGA. This structure is very similar to that of the
Altera FLEX devices. The LEs in the cyclone can operate in different modes
each of which defines different usage of the LUT inputs.
In the Stratix II devices, the basic computing unit is called adaptive logic
module (ALM) (figure 2.27). The ALM is made upon a mixture of 4-input and
Reconfigurable Architectures 39

Figure 2.25. CLB in the newer Xilinx FPGAs (Spartan 3, Virtex 4 and Virtex 5)

Figure 2.26. Logic Element in the Cyclone II

3-input LUTs that can be used to implement logic functions with variable num-
ber of inputs. This insures a backward compatibility to 4-input based designs,
while providing the possibility to implement coarse-grained module with vari-
able number (up to 8) inputs. Additional modules including flip flops, adders
and carry logic are also provided.
Altera logic cells are grouped to form coarse-grained computing elements
called Logic Array Blocks (LAB). The number of logic cells per LAB varies
from the device to device. The Flex 6000 LABs contains ten logic elements
while the FLEX 8000 LAB contains only eight. Sixteen LEs are available for
each LAB in the cyclone II while the Stratix II LAB contains eight ALMs.
40 RECONFIGURABLE COMPUTING

Figure 2.27. Stratix II Adaptive Logic Modul

4.3 FPGA structures

FPGAs consist of a set of programmable logic cells placed on the device
such as to build an array of computing resources. The resulting structure is
vendor dependant. According to the arrangement of logic blocks and the inter-
connection paradigm of the logic blocks on the device, FPGAs can be classified
in four categories: symmetrical array, row-based, hierarchy-based and sea of
gates (figure 2.28). We next explain each of the mentioned structure based on
a commercial example.

4.3.1 Symetrical Array: The Xilinx Virtex and Atmel AT40K Families
A symmetrical array-based FPGA consists of a two dimensional array of
logic blocks immersed in a set of vertical and horizontal lines. Switch elements
exist at the intersections of the vertical and horizontal lines to allow for the
connections of vertical and horizontal lines.
Examples of FPGAs arranged in a symmetrical array-based are the Xilinx
Virtex FPGA and the Atmel (figure 2.29).
On the Xilinx devices, CLBs are embedded in the routing structure that
consists of vertical and horizontal wires. Each CLB element is tied to a switch
matrix to access the general routing structure, as shown in figure 2.30(a). The
switch matrix provides programmable multiplexers, which are used to select
the signals in the given routing channel that should be connected to the CLB
terminals. The switch matrix can also connect vertical and horizontal lines,
thus making routing possible on the FPGA.
Each CLB has access to two tri-state driver (TBUF) over the switch matrix.
Those can be use to drive on-chip busses. Each tri-state buffer has its own tri-
Reconfigurable Architectures 41

Figure 2.28. The four basic FPGA structures

state control pin and its own input pin that are controlled by the logic built in
the CLB. Four horizontal routing resources per CLB are provided for on chip
tri-state busses. Each tri-state buffer has access alternately to two horizontal
lines, which can be partitioned as shown in figure 2.30(b). Besides the switch
matrix, CLBs connect to their neighbors using dedicated fast connexion tracks.
The routing is done on the Atmel chips using a set of busing planes. Seven
busing planes are available on the AT40K. Figure 2.31 depicts a part of the
plane with five identical busing planes. Each plane has three bus resources: a
local-bus resource (the middle bus) and two express-bus resources (both sides).
Repeaters are connected to two adjacent local-bus segments and two express-
bus segments. Local busses segments span four cells while an express bus
segments span eight cells. Long tri-state bus can be created by bypassing a
repeater.
42 RECONFIGURABLE COMPUTING

(a) The Xilinx Virtex II (b) Atmel’s symetrical array arrangement

Figure 2.29. Symetrical array arrangement in a) the Xilinx and b) the Atmel AT40K FPGAs

(a) CLB connexion to the switch matrix (b) Tri-state buffer connection to horizontal lines

Figure 2.30. Virtex routing resource

Figure 2.31. Local connection of Atmel Cell

Reconfigurable Architectures 43

Figure 2.32. Row based arrangement on the Actel ACT3 FPGA Family

Locally, the Atmel chip provides a star-like connection resource that allows
each cell (which is the basic unit of computation) to be connect directly to all
its 8 neighbors. Figure 2.31 depicts direct connections between a cell and its
eight nearest neighbors.

4.3.2 Row-Based FPGAs: The Actel ACT3 Family

A row-based FPGA consists of alternating rows of logic block or macro
cells and channels (figure 2.32). The space between the logic blocks is called
channel and used for signal routing.
The Actel ACT3 FPGA family (figure 2.32) is an example of row-based
FPGA. The macro cells are the C-elements and S-elements presented earlier
in this chapter. The routing is done via the in the horizontal direction using
the channels. In the vertical direction, dedicated vertical tracks are used. As
shown in figure 2.33, a channel consists of several routing tracks divided into
segments. The minimum length of a segment is the width of a module pair and
its maximum length is the length of a complete channel, i.e. the width of the
device.
Any segment that spans more than one-third of the rows length is considered
a long horizontal segment. Non dedicated horizontal routing tracks are used
to route signal nets. Dedicated routing tracks are used for the global clock
networks and for power and ground tracks. Vertical tracks (figure 2.33) are of
44 RECONFIGURABLE COMPUTING

Figure 2.33. Actel’s ACT3 FPGA horizontal and vertical routing resources

three types: input, output and long. They are also divided in more segments.
Each segment in an input track is dedicated to the input of a particular module
and each segment in the output track is dedicated to the output of a particular
module. Long segments are uncommitted and can be assigned during routing.
Each output segment span four channels (two above and two below) except
near the top and the bottom of the array. Vertical input segments span only the
channel above or the channel below. The tracks dedicated to module inputs
are segmented by pass transistors in each module row. During normal user
operation the pass transistors are inactive, which isolate the inputs of a module
from the inputs of the module above it.
The connections inside Actel FPGAs are established using antifuse. Four
types of antifuse connections exist for the ACT3: Horizontal-to-vertical (XF)
connection, Horizontal-to-horizontal (HF) connection, vertical-to-vertical (FF)
connection and fast-vertical connection (figure 2.33).

4.3.3 Sea-of-gates: The Actel ProASIC family

Like in symmetrical arrays, the macro cells are arranged on an two dimen-
sional array structure such that an entry in the array correspond to the coordi-
nate of a given macro cell. The difference between the symmetrical array and
the sea-of-gate is that there is no space left aside between the macro cells for
routing.
The interconnection wires are fabricated on top of the cells. The Actel
ProASIC FPGA family is an implementation of the sea-of-gate approach. The
ProASIC core consist of a sea-of-gates called sea-of-tiles. The macro cells
are the EEPROM-based tiles previously seen. The device uses a four level of
hierarchy routing resource to connect the logic tiles: The local resources, the
long-line resources, the very long-line resources and the global networks. The
local resources allow the output of the tile to be connected to the inputs of one
Reconfigurable Architectures 45

Figure 2.34. Actel ProASIC local routing resources

of the eight surrounding tiles (figure 2.34). The long-line resources provide
routing for longer distances and higher fanout connections. These resources,
which vary in length (spanning one, two, or four tiles), run both vertically and
horizontally and cover the entire device. The very long lines span the entire
device. They are used to route very long or very high fanout nets.

4.3.4 Hierarchical-based: The Altera Cyclone, Flex and Stratix

families
In hierarchical-based FPGAs, macro cells are hierarchically placed on the
device. Elements with the lowest granularity are at the lowest level hierarchy.
They are grouped to form the elements of the next level. Each element of a
level i consists of a given number of elements from level i − 1.
Altera FPGAs (FLEX, Cyclone II and Stratix II) have two hierarchical lev-
els. The logic cells (in the Cyclone II and FLEX) and the ALM in the Stratix II
are on the lowest level of the hierarchy. The Logic Array Blocks (LAB) build
the higher level (figure 2.35). Each LAB contains a given number of logic ele-
ments (eight for the FLEX8000, ten for the FLEX6000, sixteen for the Cyclone
and eight AMLs for the Stratix II). The LABs in turn are arranged as array on
the device.
Signal connections to and from device pin are provided via a routing struc-
ture called FastTrack in the FLEX and MultiTrack in Cyclone II and Stratix
II). The FastTract as well as the MultiTrack interconnects consist of a series of
fast, continuous row and column channels that run the entire length and width
of the device.
46 RECONFIGURABLE COMPUTING

Figure 2.35. Hierarchical arrangement on the Altera Stratix II FPGA

Figure 2.36. LAB connection on the Altera Stratix devices

Signals between LEs or ALMs in the same LAB and those in the adjacent
LABs are routed via local interconnect signals. Each row of a LAB is served
by a dedicated row interconnect, which routes signals between LABs in the
same row. The column interconnect routes signals between rows and routes
signals from I/O pin rows.
Reconfigurable Architectures 47

A row channel can be driven by an LE (or ALM in the Stratix II) or by

one of two column channels. Each column of LABs is served by a dedicated
column interconnect. The LEs in an LAB can drive the column interconnect
which can then drive another rows interconnect to route the signals to other
LABs in the device. A signal from the column interconnect must be routed to
the row interconnect before it can enter an LAB.
LEs can drive global control signals. This is helpful for distributing the
internally generated clock, asynchronous clear and asynchronous preset signals
and high-fan-out data signals.

4.3.5 Programmable I/O

Located around the periphery of the device, I/O components allow for the
communication of a design inside the FPGA with off-chip modules. Like the
logic cells and the interconnections, FPGA I/Os are programmable, which al-
lows designs inside the FPGA to configure a single interface pins either as
input, output or bidirectional.
The general structure of an I/O component is shown in figure 2.37: It con-
sists of an input block, an output block and an output enable block for driving
the tri-state buffer. Two registers that are activated either by the falling or by
the rising edge of the clock, are available in each block.
The I/Os can be parameterized for a single data rate (SDR) or a double data
rate (DDR) operation mode. While in the SDR-mode data are copied into the
I/O registers on the rising clock edge only, the DDR mode exploits the falling
clock edge and the rising clock edge to copy data into the I/O registers. On
the input, output, tri-state, one of the Double Data Rate (DDR) register can be
used. The double data rate is directly accomplished by the two registers on
each path, clocked by the rising edge (or the falling edge) from different clock
nets.
DDR input can be done using both input registers while DDR output will
use both output registers. In the Altera Stratix II, the I/O component is called
I/O Element (IOE). The IOEs are located in I/O blocks around the periphery
of the device.
The I/O blocks, which contain up to four IOEs, are used to drive the rows
and columns interconnects. They are divided in two groups: The row I/O
blocks, which drive row, column, or direct link interconnects, while the column
I/O blocks drive column interconnects.
The Xilinx Virtex I/O components are called I/O Block (IOB) and they are
provided in groups of two or four on the device boundary. The IOB can be used
independent from each other as input and/or output, or they can be combined
in group of two to be used as differential pair directly connected to a switch
matrix.
48 RECONFIGURABLE COMPUTING

Figure 2.37. General structure of an I/O component

4.4 Hybrid FPGAs

The process technology as well as the market demand is pushing manufac-
turers to include more and more pre-designed and well tested hard macros in
their chips. Resources, like memory, that are used in almost all designs can be
directly be found on the chip. This allows the designer to use well tested and
efficient modules. Moreover, hard macros are more efficiently implemented
and are faster than macro implemented on the universal function generators.
The resources often available on hybrid FPGAs are RAMs, clock managers,
arithmetic modules, network interface modules, and processors. The market
demand has pushed almost all the manufactures to include hard macros in all
their devices. Because the resources required by the users varies with their
application class, some manufacturers, like Xilinx, provided different classes
of FPGAs, each of which is suited for a given purpose. The most emerging
classes as classified by Xilinx are: the System on Chip (SoC), Digital Signal
Processing (DSP) and pure logic. The system on chip class to which the Virtex
4 FX belongs, is characterized by embedded processors directly available on
the chip, memory and dedicated bus resources. Reference designs also exist,
which efficiently use the chip resource to get the maximum performance. The
DSP class, which contains the Virtex 4 SX is characterized by the abundance of
multipliers macros and the pure logic class, to which the Virtex 4 LX belongs,
is dominated by LUT function generators.
Figure 2.38 depicts the architecture of the Xilinx Virtex II Pro, one of the
first FPGAs to include Hard macro, among which complete processors.
The Xililnx Virtex II Pro contains up to four embedded IBM Power PC 405
RISC hard core processors, which can be clocked at more than 300 MHZ.
Embedded high speed serial RocketIO transceivers with up to 3.125 Gb/s per
Reconfigurable Architectures 49

Figure 2.38. Structure of a Xilinx Virtex II Pro FPGA with two PowerPC 405 Processor blocks

channel, internal BlockRAM memory module to be used as dual-ported RAM,

embedded 18-bit x 18- bit multiplier blocks, Digital Clock Manager (DCM)
to provide self-calibrating, fully digital solution for clock distribution delay
compensation, clock multiplication and division, and coarse and fine-grained
clock phase shifting are available on the chip.

5. Coarse-Grained Reconfigurable Devices

FPGAs allow for programming any kind of function as far as this can fit
onto the device. This is only possible because of the low granularity of the
function generators (LUT and MUX). However, the programmable intercon-
nections used to connect the logic blocks reduce the performance of FPGAs.
A way to overcome this is to provide frequently used module as hard macro,
as it is the case in hybrid FPGAs, and therefore, to allow programmable inter-
connections only between processing elements available as hard macros on the
chip. Coarse-grained reconfigurable devices follow this approach. In general
those devices are made upon a set of hard macros (8-bit, 16-bit or even a 32-
bit ALU), usually called Processing Element (PE). The PEs are able to carry
few operations like addition, subtraction or even multiplication. The intercon-
nection is realized either through switching matrices or dedicated busses. The
configuration is done by defining the operation mode of the PEs and program-
ming the interconnection between the processing elements.
In the last couple of years, several coarse-grained architectures were built
by different companies, some of which went bankrupt. Despite all the hope
placed in coarse-grained reconfigurable device, they fail until now to provide
50 RECONFIGURABLE COMPUTING

a large acceptance and their future is not really easy to predict. This does not
mean that the philosophy behind coarse-grained device is wrong Companies
investigating coarse-grained reconfigurable devices must also face the FPGA
competition, dominated by large companies that provides many coarse-grained
elements in their devices according to the market need.
A wide variety of coarse-grained reconfigurable devices that we classify
in three categories will be presented in this section: In the first category, the
dataflow machines, functions are usually built by connecting some PEs in or-
der to build a functional unit that is used to compute on a stream of data. In
the second category are the network-based devices in which the connection be-
tween the PEs is done using messages instead of wires. The third category are
the Embedded FPGA devices, which consist of a processor core that cohabit
with a programmable logic on the same chip.

5.1 Dataflow Machines

Dataflow machines are the most dominating coarse-grained reconfigurable
devices. In this section, we present three of those architectures: The PACT-
XPP, the NEC-DRP and the PicoChip devices.

5.1.1 The PACT XPP device

The idea behind the PACT XPP architecture [17] [176] is to efficiently com-
pute streams of data provided from different sources such as A/D converters
rather than single instructions as it is the case in the Von-Neumann computers.
Because the computation should be done while data are streaming through the
processing elements, it is suitable to configure the PEs to adapt to the natural
computation paradigm of a given application or part of it at a given time. The
XPP (eXtreme Processing Platform) architecture of PACT consist of:
An array of Processing Array Elements (PAE) grouped in Processing Ar-
ray(PA)
A communication network
A hierarchical configuration tree
Memory elements aside the PAs
A set of I/O elements on each side of the device
One Configuration Manager (CM) attached to a local memory is responsible
for writing configuration onto a PA. The configuration manager together with
PA build the Processing Array Cluster (PAC). An XPP chip contains many
PACs arranged as grid array on the device. Figure 2.39 shows an XPP device
with four PACs, each of which contains 4 PAEs and surrounded by memory
blocks. The CMs at a lower level are controlled by a CM at the next higher
Reconfigurable Architectures 51

Figure 2.39. Structure of the PACT XPP device

level. The root CM at the highest level is attached to an external configuration

memory and supervises the whole device configuration.

The Processing Array Element (PAE). There exist two different kinds of
PAEs: the ALU PAE and the RAM-PAE. An ALU-PAE contains an ALU that
can be configured to perform basic arithmetic operations, while the RAM-PAE
is used for storing data. The Back-Register (BREG) provides routing chan-
nels for data and events from bottom to top, additional arithmetic and register
functions while the Forward-Register (FREG) is used for routing the signals
from top to bottom and for the control of dataflow using event signals. All
objects can be connected to horizontal routing channels using switch-objects.
DataFlow Register (DF-Registers) can be used at the object output for data
buffering in case of a pipeline stall. Input registers can be pre-loaded by con-
figuration data and always provide single cycle stall.
A RAM-PAE is similar to an ALU-PAE. However, instead of an ALU, a
dual ported RAM is used for storing data. The RAM generates a data packet
after an address was received at the input. Writing to the RAM requires two
data packet: one for the address and the other for the data to be written. Figure
2.40 shows an ALU-PAE. The structure is the same for a RAM-PAE, however
a RAM is used instead of the ALU.

Routing and Communication. The XPP interconnection network consists

of two independent networks: One for data transmission and the other for event
transmission (figure 2.40). These two networks consist of horizontal and ver-
tical channels. The vertical channels are controlled by the BREG and FREG
while connection to horizontal channel is done via switch elements. Beside
52 RECONFIGURABLE COMPUTING

Figure 2.40. The XPP ALU Processing Array Element. The structure of the RAM ALU is
similiar.

the horizontal and vertical channels a configuration bus exists, which allows
the CMs to configure the PAEs.
Horizontal buses are used to connect a PAE within a row while the verti-
cal buses are used to connect objects to a given horizontal bus. Vertical con-
nections are done using configurable switch objects which segment the verical
communication channels. The vertical routing is enabled using register-objects
integrated into the PAEs.

Interfaces. XPP devices provide communication interfaces aside the chip.

The number may vary from device to device. The XPP64-A1 for example
contain six external interfaces consisting of four identical general purpose I/O
interfaces on the chip corner (bottom left, upper left, bottom right and upper
right), one configuration manager interface and a JTAG compliant interface for
debugging and testing purpose.
The I/O interfaces can operate independently from each other either in RAM,
or in streaming mode. In streaming mode, each I/O element provides two bidi-
rectional ports for data streaming. Handshake signals are used for synchro-
nization of data packets to external ports. In RAM mode, each port can access
external synchronous SRAMs with 24-bit addresses and 24-bit data. Control
signals for the SRAM transactions are available such that no extra logic is re-
quired.
The configuration manager interface consists of three subgroups of signals:
code, message send and message receive. The code group provides channels
over which configuration data can be downloaded to the device while the send
and receive groups provide communication channels with a host processor.
Reconfigurable Architectures 53

Figure 2.41. Structure of the NEC Dynamically Reconfigurable Processor

5.1.2 The NEC DRP Architecture

We now present the NEC Dynamically Reconfigurable Processor (DRP),
which operates in a similar way as the PACT. However, the reconfiguration
control is not done in a hierarchical fashion like in the PACT devices. Fig-
ure 2.41 shows the overall structure of a NEC Dynamically Reconfigurable
Processor. It consist of:

An array of byte oriented processing elements.

A programmable interconnection network to connect the processing ele-
ments.
A sequencer (State Transition Controller) which can be programmed as
finite state machine to control the dynamic reconfiguration process on the
device.
Memory blocks to store configuration and computation data. The memory
blocks are arranged around the device.
Various interfaces and RAM controllers like PCI, PLL, SDRAM/SRAM.

As shown in figure 2.42, a DRP processing element contains an ALU for

byte oriented arithmetic and logic operation, a Data Management Unit to han-
dle byte selects, shift, mask and constant generation. The operand can be
fetched from the PE’s register file or collected directly from the PE’s inputs.
54 RECONFIGURABLE COMPUTING

Figure 2.42. The DRP Processing Element

The results can be stored in the PE’s register file or can be sent to the output.
The configuration is done by the State Transition Controller that set a pointer
to a corresponding instruction register according to the operation mode of the
ALU. Having many configuration registers allow for storing configuration data
directly in the proximity of the PEs and allow a fast switching from one con-
figuration to the next one.

5.1.3 The picoChip Reconfigurable Device

A picochip consists of a picoArray core and a set of different interfaces
for external connection of various module like memory and processors. The
picoArray core has a similar structure like the NEC DRP and the PACT, how-
ever, the connection between the PES is done at the column and row inter-
section. The PC102 Chip for example contains hundreds of array elements,
each with a versatile 16-bit processors with local data memory connected by
a programmable interconnect structure (figure 2.43). The architecture is het-
erogeneous, with four types of processing element all having a common basic
structure, but optimized for different tasks: the Standard AE (STAN), the Con-
trol AE (CTRL), the Memory AE (MEM) and the Function Accelerator Unit
(FAU).
A standard AE type includes multiply-accumulate peripheral as well as spe-
cial instructions optimized for CDMA operations. The memory AE contains
multiply unit and additional memory. The function accelerator unit is a co-
processor optimized for specific signal processing tasks. The control AE is
equipped with a multiply unit and larger amounts of data and instruction mem-
ory optimized for the implementation of base station control functionality.
Reconfigurable Architectures 55

Figure 2.43. Structure of the Picochip device

Multiple elements can be programmed together as a group to perform par-

ticular function. The device can be reconfigured at run-time to run different ap-
plications like wireless protocols. Several interfaces are available for seamless
external connection (figure 2.43): Each chip has four inter processor communi-
cations links that can be used to build an array of several picoChips in order to
implement larger and complex functions that cannot fit in only one picoChip.
The communication on those links is based on a Time Division Multiplexing
(TDM) protocol scheme. Besides the inter processor communication a micro-
processor interface is available for connecting an external processor that can
be used to configured the device and stream data into the device. External stor-
age (EEPROM, FLASH) can be connected as well to this interface to allow
a self reconfiguration of the device on start up. Other interfaces are provided
among which a memory interface for connection of external memory and a
JTAG interface for debugging purpose.

5.2 Network-oriented architectures

Despite the interest on networks on chip as communication paradigm has
grown recently, very few reconfigurable devices rely on message passing for
data exchange among the PEs. One of the few companies to have implemented
this concept is Quicksilver Tech, a Start-up that failed to commercialized their
technology and, recently, went bankrupt. Nevertheless, we believe that their
56 RECONFIGURABLE COMPUTING

Figure 2.44. The Quicksilver ACM hierarchical structure with 64 nodes

architecture merit more attention. We therefore decide to present some details

on it in this section.

5.2.1 The Quicksilver ACM Architecture

The Adaptive Computing Machine (ACM) is based on a the revolutionary
network on chip paradigm and is one of the very few devices that work on such
a principle.
The Quicksilver ACM consists of a set of heterogeneous computing nodes
hierarchically arranged on a device. At the lowest level, four computing nodes
are placed in a cluster and connected locally together. Many clusters at a given
level are put together to build bigger clusters at the next higher level (figure
2.44).
An ACM chip consists of the following elements:

A set of heterogeneous Processing Nodes (PN)

An homogenous Matrix Interconnect Network (MIN)

A system controller

Various I/O interfaces

The ACM Processing Node structure. An ACM processing nodes consist

of:
Reconfigurable Architectures 57

(a) The ACM node structure (b) The ACM QS2412 Resources

Figure 2.45. ACM Node and Routing resource

An algorithmic engine that defines the node type. The node type can be
customized at compile-time or at run-time by the user to match a given
algorithm. Four types of nodes exist: The PProgrammable Scalar Node
(PSN) provides a standard 32-bit RISC architecture with 32 general pur-
pose registers, the Adaptive Execution Node (AXN) provides variable word
size Multiply Accumulate (MAC) and ALU operations, the Domain Bit
Manipulation ( DBN) node provides bit manipulation and byte oriented
operations, and the External Memory Controller node provides DDRRAM,
SRAM, memory random access and DMA control interfaces for off-chip
memory access.
The node memory for data storage at node level.
A node wrapper which hide the complexity of the network architecture. It
contains a MIN interface to support communication, a hardware task man-
ager for task managements at node level, and a DMA engine. The wrapper
envelops the algorithmic engine and presents an identic interface to the
neighboring nodes. It also incorporates dedicated I/O circuitry, memory,
memory controllers and data distributors and aggregators (figure 2.45).

The Matrix Interconnect Network (MIN). The communication inside an

ACM chip is done via a MIN, which is organized hierarchically. At a given
level, the MIN connects many lower level MINs. The top level MIN, the MIN
root, is used to access the nodes from outside and to control the configura-
tion of the nodes. The communication among nodes is done via the MIN with
the help of the node wrapper. The MIN provides a diversity of services like
point-to-point dataflow streaming, real-time broadcasting, direct memory ac-
cess, random memory access. The ACM chip also contains various I/O in-
58 RECONFIGURABLE COMPUTING

terfaces accessible via the MIN for testing (JTAG) and communication with
off-chip devices (figure 2.45(b)).

System Controller. The system management is done via an embedded sys-

tem controller (figure 2.45(b)). The system controller loads tasks into the
node’s ready-to-run queue for execution, statically or dynamically sets the
communication channels between the processing nodes. Any node can be
adapted (reconfigured) at run-time by the system controller to perform a new
function, in a cycle-by-cycle manner.

5.3 Embedded FPGA

Embedded programmable logic devices, also known under the name embed-
ded FPGA usually integrate a processor core, a programmable logic or FPGA
and memory on the same chip. Strictly seen, several existing FPGAs like the
Xilinx Virtex 4 FX fall under this category, since they provides the same ele-
ments like any other embedded FPGA device. However, the processor in the
Xilinx Virtex 4 FPGAs is immersed in the the programmable logic while it
is strictly decouple from the logic in common embedded FPGAs. Two exam-
ples of such devices are the DAP/DNA from IPflex and the S500 series from
Stretch.

5.3.1 The IPflex DAP/DNA Reconfigurable Processor

The DAP/DNA consists of an integrated DAP RISC processor1 , a Distributed
Network Architecture (DNA) matrix and some interfaces (figure 2.46). The
processor controls the system, configures the DNA, performs computations in
parallel to the DNA and manages the data exchange on the device.
The DNA Matrix 2.46 is a dataflow accelerator with more than hundred dy-
namic reconfigurable processing elements. The wiring among elements can be
changed dynamically, therefore, providing the possibility to build and quickly
change parallel/pipelined processing system tailored to each application. The
DNA configuration data are stored in the configuration memory from where it
can be downloaded to the DNA on a clock-by-clock basis. Like in other recon-
figurable devices, several interfaces exist for connecting the chip to external
devices.
Large applications can be partitioned and sequentially executed on a DAP/DNA
chip. While the processor controls the whole execution process, the DNA exe-
cutes critical parts of the application.

1 That is a 32 bit RISC processor operating at 100 MHz with a data cache of 8K bite and an instruction

cache of 8K bite
Reconfigurable Architectures 59

Figure 2.46. IPflex DAP/DNA reconfigurable processor

5.3.2 The Stretch Processor

The second example of embedded FPGA that we present is the S5000 series
from Stretch. The device consists of a 32-bit Xtensa RISC Processor from
Tensilica, operating at 300 MHz and featuring a single precision floating point
unit, an Instruction Set Extension Fabric (ISEF), embedded memory and a set
of peripheral control modules.
Like the DAP/DNA, the processor controls the whole system and configure
the ISEF. The ISEF is used to augment the processor capacity by implement-
ing additional instructions that are directly accessible by a program running
on the processor. The S5 device uses 32 128 bit wide registers coupled with
128-bit wide access to memory to feed data to the ISEF. Several interfaces,
connected to a peripheral bus exist for communication with the external world.
The S5530 peripherals are divided in three categories as follows: the low speed
peripherals that includes UART, IrDA, Serial Peripheral Interface (SPI). The
second category consists of the mid speed peripherals that includes parallel
Generic Interface Bus (GIB) for Flash and SRAM connection. The last cate-
gory are the high speed peripherals that includes a 64-bit PCI/PCI-X port and
one 64-bit SDRAM port.
60 RECONFIGURABLE COMPUTING

Figure 2.47. The Stretch 5530 configurable processor

5.4 Academic Efforts

Besides commercial coarse-grained reconfigurable devices, several other
coarse-grained devices were developed in the academic world. Starting from
1990, several approaches for coarse-grained device, some of which are de-
scribed in this section, have been published [110].

5.4.1 The MATRIX Architecture

The Multiple ALU architecture with Reconfigurable Interconnect eXperi-
ment (MATRIX) [159] comprises an array of identical Basic Functional Units
(BFUs). Each BFU contains an 8-bit ALU, 256 words of 8-bit memory and
control logic. The ALU features the standard set of arithmetic and logic func-
tions and a multiplier. A configurable carry-chain between adjacent ALUs can
be used to cascade ALUs for wide-word operations. The control logic can gen-
erate local control signals from ALU output by a pattern matcher. A reduction
network can be employed for control generated from neighbour data. Finally, a
20-input, 8-output NOR block may be used as half of a PLA to produce control
signals. According to these features, a BFU can serve as an instruction mem-
ory, a data memory, a register-file-ALU combination, or an independent ALU
function. Instructions can be routed over the array to several ALUs. The rout-
ing fabric provides three levels of 8-bit buses: eight nearest neighbor and four
second nearest neighbor connections of length four, and global lines spanning
an entire row or column.
Reconfigurable Architectures 61

5.4.2 RAW: Reconfigurable Architecture Workstation

The idea of RAW [210] is to provide a simple and highly parallel com-
puting architecture composed of several repeated tiles connected among each
other by nearest neighbor connections. The tiles comprise computation facil-
ities as well as memory, thus implementing a distributed memory model. A
RAW microprocessor is a homogeneous array of processing elements called
tiles. The prototype chip features 16 tiles arranged in a 4 x 4 array. Each
tile comprises a simple RISC-like processor consisting of ALU, register file
and program counter, SRAM-based instruction and data memories, and a pro-
grammable switch supplying point-to point connections to NN. The CPU in
each tile of the prototype is a modified 32-bit MIPS R2000 processor with an
extended 6-stage pipeline, a floating point unit, and a register file of 32 general
purpose and 16 floating point registers. Both the data memory and the instruc-
tion memory consist of 32 Kilobyte SRAM. While the instruction memory is
uncached, the data memory can operate in cached and uncached mode. In-
terconnect of the RAW architecture is done over nearest neighbor connections
being optimized for single data word transfer. Communication between tiles is
pipelined over these connections and appears at register level between proces-
sors, making it different from multiprocessor systems. A programmable switch
on each tile connects the four nearest neighbor links to each other and to the
processor. The RAW architecture provides both static and a dynamic network
with wormhole routing for the forwarding of data.

5.4.3 REMARC: Reconfigurable Multimedia Array Coprocessor

REMARC [160] is a reconfigurable coprocessor that is tightly coupled to a
main RISC processor. It consists of an 8x8 array of 16-bit programmable logic
units called nanoprocessor, which is attached to a global control unit. The
control unit manages data transfers between the main processor and the recon-
figurable array and controls the execution on the nanoprocessors. It comprises
an instruction RAM with 1024 entries, 64-bit data registers and four control
registers. The nanoprocessor consists of a 32-entry local instruction RAM, an
ALU with 16-bit datapath, a data RAM with 16 entries, an instruction register,
eight data registers, four 16-bit data input registers, and a 16-bit data output
register. The ALUs can execute 30 instructions, including addition, subtrac-
tion, logical operations, shift instructions, as well as some operations often
found in multimedia applications like minimum, maximum, average, absolute
and add. Each ALU can use data from the data output registers of the adja-
cent processors via nearest neighbor connect, from a data register, a data input
register, or from immediate values as operands. The result of the ALU oper-
ation is stored in the data output register. The communication lines consist of
nearest neighbor connections between adjacent nanoprocessors and additional
horizontal and vertical buses in each row and column. The nearest neighbor
62 RECONFIGURABLE COMPUTING

connections allow the data in the data output register of a nanoprocessor to be

sent to any of its four adjacent neighbors. The horizontal and vertical buses
have double width (32-bit) and allow data from a data output register to be
broadcasted to processors in the same row or column, respectively. Further,
the buses can be used to transfer data between processors, which are not adja-
cent to each other.

5.4.4 MorphoSys
The complete MorphoSys [194] chip comprises a control processor (TinyRISC),
a frame buffer (data buffer), a DMA controller, a Context Memory (configura-
tion memory), and an array of 64 Reconfigurable Cells (RC). Each RC com-
prises an ALU-multiplier, a shift unit, and two multiplexers at the RC inputs.
Each RC also has an output register, a feedback register and a register file. A
context word, loaded from the configuration Memory and stored in the Context
Register, defines the functionality of the RC. Besides standard logic/arithmetic
functions, the ALU has other functions such as computation of absolute value
of the difference of two operands and a single cycle multiply-accumulate oper-
ation. There are a total of 25 ALU functions. The RC interconnection network
features three layers. In the first layer, all cells are connected to their four near-
est neighbors. In the second layer, each cell can access data from any other
cell in the same row or column of the same array quadrant. The third layer of
hierarchy consists of buses spanning the whole array and allowing transfer of
data from a cell in a row or column of a quadrant to any other cell in the same
row or column in the adjacent quadrant. In addition, two horizontal 128-bit
buses connect the array to the frame buffer.

5.4.5 NISC Processor: No-Instruction-Set Computer

The NISC [92] processor consists of a Controller and Datapath on which
any C program can be executed on it. The datapath consists of a set of storage
elements (registers, register files, memories), functional units (ALUs, multi-
pliers, shifters, custom functions) and a set of busses. Each component may
take one or more clock cycles to execute, each component may be pipelined
and each component may have input or output latches or registers. The en-
tire Datapath can be pipelined in several stages in addition to components be-
ing pipelined themselves. The Controller defines the state of the processor
and issues the control signals for the Datapath. The controller can be fixed
or programmable, while the Datapath can be reprogrammable and reconfig-
urable. Reprogrammable means that the Datapath can be extended or reduced
by adding or omitting some components, while reconfigurable means that the
Datapath can be reconnected with the same components. In order to speed up
the NISC pipelining, a Control Register (CR) and a Status Register (SR) has
been inserted between the Controller and the Datapath.
Reconfigurable Architectures 63

Figure 2.48. Pipeline Reconfiguration: Mapping of a 5 stage virtual pipelline auf eine 3 stage

5.4.6 Virtual Pipelines: The PipeRench

Pipeline reconfiguration was proposed by Goldstein et al [98] as a method
of virtualizing pipelined hardware application designs. The single static re-
configuration is broken into pieces that correspond to pipeline stages in the
application. The resulting configurations are then loaded into the device on a
cycle-by-cycle basis. With pipeline virtualization an application is pipelined
implemented on a given amount of virtual resources. The virtual resources are
then mapped to the physical resources in a final step. The physical resources
are computing units modules, whose functionality can be changed by reconfig-
uration.
The process is somehow similar to the implementation on parallel machines,
where virtual processors are first used to easily capture the inherent parallel
structure of an application. The virtual processors are then mapped to the phys-
ical one in the next step. Virtualization provides the designer freedom to better
explore the parallelism in the implementation, since he does not need to face
the resources constraints of a given platform.
Figure 2.48 illustrates the virtualization processing on the the mapping of a
five-stage virtual pipeline on a three-stage physical fabric. On the top part of
this figure, we see the five-stage application and the state of each of the stages
of the pipeline in the five consecutive cycles. The bottom half of the figure
64 RECONFIGURABLE COMPUTING

shows the mapping of the virtual blocks on the physical modules of the device.
Because the configuration of each single unit in the pipeline is independent
from the other, the reconfiguration process can be broken down to a cycle-by-
cycle configuration. In this way, part of the pipeline can be reconfigured while
the rest is computing.
Based on the pipeline reconfiguration concept, a class of reconfigurable de-
vices called PipeRench was proposed in [98] as co-processor in multimedia
applications. A PipeRench device consist of a set of physical pipeline stages
also called stripes. A stripe is composed of interconnect and processing ele-
ments (PE), each of which contains registers and LUT-based ALUs. The PEs
access operands from registered outputs of the previous stripe as well as reg-
istered or unregistered outputs of the other PEs in the stripe. All PipeRench
devices have four global busses, two of which are dedicated to storing and
restoring stripe state during hardware virtualization (configuration). The other
two are used for input and output.

5.4.7 RaPiD
The last work that we would like to considerer in this section is the one of
Ebeling et al [100] whose goal was to overcome the handicap of FPGA2 . A
one dimensional coarse-grained reconfigurable architecture called RaPiD, an
acronym for Reconfigurable Pipelined Datapath is proposed for this purpose.
The structure of RaPiD datapaths resembles that of systolic arrays. The
structure is made upon linear arrays of functional units communicating in
mostly a nearest neighbour fashion. This can be used for example, to con-
struct a harware module, which comprises different computations at different
stages and at different times resulting in a linear array of functional units that
can be configured to form a linear computational pipeline. The resulting ar-
ray of functional units is divided into identical cells which are replicated to
form a complete array. A RaPiD-cell consists of an integer multiplier, two in-
teger ALUs, and six general purpose registers and three small local memories.
Interconnections among the functional units are realized using a set of ten seg-
mented busses that run the length of the datapath. Many of the registers in a
pipelined computation can be implemented using the bus pipeline registers.
Functional unit outputs are registered. However, the output registers can
be bypassed via configuration control. Functional units may additionally be
pipelined internally depending on their complexity.
The control of the datapath is done using two types of signals: the static
control signals that are defined by the configuration memory as in ordinary FP-
GAs, and dynamic control that must be provided on every cycle. To program

2 In this case, the large amount of resources deployed to build macro instructions and the difficulty in

programming the FPGAs

Reconfigurable Architectures 65

an application on the RaPiD a mapping of functional blocks to computing ele-

ments of the datapath must be done, which result on the generation of a static
programming bitstream. This will be used to construct the pipeline and the
dynamic programming bits are used to schedule the operations of the compu-
tation onto the datapath over time. A controller is programmed to generate the
dynamic information needed to produce the dynamic programming bits.

5.5 Capacity / Chip Size

The last point we would like to deal with in the architecture is the defini-
tion of a measure of comparison between the RPUs. The most used measure
of comparison is the device capacity, which is usually provided by the man-
ufacturer as the number of gates that can be used to build a function into the
device. On fine-grained reconfigurable devices, the number of gates is often
used as unit of measure for the capacity of an RPU. A gate equivalent corre-
sponds to a two-inputs NAND gate, i.e. a circuit that performs the function
F = A · B. The gate density defines the number of gates per unit area. Beside
the capacity of an RPU, others factors like the number of pins and the device
speed may also play a big role in the comparison.
Coarse-grained reconfigurable devices doest not have any established mea-
sure of comparison. However, the number of PEs on a device as well as their
characteristics like the granularity, the speed of the communication link, the
amount of memory on the device and the variety of peripherals may provide
an indication on the quality of the device.
While the capacity of the device defines the "amount of parallelism" which
can be implement in the device, the speed gives an indication on the throughput
of data in the device. Designers should however bear in mind that the real speed
and real size that a design can achieve, depends on the implementation style
and the compilers used to produce the designs.

6. Conclusion
Our goal in this chapter was not to provided all possible details on the single
architectures of the reconfigurable computer chips presented. We rather focus
on the main characteristics in the technology as well as the operations. More
details on the single devices can be found in the corresponding datasheet. De-
spite the large amount of architecture presented in this section, the market is
still dominated by the FPGAs, in particular those from Xilinx and Altera. The
coarse-grained reconfigurable device’s market has not really taken-off so far,
despite the large amount of concept and prototypes developed in this direction.
Research and development in the architecture of reconfigurable computing
systems is very dynamic. By the time this book will be published, some con-
cept presented in this section will probably be no more actual. However, as
66 RECONFIGURABLE COMPUTING

we have seen with the FPGAs, the basic structure of reconfigurable device will
still remain the same in the future. In the case of FPGA, we will experience
some changes in the amount of input of the LUT, modifications in the I/O el-
ements and also the inclusion of various coarse-grained element in the chip.
However, LUTs will still be used as computational connected to each other
using programmable crossbar switches. I/O-elements will still be used for ex-
ternal communication. Coarse-grained device will still have the same structure
consisting of ALU-computing elements, programmable interconnections and
I/O-components. Understanding the concept presented here will therefore help
to better and faster understand the changes that will be made on the devices in
the future.
Chapter 3

IMPLEMENTATION

In the first part of this chapter, we present the different possibilities for the
use of reconfigurable devices in a system. According to the way those devices
are used, the target applications and the systems in which they are integrated,
different terminologies will be defined for the systems. We follow up by pre-
senting the design flow, i.e. the steps required to implement an application on
those devices. Because the programming of coarse-grained reconfigurable de-
vices is very similar to that of processors, we will not focus on the variety of
tool that exists for this purpose. The implementation on FPGA devices is rather
unusual in two points. First the programming is not aimed at generating a set
of instructions to be executed sequentially on a given processor. We seek the
generation of the hardware components that will be mapped at different time
on the available resources. According to the application, the resources needed
for the computation of the application will be built as components to be down-
loaded to the device at run-time. The generation of such components is called
logic synthesis. It is an optimization process whose goal is to minimize some
cost functions aimed at producing, for instance, the fastest hardware with the
smallest amount of resources and the smallest power consumption. The map-
ping of the application to the FPGA resources is a step of the logic synthesis
called technology mapping. The second unusual point with FPGAs is that the
technology mapping targets look-up tables rather than NAND-Gate as it is the
case with many digital devices. In the last part of the chapter, we will therefore
shortly present the steps required in logic synthesis and focus in more details
on technology mapping for FPGAs.

1. Integration
Reconfigurable devices are usually used in three different ways:
68 RECONFIGURABLE COMPUTING

Rapid prototyping: In this case, the reconfigurable device is used as an

emulator for another digital device, usually an ASIC. The emulation pro-
cess allows for functionally test the correctness of the ASIC device to be
produced, sometimes in real operating and environmental conditions, be-
fore production. The reconfigurable device is only reconfigured to emulate
a new implementation of the ASIC device.
Non-frequently reconfigurable systems: The reconfigurable device is in-
tegrated in a running system where it is used as an application specific pro-
cessor. These systems are usually stand alone systems. The reconfiguration
is used for test and initialization at start-up, and for upgrading purpose. The
device configuration is usually stored in an EEPROM of Flash from which
it is downloaded at start-up to reconfigure the device. No configuration
happens during operation.
Frequently reconfigurable systems: This third category comprises sys-
tems, which are frequently reconfigured. Those systems are usually cou-
pled with a host processor, which is used to reconfigure the device and
control the complete system.
With the increasing size and speed of reconfigurable processor1 it is possible
to implement many large modules on a reconfigurable device at the same time.
Moreover, for some reconfigurable devices, only a part of the device can be
configured while the rest continues to operate. This partial reconfiguration ca-
pability enables many functions to be temporally implemented on the device.
Depending on the time at which the reconfiguration sequence are defined, the
computation and configuration flow on a reconfigurable devices can be classi-
fied in two categories:

Compile-time reconfiguration: In this case, the computation and config-

uration sequences as well as the data exchange are defined at compile-time
and never change during a computation. This approach is more interesting
for devices, which can only be full reconfigured. However it can be applied
to partial reconfigurable devices that are logically or physically partitioned
in a set of reconfigurable bins.
Run-time reconfiguration: The computation and configuration sequences
are not known at compile-time. Request to implement a given task is known
at run-time and should be handled dynamically. The reconfiguration pro-
cess exchanged part of the device to accommodate the system to changing
operational and environmental conditions. Run-time reconfiguration is a
difficult process that must handle side effect factors like defragmentation of

1 Today leading edge reconfigurable devices contain million of gates with few hundreds MHz
Implementation 69

the device and communication between newly placed modules. The man-
agement of the reconfigurable device is usually done by a scheduler and a
placer that can be implemented as part of an operating system running on
a processor (figure 3.1). The processor can either resides inside or outside
the reconfigurable chip.

Figure 3.1. Architecture of a run-time reconfigurable system

The scheduler manages the tasks and decides when a task should be exe-
cuted. The tasks, which are available as configuration data in a database are
characterized through their bounding box and their run-time. The bounding
box defines the area that a task occupies on the device. The management of
task execution at run-time is therefore a temporal placement problem that
will be studied in details in chapter 5. The scheduler determines which task
should be executed on the RPU and then gives the task to the placer which
will try to place it on the device, i.e. allocate a set of resources for the
implementation of that task. If the placer is not able to find a site for the
new task, then it will be sent back to the scheduler which can then decide
to send it later and to send another task to the placer. In this case, we say
that the task is rejected.

No mater if we are dealing with a compiled-time or run-time reconfigurable

system, the computation and reconfiguration flow is usually the one shown on
Figure 3.2. The host CPU is used for device configuration and data transfer.
Usually, the reconfigurable device and the host processor communicates via a
70 RECONFIGURABLE COMPUTING

bus that is used for the data transfer between the processor and the reconfig-
urable device. In supercomputing, systems are made upon a high speed pro-
cessor from Intel or AMD and an FPGA-board attached to a bus like the PCI.
Several systems (Cray XD1, SGI, Nallatech SRC Computers) in which FPGAs
cohabits and communicate using a dedicated high speed bus, are available to
buy. In embedded systems, the processors are more and more integrated in the
reconfigurable devices and are heavily used for management purpose rather
than for computation. The RPU acts like a coprocessor with varying instruc-
tion sets accessible by the processor in a function call. The computation flow
can be summarized as shown in algorithm 2.

Figure 3.2. A CPU-RPU configuration and computation step

At the beginning of a computation, the host processor configures the recon-

figurable device2 . Then it downloads the segment of the data to be processed
by the RPU3 and give the start signal to the RPU. The host and the RPU can
then process in parallel on their segments of data. At the end of its computation
the host reads the finish signal of the RPU. At this point the data (computation
result) can be collected from the on RPU memory by the processor. The RPU

2 It is already possible to trigger the reconfiguration of the Xilinx FPGAs from within the device using

their ICAP-port. This might allow a self reconfiguration of a system from a processor running within the
device
3 The data might also be collected by the RPU itself from an external source
Implementation 71

Algorithm 2 CPU-RPU configuration and computation steps

1: Start
2: Initialize the RPU
3: while (1) do
4: Configure the RPU to implement a new task
5: Download Data for RPU computation into RPU-memory
6: Computes in parallel with the RPU if necessary
7: Upload the data computed by the RPU from the RPU-memory
8: end while
9: Stop

can also send the data directly to an external sink. In the computation flow
presented above, the RPU is configured only once. However, in frequently re-
configured systems several configurations might be done. If the RPU has to be
configured more than once, then the body of the while loop must be run again
according to the number of reconfigurations to be done before.

The design flow of a dynamic reconfigurable system is primary a hard-

ware/software partitioning process in which:
The part of the code to be executed on the host processor is determined.
This part is usually control-dominated.
The parts of the code to be executed on the reconfigurable device are iden-
tified. Those are usually data-dominated parts for which efficient dataflow
computation modules are required.
The interface between the processor and the reconfigurable device is im-
plemented.
The implementation of the control part on the CPU is done in software using
the common development languages and tools. It is usually a C-program with
system calls to access the device for reconfiguration. We will not focus on the
software development details in this book, since enough material that cover
software development exist.
No mater if the device is partially reconfigurable or not, the development
of the module to be downloaded later on the reconfigurable device follows the
same approach. The goal is to generate a dataflow computing block that best
matches the inherent parallelism of the part of the application to be imple-
mented as module.
Almost all manufacturers of coarse-grained reconfigurable devices provide
proprietary language and tools to implement such modules. It is usually a C-
like language with an appropriate scheduler that defines the function of the
72 RECONFIGURABLE COMPUTING

processing elements as well as their interconnections as function of the time.

Table 3.1 gives an overview of the languages and tools used by some manufac-
turers.

Manufacturer Language Tool Description

PACT NML (Structural) XPP-VC C into NML and configuration
Quicksilver SilverC InSpire SDK SiverC into Configuration
NEC-DRP C DRP Compiler C into configuration
picoChip C Picochip Toolchain C into configuration
IpFlex C/MATLAB DAP/DNA FW C/MATLAB into configuration

Table 3.1. Language and tools overview for coarse-grained RPUs

We will not further consider the programming on coarse-grained devices,

but rather refer to the corresponding manufacturers for more details on their
implementation languages and tools. In the next section, we will briefly present
the standard design flow for implementing digital circuits like ASIC. We will
focus in much detail on technology mapping for FPGA because it differs from
that of other digital devices.

2. FPGA Design Flow

The standard implementation methodology for FPGA designs is borrowed
from the ASIC design flow. The usual steps are presented in Figure 3.3. Note
that those steps concern only the modules of the application that has been iden-
tified by the hardware/software co-design process to be executed on the FPGA.
The software part is implemented using standard software development ap-
proaches, which are well covered in several textbooks.

2.1 Design Entry

The description of the function is made using either a schematic editor, a
hardware description language (HDL), or a finite state machine (FSM) editor.
A schematic description is made by selecting components from a given library
and connecting them together to build the function circuitry. This process has
the advantage of providing a visual environment that facilitates a direct map-
ping of the design functions to selected computing blocks. The final circuit is
built in a structural way. However, designs with very large amount of function
will not be easy to manage graphically. Instead, a Hardware Description lan-
guage (HDL) may be used to capture the design either in a structural or in a
behavioral way. Besides VHDL and Verilog, which are the most established
HDLs several C-like languages, mostly tied to just one compiler manufacturer
exist to describe hardware. Here we can cite the languages Handel-C [125]
and ImpulseC [175] and, in some extend, SystemC [170].
Implementation 73

Figure 3.3. The FPGA design flow

2.2 Functional Simulation

After the design entry step, the designer can simulate the design to check
the correctness of the functionality. This is done by providing test patterns to
the inputs of the design and observing the outputs. The simulation is done in
software by tools which emulate the behavior of the components used in the
design. During the simulation, the inputs and outputs of the design are usually
shown on a graphical interface, which describes the signal evolution in time.

2.3 Logic Synthesis

After the design description and the functional simulation, the design can
be compiled and optimized. It is first translated into a set of Boolean equa-
tions. Technology mapping is then used to implement the functions with the
available modules in function library of the target architecture. In case of FP-
GAs, this step is called LUT-based technology mapping, because LUTs are the
modules used in the FPGA to implement the boolean operators. The result of
the logic synthesis is called the netlist. A netlist describes the modules used to
implement the functions as well as their interconnections. There exist different
netlist formats to help exchange data between different tools. The most known
are the Electronic Design Interchange Format (EDIF). Some FPGA manufac-
turers provide proprietary formats. This is the case the Xilinx Netlist Format
(XNF) for the Xilinx FPGAs.
74 RECONFIGURABLE COMPUTING

2.4 Place and Route

For the netlist generated in the logic synthesis process, operators (LUTs,
Flip-Flopss, Multiplexers, etc...) should be placed on the FPGA and connected
together via routing. Those two steps are normally achieved by CAD tools
provided by the FPGA vendors. After the placement and routing of a netlist,
the CAD tools generate a file called a bitstream. A bitstream provides the
description of all the bits used to configure the LUTs, the interconnect ma-
trices, the state of the multiplexer and I/O of the FPGA. The full and partial
bitstreams can now be stored in a database to be downloaded later according
to the paradigm described in section 1

2.5 Design Tools

The design entry, the functional simulation and the logic synthesis are done
using the CAD tools from Xilinx, Synopsys, Synplicity, Cadence, ALTERA
and Mentor Graphics. The place and route as well as the generation of con-
figuration data is done by the corresponding vendors tools. Table 3.2 provides
some information on the tool capabilities of some vendors.

Manufacturer Tool Description

Synopsys FPGA Compiler synthesis
Mentor FPGA synthesis, place and route
Synplicity Sinplify Synthesis, place and route
Xilinx ISE Synthesis, place and route (only Xilinx products)
Altera Quartus II Synthesis, place and route (only Altera products)
Actel Libero Synthesis, place and route (only Actel products)
Atmel Figaro Synthesis, place and route (only Atmel products)

Table 3.2. Overview of FPGA Manufacturers and Tool Providers

The logic synthesis step is an important part in FPGA design. It is done by

tools, which solve a given number of optimization problems on the path from
the design entry to the generation of the configuration data. In the next section,
we will take a look inside those tools in order to understand how they work.
Since the technology mapping process of FPGA differs from that of ASIC, we
will pay more attention to this step and present in detail some of technology
mapping algorithms, developed for FPGA devices. Each of the algorithms is
best adapted for the optimization of a given cost function.

3. Logic Synthesis
A function, assigned to the hardware in the hardware/software co-design
process, can be described as a digital structured system. As shown in Figure
3.4, such a digital structured system consists of a set of combinatorial logic
modules (the nodes), memory (the registers), inputs and outputs.
Implementation 75

Figure 3.4. A structured digital system

The inputs provide data to the system while the outputs carry data out of the
system. Computation is performed in the combinatorial parts and the results
might be temporally stored in registers that are placed between the combinato-
rial blocks. A clock is used to synchronize the transfer of data from register to
register via combinatorial parts. The description of a design at this level is usu-
ally called register transfer description, due to the register to register operation
mode previously described.
For such a digital system, the goal of the logic synthesis is to produce an
optimal implementation of the system on a given hardware platform. In the
case of FPGA, the goal is the generation of configuration data which satis-
fies a set of given constraints like the maximal speed, the minimum area, the
minimum power consumption, etc. In a structured system, each combinatorial
block is a node that can be represented as a two-level function or as multi-level
function. Depending on the node representations, the two following synthesis
approaches exist:

Two-Level Logic Synthesis: Two-level synthesis deals with the synthesis of

designs represented in two-level logic. Those are representations in which
the longest path from input to output, in term of number of gates crossed on
the path, is two. Two-level logic is the natural and straightforward approach
to implement a Boolean function, because each Boolean function can be
represented as a sum of product terms. In the first level, the products are
built using the AND-primitives. The sums of the resulting products are
built in the second level with the OR-primitives.
76 RECONFIGURABLE COMPUTING

Multi Level Logic Synthesis: In the multi-level synthesis, functions are rep-
resented using a multi-level logic. Those are circuits in which the longest
path from input to output goes through more than two gates.

Most of the circuits used in practice are implemented using multi-level logic.
Multi-level circuits are smaller, faster in most cases and consume less power
than two-level circuits. Two-level logic is most appropriate for PAL and PLA
implementations while multi-level is used for standard cell, mask-programmable
or field-programmable devices.
We formally represent a node of the structured system as a Boolean network,
i.e. a network of Boolean operators that reflects the structure and function of
the nodes. A Boolean network is defined as a directed acyclic graph (DAG)
in which a node represents an arbitrary Boolean function and an edge (i, j)
represents the data dependency between the two nodes i and j of the network.

Figure 3.5. Example of a boolean network with: y1 = x1 + x2 , y2 = x3 · x4 , y3 = x5 · x6 ,

y4 = y1 + y2 , z1 = y1 + y4 , and z2 = y2 ⊕ y3

Figure 3.5 shows a Boolean network example for the functions.

3.1 Node representation

The representation of a node is of great importance in the synthesis process.
In two-level logic, the question on how to represent the node of a Boolean net-
work is not relevant, since the final representation is the same sum of products
as the initial representation. Although any valid representation is allowed, in
multi-level logic, representation are seek, which efficiently use the memory
and for which a correlation with the final representation exists. Furthermore,
the representation should be easy to manipulate.
Logic synthesis can be done in two different approaches: the technology
dependant synthesis and the technology independent. In the first case, only
valid gates chosen from the target library are used in the node representation.
Implementation 77

The final implementation matches the node representation. In the second case,
the representation is technology independent, i.e the design is not tied to any
library. A final mapping must be done to the final library in order to have
an implementation. The technology independent method is most used, due to
the large set of available optimization methods available. With a technology
independent representation, synthesis for FPGA devices is done in two steps.
In the first step, all the Boolean equations are minimized, independent of the
function generators used. In the second step, the technology mapping process
maps the parts of the Boolean network to a set of LUTs.
In general, the following choices are made for the representation of a node:
Sum of products form: A Sum of Product ( SOP) is the most trivial form
to represent a Boolean function. It consists of a sum of product terms and
it is well adapted for two-level logic implementation on PALs and PLAs.
Example: f = xyz + xyz + wxy.
This representation has the advantage that it is well understood and it is
easy to manipulate. Many optimization algorithms are available (AND,
OR, Tautology, two-level minimizers). The main disadvantage is the non
representativity of the logic complexity. In fact, designs represented as
sum of products are not easy to estimate as the complexity of the design
decreases through manipulation. Therefore estimation of progress during
logic minimization on SOPs is difficult.
Factored form A factored form is defined recursively either as a single
literal or, as a product or a sum of two factored forms: a product is either
a single literal or the product of two factored forms and a sum is either a
single literal or the sum of two factored forms. c(a + b(d + e)) is a product
of the factored forms c and a + b(d + e), and a + b(d + e) is a sum of the
factored forms a and b(d + e).
Factored forms are representative of the logic complexity. In many design
styles, the implementation of a function corresponds to its factored form.
Therefore, factored forms are good estimation of complexity of the logic
implementation. Their main disadvantage is the lack of manipulation algo-
rithms. They are usually converted in SOPs before manipulation.
Binary decision diagram (BDD) A binary decision diagram (BDD) is a
rooted directed acyclic graph used to represent a boolean function. Two
kinds of nodes exist in BDDs: variable and constant nodes.
– A variable node v is a non terminal having as attribute its argument
index4 index(v) ∈ {1, ..., n} and its two children low(v) and high(v).

4 An index i defines a variable xi .

78 RECONFIGURABLE COMPUTING

– A constant node v is a terminal node with a value value(v) ∈ {0, 1}.

A BDD in which an ordering relation among the nodes exists is called a
Ordered BDD (OBDD). The non terminal nodes are ordered from the root
to the terminal nodes. Formally, for each non terminal node v, if low(v) is
non terminal, then index(low(v)) < index(v). Similarly, if if high(v) is
non terminal, then index(high(v)) < index(v).
The correspondence between a BDD and a Boolean relation is define as
follow: A BDD with root v denotes a function fv .
– If v is a terminal node, then if value(v) = 1, then fv = 1, else
(value(v) = 0) fv = 0.
– If v is a non terminal node with index(v) = i , the Shannon expan-
sion theorem is used to express the function fv as fv = xi flow(v) +
xi fhigh(v) , where flow(v) respectively fhigh(v) denote the function rooted
at low(v) respectively high(v). The value of fv for a given assignment
is obtained by traversing the graph from the root terminal according to
the assignment values of the nodes.
A BDD G is a Reduced Ordered BDD (ROBDD) if the following holds:
– low(v) 6= high(v), ∀v in G
– ∀v, v ′ ∈ G, the subtrees rooted at v and the subtree routed at v ′ are not
isomorphic5
Figure 3.6 shows the optimal BDD-representation of the function f =
abc + bd + bcd.
Basically, BDDs are not canonical, i.e. there might exist several BDD rep-
resentations of the same Boolean function. Reduced ordered BDDs are
canonical and compact, thus they are good replacements of truth tables.
For a good ordering, reduced ordered BDDs remain reasonably small for
complicated functions. Manipulations of BDDs are well defined and effi-
cient.

3.2 Node manipulation

Having represented the node in one of the formats previously presented op-
timization algorithms can be applied to the Boolean network. The goal is to
generate an optimized network with either a less amount of gates or the lowest

5 Two BDDs G and G are isomorph ⇐⇒ there exists a bijective function σ from G in G such
1 2 1 2
that: 1) for a terminal node v ∈ G1 , σ(v) = w is a terminal node in G2 with value(v) = value(w); 2)
for a non terminal node v ∈ G1 , σ(v) = w is a non terminal node of G2 with index(v) = index(w),
σ(low(v)) = low(w) and σ(high(v)) = high(w)
Implementation 79

Figure 3.6. BDD-representation of the function f = abc + bd + bcd

depth of the gates. A given number of transformations can be applied to the

network for this purpose. Among them we can cite the following:
Decomposition: The decomposition takes a single Boolean function and
replace it with a collection of new expressions. We say that a Boolean
function f (X) is decomposable if we can find a function g(X) such that
f (X) = f ′ (g(X), X).
The function f = abc+abd+acd+bcd for example operates on 12 literals.
Using decomposition, f can be written as: f = ab(c + d) + (a + b)cd =
ab(c + d) + ab(c + d) = XY + XY with X = ab and Y = c + d. This
decomposition reduces f the operation of f to only 8 literals.
Extraction: The extraction is used to identify common intermediate sub-
functions of a set of given functions in order to avoid redundancy.
Example: f = (a + bc)d + e and g = (a + bc)e can be rewritten as
f = xd + e and g = xe with x = a + bc. The output x of the common part
of the circuit will be used as common input for f and g.
Factoring: Factoring is the transformation of SOP-expressions in factored
form. For example the function f = ac + ad + bc + bd + e can be rewritten
as f = (a + b)(c + d) + e
Substitution: Substitution replace an expression e within a function f with
the value of an equivalent function g(X) = e. For example the function
f = (a + bc)(d + e) can be rewritten as f = g(d + e), with g = a + bc.
Collapsing: Also called elimination, collapsing is the reverse operation of
the substitution. Collapsing is used to eliminate levels in order to meet the
80 RECONFIGURABLE COMPUTING

timing constraints. The function f = ga + gb for example will be replaced

by f = ac + ad + bcd with g = c + d.

3.3 LUT-based Technology Mapping

The technology independent optimization phase ends with a reduced Boolean
network in which the fanin and fanout of gates vary. The next step consists of
allocating LUTs, which are the FPGA library elements for the implementa-
tion of the different nodes. Several goals might be followed here. If the goal
is to minimize the chip area used by the circuit, then the mapping will try to
allocate the less possible amount of LUTs. If the goal is to minimize the de-
lay of the final circuit, then the mapping will try to minimize the depth of the
LUTs used. Other goals like testability and low power might also be followed.
Several LUT technology mapping algorithms have been developed in the past.
Depending on their optimization goals, those algorithms can be classified in
three categories:
The first category contains the algorithms, whose goal is to minimize the
area. This category includes the Chortle-crf [87][86], the MIS-fpga [163]
and the Xmap [133]. Due to its popularity and the proof of area optimality
for LUTs with less than 5 inputs, the Chortle algorithm will be presented
in this section.
The algorithms in the second category target the delay minimization. Al-
gorithms in this category include the FlowMap [53], the Chortle-d[86],
the DAG-map[46] and the MIS-pga-delay[164]. The FlowMap algorithm
was a breakthrough in delay minimization because the authors were able
to present an optimal algorithm for LUT-technology mapping with delay
minimization as well as a proof of the polynomial-time complexity of their
algorithm. We present the FlowMap in detail in this section.
The third category contains algorithms which focuses on maximizing the
routability. This category includes the Bhat and Hill [26] work as well as
the Schlag[54], Kong and Chang approach[54]. None of those methods will
be presented here.
Besides the algorithms presented in this section, a large variety of high quality
LUT-technology mapping algorithms have been developed in the last couple of
years. Also research on LUT-based technology mapping keeps going on. All
those algorithms cannot be presented here. We therefore referred the interested
readers to the large amount of available publications. A starting point is the
survey provided by Cong et al [54]. We will first present some definitions
needed to better understand the algorithms that we present. We start providing
a formal definition of the problem that the algorithms we introduce intend to
solve. The LUT-technology mapping problem can be stated as follow:
Implementation 81

Definition 3.1 (LUT-Based Technology mapping problem) Given

a Boolean network representing a function f and an integer k. Find an imple-
mentation of f using only k-inputs LUTs, such that:
1 The amount of LUTs use is minimal or
2 The delay of the resulting circuit is minimal.

The LUT-based technology mapping is the problem of covering a given graph

(the Boolean network) with a set of k-input LUTs. Because the area of the final
implementation is defined through the amount of LUT used, the first condition
is equivalent to having a covering with a minimal amount of LUTs. The delay
in an FPGA is influenced by two main factors: The delay in LUTs and the
interconnection delay. While the delay in LUT is known, the interconnection
delay can only be accurately known after the place and route phase. Delay
estimation at this stage is done using the depth of LUTs in the design, thus
assuming the interconnection delay to be one for each wire. The more LUTs
are available on a path from the input to the outputs the higher the delay in the
circuit.

Definition 3.2 (Primary Input, Primary Output, Node Level,Node Depth, Fan-in,
Given a boolean network G, we define the following:
1 A primary input (PI) node is a node without any predecessor.
2 A primary output (PO) node is a node without any successor.
3 The level l(v) of a node v is the length of the longest path from the primary
inputs to v.
4 The depth of a network G is the largest level of a node in G.
5 The fan-in of a node v is the set of gates whose outputs are inputs of v.
6 The fan-out of v is the set of gates that use the output of v as input.
7 Given a node v ∈ G , input(v) is defined as the set of node of G, which
are fan-in of v, i.e. the set of predecessors of v.

Definition 3.3 (Trees, Leaf-Dag) 1 A tree or fan-out-free circuit is

one in which each node has a maximal fan-out of one.

2 A leaf-DAG is a combinational circuit in which the only gates with a fan-in

greater than one are the primary inputs.

Definition 3.4 (K-Bounded network) Given a Boolean network G

and a subgraph H of G.
82 RECONFIGURABLE COMPUTING

With input(H), we denote the set of all nodes not included in H, which are
predecessors of some nodes in H.

G is K-bounded if |input(v)| ≤ K for all nodes of G.

A K-bounded Boolean network can be directly mapped to a set of k-inputs
LUT by assigning a LUT for each node. However this straightforward ap-
proach may not produce the expected optimal results.
Definition 3.5 (Cone at a node) Given a Boolean network G,
1 A cone Cv at a node v is the tree with root v which spans from v to its
primary inputs.

2 The cone Cv is K-feasible if:

input(Cv ) ≤ K
any path connecting two nodes in Cv to v lies entirely in Cv
An illustration of K-feasible cone at a node v is given in figure 3.7

Figure 3.7. Example of a K-feasible cone Cv at a node v

With the previous definition of K-feasible cones, the LUT technology map-
ping becomes the problem of covering the graph with a set of K-feasible cones
that are allowed to overlap. The technology mapping results in a new DAG in
which nodes are K-feasible cones and edges represent communication among
the cones. Figure 3.8 shows the covering of a graph with 3-feasible cones and
the resulting LUT-mapping to 3-input LUTs.
Next, we present some existing LUT-technology mapping algorithms and
explain their advantage as well as their drawbacks.

3.3.1 The Chortle Algorithm

The chortle algorithm was developed by Francis et al [87][86] at the Uni-
versity of Toronto in 1991 with the aim of minimizing the amount of LUTs in
Implementation 83

3−LUT 3−LUT

3−LUT

Figure 3.8. Example of a graph covering with K-feasible cone and the corresponding covering
with LUTs

the implementation of a given circuit. It operates in two steps: In the first step,
the original Boolean network is partitioned into a forest of trees that are then
separately mapped into circuits of K-input LUTs. The second step assembles
the circuits implementing the trees to produce the final circuit.
The transformation of the original network into a forest is done by partition-
ing each fan-out node v. Therefore, sub-network rooted at v is duplicated for
each input triggered by the fan-out nodes of v. The resulting sub-networks are
either trees or leaf-DAGs. The leaf-DAGs are converted in trees by creating a
unique instance of a primary input for each of its fan-out edges.

Mapping the trees. The strategy used by Chortle to map a tree is a combina-
tion of bin packing and dynamic programming. Each tree is traversed from the
primary inputs to the primary outputs. At each node v, a circuit referred to as
the best Circuit, implementing the cone at v extending from the node to the pri-
mary inputs of the network is constructed. The best circuit is characterized by
two main factors: The tree rooted at v and represented by a cone must contain
the minimum number of LUTs and the output LUT (the root-LUT) implement-
ing v should contain the maximum number of unused input pins. For a primary
input p the best circuit is a single LUT whose function is a buffer. Using the
dynamic programming, the best circuit at a node v can be constructed from
its fan-in nodes, because each of them is already optimally implemented. The
procedure enforces the use of the minimum number of LUTs at a given node.
The best-circuit is then constructed from the minimum number of LUTs used
to implement its fan-in nodes. The secondary goal is to minimize the number
of unused inputs of the circuit rooted at node v.
The construction of the tree is done in two steps. First a two-level decompo-
sition of the cone at v is constructed and then this decomposition is converted
into a multi-level decomposition.
84 RECONFIGURABLE COMPUTING

Two-level decomposition. The two-level decomposition consists of a single

first level node and several two-level nodes (figure 3.9(a)). Each second level
node implements the operation of the node being decomposed over a subset of
the fan-in LUTs. The first level node is not implemented at this stage. This will
be done in the second phase of the algorithm where the two-level representation
is converted into a multi-level implementation.

(a) Fan-in LUTs at a node (b) The two-level decomposition

Figure 3.9. Chortle two-level decomposition

The two-level decomposition is constructed using a bin packing algorithm

approach. In the traditional bin packing algorithm, the goal is to find the min-
imum number of bins with a given capacity, into which a set of boxes can be
packed. In the Chortle algorithm, the bins are the second-level LUTs and the
boxes are the fan-in LUTs. The capacity of each bin is the number k of LUT
inputs. The packing at this stage consist of combining two fan-in LUTs, LU T1
that realizes the function f1 and LU T2 that realizes the function f2 into a new
LUT LU Tr that implements the function f1 ⊘ f2 , where ⊘ is the operation
implemented by the fan-out node. Figure 3.9(a) shows an original graph and
its decomposition is shown in 3.9(b). In example, the ⊘ is the OR function.
The pseudo code of the two-level decomposition algorithm is given in algo-
rithm 3.
This algorithm uses a First-Fit-Decreasing (FFD) method, which places a
fan-in LUT to be packed into the first LUT it will fit within. However, a Best-
Fit (BF) approach that packed the fan-in LUTs into the LUT they best fit in,
can also be used.

Multi-level Decomposition. In the second step, the first-level nodes are im-
plemented using a tree of LUTs. The number of LUTs used is minimized by
using second level LUTs that have unused pins to implement a portion of the
first-level three as shown in figure 3.10. The detailed procedure for converting
a two-level decomposition into a multi-level decomposition is given in algo-
rithm 4 and figure 3.10 provides an illustration of this process.
Implementation 85

Algorithm 3 Chortle’s two-level decomposition

Start with an empty list of LUTs
while there are unpacked fan-in LUTs do
if the largest unpacked fan-in LUT will
not fit within any LUT in the list then
create an empty LUT and add
it to the end of the list
end if
pack the largest unpacked fan-in LUT
into the first LUT it will fit within
end while

Algorithm 4 Chortle’s multi-level decomposition

while there is more than one unconnected LUT do
if there are no free inputs among
the remaining unconnected LUTs then
create an empty LUT and add
it to the end of the LUT list
end if
connect the most filled unconnected LUT
to the next unconnected LUT with a free input
end while

The fact that the most filled unconnected LUTs are always selected pushes
less filled LUTs to the root of the tree being built. Therefore, the LUTs with
the most unused inputs will be found near the root of the tree.

Improvement of the Chortle algorithm. Before applying the decomposi-

tion on the trees in the forest, the Chortle algorithm performs a pre-processing
step in which De Morgan’s Theorem as well as the associativity rules are ap-
plied in order to insure that:
the only inverted edges in the trees are those originating from the leaf nodes.
no consecutive OR and no consecutive AND exist in the trees.
Subject to this Francis et al. [87] could prove that the Chortle algorithm con-
struct an area optimal algorithm for LUTs with less than or equal five inputs.

Exploiting the reconvergent paths. A further optimization of the Chortle

consists of optimizing the reconvergent paths to improve the implementation
at a given node. A reconvergent path is caused by a leaf node with a fan-out
greater than one. This produces two paths in the leaf-DAG that terminate at
86 RECONFIGURABLE COMPUTING

Figure 3.10. Example of multi-level decomposition

a given node. The goal of the optimization is to pack the reconvergent paths
caused by a given input into just one LUT.
This strategy is illustrated in figure 3.11. To see are two paths going through
the blocks g and h and placed in the same LUT, thus reducing the number of
LUTs used from five to four.

Figure 3.11. Exploiting reconvergent paths to reduce the amount of LUTs used
Implementation 87

If more than one pair of fan-in LUTs share inputs, there will be several
pairs of reconvergent paths. To determine which one should be packed in the
same LUT, two approaches exist in the Chortle algorithm. The first one is an
exhaustive approach that first finds all pairs of fan-in LUTs that share inputs.
Then every possible combination is constructed by first merging the fan-in
LUTs and then proceeding with the FFD bin packing algorithm. The two-level
decomposition that produces the fewest bins and the smallest least filled bins
is retained as the solution.
A large amount of pairs of fan-in LUTs sharing the same inputs cause the
algorithm to be impracticable. To overcome this limitation, a second heuristic
called the Maximum Share Decreasing (MSD) can be used. The goal is to
maximize the sharing of inputs when fan-in LUTs (boxes) are packed into bins.
The MSD iteratively chooses the next box to be packed into bins according to
the following criteria: 1) the box hast the greatest number of inputs, 2) the box
shares the greatest number of inputs with any existing bin and 3) it shares the
greatest number of inputs with any of the remaining boxes.
The first criterion insures that the MSD algorithm works like the FFD if no
reconvergeant path exists. The second and third criteria help to place boxes
that share inputs into the same bins. The chosen box is then packed into the
bins with which it shares the most input without exceeding the bin capacity,
i.e. the number of inputs. If no more bins exist for packing the chosen bin,
then a new bin is created and the chosen box is packed into the new bin.

Logic replication at fan-out LUTs. The replication of the logic at fan-out

nodes can also help to reduce the amount of LUTs used. As stated earlier,
the Chortle technology mapping first decompose the Boolean network into a
forest, then maps each tree separately and finally assembles the final circuit
from the mapped threes. In the assembling phase logic replication is used to
reduce the amount of LUTs in the final circuit as illustrated in figure 3.12.

Figure 3.12. Logic replication at fan-out nodes to reduce the number of LUTs used
88 RECONFIGURABLE COMPUTING

The chortle algorithm exists in two versions. One with the replication of all
nodes and the other without replication. The solution that produces the less
amount of LUT is retained.
The Chortle algorithm presented in this section is known in the literature
[87] as the Chortle-crf 6
As seen in figure 3.10, the path generated by the Chorlte algorithm can be-
come very long if no effort is spent in reducing the delay. This problem is
addressed in another version of the Chrotle algorithm, the chortle-d7 . Instead
of using the bin packing to minimize the amount of LUTs used, the Chortle-d
focus in the bin packing strategy, on the reduction of the number of levels in
the final design. Because we investigate a delay optimal algorithm in the next
section, the Chortle-d algorithm will not be considered futher. The FlowMap
algorithm that we next present focuses on delay minimization using a network
flow approach, a technique that was also use in the MIS-pga algorithm of Mur-
gai et al[163].

3.3.2 The FlowMap Algorithm

Cong and Ding proposed a polynomial time algorithm for LUT-based tech-
nology mapping with the goal of delay minimization [53]. Their algorithm, the
FlowMap uses the notion of cut in a network to construct an optimal solution
in polynomial time.
Because the delay of the final circuit is determined by the delays in the LUTs
as well as those of the interconnections, an algorithm seeking to minimize the
delay should consider those two factors. However, since the components are
not placed and routed on the chip before the technology mapping phase, no real
accurate estimation of the interconnection delays can be done. The technology
mapping with delay minimization as goal can be done only on the basis of the
LUT delays. Delay minimization is equivalent to the minimization of the depth
of the resulting DAG, i.e. the minimization of the number of LUTs on a path
from the primary inputs to the primary outputs of the final circuit.
In contrast to most existing algorithms that first partitioned the Boolean net-
work into a forest of trees and then map each tree separately, the FlowMap
algorithm can be directly applied to the Boolean network. The precondition
is that the network must be k-bounded. If this is not the case, then a pre-
processing must be performed on the network to transform it to a K-bounded
one. This can be done for example with the same approach used in the Chortle
to generate the trees.
The FlowMap algorithm is a two step method, which first determines the
labels of the nodes in the first phase and then, in the second phase, assembles

6c is for the constructive bin packing, r for the reconvergent path and f for the logic replication.
7d stays for delay
Implementation 89

the nodes in the LUTs according to their level number. We present the two
steps in the next paragraphs.

First Phase: Node Labeling. Labeling is based on the notion of cut in a

network. We therefore first recall some definitions related to the notion of
network and flow in a network.

Definition 3.6 (Network, cut, label, height, volume) Given

a network N = (V, E) with a source s and a sink t, a cut is a partition
(X, X) of the graph N such that s ∈ X and t ∈ X).

The cut-size n(X, X) of a cut (X, X) is the number of nodes in X adjacent

to some nodes in X

A cut (X, X) is K-feasible if n(X, X) ≤ K

The edge cut-size e(X, X) of (X, X) is the sum of the crossing edges ca-
pacities.

The label l(t) of a node t is defined as the depth of the LUT, which contains
t in an optimal mapping of the cone at t.

The height h(X, X) of a cut (X, X) is the maximum label in X.

h(X, X) = max {l(x) : x ∈ X}

The volume vol(X, X) of a cut (X, X) is the number of nodes in X.

vol(X, X) = X

The level of the K-LUT containing the node t in an optimal mapping of N is

at least l(t) and the label of all primary outputs of N is the depth of the optimal
mapping of N .
The first phase of the algorithm computes the label of the nodes in a topolog-
ical order, starting with the primary inputs of the graph. The topological order
guarantees that each node is processed after all its predecessors have been pro-
cessed. The labeling is done as follows: Each primary input u is assigned the
label l(u) = 0. For a given node to be processed at a given level, the subgraph
Nt representing the cone at node t is transformed into a network Nt by insert-
ing a source node s which is connected to all inputs of Nt as shown in figure
3.13. For the sake of simplicity the resulting network is still denoted Nt
Assuming that t is implemented in the K-LUT LU T (t) in an optimal map-
ping of Nt , the cut (X(t), X(t)), where X(t) is the set of nodes in LU T (t)
and X(t) = Nt − X(t), is a K-feasible cut between s and t and the level of
LU T (t) is l(u) + 1, where u is the node of X(t) with the maximum level.
90 RECONFIGURABLE COMPUTING

Figure 3.13. Construction of the network Nt from the cone Ct at node t

With the previous reflection, the K-LUT mapping with minimal delay is
reduced to the problem of finding a K-feasible cut with minimum height8 for
each node in the graph. The level l(t) of the node t is therefore given by the
following formula:

l(t) = min{(X,X) is K−f easible} (h(X, X) + 1) (3.1)

The following Lemma results from the previous discussion and therefore it will
be given without proof.

Lemma 3.7 The minimum depth of any mapping solution of Nt is given by:

l(t) = min{(X,X) is K−f easible} (h(X, X) + 1)

Figure 3.14 illustrates the FlowMap labelling method. Because there is a min-
imum height 3-feasible cut in Nt , we have l(t) = 2 and the optimal 3-LUT
mapping solution for Nt is given in the figure.
The goal of minimizing the delay can be reduced to efficiently compute
the minimum height K-feasible cut for each node visited in the graph. The
FlowMap algorithm constructs a mapping solution with minimum delay in

8 It is assumed here that no cut (X, X) is computed with primary input nodes in X.
Implementation 91

Figure 3.14. Minimum height 3-feasible cut and node mapping

time O(Km), where m is the number of nodes in the network. Further trans-
formations are required on the networks Nt in order to reach this goal. The
node labels defined by the FlowMap scheme satisfy the following property.

Lemma 3.8 Let l(t) be the label of node t, then l(t) = p or l(t) = p + 1,
where p is the maximum label of the nodes in input(t).

Proof Let t′ be any node in input(t). Then for any cut (X, X) in Nt , either

1 t′ ∈ X or
2 (X, X) also determines a K-feasible cut (X ′ , X ′ ) in Nt′ with
h(X ′ , X ′ ) ≤ h(X, X), where X ′ = X ∩ Nt′ and X ′ = X ∩ Nt′ . Those two
cases are lillustrated in figure 3.15

In the first case, we have l(t) = h(X, X) + 1 ≥ l(t′ ) + 1 and therefore

l(t) ≥ l(t′ )
In the second case we have l(t′ ) − 1 ≤ h(X ′ , X ′ ) ≤ h(X, X) = l(t) − 1,
which implies l(t) ≥ l(t′ ). Therefore, l(t) ≥ p. This part of the proof shows
that the label of a node cannot be smaller than those of its predecessors.
Because the network Nt is K-bounded, input(t) ≤ K. Therefore (Nt −
t, {t}) is a K-feasible cut. Since each node in Nt − {t} is either in input(t)
or is a predecessor of some node in input(t), the maximum label of the nodes
in Nt − {t} is p. Therefore, h(Nt − {t}, {t}) = p, i.e l(t) ≤ p + 1.
92 RECONFIGURABLE COMPUTING

Figure 3.15. Illustration of the two cases in the proof of Lemma 3.8

According to Lemma 3.8, a delay optimal mapping algorithm could first

checks if there is a K-feasible cut (X, X) of height p − 1 in Nt . If such a cut
exists, then l(t) is assigned the value p and the node t will be packed in the
second phase in a common LUT with the nodes in X. If such a cut does not
exist, then the minimum height of the K-feasible cuts in Nt is p and Nt −
{t} , {t} is such a cut. The value of l(t) is set to p + 1 in this case and a new
LUT will be used for t in the second phase of the algorithm. This is the way
the FlowMap algorithm works.
We now face the next problem, namely finding out if a network has a K-
feasible cut with a given height h. In order to provide an answer to this ques-
tion, further transformations are required. The first one transforms Nt into N ′ t
as described in Lemma 3.9
Lemma 3.9 Let N ′ t be the graph obtained from Nt by applying a transfor-
mation which collapses all the nodes in Nt with maximum label p − 1 together
with t in a new node t′ . Nt has a K-feasible cut of height p − 1 iff N ′ t has a
K-feasible cut.

Proof Let Ht denote the set of node in Nt that are collapsed into t′ .
⇐ If N ′ t has a K-feasible cut(X ′ , X ′ ), let X = X ′ and X = (X ′ − {t′ } ∪
Ht ), then (X, X) is a K-feasible cut of Nt . Because no node in X ′ (= X) has a
label p or larger, we have h(X, X) ≤ p − 1. Furthermore according to Lemma
3.8, l(t) ≥ p, which implies h(X, X) ≥ p − 1. Therefore h(X, X) = p − 1.

⇒ If Nt has a cut (X, X) of height p − 1, then X cannot contain any node

of label p or higher. Therefore, Ht ⊆ X, meaning that (X, (X − Ht ) ∪ {t′ })
forms a K-feasile cut of Nt′ .
Now that the problem of finding a minimum height K-feasible cut in the
network Nt is reduced to the problem of finding a K-feasible cut in Nt′ , exist-
ing network flow algorithms can be used to compute cuts in Nt′ and then check
if they are feasible. Unfortunately, for those algorithms the cuts are defined
Implementation 93

Figure 3.16. Transforming Nt into N ′ t by node collapsing

on crossing edges. Moreover, it is difficult to establish a relation between the

number of crossing edges and the number of adjacent nodes in the original net-
work. A second transformation, called node splitting is therefore applied on
Nt′ . The goal of this second step is to reduce the node cut-size constraint in Nt′
to an edge cut-size constraint in the new graph Nt′′ and then, use well known
existing edge-cut computation algorithms.
The node splitting transforms the graph Nt′ into a graph Nt′′ as follow:
For each node v in Nt′ other than s and t′ , two new nodes v1 and v2 are
introduced and connected by a bridging edge (v1 , v2 ).
The source s and sink t′ are also introduced in Nt′′ . For each edge (s, v)
respectively (v, t′ ) in Nt′ , an edge (s, v1 ) respectively (v2 , t′ ) is introduced
in Nt′′ .
For each edge (u, v) in Nt′ with u 6= s and v 6= t′ , an edge (u2 , v1 ) is
introduced in Nt′′ . The capacity of each bridging edge is set to one and
those of the non bridging edges is set to infinity.
The constructions of the graphs Nt′ and Nt′′ are illustrated on figures 3.16 and
3.17.
The transformation of the graph Nt′ into Nt′′ insures that if a cut exists in
′′
Nt with capacity less than K, then no edge with infinite capacity will be a
crossing one. The only edges that would be crossing the cut are the bridging
ones. Because each bridging edge represents a node of Nt′ and the capacity of
94 RECONFIGURABLE COMPUTING

Figure 3.17. Transforming the node cut constraints into the edge cut ones

each bridging edge is one, the edge size cut in Nt′′ is equivalent to the node cut
size in Nt′ . Based on this observation, the following Lemma can be stated.

Lemma 3.10 Nt′ has a K-feasible cut if Nt′′ has a cut whose edge cut size is
no more than K.

The Ford and Fulkerson method [84] [55] can be used to check if a cut
with edge cut-size smaller or equal K exists. We first briefly provide some
more background on networks, which are important to understand the testing
procedure.
In a network N with source s and sink t a flow can be seen as a streaming
of data on different direction on the edges of the network. A node might have
data coming in and data going out through its edges, each of which has a given
capacity. Data streaming in the s-direction (resp. in the t-direction) caused a
negative (resp. positive) value of the flow at a given node. The value of the
flow in the network is therefore the sum of the flows on the network edges.
Formally, a flow f in N can be defined as a function of N × N in .
A residual value exists on an edge if, the flow at that edge has a lower value
than the edge capacity. The residual value is then the difference between the
capacity and the flow’s value. This can be added to the flow in order to saturate
the edge. The capacity of a cut in the network is defined as the sum of all
positive crossing edge capacities. Negative crossing edges do not influence the
capacity of a cut. An edge not saturated by the flow is called a residual edge.
The residual network of a network is made upon all the residual edges as well
Implementation 95

as the nodes that they connect. An augmenting path in the residual network is
a path from source s to the sink t that contains only residual edges, i.e. a path
in the residual network.
A very important relationship between a flow and a cut in a network is given
by the following corollary:

Corollary 3.11 The value of any flow in the network is bounded from
above by the capacity of any cut in the network.

This implies the following: the maximum flow value in the network is bounded
from above by the minimum cut of the network.
Based on the above observation, the notion of cut and that of residual net-
work, the famous max-flow min-cut theorem [55] of Ford and Fulkerson states
that a flow is maximum in the network if the residual network does not contain
any augmenting path. The value of the flow is then equal to the capacity of the
minimum cut.
Applying the max-flow min-cut theorem to our problem, we can state the
following: If a cut with edge cut-size smaller or equal K exists in Nt′′ , then the
maximum value of any flow between s and t′′ in Nt′′ will be less than K.
Because we are only interested in testing if the value of the cut is smaller
than K, the augmenting path of Ford an Fulkerson can be applied to compute
the maximum flow. The approach starts with a flow f , whose value is set to
0 and then, iteratively find some augmenting path P in the residual network
and increase the flow on P by the residual capacity cf (P ), that is the value
by which the flow on each edge of P can be increased. If no path from s to t
exists, then the computing stops and return the maximum flow value.
Because each bridging edge in Nt′′ have a capacity of one, each augmenting
path in the flow residual graph of Nt′′ from s to t′′ increases the flow by one.
Therefore, the augmenting path can be recursively used to check if the maxi-
mum value for a flow associated to a cut is less than K. For a given cut, if a
K + 1 augmenting paths could be found, then the maximum flow in Nt′′ has a
value more than K. Otherwise, the residual graph will not contain a (K + 1)th
path.
Testing if Nt′′ has a cut, whose value is no more than K can therefore be
done through a depth first search starting at s and including in X ′ all nodes
reachable by s. The run-time of the Ford and Fulkerson method is O(m|f ∗ |,
where |f ∗ | is the value of the maximal flow computed. In the F lowM ap al-
gorithm, this value is K, which corresponds to the number of iterations that
were performed to find the K augmenting paths. Since finding an augmenting
path takes O(m) (m being the number of edges of Nt′′ ), testing if a cut with
edge cut-size less or equal K exists can can be determined in time O(Km).
The resulting cut (X ′′ , X ′′ ) in Nt′′ induces a cut (X ′ , X ′ ) in Nt′ which in turn
induces a K-feasible cut (X, X) in Nt .
96 RECONFIGURABLE COMPUTING

Because the above procedure must be repeated for each node in the original
boolean network, we conclude that the labeling phase, i.e. computing the label
of all edges in the graph can be done in O(Kmn) where n is the number of
nodes and m the number of edges in N .

Node Mapping. In its second phase, the FlowMap algorithm maps the nodes
into K-LUTs. The algorithm works on a set L of node outputs to be imple-
mented in the LUTs. Each output will therefore be implemented as a LUT-
output.
Initially L contains all primary outputs. For each node v in L, it is assumed
that the minimum K-feasible cut (X, X) in Nv has been computed in the first
phase. A K-LUT LU T v is then generated to implement the function of v as
well as that of all nodes in X. The inputs of LU T v are the crossing edges from
X to X which are less than K, since the cut is K-feasible. L is then updated
to be (L − {v}) ∪ input(X). Those nodes w belonging to two different cut-set
X u and X v will be automatically duplicated. Algorithm 5 provides the pseudo
code that summarizes all the steps of the FlowMap presented here.
Improvements performed in the FlowMap algorithm have the goal of re-
ducing the amount of LUTs used, while keeping the delay minimal. The first
improvement is used during the maping of the nodes into the LUTs. The al-
gorithm tries to find the K-feasible cut with maximum volume. This allows
the final LUTs to contain more nodes, thus reducing the area used. The second
possibility is the used of the so called flow-pack method, which generalizes the
predecessor packing used in the DAG-Map [46]. The idea of predecessor pack-
ing is to pack a K-inputs LUT v in the same K-inputs LUT u, if v is a fan-out
free fan-in LUT of u and if the total amount the inputs of u and v is less than
K. The flow-pack method generalizes the predecessor packing method to the
set of predecessors Pu of u (including u) of the node u, provide that Pu has a
number of inputs less or equal to K.
Figure 3.18 illustrates the predecessor packing approach. The gate decom-
position presented in the Chortle algorithm can be used as well for the reduc-
tion of the LUT amount.

Figure 3.18. Improvement of the FlowMap algorithm through efficient predecessor packing
Implementation 97

Algorithm 5 Pseudo code of the FlowMap algorithm

/* phase 1: node labeling */
for each P I node v do
l(v) := 0
end for
T := list of non-PI nodes in topological order
while T is not empty do
remove the first node t from T
construct the network Nt
let p = max(l(u)) : u ∈ input(t)
transform Nt into Nt′ using collapsing
transform Nt′ into Nt′′ using node splitting
compute a cut (X ′′ , X ′′ ) in Nt′′ with e(X ′′ , X ′′ ) ≤ K
using the augmenting path method
if (X ′′ , X ′′ ) with e(X ′′ , X ′′ ) ≤ K is not found in Nt′′ then
Xt := {t}
l(t) := p + 1
else
Induce a cut (X, X) in Nt from the cut (X ′′ , X ′′ ) in N ′′ t
Xt := X
l(t) := p
end if
end while
/* phase 2: mapping nodes to LUTs */
L := list of PO nodes
while L contains non-PI nodes do
take a non-PI node v from L
generate a K-LUT LU T (v ′ ) to implement the function of v
such that input(v ′ ) = input(X)
L := (L − {v}) ∪ input(LU T (v ′ ))
end while

4. Conclusion
In this chapter, the general use of FPGAs in different systems was explained
as well as the different steps needed to synthesize the hardware modules that
will be executed in the FPGAs at run-time. We have taken a brief tour in
logic synthesis with the focus placed on LUT technology mapping, which is
the step by which FPGAs differ from other logic devices. Our goal was to
present two algorithms that best match the optimization requirements, namely
minimal delay and minimal area. Research in FPGA synthesis keep on going
and we several LUT-based technology mapping algorithms are expected to be
98 RECONFIGURABLE COMPUTING

developed in the future. The dynamic development in FPGAs introduce new

challenges for the synthesis tools, which has to deal with new elements like em-
bedded arithmetic modules, embedded memory, different clocking schemes,
etc... Therefore good technology mappers are required to cope with the on-
going innovations. Our goal in this chapter was to provide some background
to better understand the challenges and provide better solutions for the next
generations of FPGAs.
Chapter 4

HIGH-LEVEL SYNTHESIS FOR RECONFIGURABLE

DEVICES

One of the key points for the success of microprocessors is the ease in pro-
gramming such systems. This is in part due to the maturity of compilers as
well as the operation mode of microprocessors, the Von Neumann paradigm
that allows any sequential program to be executed on the underlying hardware.
In almost three decades of progress in compilers and microprocessors, a huge
amount of algorithms have been developed, coded and deployed in high-level
languages like FORTRAN, C, C++, Java, etc... Most of those existing pro-
grams can be executed with low modifications and low porting efforts on new
platforms. With the very attractive nature of software programming, a very
large community of programmers has grown, thus providing software code for
the most existing problem. The consequence of this development is that high
expectation in programmability is placed on new hardware platforms. A new
and highly innovative hardware computing platform, providing the best archi-
tectural organization to speed-up applications will certainly fail to be adopted,
if its programmability is poor. A less competitive hardware platform in turn
will certainly succeed if the portability of existing algorithms is shown to be
easy with a small increase in the computation time. This constat has led to the
development of languages and frameworks, which allows for the compiler to
be generated for a specified hardware description [168] [187].
While a certain degree of maturity has been reached in software program-
ming, or better said in sequential programming languages and compilers, par-
allel programming has not experienced the same advancements. The parallel
implementation of a given application requires two main steps. First the ap-
plication must be partitioned in order to identify the dependencies among the
different parts and set the parts that can be executed in parallel. This partition-
ing step requires a good knowledge on the structure of the application. After
partitioning, a mapping phase is required to allocate the independent blocks
100 RECONFIGURABLE COMPUTING

of the application to the computing hardware resources. For this purpose, a

good understanding of the underlying hardware structure is required. As con-
sequence, the community of programmers is not very enthusiastic whenever
parallel programming is concerned.
Reconfigurable computing also faces the disadvantages of parallel comput-
ing. In order to program most of the existing systems, one must write code in
a given hardware description language like VHDL or Verilog and some times
in a C-like language, which syntax is not so far from VHDL or Verilog. It is
therefore not easy to decouple the implementation process from the hardware.
The consequence is that FPGA programming is still considered a close field
for electrical engineers, despite all the benefits in performance and flexibility
that they provided.
High-level synthesis has been introduced for more than a decade as an
attempt to describe electronic systems using a high-level language that can
be compiled down to the hardware without manual intervention of the user.
Having such systems, software developers could be encouraged to develop
more applications for reconfigurable systems and existing applications could
be ported on such systems with low effort, thus increasing the popularity of
such systems.
High-level synthesis for reconfigurable devices is slightly different from that
of other electronic devices in that the resources on which part of the application
must be mapped are not fixed. This provides a greater flexibility during the
temporal mapping of code segments to computing blocks.
In this chapter, we present the high-level synthesis problem in general and
focus on the high-level synthesis for reconfigurable devices, also known as
temporal partitioning. Afterwards, we will investigate the algorithms that were
developed for this purpose.

1. Modelling
High-level synthesis (HLS) deals with the specification of a given applica-
tion at a very high level of abstraction as well as its implementation on a given
platform. The starting point is to specify the application in a given high-level
language or tool. For this modelling step, several possibilities exist for cap-
turing the behaviour of a systems, from the very simple Finite State Machines
(FSM) and their extensions like State charts, Control Dataflow Graphs, to very
complex tools like the Petri Nets, each of which has a different level of pow-
erfulness. We first present the dataflow graph that is used as model in this
chapter. We also consider two extensions of dataflow graphs, the sequencing
graphs and the finite state machine with datapath and explain why those two
models are not used in this context.
High-Level Synthesis for Reconfigurable Devices 101

1.1 Dataflow Graphs

A Dataflow Graph (DFG) provides a means to describe a computing task
in a streaming mode. Given a code segment in a high-level language like C
or C++, each operator represents a node in the dataflow graph. The inputs
of the nodes are the operand on which the corresponding operator is applied.
The output of a node represents the result of the operation on that node. The
output of a node can be used as input to other nodes, thus defining a data
dependency in the graph. Nodes that they depend on others are called sucessors
of the nodes on which they depend. Nodes on which other nodes depends
are the predecessors of the nodes that depend on. Some nodes in the graph
have no predecessors and others have no successors. Dataflow graphs might
be normalized by inserting in the graph two fake nodes, whose operation has
no effect on the dataflow computation. The first fake node is connected as
predecessor of all nodes without any predecessor and the second fake node is
connected as successor of all nodes that have no successors.

Example 4.1 As example of dataflow graph, consider the computation of

1
2
the quadratic root using the formula x = (b −4ac) −b
2
2a . The corresponding
dataflow graph is shown in figure 4.1. Data are streamed through the inputs
a, b, c, d, 4, 2 and the result is collected at the output of the graph.

Fromaly, we define a dataflow graph as follows:

Definition 4.2 (Dataflow Graph) Given a set of tasks T = {T1 , ...., Tk },

a dataflow graph (DFG) is a directed acyclic graph G = (V, E), where
V = T is the set of nodes representing operators and E is the set of edges.
An edge e = (vi , vj ) ∈ E is defined through the (data)dependency between
task Ti and task Tj .

We assume that for each task, an equivalent hardware implementation exists,

which occupies a rectangular area on the chip. Therefore, the nodes as well
as the edges in a DFG possess some characteristics like height, length, area,
latency and width, that are derived from the hardware resources used later to
implement those tasks. Those values are formally defined as follows:

Definition 4.3 (Latency, length, height, area, weight of nodes and edges)
Given a node vi ∈ V and its implementation Hvi as rectangular shape module
in hardware.
1 li denotes the lenght and hi the height of Hvi
2 ai = li × hi denotes the area Hvi
102 RECONFIGURABLE COMPUTING

Figure 4.1. Computational Graph for Quadratic Root

3 The latency ti of vi is the time it takes to compute the function of vi using

the module Hvi
4 For a given edge eij = (vi , vj ), which defines a data dependency between
vi and vj , we define the weight wij of eij as the width of bus connecting
two components Hvi and Hvj .
5 The latency tij of eij is the time needed to transmit data from Hvi to Hvj .

For sake of simplicity, we will just use the notation vi to denote a node of the
graph as well as its hardware implementation Hvi .
Any program written in a high-level language can be compiled into a dataflow
graph, provide that the program is free of loops and branching instructions.
This restriction does not match with the reality, since loops and branch instruc-
tions are available in the most of the programs, for the evaluation of branching
conditions at run-time, and to decide on the segment to be executed according
to the value of the condition variables. In case of non nested loops, the body of
a loop is always a set of instructions that can be represented using a dataflow
data structure. Several extension of dataflow graph exist to capture program
with control structures and loop. We will consider two of them in this chapter:
the sequencing graphs and the finite state machines with datapath.
High-Level Synthesis for Reconfigurable Devices 103

1.2 Sequencing Graph

A sequencing graph [156] is a hierarchical dataflow graph with two differ-
ent types of nodes: The operation nodes corresponding to normal "task nodes"
in a dataflow graph and the link nodes or branching nodes that point to another
sequencing graph in a lower level of the hierarchy. Linking nodes evaluate
conditional clauses. They are placed at the tail of alternative paths correspond-
ing to possible branches. Loops can be modeled by using a branching as a tail
of two paths, one for the exit from the loop and the other for the return to the
body of the loop, which is the sub sequencing graph associated with the link
node.

Example 4.4 Figure 4.2 shows an example of sequencing graph with a branch-
ing node BR containing two branching paths.
According to the conditions that node BR evaluates, one of the two sub-
sequencing (1 or 2) graphs can be activated. In order to implement a loop, only
one sub-sequencing graph in which the body of the loop is implemented, will be
considered. The node BR will then evaluate the exit condition and branch to the
next node of the hierarchy level, if the condition holds. Otherwise the body of
the loop is re-entered by reactivating the corresponding sub-sequencing graph.

Figure 4.2. Sequencing graph with a branching node linking to two different sub graphs

1.3 Finite State Machine with Datapath

Another extension can be done on a dataflow by integrating a finite state ma-
chine in the model, in order to control the execution on a datapath defined by
the dataflow graph. The result is the so called Finite State Machine with Data-
104 RECONFIGURABLE COMPUTING

path (FSMD)1 We adopt in this section the FSMD definitions and terminology
from Vahid et al. [203]. Thereafter, a finite Finite State Machine with Datapath
(FSMD) can be formally defined as a 7-tuple < S, I, O, V, F, H, s0 > where:
S = {s0 , s1 · · · , sl } is a set of states,
I = {i0 , i1 · · · , im } is a set of inputs,
O = {o0 , o1 · · · , on } is a set of outputs,
V = {v0 , v1 · · · , vn } is a set of variables,
F : S × I × V → S is a transition function that maps a tupel (states, input
variable, output variable) to a state,
H : S → O + V is an action function that maps the current state to output
and variable,
s0 is an initial state.
FSMD have some fundamental differences with traditional finite state ma-
chines. First, the transition function operates on arbitrary complex data type
like in high-level programming language; second the transition and action
functions may include arithmetic operations rather than just Boolean opera-
tions. The arithmetic operations and the complex data types implicitly define
a datapath structure in the specification.
The transformation of a program into a FSMD is done by transforming the
statements of the program into FSMD states. The statements are first classified
in three categories:
the assignment statements: For an assignment statement, a single state is
created that executes the assignment action. An arc connecting the so cre-
ated state with the state corresponding to the next program statement is
created.
the branch statements: For a branch statement, a condition state C and
a join state J both with no action are created. An arc is added from the
conditional state to the first statement of the branch. This branch is labeled
with the first branch condition. A second arc, labeled with the complement
of the first condition ANDed with the second branch condition is added
from the conditional state to the first statement of the branch. This process
is repeated until the last branch condition. Each state corresponding to the
last statement in a branch is then connected to the join state. The join state

1 In the original definition from Gasky[91], the extension is done on a finite state machine FSM in order

to support more complex data types and variables as well as complex operators.
High-Level Synthesis for Reconfigurable Devices 105

is finally connected to the state corresponding to the first statement after the
branch.

and the loop statements: For a loop statement, a condition state C and a join
state J, both with no action are created. An arc, labeled with the loop con-
dition and connecting the conditional state C with the state corresponding
to the first statement in the loop body is added to the FSMD. Accordingly
another arc is added from C to the state corresponding to the first statement
after the loop body. This arc is labelled with the complement of the loop
condition. Finally an edge is added from the state corresponding to the last
statement in the loop to the join state and another edge is added from the
join state back to the conditional state.
The transformation steps of a given program to a FSMD is illustrated in
figure 4.3.

Figure 4.3. Transformation of a sequential program into a FSMD

Example 4.5 Let us model the greatest common divisor GCD of two num-
bers, using an FSMD for as, explained [203]. The sequential version of the
GCD is given in algorithm 6 and the corresponding FSMD is shown in figure
4.4.
The loop state (C1) and the branching state within the loop are white and
other states are grey. Also we labeled the states with the action to be taken in
106 RECONFIGURABLE COMPUTING

Algorithm 6 The greatest common divisor sequential algorithm

1: variable a, b, gcd: integer;
2: done := FALSE;
3: while (!done) do
4: if (a > b) then
5: a := a - b;
6: else
7: if (b > a) then
8: b := b - a;
9: else
10: done = TRUE;
11: end if
12: end if
13: end while
14: gcd := a;

those states. An additional label corresponding to the line of the instruction in

the program is placed on each state.
After the program transformation into an FSMD, a datapath must be created
that will be controlled by a finite state machine derived from the FSMD. The
process of creating the datapath is straightforward. First a register must be
instantiated for each variable in the program. An output port implicitly declares
a variable for which a register must be created. In the second step, a functional
unit will be created for each arithmetic operation in the state diagram. The
third step consists of connecting the ports of the functional unit with those of
the variable. In the fourth step, a unique identifier is created for each control
input and output of the functional units in the datapath. Sharing the operator
can be done using multiplexers.
Figure 4.5 shows the datapath of the greatest common divisor program to-
gether with control finite state machine resulting from the FSMD. The control
signals are LDA and LDB to load the two registers A and B with the values
coming from the subtractor, the SELL and SELR to select the corresponding
value from the multiplexer. Those signals are controlled by the FSM that set
them according to its current state. The status signals AGTB and ALTB are
used by the FSM to decide about the next state to move to, depending on the
value of the comparison between A and B.
Having specified a system using one of the modelling tools previously pre-
sented, the next step will consists on the compilation of the specification into
a set of hardware components. This compilation step, also known as synthesis
is usually done in three different steps. The first one is the allocation, which
defines the type of resources required by the design, and for each type the
High-Level Synthesis for Reconfigurable Devices 107

Figure 4.4. Transformation of the greatest common divisor program into a FSMD

number of instances. In the GCD case, the types of resources needed are com-
parators, subtractors, registers and multiplexers. Two registers are instantiated
and two multiplexers are instantiated while only one subtractor and one com-
parator are used. Instead of using just one subtractor and two multiplexers,
we could choose to use two subtractors and no multiplexers. The next step
after the allocation is the binding. In this step, each operation is mapped to
an instance of a given resource. Since many operators can be mapped to the
same instance of a resource, a schedule is used in the third step to decide on
which operator should be assigned a given resource at a given period of time.
Formally, allocation binding and scheduling can be defined as follow:

Definition 4.6 (Allocation) For a given specification with a set of op-

erators or task T = {t1 , t2 , · · · , tn } to be implemented on a set of resource
types R = {r1 , r2 , · · · , rt }. An Allocation is a function α : R → + , where
α(r) = zi denotes the number of available instances of resource type ri
108 RECONFIGURABLE COMPUTING

Figure 4.5. The datapath and the corresponding FSM for the GCD-FSMD

Definition 4.7 (Binding) For a given specification with a set of opera-

tors or task T = {t1 , t2 , · · · , tn } to be implemented on a set of resource types
R = {r1 , r2 , · · · , rt }. A binding is a function β : T → R × + , where
β(ti ) = (ri , bi ), (1 ≤ bi ≤ α(ri )) denotes the instance of the resource type ri
on which ti is mapped to.

Definition 4.8 (Schedule) For a given specification with a set of oper-

ators T = {t1 , t2 , · · · , tn }, a schedule is a function ς : V → + , where ς(ti )
denotes the starting time of the tast ti .

1.4 Fundamental differences in HLS for reconfigurable

computing
The high-level synthesis for reconfigurable computing differs from that of
other architectures2 in two main points. First the fundamental problems that
must be solved in common high-level synthesis are the binding and scheduling.
The allocation is usually not necessary, in particular in those systems where the
resource types and their numbers are already fixed. The binding will just map
operator to resources and the scheduling will decide which operator own the

2 Also known as architectural synthesis

High-Level Synthesis for Reconfigurable Devices 109

resource at a given time. This is not the case in reconfigurable computing,

where the architectural resources are created on the reconfigurable device ac-
cording to the resource types needed to map the operators at a given time. As
long as basic computing resources are available on the device, new resources
can be built. With the uniformity of resources available on a reconfigurable
device, it is possible to implement any task on a given part of a reconfigurable
device, provide that the available resource are enough. The problem of binding
an operator to a given resource then becomes an area assignment problem.

Example 4.9 Let’s illustrate this major difference with using dataflow graph
for the computation of the functions x = ((a×b)−(c×d))+((c×d)−(e−f ))
and y = ((c × d) − (e − f )) − ((e − f ) + (g − h)) as shown on figure 4.6.

Figure 4.6. Dataflow graph of the functions: x = ((a × b) − (c × d)) + ((c × d) − (e − f ))

and y = ((c × d) − (e − f )) − ((e − f ) + (g − h))

We consider the implementation on a common architecture. Assume an al-

location that selects one instance of each resource type (multiplier, adder and
subtractor). Consider that the multiplication needs 100 basic resource units.
This can be 100 NAND-Gate for example, but we would like to express the ba-
sic unit of computation in look-up table (LUT). The multiplier therefore needs
100 LUTs, the adder and the subtractor need 50 LUTs each. Because only
one instance for each resource type is available, the two subtractions in the
first level of the graph corresponding to the nodes 3 and 4 cannot be executed
110 RECONFIGURABLE COMPUTING

in the same step, although enough basic resources are available. But, those
resources were used to implement one adder and one subtractor instead of two
subtractors. The adder cannot be used in the first step, because it depends on
a subtractor that must be first executed. The minimum execution delay is four
steps if we use chaining3 in the schedule (figure 4.7).

Figure 4.7. HLS of the graph in figure 4.6 on a an architecture with one instances of the
resource types +, ∗ and −

Consider now the implementation on an reconfigurable device having the

same amount of basic resources, namely 200 LUTs. Because the basic re-
sources are not bounded to any operator, they can be configured according to
our need. We are therefore able to map one multiplier and the two subtractors
in the first step. In the second step, we map one multiplier, one adder and one
subtractor. In the third step, the rest of operators are map to onto the basic
resources. The resulting implementation is shown on figure 4.8 This imple-
mentation is one step faster than the previous one, although they use the same
amount of resource.

3 Chaining operations means sequentially executing a set of operations in the same time slot, provided

that their overall delay is less that the clock delay

High-Level Synthesis for Reconfigurable Devices 111

Figure 4.8. HLS of the graph in figure 4.6 on a reconfigurable device

The second major difference concerns the control of the computation steps
of a given application. In general high-level synthesis, the application is spec-
ified using a structure that encapsulates a datapath (set of computational re-
sources and their interconnections) and a control part. The synthesis process,
then allocates the resources to operators at different time according to a com-
puted schedule. This temporal resource assignment is controlled by a separate
part of the system for which a synthesis is required. In reconfigurable devices,
a set of hardware modules implemented as datapath normally compete for exe-
cution on the chip. Instead of a separate control modules to set the control lines
of the datapath modules, a processor is used to control the selection process of
the hardware modules by means of reconfiguration. The same processor is
also in charge of activating the single resources in the corresponding hardware
accelerators. For the modelling of the datapath, a dataflow graph is usually
sufficient, since loops and branch control is left to the processor. While the
general high-level synthesis has to control each single operator, the high-level
synthesis for reconfigurable devices deals with a set of operators with differ-
ent execution delay at a time. This means that for a given application, the
resources on the device are not allocated to only one operator, but to a set of
operators that must be placed at the same time and remove at the same time.
With this, an application must be partitioned in sets of operators. The parti-
tions will then be successively implemented at different time on the device.
This process, called temporal partitioning allows an application to be sequen-
tially computed, by allowing a temporal sharing of resource among different
112 RECONFIGURABLE COMPUTING

sets of operators. We next present the temporal partitioning problem and some
of the solution approaches developed to solve it.

1.5 Temporal Partitioning

We assume that for each task ti ∈ T = {t1 , t2 , · · · , tn } a core4 Ci is avail-
able in a given library. This can be a simple description of the task ti in a
hardware description language or the final implementation of the task, i.e. the
final partial bitstream for the given task. Recall that the implementation of a
given task for a given device describes all the resources used to implement the
given task in the device.
In temporal partitioning, complete partitions are downloaded onto the de-
vice rather than just a single component. Whenever a new partition is down-
loaded, it is assigned all the resources onto the device and the partition that was
running is destroyed. Two possibilities exist to build the implementation of a
given partition. The first one is to compose the partition from the implemen-
tations of the single components in the partition. This exercise requires some
knowledge about the structure of the bitstream in order to place the component
at the correct locations and perform the right interconnections. The second
possibility is to instantiate the HDL description of each single component in
the HDL description of the partition and let the compiler do the rest of the job.
Preserving the boundary of each component is usually not required. Therefore,
the complete design is usually flattened and optimized across the component
boundaries before the technology mapping, the placement and routing phases.
Each task is characterized by its latency, the height, the length and area of
its bounding box. Those values are estimated values on the basis of the im-
plementation of the task in the region defined by its bounding box. It may not
be the same in the final implementation if the complete partition design is first
flattened before optimization. The flattening of a design before optimization
provides more room for optimization. The resulting design is usually smaller
than the non flattened. However, placement algorithms tend to place compo-
nents belonging to the same group in the same region. If some modules in
a given component are tied to other modules in another component, then the
placer will tend to placed tied components in the same group. This may result
in an increase in the latency of the two involved components. We next provide
some definitions before stating the temporal partitioning problem.
Definition 4.10 (Configuration) Given a reconfigurable processing
unit H and a set of tasks T = {t1 , ...., tn } available as cores C = {c1 , ...., cn },
we define the configuration ζi of the RPU at time ti to be the set cores {ci1 , ..., cik } ⊆
C running in H at time ti . We set ζi = {ci1 , ..., cik }

4 Hardware module.
High-Level Synthesis for Reconfigurable Devices 113

We consider a model in which the execution control is left to the processor

that is also in charge of controlling the reconfiguration process on the chip.
We therefore use the dataflow model as entry point for the part of the design
to be implemented in the reconfigurable device. Since the control loops are
no more available in the design, cycles can be only available in the dataflow
graph, if the outputs of some components are used as input to some of their
predecessors. This is for instance the case with the GCD-datapth in figure 4.5,
where the output of the subtractor is used as inputs for the two registers a and
b. In this case, the part of the datapath containing the loop is encapsulated into
a single component that will be executed for a given number of clock cycles
fixed by the processor. We can also imagine a coarse-grained structure in which
a smaller controller supervises the execution process until completion. Upon
completion of the execution, status signals are used to notify the processor.
If a coarse-grained element containing loops fits into the device, then it may
be entirely placed into a partition with other component. However, we may
also face the situation, where a coarse-grained element containing a loop is
to large to fit onto the device. In this case, we will partition only this coarse
grained element in parts that will be sequentially implemented in the chip. The
number of iterations will then be done in the partitioned version as required
in the original version. This approach is illustrated in figure 4.9, where the
coarse-grained node C with a loop in its implementation is partitioned in two
part P1 and P2 . P1 will be downloaded before P2 in one loop iteration.

Figure 4.9. Partitioning of a coarse-grained node in the dataflow graph

The rest of the chapter is based on the assumption that an underlying dataflow
graph model is available. We therefore redefine the schedule on the basis of
the dataflow graph.

Definition 4.11 (Schedule and Ordering Relation) Given a DFG

G = (V = {v1 , v2 , · · · , vn }, E),
114 RECONFIGURABLE COMPUTING

a schedule is a function ς : V → + , where ς(vi ) denotes the starting time

of the node vi that implements a task ti .
A schedule ς is feasible if: ∀eij = (vi , vj ) ∈ E, such that eij defines a data
dependency between tasks ti and tj , ς(tj ) ≥ ς(ti ) = ti +tij (tij denotes the
latency of the edge eij and ti is the time it takes to the node vi to complete
execution).
We define an ordering relation ≤ among the nodes of G, such that: vi ≤
vj ⇐⇒ ∀ schedule ς, ς(vi ) ≤ ς(vj ).
≤ is a partial ordering, since it is not defined for all pairs of nodes in G.
Definition 4.12 (Partition) A partition P of the graph S G = (V, E) is
its division into some disjoint subsets P1 , ..., Pm such that m k=1 Pk = V
A partition is feasible in accordance to a reconfigurable device H with area
a(H) and pin count p(H) if:
P
∀Pk ⊆ P : a(Pk ) = ( vi ∈Pk ai ) ≤ a(H)
P
1/2( {eij ∈E:(eij ∩Pk )6=∅ and (eij −Pk )6=∅} (wij )) ≤ p(H)
The run-time t(Pi ) of a partition Pi is the time it takes to complete the exe-
cution of that partition. It its the maximum time from the input of the data to
the output of the result. The definition of the run-time is limited to only one
execution step of the complete partition. If a partition is use in a loop, then we
must multiply the run-time of the partition by the number of iterations.
Each partitioning is submitted to the restriction that the total amount of re-
sources assigned to it, i.e. the sum of the resources of all its components, does
not exceed the available amount of resources. The total area of a partition must
therefore be smaller than the device area and the weighted sum of all crossing
edges5 is less than the number of terminals or pins. This restriction on the
crossing edges can be relaxed by multiplexing the use of input output pins in
order to allow a design with a greater amount of pins to communicate with the
external world.
Definition 4.13 (Ordered Partitions) We can extend the ordering
relation ≤ to partitions as follows: Pi ≤ Pj ⇐⇒ ∀vi ∈ Pi , ∀vj ∈ Pj ,
either vi ≤ vj or vi and vj are not in relation. Thereafter, a partitioning P is
ordered ⇐⇒ an ordering relation ≤ exists on P .
An ordered partitioning is characterized by the fact that for a pair of par-
titions, one can always be implemented after the other with respect to any
scheduling relation.

5 A crossing edge is an edge which a connect one component in a partition with another component out

of the partition
High-Level Synthesis for Reconfigurable Devices 115

Definition 4.14 (Temporal Partitioning) Given a dataflow graph

G = (V, E) and a reconfigurable device H, a temporal partitioning of G on
H is an ordered partitioning P of G with respect to the reconfigurable device
H.
Applying a temporal partitioning on a dataflow graph generates another
graph that we call configuration graph. With the necessary controll part of
an application, a final schedule must be applied on the configuration graph in
order to compute the original application. Formally, we define a configuration
graph as follows:

Definition 4.15 (Configuration Graph) Given a DFG G = (V, E)

and a temporal partitioning P = {P1 , ..., Pn } of G, we define a Configuration
graph of G relative to the partition P , with notation Γ(G/P ) to be the graph
Γ(G/P ) = (P, E P ) in which the nodes are partitions in P . An edge eP =
(Pi , Pj ) ∈ E P ⇐⇒ ∃e = (vi , vj ) ∈ E with vi ∈ Pi and vj ∈ Pj .

For a given partition P , each node Pi ∈ P has an associated configuration ζi

that is the implementation of Pi for the given device H. Because two partitions
are never running at the same time, communication data between the partitions
must be stored in a memory external to the device. We call this memory that is
used for communication between partitions the communication memory. Phys-
ically, the communication memory might not be on a different chip than the
main system memory. It can be a reserved part of the main memory.
The computation steps for two partitions Pi and Pj such that Pi ≤ Pj is
done as follows: The corresponding configuration for Pi is first downloaded
into the device. Pi copies all the data it needs to send to other partitions into
the communication memory. Upon completion, the device is reconfigured to
implement the partition Pj , which can then access the communication mem-
ory and collect the data that was sent to it. By inter configuration registers,
we denote those registers that are usually used at the boundary of the device
to hold the input and output values. Inter configuration registers are usually
mapped into a processor address space that can be use by the processor for
communication with the device. An example of configuration graph with in-
terconfiguration registers is given in figure 4.10.
According to the goal to be reached, several objective functions can be de-
fined for the temporal partitioning problem. One objective could be the mini-
mization of the number of partitions in order to reduce the overall reconfigu-
ration overhead. Another objective could be the minimization of the compu-
tation time. This can be expressed for example through the minimization of
the maximum computational delay across all the partitions. A third objective
could be the minimization of the overall amount of wasted resources on the
chip. One of the major drawbacks of temporal partitioning is the non efficient
116 RECONFIGURABLE COMPUTING

Figure 4.10. Example of configuration graph

use of the device area. Temporal partitioning normally targets non-partial re-
configuration devices. It has the advantage that the resulting partition can be
implemented in just one circuit. Therefore, floorplanning efforts are left to the
synthesis tools. However, when components with shorter run-times are placed
in the same partition with other components with longer run-time, those with
the shorter components remains idle for a longer period of time, resulting in a
waste of the device resources. The wasted resource of a given partition can be
formally defined as follows:

Definition 4.16 (Wasted Resource) Given a dataflow graph G = (V, E)

and a temporal partitioning P = {P1 , ..., Pk } of G:

The wasted resource wr(vi ) of a node vi is the unused area occupied by

the node vi during the computation of a partition. It is the area occupied
by the node times the idle time of the node.
wr(vi ) = (t(Pi ) − ti ) × ai , where t(Pi ) is the run-time of the partition. The
idle time t(Pi ) − ti of vi is the run-time of the partition minus the running
time of the component.

The wasted resource wr(Pi ) of a partition Pi = {vi1 , .., vin } is defined as

the sum of P
wasted resources of its components:
wr(Pi ) = nj=1 (t(Pi ) − (tij ) × aj

The wasted resource of a partitioning P is the sum of the wasted resource

of all it’s partitions.
P
wr(P ) = ki=1 wr(Pi )
High-Level Synthesis for Reconfigurable Devices 117

Figure 4.11 graphically illustrates the waste resource of a partition. The run-
time of the partition is determined by the component v1 . The shaded area
defines the overall wasted resource of the partition.

Figure 4.11. Wasted resources

The minimization of the overall run-time is difficult to capture in an equa-

tion, since it can be affected by many factors. The reduction of the number of
partition automatically reduce the overall configuration time, but not necessary
the overall computation time. If the configuration time is smaller than the com-
putation time of some components, an additional packing of those components
in a partition will reduce the number of partition on one side. However, the
run-time of the partitions in which the components were packed will increase
by a factor greater than the reconfiguration time. Fortunately, this is not the
case with the existing technology, where the reconfiguration time is usually
higher than the computation time of single components.
The execution time does not only depend on the partition, but also on the
processor used for reconfiguration and the speed of data exchange. Recall that
fully reconfigurable devices are usually coupled with a processor, which man-
ages the complete system. The processor reconfigures the device and manages
the data exchange. One of the most important goal to considered in temporal
partitioning is the communication overhead. This goal can be capture in sev-
eral ways. We present here a simple model to capturing the communication
cost during temporal partitioning: the graph connectivity.

Definition 4.17 (Graph Connectivity) Given a dataflow graph G =

(V, E), we define the connectivity, con(G) = |V2×|E|
|2 −|V |
, of G as the rapport be-
118 RECONFIGURABLE COMPUTING

tween the number of edges in E over the number of all edges which can be
built with the nodes of G.
For a given subset V ′ of V , the connectivity of V ′ is defined as the relation
between the number of edges connecting the nodes of V ′ over the set of all
edges which can be built with the nodes of V ′ .
The connectivity of a set provides a means to measure how strongly the
components of a set are connected. High connectivity means a strongly con-
nected set, while low connectivity reflects a graph in which many modules are
not connected together. The connectivity may be used to define how good a
partitioning algorithm, whose goal is the minimization of the communication
cost, has performed. We formally define the quality of a partitioning as fol-
lows:
Definition 4.18 (Quality) Given a dataflow graph G = (V, E) and
Ppartitioning
a P = {P1 , ..., Pn } of G, we define the quality Q(P ) = n1 ×
n
i=1 (con(Pi )) of P as the average connectivity over all the partitions Pi , 1 ≤
i≤n.
Most of the target architectures for which temporal partitioning is used are
made upon a reconfigurable device connected to a host processor by a bus
like, for instance the PCI6 . The number of bus lines is limited. Therefore, the
communication has to be time-multiplexed on the bus, if many data have to
be transported on the bus. This happens for example when a partition has to
be replaced. Recall that the communication between the partitions is done by
a set of registers inside the reconfigurable devices. All the temporary data in
those registers have to be saved in the communication memory before recon-
figuration. The device is then reconfigured and the data are copied back into
the reconfigurable device registers. Minimizing the communication overhead
can be done by minimizing the weighted sum of crossing edges among the
partitions. This will also minimize the set of registers needed to communicate
between the generated partitions, thus reducing on one side the size of of the
communication memory and the communication time on the other side. This
goal is likely to be reached if highly connected components are placed in the
same partition.
After a partitioning by a given algorithm, the quality of the partition will
determine if the algorithm performed well or not. If a graph is highly connected
and the partitioning algorithm performs with low quality, then there will be
more edges connecting different partitions and therefore more data exchange
among the partitions. But if the graph is highly connected and the partitioning
algorithm performs with high quality, then the components in the partitions

6 Peripheral Component Interconnect

High-Level Synthesis for Reconfigurable Devices 119

are highly connected and therefore there will be fewer edges connecting the
partitions.
Example 4.19 Figures 4.12 and 4.13 illustrate the connectivity of a graph
and the quality of the algorithm that was used for the partitioning. In figure
4.12 a graph with a connectivity of 0.24 is partitioned by a first algorithm,
which produces a quality of 0.25. The same graph is partitioned by another
algorithm in figure 4.13 with a quality of 0.45. In the first case, we have 6
edges connecting the two partitions, while there are only two edges connecting
the partitions in the second case. The second case is, therefore, better than the
first case for data communication between the partitions.

G: P0 P1
8
1 2 2
3
1
3 7 10
9 7
4 8 6
4 6
5 9
10
5

Figure 4.12. Partitioning of the graph G with connectivity 0.24 with an algorithm that pro-
duces a quality of 0.25

P0 P1
G:
8 8
1 2 1 2

3 7 3 7
9 4 9
4 6 6
5
10 10
5

Figure 4.13. Partitioning of the graph G with connectivity 0.24 with an algorithm that pro-
duces a quality of 0.45

In the next section, we present contributions of authors who developed var-

ious techniques and methodologies for temporal partitioning. The methods
120 RECONFIGURABLE COMPUTING

can be grouped in four different categories. In the first category are the list-
scheduling based methods. The second category encounters exact methods,
which use integer linear programming equations for capturing the optimiza-
tion problem. The solution of the equations provides the exact solution of the
temporal partitioning problem. The third category covers network flow based
algorithms while in the fourth category, we have spectral method based on
the computation of eigenvalues of a matrix derived from the dataflow graph.
The methods in the first, third and fourth categories can also be grouped un-
der the umbrella of recursive bipartition, because the partitions are iteratively
generated through the bipartition of a remaining set of components.

2. Temporal Partitioning Algorithms

Several heuristics like the list scheduling algorithm use the mobility range of
components to perform further optimizations. This range can be determined on
the basis of the earliest and the latest starting time of the corresponding node
resulting from the solution of the unconstrained scheduling using the ASAP
and ALAP methods. We therefore first present the unconstrained scheduling
before focussing on the temporal partitioning methods.

2.1 Unconstrained Scheduling

In general high-level synthesis, unconstrained scheduling [156] is the allo-
cation of starting time to components under the assumption that an unlimited
amount of resources is available. In the case of reconfigurable devices, this
corresponds to placing modules in a device with unlimited size that allows any
partition to be implemented. Because the devices in reality have only a limited
amount of resources, unconstrained scheduling cannot be used as such. Instead
it is usually used as pre-processing step for other algorithms. It can be used
for instance for the computation of the upper and lower bounds on the starting
time of operations in a dataflow graph. While the lower bound provides the
earliest time at which a module can be scheduled, the upper bound defines the
latest time at which a module can be started. The difference between the upper
and the lower bound of a module is its mobility range. The mobility interval
is given by [ASAP (v) ALAP (v)], where ASAP (v) resp. ALAP (v) is the
starting time of v computed with the ASAP resp. ALAP-algorithm, that we
present later in this section. The starting time of a node cannot be out of its
mobility range.
In order to compute the earliest starting time of each component, the ASAP
(As Soon As Possible)-Algorithm can be used. The ASAP-Algorithm traverses
the dataflow in topological order starting from the primary inputs. Each pri-
mary inputs is assigned the starting time 0. A node vi is processed only when
all its predecessors have been processed. The starting time of the node vi is the
High-Level Synthesis for Reconfigurable Devices 121

highest end time among all its predecessors. The ASAP algorithm idealizes
the binding process by assuming an unlimited amount of available resource,
and assigns each operation as soon as of is ready to be executed, therefore pro-
viding the lowest starting time of the tasks in the graphs. The pseudo code for
the ASAP-algorithm is provided in algorithm 7 for the resulting schedule ς.

Algorithm 7 ASAP-Algorithm for unconstrained scheduling

1: for each node v ∈ V do
2: if v has no predecessors then
3: ς(v) := 0
4: V := V − v
5: end if
6: end for
7: while V 6= ∅ do
8: select a vertex vi ∈ V whose predecessors are all scheduled
9: schedule vi by setting ς(vi ) := max(vj ,vi )∈E (ς(vj ) + tj )
10: V := V − vi
11: end while

Example 4.20 Assuming a latency of 100 clocks for the multiplication and
50 clocks for the addition as well as for the subtraction, an example of ASAP-
scheduling is provided in figure 4.14. The nodes are labeled with their number
as well as their starting time as computed by the algorithm. The label on
the edges represent the delay computation delay of the previous node. The
data transmission delay is neglected. The number on the nodes represent their
starting times.

Contrary to the ASAP-algorithm, the ALAP (As Late As Possible)-algorithm

traverses the graph from the primary outputs to the primary inputs. A node is
processed only if all its successors have been processed. The primary outputs
are assigned the maximum starting time, which is an upper bound on the com-
putation time of the dataflow graph. In the simple case, this upper bound is the
length (latest starting time - earliest starting time) of the schedule computed by
the ASAP-algorithm. The starting time of the node vi is assigned the minimum
value among all the staring time of its predecessor minus the latency of vi . Al-
gorithm 8 provides the pseudo code of the ALAP-algorithm for the resulting
schedule ς and an upper bound λ on the overall computation time.

Example 4.21 With the previous assumption on the module latencies, we

schedule the example of figure 4.14 using the ALAP-algorithm with upper
bound, thus providing the result of figure 4.15.
122 RECONFIGURABLE COMPUTING

Figure 4.14. Scheduling example with the ASAP-algorithm

Algorithm 8 ALAP-Algorithm for unconstrained scheduling

1: for each node v ∈ V do
2: if v has no successors then
3: ς(v) := λ
4: V := V − v
5: end if
6: end for
7: while V 6= ∅ do
8: select a vertex vi ∈ V whose successors are all scheduled
9: schedule vi by setting ς(vi ) := min(vi ,vj )∈E (ς(vj ) − ti )
10: V := V − v
11: end while

The constrained scheduling problem considerers the restriction on the ar-

chitecture. If there are many tasks competing for a given resource at a given
time, one of them must be chosen according to a given criteria and the rest
will be scheduled later. Sometimes, a task which could be executed at a given
time cannot be run because the resource type needed to implement that task is
not available at that time. This may delay the execution of the complete set of
remaining tasks, if the rest of the tasks depends on that task.
High-Level Synthesis for Reconfigurable Devices 123

Figure 4.15. ALAP Scheduling example

2.2 The List Scheduling Approach

List Scheduling. One of the most used approaches used to solve the temporal
partitioning problem is the list scheduling (LS) method. List scheduling was
first used as method for microcode compaction in order to generate efficient
microcode from high-level languages [65]. The idea behind list scheduling
in architectural synthesis is to first sort the nodes in topological order and then
assign some priority to the node of a dataflow graph. There are several methods
to assign a priority to a node. A common strategy is to use the latency weighted
depth of the node as its priority [96][143]. The depth of a node v is defined
as the length (number of nodes) of the longest path from an input to v. The
latency weithed depth is the same as the latency depth, however the path from
the input to v are weighted using the latency of the operation to be executed
by the nodes on the path. Also the mobility of a given node, i.e. the difference
between its ALAP-value and its ASAP-value, can be used as its priority.
At any time step t, the so called ready set, that is the set of operations ready
to be be scheduled, is constructed. The ready set contains operations whose
predecessors have already been scheduled early enough to complete their ex-
ecution at time t. The algorithm checks if there are enough resources of a
given type k to implement all the operations of type k. If so, the operations
are assigned the resources, otherwise, higher priority tasks are assigned the
available resources and the rest of the operations will be scheduled later, when
some resources will be available. If the mobility of a node is used as priority
124 RECONFIGURABLE COMPUTING

criteria, it is possible that all operators in the ready list are on a critical paths,
which means that their mobility is zero. As consequence, the complete depth
of each operators is increased by one, thus increasing the latency of the graph’s
execution.
Example 4.22 We consider the graph of figure 4.16 on the left side, in which
each node is labelled with its priority. The nodes must be scheduled on a
resource set made upon an adder and a multiplier. With the priority of a node
defined as its depth, the list scheduling is shown on the right side.

Figure 4.16. An example of list scheduling using the depth of a node as priority

Forced-Directed List Scheduling. The Force-Directed List Scheduling (FDLS)

from Paulin and Knight [174] is an extension of the list-scheduling algorithm.
Instead of computing the priority for each node at the beginning, priorities are
dynamically computed at each step. Force-directed list scheduling is a resource
constrained scheduling method aimed at finding a minimum latency schedule
for a given resource set. The concept of force is used to select the operation
to be assigned a given resource. First the possible time frame, i.e. the mobil-
ity interval of each node is computed using the ALAP and ASAP scheduling.
Next, the probability p(v, t) of a node v to be scheduled at step t is computed
High-Level Synthesis for Reconfigurable Devices 125

for each node v of the graph. The probability p(v, t) is zero outside the mo-
bility interval of v and equal the reciprocal of the mobility within the mobility
interval. Formally, we have:
(
1
if t ∈ [ASAP (v) ALAP (v)]
f (n) = mob(v)+1 (2.1)
0 otherwise
where mob(v) = ALAP (v) − ASAP (v).
A distribution function d(t, k) is calculated as the sum of probabilities of all
the operations with a resource type k. Formally, we have:
X
d(t, k) = p(v, t) (2.2)
{v∈V,α(v)=k}

This can be plotted into a graph called the distribution graph that indicates
the concurrency of similar operations over the schedule steps. Whenever many
operators compete for a fewer amount of resources, the operations that produce
a global increase of concurrency in the graph are selected and assigned the
resources.

Example 4.23 Consider once again the graph of figure 4.6 whose ASAP
and ALAP time is computed in figures 4.14 and 4.15. Nodes v1 , v2 , v5 , and
v8 have each a mobility of zero. Therefore p(v1 , i) = p(v2 , i) = p(v5 , i) =
p(v8 , i) = 0, ∀i ∈ [0 , 200].

As stated earlier, the scheduling is done on the basis of the force concept,
whose role is to attract or repel operators in that scheduling step. Two types
of forces are defined: the self-force and the predecessor-successor force. The
self-force is the one that relate an operation to different control steps in which
it can be scheduled. For a given node v the self-force is formally defined as
follows:
ALAP (v)
1 X
self-force(v, t) = d(t, k) − d(m, k) (2.3)
mob(v) + 1
m=ASAP (v)

The selection of an operator influences the mobility interval of other opera-

tors. This influence can be expressed as a force, called predecessor-successors
force that we will not present in detail here. The predecessor-successor force
is added to the self-force to build the total forces that are used for the selection
process. More details on the computation of forces can be found in [174].

List Scheduling for Reconfigurable Devices. In high-level synthesis for re-

configurable devices, the resource type does not play a big role. Only the total
126 RECONFIGURABLE COMPUTING

amount of basic resources is important. We don’t have operators competing

for a resource type but for a given surface on the device. Furthermore, we
are interested in building complete partitions describing a set of computational
steps that will be performed at the same time. Despite the fact that any com-
ponent has a different starting time and a different end time, only the starting
time and the end time of the complete partition is usually considered. Several
authors [173],[41], [171], [45], [202], [181] have used this fact to adapt known
architectural synthesis algorithm to reconfigurable devices.
The list-scheduling algorithm in reconfigurable devices works in the same
way as the common list-scheduling algorithm, by building and updating a list
of ready operators. Components are then removed from the ready list and
assigned to partitions. The only assignment criterion is that there should be
enough places left on the device to accommodate the new component. If this
is the case, then the component is placed on the partition currently being built.
Otherwise, a new partition is built and the process is repeated until all the
nodes of the graph are placed in partitions. The pseudo code for this approach
is provided in algorithm 9

Algorithm 9 List-scheduling scheduling algorithm for reconfigurable devices

sort the nodes of v according to their priorities
P0 := ∅
while V 6= ∅ do
select a vertex v ∈ V with highest priority and whose predecessors are all
placed
if (a partition Pi exists with s(Pi ) + s(v) ≤ s(H)) then
Pi = Pi ∪ {v}
else
create a new partition Pi+1 and set Pi+1 = {v}
end if
end while

Because the method is an iterative heuristic, several improvements can be

done during the construction in order to generate optimal solutions.
Optimization usually considers the minimization of the overall computation
latency of the dataflow graph. The two main factors that influence the overall
latency are the latency of each segment and the configuration overhead. If k
segments are generated by an algorithm, then the overall computation time can
be formally written as shown in equation 2.4.

n
X
tDF G = k × CH + (tPi ) (2.4)
i=1
High-Level Synthesis for Reconfigurable Devices 127

where tDF G is the overall computation latency for a dataflow graph DF G, CH

is the reconfiguration time of the device H and tPi is the computation time of
partition i.
The minimization of the overall computation time of the given function can
be done by tuning with the different parameters in equation 2.4. If the re-
configuration time of the device is too big, then the optimization will tend to
minimize the number of partitions in order to avoid a lot for reconfigurations.
However, if the reconfiguration time can be neglected, then only the delay in
partitions will be of great interest and a good algorithm will tend to avoid long
paths in partitions.
The main advantage of the list scheduling method is it’s linear run-time in
the number of nodes in the graph. Furthermore the method allows for local
optimizations while selecting the nodes to be placed in partitions. This fact
is exploited in [173] and [30] to compute a minimum set of resources used
by consecutive partitions. In [202], a pair wise interchange is done between
adjacent segments in order to minimize a cost function defined as the sum of
squares of the number of nets in each segments.
List scheduling has a major drawback called the levelization. Levelization
means that the partitions are built on the basis of the level number of the com-
ponents in the dataflow graph. Modules are assigned to partitions based more
on their level number, rather than their interconnectivity with other compo-
nents already placed in the partition. This may have a bad effect on the quality
of the partitioning as result of a bad connectivity inside the partitions and be-
tween the partitions. With this, the goal of minimizing the number of nets
connecting two different partitions, i.e. the minimization of the data exchange
among partitions becomes difficult to reach.

Example 4.24 Consider the circuit of figure 4.17 partitioned by a list-scheduling

algorithm that produces three partitions. While the components are less con-
nected inside the partitions, the partitions have more edges among each other,
thus leading to a poor quality. A better partitioning is shown in figure 4.18
for the same graph. The connectivity is better preserved inside and outside the
partitions.

Despite its drawback in connectivity, the list scheduling algorithm remains,

thanks its linear computation time, a good temporal partitioning candidate, in
particular for low-connected graphs. Also, it can be used to construct initial
solutions that can be improved by more complex iterative improvement opti-
mization heuristics.
128 RECONFIGURABLE COMPUTING

Figure 4.17. Levelizing effect on the list-scheduling on a dataflow graph

Figure 4.18. Partitioning with a better quality than the list-sceduling

High-Level Synthesis for Reconfigurable Devices 129

2.3 Integer Linear Programming

As defined in [49], an instance of an Integer Linear Programming (ILP)
problem is defined through the equation 2.5.
min cT x (2.5)
Ax = b
x≥0

with A, b and c and x being matrices of integers. It consists of finding a vari-

able x, which minimizes cT x under the side constraints Ax = b and x ≥ 0.
The ILP method provides an exact approach for solving the general architec-
tural synthesis problem and therefore can be used for the temporal partitioning
problem as well.
The approach consists of formulating the temporal partitioning problem as
an ILP problem instance and then to use well known tools to compute the
solution to the equation. All problem related constraints are defined as a set
of equations involving only integer numbers. The equations must be solved
subject to a given goal to be minimized.
The most important constraints that need to be captured in an ILP problem
instance are the unique assignment constraint, the precedence constraint and
the resource constraint. The integer requirement in the equation vector entries
is due to the fact that we deal with indivisible variables and therefore no real
number must be involved.
Before we present the formal description of the constraints, we would like
to provide the definitions of the variables that will be used in the equations.
Definition 4.25 (0-1 variable) Given a dataflow graph G = (V, E)
and a partitioning P = P1 , ..., Pn of G sorted in order of precedence. We
defined 0-1 variable y and w as follow:
yvi = 1 ⇔ v ∈ Pi
wuv = 1 ⇔ ((u, v) ∈ E) ∧ (u ∈ Pi ) ∧ (v ∈ Pj ) ∧ (Pi 6= Pi )
The role of the y variables is to define the memberships of a given node
to a partition. A value of one means that the node is in the corresponding
partition, while a value of zero means that the node is not in the partition. In
the same way a 0-1 variable for an edge determines if that edge connects two
different partitions. The 0-1 variables provide the basis for defining the set of
ILP equations, that we next formulate.
1 Unique assignment constraint: This constraint is used to control the as-
signment of a node to a single partition. Each node must be assigned exactly
130 RECONFIGURABLE COMPUTING

to one partition. The unique assignment constraint is Formally defined by

equation 2.6.
n
X
∀v ∈ V, yvi = 1 (2.6)
i=1

The sum of the assignment index of a given variable over all partitions is
exactly one, which means that there is exactly one value i for which the y
variable is one. This is the index i of the partition in which v is placed.
2 Precedence constraint: This constraint is used to control the assignment
of data dependent nodes of the dataflow graph to partitions. A node v that
is data dependent from a node u must be placed in a partition with index
bigger or equal than that of the partition into which the node u is placed.
This constraint is captured by the following equation:
n
X n
X
∀(u, v) ∈ E, i × yui ≤ i × yvi (2.7)
i=1 i=1

The two sums define the index of the partition in which the components are
placed.
3 Resource constraint: The resource constraint defines the constraint on the
architecture used. This can be limited to the reconfigurable device or to a
complete system into which the reconfigurable device is integrated. For a
reconfigurable device, the area constraint as well as the constraint on the
number of terminals must be specified.
The area constraint states that the total amount of resources assigned to a
given partition must not exceed the amount of available resource on the
chip. Recall that the computation resources in a reconfigurable device are
defined in term of area occupy on the device. The constraint can therefore
be expressed in term of device area: the total area assigned to a given parti-
tion must not exceed the device area. This is formally defined by equation
2.8 X
∀Pi ∈ P, a(v) ≤ a(H) (2.8)
v∈Pi
The terminal constraint defined in equation 2.9 states that the total number
of input and output in a partition must not exceed the total number of pins
on the device.
X X
∀Pi ∈ P, wuv + wuv ≤ p(H) (2.9)
(u∈Pi )∧(v ∈P
/ i) (u∈P
/ i )∧(v∈Pi )

where p(H) is the number of terminals (pins) of the device H.

High-Level Synthesis for Reconfigurable Devices 131

4 communication memory constraint: The total amount of data to be tem-

porally saved must not exceed the size of the communication memory used.
This constraint is captured by the following equation:
X
wuv ≤ Ms (2.10)
(u,v)∈E and (u∈Pi )∧(v∈Pj(j>i) )

where wuv is the width of the edge (u, v) and Ms is the size of the commu-
nication memory.

Having formulated the constraint as ILP equations, commercial solvers can

be invoked to compute the solutions of the equations. ILP in temporal parti-
tioning were investigated in [76] [135, 136], [93] and [141].

Example 4.26 Considering the partitioning of figure 4.18, let us check if

the ILP-constraints hold.
Unique assignment constraint:
Partition P1 : y22 = y23 = 0, y21 = 1 and y32 = y33 = 0, y31 = 1 and
y42 = y43 = 0, y41 = 1
Partition P2 : y11 = y13 = 0, y12 = 1 and y51 = y53 = 0, y52 = 1 and
y61 = y63 = 0, y62 = 1
Partition P3 : y71 = y72 = 0 and y73 = 1
P P P
Precedence constraint: 1 = ni=1 y4i ≤ ni=1 y5i = 2 and 2 = ni=1 y6i ≤
P n
i=1 y7i = 3

Resource constraint assuming a device with a size of 200 LUTs, and 100
LUTs for the multiplication, 50 LUTs each for the addition, the comparison
and the multiplexer,
P we have:
Partition P1 : 7u=1 yu1 = (100 + 50 + 50) ≤ a(H) = 200
P
Partition P2 : 7u=1 yu2 = (100 + 50 + 50) ≤ a(H) = 200
P
Partition P3 : 7u=1 yu3 = (100) ≤ a(H) = 200

Communication memory constraint: let’s assume that a memory with 50

bytes is available for communication and each data has a 32 bits width. We
have communication between partition P1 and P2 and between P2 and P3 .
For P1 to P2 : w45 = 32Bytes ≤ Ms = 8 × 50
For P2 to P3 : w67 = 32Bytes ≤ Ms = 8 × 50

The general problem of the ILP approaches for partitioning a DFG is the
computation time, which grows drastically with the size of the problem in-
stance. The algorithm can be applied only to small examples. Branch and
bound strategies are some times used to limit the search space. In this case,
132 RECONFIGURABLE COMPUTING

the bounds can be computed using the ASAP/ALAP approach previously de-
scribed. To overcome this problem, some authors reduce the size of the model
by reducing the set of constraints in the problem formulation, but the num-
ber of variables and precedence constraints to be considered still remains high.
Reconfigurable devices are no longer those tiny devices which could not hold
more than four multipliers. Their sizes have increased very fast in the past
and this will continue in the future. Temporal partitioning algorithms should
therefore be able to partition very large graphs (graphs with thousands of
nodes). Trying to formulate all the precedence constraints with the ILP ap-
proach can drastically increase the size of the model, thus making the algo-
rithm intractable.

2.4 Network Flow

The next method that we would like to consider is the network flow approach
for solving temporal partitioning. The network flow methodology is based on
the Ford and Fulkerson min-cut max-flow theorem [84]. This method has been
used in circuit partitioning [231] [123] and also in LUT-technology mapping
by the FlowMap method presented in section 3.3.2. The use of the network
flow approach to solve temporal partitioning problems was first proposed in
[149], [148].
The method is a recursive bipartition approach that successively partitions
a set of remaining nodes in two sets, one of which is a final partition, while a
further partition step must be applied on the second one.
The goal to be reached during partitioning is the minimization of the com-
munication overhead among the partitions, which also means the minimization
of the communication memory. The goal is formulated as the minimization of
the overall cut-size among the partitions. A little cut-size among the partitions
means fewer edges connecting the partitions, less communication, and there-
fore a good partitioning quality.
Formally, we define P̃i as the set of nodes to be partitioned at step i. We
call P̃i the restgraph of the original dataflow graph. P̃0 = G is the complete
graph G. The partitions P0 , P1 , ..., Pk are successively generated using recur-
sive bisection of the sets P̃i until the set P̃k fulfills the device size and terminal
limitations, i.e. the a(P̃k ) ≤ a(H) and p(P̃k ) ≤ p(H).
We consider partitions that generate cycle free configuration graphs. Con-
figuration graph with cycles can be treated in the same way, however the im-
plementation of the feedback loops is left to the processor that controls the
reconfiguration and the data streaming on the datapaths. With this, the bi-
partition of the set P̃i produces two new sets Pi and P̃i+1 with the following
characteristics:

1 s(Pi ) ≤ s(H) and p(Pi ) ≤ p(H), i.e. Pi must not be further partitioned.
High-Level Synthesis for Reconfigurable Devices 133

2 With the ordering the relation of definition 4.13, only one of the following
condition should hold:

(Pi ≤ P̃i+1 )
(P̃i+1 ≤ Pi )
Pi and P̃i+1 are not in relation

Those conditions insure that no cycles exists between Pi and P̃i+1 The net-
work flow method is used in each step of the recursive bipartitioning process
to compute a cycle free bipartition. At each step, the following processing
operations are applied to the rest graph P̃i
1 P̃i is first transformed into a network graph P̃i′ by introducing two new
nodes into the graph P̃i . The first one is an input free source node, whose
outputs are connected to all the primary inputs. In the same way, a sink node
is inserted in the dataflow graph and all the primary outputs are connected
as input to it.
2 A second transformation is done on the resulting network P̃i′ in order to
generate the graph P̃i2 .
′ ′
For an edge (v1 , v2 ) ∈ P̃i × P̃i two edges e1 = (v1 ′ , v2 ′ ), with a
capacity of c1 = 1 and e2 = (v2 ′ , v1 ′ ) with a capacity of c2 = ∞, are
added to P̃i2 .
′ ′ 2
For a multi-terminal edge in P̃i × P̃i , a bridging node is added to P̃i .
An edge weighted with 1 connects the source node with the bridging
2 2
node in P̃i × P̃i . For each sink node in the multi terminal net, an
edge weighted with ∞ is added between the bridging edge and the sink
nodes and between the sink nodes and the source node.

This transformation is similar to the one done by the FlowMap algorithm

in section 3.3.2 before applying the min-cut maxflow theorem. Figure 4.19
illustrates the transformation previously described.
Having transformed the rest graph into a network graph, the maximum flow
on the resulting network is computed, leading to a min-cut (X 2 , X 2 ) of the
network P̃i2 , all the forward edges (from X 2 to X 2 ) must be saturated (flow
equal to the capacity) and all backward edges (from X 2 to X 2 ) should have
no flow. If a net is cut, then only bridging edges can connect X 2 to X 2 , thus
preserving the precedence constraints. The computation of the maximal flow
is done using the augmenting path method described in [186]
For a computed min-cut (X 2 , X 2 ) in P̃i2 , the corresponding cut (X, X) is
induced in P̃i by inserting the equivalent nodes from (X 2 , X 2 ) in (X, X). The
134 RECONFIGURABLE COMPUTING

Figure 4.19. Dataflow graph transformation into a network

four steps for transformation and partitioning using the network flow approach
are illustrated on figure 4.20.
The min-cut max-flow theorem of Ford and Fulkerson is a powerful tool to
minimize the communication in a cut in polynomial time. However, the model
is constructed by inserting a great amount of nodes and edges in the original
graph. The resulting graph P̃i2 may grow too big. In the worst case, the number
of nodes in the new graph can be twice the number of the nodes in the original
graph. The edges number of additional edges also grows dramatically and
become difficult to handle.

2.5 Spectral Methods

In the case where the goal in temporal partitioning is the minimization of
communication between partitions, the spectral approach [30][29][28] can be
used for temporal partitioning.
The connectivity of a graph can be minimized by placing components in
an n dimensional space in such a way that the sum of the distance between
component pairs is minimized. This approach is called the wire length model.
Since the sum of the distances between component pairs can be minimized if
connected components are placed close to each other, the wire length model is
likely to provide an optimal placement of the components in an n dimensional
space. Using the wire length model to solve the temporal partitioning with the
minimization of communication among the partitions can be done by solving
the following two sub problems:

1 Placement of the components in an n dimensional vector space such as to

minimize the sum of the distances between the components. This procedure
is known as an n-dimensional spectral embedding [5] [7] [6]
High-Level Synthesis for Reconfigurable Devices 135

Figure 4.20. Transformation and partitioning steps using the network flow approach

2 Derivation of a partition from an optimal placement which minimizes the

sum of the distance between the components.
136 RECONFIGURABLE COMPUTING

We start with the first problem by considering the one dimensional7 version
as defined by Hall in [107] and illustrated in figure 4.21 [195]. (The subfig-
ure (a) shows an example of a directed graph with four nodes and three edges
borrowed from [195], while subfigure (b) shows the one-dimensional place-
ment on a line. The small black squares represent the centers of the logic cells.
Subfigure (c) shows the two-dimensional placement of the same nodes. The
method does not take account of sizes of the logic cells or actual location of
logic cell connectors. The scaling of the computed position (by the width and
height of the node bounding box). In subfigure (d) a complete layout is made
by placing the logic cells on valid locations.
−0.8 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8
1 2
3 2 4 1
Node
3 4
b)
a)
0.6

3 1
0.4
0.2
0.0

Node
Bounding box
3 1
−0.6 −0.4 −0.2

2 4 2 4
−1 −0.8 −0.6 −0.4 −0.2 0.0 0.2 0.4 0.6 0.8 1

c) d)

Figure 4.21. 1-D and 2-D spectral-based placement of a graph

The spectral embedding problem can be stated as follows: Given a dataflow

graph G = (V, E), find locations for the |V | nodes which minimize the weighted
sum of squared distances between the nodes. If xi denotes the X-coordinates
of node vi ∈ V and r denotes the weighted sum of squared distances be-
tween the nodes, then the 1-dimensional problem is to find the row vector
X T = (x1 , x2 , ..., xn ) which minimizes:
n X
X n
r= (xi − xj )2 wij (2.11)
i=1 j=1

7 Placement on a line
High-Level Synthesis for Reconfigurable Devices 137

To avoid the trivial case in which xi = 0 for all i, we impose the following
condition (normalization):

XT X = 1 (2.12)
We assume that the non interesting solution xi = xj (for all i, j ∈ {1, .., n})
is to be avoided. This leads to all components placed at the same location.
Next, we define the connection matrix, the degree matrix and the Laplacian or
disconnection matrix of G as follows:

Definition 4.27 (Connection Matrix, Degree Matrix, Laplacian Matrix)

Given a DFG G = (V, E), we define:
The connection matrix of G as the symmetric matrix C = (ci,j ) with (1 ≤
i, j ≤| V |) and ci,j = 1 if (vi , vj ) ∈ E and ci,j = 0 otherwise.
The degree matrix of G as the diagonal matrix D = (di,j ), (1 ≤ i, j ≤| V |
P|V |
) with i 6= j → di,j = 0 and di,i = j=1 ci,j .
The Laplacian matrix of G as the matrix B = D − C.

For two nodes vi and vj of the DFG connected by an edge the matrix the
connection matrix will have an entry one in line i and column j. The degree
matrix is a diagonal matrix. An entry in the diagonal (line i, column i) cor-
responds to the number of nodes adjacent to vi . The Laplacian matrix is the
difference between the degree and the connection. Hall has prooved in [107]
that:
r = X T BX (2.13)
Since B is positive semi-definite (B ≥ 0) and B is of rank |V | − 1, whenever
G is connected [107], the initial problem is now reduced to the following:
(
minimize r = X T BX with B ≥ 0
(2.14)
subject to X T X = 1

This is a standard constraint optimization problem which can be solved using

the method of the Lagrange multipliers, which is a standard method used to find
the extrema of a function f (x1 , ..., xn ) subject to a constraint g(x1 , ..., xn ) =
0. An extrema exists if equations (2.15) and (2.16) are satisfied.
∂f ∂f
df = dx1 + · · · + dxn = 0 (2.15)
∂x1 ∂xn
∂g ∂g
dg = dx1 + · · · + dxn = 0 (2.16)
∂x1 ∂xn
138 RECONFIGURABLE COMPUTING

Multiplying (2.16) by a parameter λ to be determined and subtracting the result

from (2.15), we obtain equation (2.17):
∂f ∂g ∂f ∂g
( −λ )dx1 + · · · + ( −λ )dxn = 0 (2.17)
∂x1 ∂x1 ∂xn ∂xn
Because the differentials are all independent, we can set any combination equal
to zero and the remainder must still be zero. Therefor,
∂f ∂g
( −λ )dxk = 0 , ∀k ∈ 1, ..., n (2.18)
∂xk ∂xk
The constant λ to be computed is called the Lagrange multiplier.
To solve problem 2.14, we apply the method of the Lagrange multipliers
with f = X T BX and g = X T X − 1. We introduce the Lagrange multiplier λ
and form the Lagrangian L = X T BX − λ(X T X − 1). Taking the first partial
derivative of L with respect to X and setting the result equal to zero yields to
equation 2.19

2BX − 2λX = 0 ⇐⇒ (B − λI)X = 0 (2.19)

Multiplying equation (2.19) by X T and applying the constraint (2.12), equa-

tion (2.19) yields a non trivial solution if and only if X is the eigenvector
of B which minimizes r and λ(= r) is the corresponding eigenvalue. Be-
cause the minimump Eigenvalue λ0 = 0 yields the non interesting solution
X T = (1, 1, ..., 1)/ (n), the second smallest eigenvalue λ1 should be cho-
sen. The Eigenvector X1 related to the eigenvalue λ1 is the solution to the one
dimensional problem (2.14).
For a placement in a k-dimensional vector space, the problem is formulated
similar to the one dimensional version. The entire dimensions involved have
to be considered. The following equation must therefore hold.

(
minimize R = X1 T BX1 + X2 T BX2 + ... + Xk T BXk
(2.20)
subject to X1 T X1 = X2 T X2 = ... = Xk T Xk = 1

(Xi defines the coordinates of the nodes of V in the i-th dimension) has to be
solved. Analog to the 1-dimensional case, the Lagrange multiplier method will
be applied with the k (each for one dimension) Lagrange multipliers λ1 , λ2 , ...., λk .
The solution are the eigenvectors associated to the k smallest non zero Eigen-
values λ1 , λ2 , ...., λk . This approach is known in the literature as spectral
method. Spectral methods have been widely used in the past for partitioning
and placement [6] [7] [118] [44] [74] [73]. It’s run-time is dominated by the
computation of the eigenvalues, which can be done using various methods on
High-Level Synthesis for Reconfigurable Devices 139

different architectures. The most used algorithm for computing the eigenval-
ues of a matrix is the Golub-Kahan method [99] that needs O(n3 ) on a single
processor for an n by n matrix. Using O(n) processors, the eigenvalues can
be compute with a parallel version of the Hestenes method [200] [172] [119]
[38] in O(n2 S) where S is the number of so called sweeps [119][38]. Brent
and Luk [38] conjectured that S = log(n) and therefore S ≤ 10 in general.
For sparse and quadratic matrices, the eigenvalues can be computed in O(n1.4 )
using the more efficient Lanczos method [99, 6].

2.6 Application to the Temporal Partitioning Problem

So far we have shown that the spectral method can be used to solve the first
part of our problem. The second part, the generation of partitions from a place-
ment with minimum wire length, has to be considered yet. One approach is to
consider the placement of the components in a three dimensional vector space
in which the X axis and Y axis represent the device surface and the Z axis rep-
resent the time at which each component should be mapped inside the device.
In the time dimension, the precedence constraints, ∀e = (vi , vj ) ∈ E, zi ≤ zj
between the components of the dataflow graph must hold. A possible solu-
tion will consist of first placing the component in the space without taking the
precedence constraints into consideration. Thereafter, the partition can be built
using an iterative approach in which the constraint can be checked.
Assuming that three dimensional spectral placement of the component is
computed, the partitions P0 , P1 , ... and Pk can incrementally be generated us-
ing a recursive bisection of the sets P̃i , which is the set of the modules that still
need to be placed in partitions. Initially we set P̃0 = V . At step i, the Partition
P̃i is divided into two partitions Pi and P̃i+1 . This process is repeated untill
P̃i+1 = ∅.
The partition Pi is built at step i by picking components along the Z-axis and
placing them in Pi until the size of Pi reaches a given limit (the device size).
This process creates a bipartition (Pi , P̃i −Pi = P̃i+1 ) of the set of components
not yet assigned to a partition. For each computed bipartition (Pi , P̃i+1 ), we
seek exactly one of the following situations: either (Pi ≤ P̃i+1 ), or (P̃i+1 ≤
Pi ), or there is no relation between Pi and P̃i+1 . This means that either no
edge exists that connects a component of one partition with a component of
another partition, or all the edges connecting components between the two
partitions have the same direction. This is not always true, since we have used
an undirected graph for the placement, and we did not consider the precedence
constraint. The bisection has to be improved to fit the cycle free constraint.
We do this by moving nodes from one side of the bisection to the other until a
strict order is reached between Pi and P̃i+1 .
140 RECONFIGURABLE COMPUTING

Example 4.28 To illustrate our temporal partitioning using the spectral

method, we consider the graph of figure 4.22 whose laplacian matrix is given
in matrix 4.1.

Figure 4.22. Dataflow graph of f = ((a + b) ∗ c) − ((e + f ) + (g ∗ h))

 
1 0 0 0 0 0 0 0 0 −1 0 0 0 0
 0 1 0 0 0 0 0 −1 0 0 0 0 0 0 
 
 0 0 1 0 0 0 0 0 0 0 0 −1 0 0 
 
 0 0 0 1 0 0 0 0 0 0 −1 0 0 0 
 
 0 0 0 0 1 0 0 0 0 0 −1 0 0 0 
 
 0 0 0 0 0 1 0 0 0 0 0 −1 0 0 
 
 0 0 0 0 0 0 1 0 0 0 0 0 −1 0 
 
 0 0 0 0 0 0 0 3 −1 −1 0 0 0 0 
 
 0 0 0 0 0 0 0 −1 3 0 −1 −1 0 0 
 
 −1 0 0 0 0 0 0 −1 0 3 0 0 −1 0 
 0 0 0 −1 −1 0 0 0 −1 0 3 0 0 0 
 
 0 0 −1 0 0 −1 0 0 −1 0 0 3 0 0 
 
 0 0 0 0 0 0 −1 0 0 −1 0 0 3 −1 
0 0 0 0 0 0 0 0 0 −1 0 0 −1 1

Table 4.1. Laplacian matrix of the graph of figure 4.22

The three smallest eigenvalues related to them are λ6 = 0.112586, λ3 =

0.267949, and λ1 = 0.438447 and the related eigenvectors are (0.0410593, −0.152761, 0.225048, −0.241
(0.0, 0.0, 0.0, 0.325058, −0.325058, 0.0), (0.34188, 0.0749482, 0.191984, −0.079481, −0.079481, −0.1
High-Level Synthesis for Reconfigurable Devices 141

Figure 4.23. 3-D spectral placement of the DFG of figure 4.22

In this example, the primary inputs and primary outputs were taken into ac-
count in the building of the matrices. This is useful in that it allows to consider
the placement of components in the vicinity of the pins that they use. How-
ever the eigenvectors only define the position of the components in the different
dimensions.

Using those three eigenvectors we computed the 3-D placement of figure

4.23.
By picking the component along the Z-axis in increasing order of their Z-
coordinates, we constructed the two partitions P0 and P1 of figure 4.24.
This partition is not ordered since the edges (3, 6) and (4, 6) have their
source in P0 and their destination in P1 , while the edge (5, 3) has it’s source in
P1 and it’s destination in P0 . This produces a configuration graph with a cycle.
By exchanging the nodes 3 and 5 from P0 to P1 and vice versa, the partitioning
becomes an ordered one.

2.7 Elimination of Cycles in the Configuration Graph

Elimination of cycles in the configuration graph is important in methods
that perform an iterative bisection of a rest graph. This ist the case for the list-
scheduling, the network flow method and the spectral method. The approach
that we present here is more general than just limited to one particular method.
The cycle elimination method is based on the well known iterative improve-
ment procedure of Kernighan and Lin (KL) described in [137] [83] [144].
Since the KL-algorithm deals only with undirected graphs and because we
142 RECONFIGURABLE COMPUTING

Figure 4.24. Derived partitioning from the spectral placement of figure 4.23

target directed graphs in this work, some modifications have to be done on the
original KL algorithm.
The fundamental idea behind the KL-algorithm is the definition of a cut of
a bisection as well as the notion of the gain of moving a vertex from one side
of the bisection to the other. For an undirected graph, a cut is defined as the
weighted sum of all the edges crossing from one partition to another. By mov-
ing a node from one partition into the other, the number of crossing edges is
also modified and the value of the cut changed. The KL-algorithm allows a
series of moves, which reduce the bisection cut. If the gain of moving a vertex
is positive, then making that move will reduce the total cost of the cut in the
partition. During one iteration of the KL-algorithm, nodes are moved from
one side of the bisection and locked on the other side. The cost of swapping
unlocked nodes in opposite parts is then computed and the nodes with the best
gain (greatest decrease or less increase of the cut) are swapped. If all the nodes
are locked, the lowest cost partition is set to the current computed partition, if
it improves the cost of the cut. One iteration of the KL-algorithms is called
a pass. After one pass, all the nodes are unlocked and a new pass is com-
puted. The iteration terminates if a pass produces no further improvement on
the cut. For a more detailed description of the KL-methods and it’s extension
by Fiduccia and Mattheyses, refer to [137] [83] [144].
The KL-method works on undirected graphs and does not differentiate be-
tween the the direction of edges. It does not matter if an edge crosses from
the first partition to the second partition or vice versa. In temporal partition-
ing, the graph are directed due to precedence constraints, thus the original
KL-algorithm has to be modified to fit our needs. To better explain how the
High-Level Synthesis for Reconfigurable Devices 143

modification can be done on the original KL-method done, we first provide

more definitions.

Definition 4.29 Let G = (V, E) be a dataflow graph. For two nodes vi

and vj in V and a bipartition {P, Q} of G, we have:
P
EP (vi ) = {(vi ,vj )|vj ∈P } wij is the weighted sum of the edges from vi to
P , i.e. edges connecting vi to nodes in P (figure 4.25).
P
IP (vi ) = {(vj ,vi )|vj ∈P } wji is the weighted sum of the edges from P to
vi , i.e. edges connecting nodes in P to vi (figure 4.25) and
EP,Q = {(vi , vj ) ∈ E | vi ∈ P and vj ∈ Q} is the set of edges crossing
from partition P to partition Q, i.e. edges having their sources in P and
their destination in Q.

Figure 4.25. Internal and external edges of a given nodes

At step i of the temporal partitioning, a bipartition (Pi , P̃i+1 ) is generated.

The precedence constraint (Pi ≤ P̃i+1 or P̃i+1 ≤ Pi ) can be captured through
the following equation: EPi ,P̃i+1 = ∅ or EP̃i+1 ,Pi = ∅. Intuitively we would
like to combine the two variables EPi ,P̃i+1 and EP̃i+1 ,Pi in the definition of
our cut and try to decrease one of them to zero using the KL-algorithm. But
decreasing one variable could have the negative effect of increasing the second
144 RECONFIGURABLE COMPUTING

one, thus producing a cycle of improvement and alteration on the cost of the
cut. To avoid this we apply two instances of the KL-algorithm on the same
bisection in parallel. The objective is to have EPi ,P̃i+1 = ∅ on the first compu-
tation path and EP̃i+1 ,Pi = ∅ on the second one. Obviously the cost of the cut
is set to |EPi ,P̃i+1 | on the first path and |EP̃i+1 ,Pi | on the second path. After one
pass on each path, we check if the objective has been reached on one path. If
this is the case, then the result is set to be the partition generated on that path.
Otherwise, a new pass is computed on the two computation paths. The gain of
moving a node is defined differently on the two computation paths.
On the first path where the goal is to have |EPi ,P̃i+1 | = 0, the gain of
moving a node j from Pi to P̃i+1 is IPi (vj ) − EP̃i+1 (vj ) and the one of
moving a node k from P̃i+1 to Pi is EP̃i+1 (vk ) − IPi (vk ).
On the second path where to goal is to have |EP̃i+1 ,Pi | = 0, the gain of
moving a node j from Pi to P̃i+1 is EPi (vj ) − IP̃i+1 (vj ) while the gain of
moving a node k from P̃i+1 to Pi is IP̃i+1 (vk ) − EPi (vk ).
The gain defined on each computation path is the same like the one defined
in the original KL-algorithm. The modified version of the KL-algorithm pre-
sented here will produce the desired result on one path. Because the targeted
graphs are acyclic dataflow graphs, there exists a partition in which all the
edges cross from the first one to the second. Such a partition is provided for
example by a list-scheduling algorithm.

2.8 Local optimization: Context Switching, Time

Multiplexing and Configuration Switching
Reconfiguration time has always been a major problem in reconfigurable
computing. Switching from one configuration to the next can take up to a frac-
tion of a second in large FPGAs, according to the technical realization of the
system. With millions of iterations, the reconfiguration overhead becomes too
high to allow a use of the device in real-life problem solving. In order to de-
crease reconfiguration overhead, the architectural concepts of time-multiplex
and context switching FPGAs have been proposed [202] [188]. The general
idea behind those two concepts is to store a given number of configurations
directly on the chip. This avoids the downloading of a large amount of recon-
figuration data when required.
Trimberger [202] proposed an FPGA architecture in which designs are mod-
eled as Mealy state machines. Micro registers are used to temporally hold
computation data when switching from one configuration to the next. The
combinational logic receives its inputs from the device inputs and from the flip
flop outputs, and the device outputs come from combinational logic and from
High-Level Synthesis for Reconfigurable Devices 145

flip flops outputs. In logic engine mode, the device emulates a large design in
many microcycles. In each microcycle, resources are allocated to a new con-
figuration. A similar architecture has been proposed by Scalera et al [188].
Data pipes are used here to exchange data between different configurations
also called contexts. A data pipe contains a plurality of Context Switching
Logic Arrays (CSLA) which can be used to process two 16-bit words. An in-
coming context can then pick its input data where its predecessor left off by
acquiring the intermediate data deposited on the rightmost portion of the pipe
and processing it in a pipeline from right to left. Unfortunately, the methods
developed have remained in a conceptual stage. Neither time multiplexing nor
context switching FPGAs have ever been commercialized.
A more practical approach have been proposed in [30] to reduce the recon-
figuration time. It is based on the observation that traditional list scheduling
based temporal partitioning algorithms can produce a series of configurations
based on the same set of operators if the components are "well ordered"8 in
the list. For example, for two consecutive configurations ζi = {C1 , ..., Cki }
and ζi+1 = {C1 ′ , ..., Cki+1 ′ } representing two partitions Pi and Pi+1 , one can
be the subset of the other one, i.e either ζi ⊆ ζi+1 , or ζi+1 ⊆ ζi . If one
of those two situations arises, then the reconfiguration overhead can be re-
duced by implementing the two partitions Pi and Pi+1 in one configuration
ζnew = ζi ∪ ζi+1 . The components of ζnew will the be shared among the two
partitions Pi and Pi+1 . That means, the modules required for both configu-
rations are placed on the device and wired in two different ways. Each way
corresponds to one configuration. Figure 4.26 shows a partitioning of a graph
in two partitions P0 and P1 . The set of components required to implement P1
is a subset of the set of components required to implement P0 .
With the use of multiplexers on the inputs of the operators in and the use
of selection signals to connect the corresponding signals to the module inputs,
it is then possible to implement the connections defined in configuration ζi as
well as those defined in ζi+1 together. Switching from the configuration ζi
to the configuration ζi+1 can be done by setting the corresponding value of
the selection signals. The device is therefore "reconfigured" without changing
the physical configuration. We call this process configuration switching. With
configuration switching there is no need to save the registers of the FPGA in
the processor address space during reconfiguration because no register is al-
tered. With this, configuration switching will help to reduce the data exchange
between the processor and the reconfigurable.
Configuration switching can be extended to a series of configurations ζ1 , ..., ζr
by extracting the smallest amount of common operators needed to implement

8 A well order defines the way components with the same level number should be place in the list before

partitioning.
146 RECONFIGURABLE COMPUTING

Figure 4.26. Partitioning of a graph in two set with a common sets of operators

Figure 4.27. Logical partitioning of the graph of figure 4.26

all the configurations and implement the configuration switching on ζ1 , ..., ζr .

But the amount of multiplexers needed to assign the operators to a particu-
lar configuration as well as the difficulty to route all the possible configura-
High-Level Synthesis for Reconfigurable Devices 147

tions can jeopardize the implementation of configuration switching for a high

number of configurations. The amount of partitions to be implemented using
configuration switching must be kept small.
Example 4.30 Figures 4.28 shows the implementation of configuration switch-
ing on a device for the two partitions of figure 4.27.
During reconfiguration, the corresponding values for the selection signal
are set on the multiplexers. The output of the corresponding partitions are
then selected and sent to the operators.

Figure 4.28. Implementation of configuration switching with the partitions of figure 4.27

Although configuration switching can save time and reduce the data trans-
fer between the processor and the reconfigurable device, its implementation
usually requires additional resources. If the amount of additional resource be-
comes too big, then it makes no sense to have a smaller number of components
implemented, and a larger logic area on the device just to realize the switch.
The tradeoffs between the numbers of additional resources and the number of
configurations to be implemented has to be chosen carefully. The search for
a tradeoffs between the number of configurations to reside in the device and
the complexity of the final design is an optimization problem that we do not
address in this book, namely the exploration of the design space.
148 RECONFIGURABLE COMPUTING

3. Conclusion
High level synthesis for reconfigurable devices has benefit from the various
method provided in the past for the general high-level synthesis problem. The
main difference between the two approaches is on the freedom of choice on
the operator type in the case of reconfigurable device. However, one must deal
with temporal partitioning, which allows the reuse of chip resource over time.
Time multiplex reuse of resources was a great matter of concerns for a decade,
as the capacity of FPGAs was very small to accommodate one the complete
functions to be implemented. Meanwhile, the size of FPGAs has grown a lot
and it is quite difficult to find those functions to be temporal partitioned in or-
der to exploit chip resources. The interest on temporal partitioning methods
has considerably decrease as consequence of this growth in the size of FPGAs.
This does not mean that temporal partitioning is useless. We still have areas
of computation like rapid prototyping where reconfigurable device can provide
a very cheap alternative to the very expensive existing systems. Rapid proto-
typing systems are usually done with expensive machines containing a set of
FPGAs. Applications are then partitioned among the FPGAs, which compute
each part of the function to be implemented. With temporal partitioning, a few
number of FPGAs can be used by several parts of the applications that must
not be computed at the same time.
Chapter 5

TEMPORAL PLACEMENT

In the last chapter, we presented the high-level synthesis problem for recon-
figurable devices and some solution approaches. The result is a set of partitions
that are used to reconfigure the complete device. While the implementation of
single partitions is easy, the amount of waste resources in partitions can be very
high. Recall that the waste resource of a component is the amount resources
occupied by that component multiplied by the time where the component is
idle on the device.
Wasting resources on the chip can be avoided if any single component is
placed on the chip only when its computation is required and remains on the
device only for time it is active. With this, idle components can be replaced by
new ones, ready to run at a given point of time. Exchanging a single compo-
nent on the chip means reconfiguring the chip only on the location previously
occupied by that component. This process is called partial reconfiguration
in contrast to full reconfiguration where the full device must be reconfigured,
even for the replacement a single component. In order to be exploited, the par-
tial reconfiguration must be technically supported by the device, which is not
the case for all available devices. While most of the existing devices support
full reconfiguration, only few are able to be partially reconfigured.
For a given set of operations to be executed, the resource allocation on the
device is a time dependant process in which not only the placement of the com-
ponents on the device is defined, but also the time slot in which the execution
of the task must be performed. The time dependant placement of task on the
device is called temporal placement.
Temporal placement can be graphically illustrated through an arrangement
of rectangular boxes in a 3-dimensional container whose base is defined by
the surface of the device and the high is the time axis. Each box represents a
150 RECONFIGURABLE COMPUTING

components with a given surface placed at a given location on the device at a

time defined on the time axis.

Example 5.1 In figure 5.1, a temporal placement of a set of components

{v1 , v2 , v3 , v4 , v5 , v6 , v7 , v8 } is given. Component v1 for example occupies a

Figure 5.1. Temporal placement as 3-D placement of blocks

surface on the defined by its length and height for a time slot 0 to 55, which
corresponds to its computation latency. Component v5 starts its execution at
time 90 and occupies among others, part of the device that was perviously
occupied by component v4 . Component v8 occupies a part of the device for the
whole computation time.

The computation of a temporal placement for a given set of components rep-

resenting some tasks to be implemented can be done off-line, at compile-time
or on-line, while the device is executing. In the first case, for an application
specified as dataflow graph that must be computed on the device, the com-
putation sequence is defined at compile- time and remains fixed for the given
function. In on-line temporal placement or simple said, on-line placement, the
computation sequence is not known in advance. The creation of task that must
be placed on the device is done dynamically at run-time. It is therefore not pos-
sible to capture the specification of an algorithm for on-line placement through
a given model at compile-time.
Temporal Placement 151

Temporal placement has the advantage of being highly flexible and efficient
in term of device utilization, however it also has a major drawback. Efficient
temporal placement algorithms are difficult and cost intensive. On-line place-
ment requires to solve some computational intensive problems at run-time in a
fraction of millisecond. Those are: the efficient management of the free space,
the selection of the best site for a new component and the management of com-
munication. For a new module to be placed at run-time, it is not sufficient to
only compute the best placement site. The communication between the mod-
ules running on the chip must also be considered. He communication between
modules running on the chip and the external world must be take in to account
as well. A reconfigurable device must provide a viable on-line communication
mechanisms to help establish the communication among modules on the chip
at run-time. This problem is not easy to solve and requires most of the time,
some prerequisites on the device architecture.
Although communication aspects like the distance among components are
considered in some of the methods provided in this chapter, we do not deal
with the technical realization of communication here. Chapter 6 is devoted
to this topic. We assume that communication among modules placed on the
device is somehow possible.
The first part of this chapter address the off-line temporal placement while
the second part deals with the on-line temporal placement. In both cases we
start with some definitions related to the part being addressed and then we
present existing approaches to solve the corresponding problem.

1. Off-Line Temporal Placement

We first provide a formal definition of the temporal placement problem be-
fore investigating existing solution approaches. For sake of simplicity and
compatibility, we use the definitions provided in [199] [82].
Definition 5.2 ((Off-line) Temporal placement) Given a dataflow
graph G = (V, E) and a reconfigurable device H with lenght (Hx and width
Hy ), a temporal placement is a 3-dimensional vector function:
p = (px , py , pt ) : V → 3 , where pt defines a feasible schedule.

For a given node vi , the values px (vi ), py (vi ) and pt (vi ) denote the coordi-
nates of the node vi in a 3 dimensional vector space. px (vi ) and py (vi ) are the
coordinates of vi on the device H, while pt (vi ) defines the starting time of vi
on H.
We next present some solution approaches for the off-line temporal place-
ment problem. We first introduce a simple incremental method based on the
first-fit/best-fit concept. In the second method, a clustering of components in
blocks that can be placed together on the chip is done. Finally, we introduce an
152 RECONFIGURABLE COMPUTING

exact method that characterized valid partitioning as a set of constraint equa-

tions, used to describe the 3-D packing of boxes in containers.

1.1 First-Fit and Best-Fit Placement

Given a dataflow graph for which a temporal placement must be computed,
a simple approach will consist of continuously keeping track of free locations
on the chip. The nodes to be placed will then be selected according to their
readiness. A node is said to be ready to be placed, if all it predecessors are
already placed. For each selected node, one of the free locations on the chip is
selected and the node is placed at that location, provides that the space is large
enough to accommodate the selected node. If the selection procedure chooses
the first free place on the chip to place the new component, then we call it a
firs-fit.
The firt-fit procedure is fast and run in linear time in the number of free
locations. However, if we assume that each selected location can accommodate
only one component, the amount of unused resource created by the component
can be very high, if a location is selected whose surface is bigger than that of
the component. To void this situation, we may choose a best fit strategy in
which the component is placed on the location whose surface is the closest to
that of the component. We call this approach a best-fit. Best fit is much time
consuming than first-fit, because the complete list of free locations must be
searched for the best location for the component to be placed.
The definition of a location has a great importance in the firs-fit and best-fit
strategy. What is a location? A simple rectangle? or is it a shape with an arbi-
trary contour? For sake of simplicity, we assume that a location is a rectangle,
in which case the test of inclusion of a given component in a free location is
easy. What happens with the new free locations that results from the place-
ment of a component on a selected free area of the device? Just ignoring those
free fragmented locations can be inefficient. In many case the total amount of
fragmented area can be enough to hold some components. If we don’t keep
track of the fragmented area and if those fragmented area are not consecutive,
then many components cannot be placed on the device, although enough free
space exist to accommodate those components. Managing the fragmented lo-
cations in turn requires a very large book keeping efforts. In section 2 we will
present some algorithms that deal with the fragmentation issues in the context
of on-line temporal placement.
A possibility to keep the first-fit and the best-fit approaches simple is to
segment the device in sectors, each of which is defined as a locations. Such
a location is able to hold a given amount of components, whose total area
cannot exceed the size of that sector. This approach is attractive for many
devices in which the partial reconfiguration is done with some restrictions, as
it is the case for example with the Xilinx Virtex and Virtex II FPGA, where the
Temporal Placement 153

reconfiguration can be done only column wise. Whenever a component has to

be exchanged, all the components sharing some columns with that component
are affected. The situation is better in the Virtex 4 and Virtex 5 families, where
the reconfiguration does not affect the complete device column, but only the
complete height of a rectangular block within the chip. By defining the location
to span the complete device column in the case of the Virtex and Virtex II
and the block height for the Virtex 4 and Virtex 5, we avoid to disturb the
components already running on the device during the reconfiguration process.
Given a dataflow graph and a deviced partitioned in locations, we would
then like to place several components at the same time at a given location on
the device. The dataflow graph must then be partitioned in sets of components
such that each set can fit into a location. Several partitioning strategies exist in
the literature. However, most of them does not take the precedence constraints
into account. Partitioning strategies are not presented in this section. We refer
to the temporal partitioning algorithms presented in the previous chapter that
can also be used here. In this case, the size constraint must se to be the size
of a location. In order to limit the amount of wasted resources on the chip,
a clustering algorithms that tries to place components in partitions according
to their finisching time was presented in [30]. This has the advantage that all
the components finish at the same time, thus minimizing the wasted amount of
resources.
Let now assume that for a given a dataflow graph G = (V, E) a partition-
ing or better say a clustering C = {C1 , ..., Cn } is computed such that all the
clusters fulfill the size constraints. Algorithm 10 computes a temporal place-
ment of C in a first-fit manner. For a new cluster Cact to be placed on the
device, the algorithm checks for the cluster Ctop with the minimum run-time
among the clusters allocated to the device for which the precedence constraint
Ctop ≤ Cact between Ctop and Cact holds. The cluster Cact is placed on top
of Ctop . The placement of cluster Cact on top Ctop simply means that the re-
sources allocated to Ctop will be allocated to Cact after Ctop has completed
he’s execution. The algorithm stops when all the clusters have been placed.

Algorithm 10 First-fit temporal placement of clusters

while while all the clusters are not placed do do
Select one cluster Cact from the list of ready to run clusters
From the clusters already placed, select the cluster Ctop
with the smallest finishing-time such that Ctop ≤ Cact
place the cluster Cact on top of Ctop
end while

Example 5.3 Figure 5.2 illustrates the firs-fit temporal cluster placement
on a set of ten clusters. We assume that the precedence constraints are sat-
154 RECONFIGURABLE COMPUTING

Figure 5.2. First-fit temporal placement of a set of clusters

isfied during the placement. At the beginning, the clusters C0 , C1 , C2 , C3 to

C4 are completely placed on the device. In order to place cluster C5 , clus-
ter C0 with the earliest finish time is selected and cluster C5 is placed on
top of C0 . Cluster C6 occupies the space freed by cluster C2 and cluster C7
is placed on top of C3 . C8 is then placed on top of C4 , C9 on top of C1
and C10 on top of C5 . The configuration sequence produced by this exam-
ple is ({0, 1, 2, 3, 4} → {5, 1, 2, 3, 4} → {5, 1, 6, 7, 4} → {5, 1, 6, 7, 8} →
{5, 9, 6, 7, 8} → {10, 9, 6, 7, 8}).

The first-fit and best-fit approaches provide fast method to compute a tem-
poral placement solution. However, no effort is spent on the efficiency of the
space management. Integer linear programming can be used as exact method
to solve the temporal placement problem. In this case, a set of constraint equa-
tions that must be fulfilled by each solution must be formulated. Solving the
equations will then provide a solution to the temporat placement problem. Be-
cause the computation of a ILP-solution is computational intensive, integer lin-
ear programming is usually suitable only for for small size problems. An exact
method that we next presented was proposed in [199] [82] to solve the tempo-
ral placement for large size problems. Equations are formulated in a similar
way as in integer linear programming. However, a branch and bound strategy
is used in order to reduce the search space and allow much large problems to
be solved.
Temporal Placement 155

1.2 Packing Approach for Temporal Placement

The idea of modeling the temporal placement problem as a three-dimensional
packing problem, which is the problem of deciding if a given set of boxes can
be placed within a given container of size (x, y, h) representing the length, the
width and the height of the container, was first presented in [199] [82]. Ac-
cording to the goals seeked, several variations can be made in the formulation
of the problem. A goal can be for example the minimization of the overall exe-
cution time or the minimization of the amount of wasted resources for a given
algorithm. Let us consider two possible variations of the packing problem: the
Base Minimization Problem (BMP) and the the Strip Packing Problem (SPP)

1 Base Minimization Problem (BMP): Given a set of boxes B and a height h,

the problem of finding a container with minimal size (x, y, h) that can acco-
modate the set of boxes B is called the base minimization problem (BMP).
The corresponding goal in tempoaral placement is to find the device with
minimum size (x, y) on which a set of components can be implemented
given an overall run-time t = h.

2 Strip Packing Problem (SPP): Given a set of boxes and a base (x, y), the
problem of finding the minimum heigh h container with size (x, y, h) that
can hold all the boxes is called the the Strip Packing Problem (SPP). The
analogy with the temporal placement is to find the minimum run-time t = h
for a set of components given a reconfigurable device with size (x, y).

With the optimal solution of the BMP, one can select the best reconfigurable
device, in term of size, on which a set of tasks can be computed in a limited
amount of time. This device is optimized for the given set of task and for the
given run-time. In reconfigurable computing, the main task is not necessary to
choose an optimal device for a fixed set of tasks, but to have a fixed device on
which various set of task can be optimally implemented. The BMP is therefore
not the right tool to be used.
The SPP that we next focus on, best matches the requirements of having
only one device on which different set of tasks can be implemented at different
time. In its original version, the goal of the SPP was to minimize the overall
computation delay. This may be changed according to the objectives seek.
As state earlier in this section, integer linear programming can be used to
formulated all constraint that must be fulfilled for a given soltion, as a set of
equation to be solved. However, this approach can be used only for small size
problems. The concept of packing classes that we next present, was used in
[199] [82] as a mean to define a feasible solution for the SPP.
156 RECONFIGURABLE COMPUTING

1.2.1 Packing Classes

The solution of an instance of the temporal placement problem can first be
characterized through a valid packing, using the concept of packing classes.
This approach is limited to the pure packing of boxes in a container and does
not capture the precedence constraints in the dataflow graph.
providing an orientation to the packing classes can then be done in a sec-
ond step to capture the precedence constraints of the dataflow graph in a valid
solution. The description provided in [199] [82] considerers packing in arbi-
trary dimension. We are interested only in a three dimensional packing and
will therefore limit our description accordingly. We start by providing some
definitions of the terms needed in this section.
Definition 5.4 (Interval Graph) Given a dataflow graph G = (V, E),
a reconfigurable device H and a packing of V into H, i.e a 3-D placement of
the components of V on the device H. An interval graph of G is a graph
Gi = (V, Ei ), ∀i ∈ {1, 2, 3} such that: (vk , vl ) ∈ Ei ⇐⇒ the projections of
the nodes vk and vl overlap in the i-th dimension, i.e on the i-th axis.
Definition 5.5 (Complement Graph) Given a dataflow graph G =
(V, E), a reconfigurable device H and a packing of V into H, The complement
graph of an interval graph Gi = (V, Ei ) is the graph Gi = (V, E i ), ∀i ∈
{1, 2, 3}.
The strategy presented here consists of first providing a better characteri-
zation of a solution. With this, potential solutions can be checked for validity.
The second step will then consists of constructing arrangements, each of which
is tested for its validity as solution. The validity of a solution can be enforced
with the two following conditions that must be fulfilled by the corresponding
interval graphs of a given arrangement:
P
Any independant1 set S ⊆ G is i-admissible. I.e. v∈S pr i (v) ≤ hi ,
where pri (v) is the lenght of projection of component vi on the i-th dimen-
sion and hi is the lenght of the device in the same dimension.
T3
i=1 Ei = ∅

The first restriction states that in a direction i, the total length of components
placed on the device and pairwise overlapping in the perpendicular dimension
to i, should not exceed the length of the device in dimension i. To better un-
derstand this restriction consider a perpendicular line to direction i that run
through the device. All the components that are cut by the line pairwise over-
lap in the perpendicular direction to i. The total sum of their width in the i

1 ∀v, w ∈ S, v overlaps with w in dimension (3 − i), i ∈ {1, 2}

Temporal Placement 157

direction should therefore not exceed the line segment which is in the device,
i.e. the size of the device in dimension i. Placing many such lines at the dif-
ferent coordinate points in direction i of the device will help to capture all the
restrictions stated by the first condition.
With the second restrictions, we avoid placements in which components
overlap in all dimensions. A placement can only be valid if for any pair of
components, at least one dimension must exists in which the two components
do not overlap.

Definition 5.6 (packing Class) A 3-tuple of interval graphs that sat-

isfy the two previous defined restriction is called a packing class.

Example 5.7 The two dimensional placement of figure 5.3 is a valid packing
with the corresponding packing classes G1 and G2 . In figure 5.4, we have an

Figure 5.3. Valid two dimensional packing

example of invalid packing for two reasons. First, part of the component v5 is
placed with a part out of the device and second, components v3 and v4 overlap
in all directions.

Using the pure packing classes as previously defined help to defined the
placement aspect of the temporal placement problem, without caring about the
constraints in the dataflow graph. Because those constraints must be preserved
in a valid temporal placement, a modification of the packing classes must be
done in order to insure that a component depending on another one starts its
execution only after the component on which it depends has completed he’s
158 RECONFIGURABLE COMPUTING

Figure 5.4. A non valid two dimensional packing

execution. The modification that we present here are the orientation of the
packing classes.

1.2.2 Orientation of Packing Classes

An undirected edge {v1 , v2 } in a component graph Gi corresponds to an
intersection between the projections of component v1 and v2 in the i-th dimen-
sion. However it does not provide any information on the relative position of
component v1 compare to v2 . The two nodes v1 and v2 can be compared ac-
cording to their placement in one of the three dimension. In the x-dimension
we have the relation "left to" and "upper" in the y dimension. However, we
are only interested in the 3-th dimension, i.e the time axis, where we define
the relation "place after". For each pair of component v1 and v2 , either the
component v1 is placed after v2 (pt (v1 ) + t1 ≤ pt (v2 )) or vice versa, if their
placement overlaps in the 3-rd dimension. We use this fact to define the com-
parability graph of an interval graph.

Definition 5.8 (Comparability Graph) Given a dataflow graph G =

(V, E), a reconfigurable device H and a packing of V into H.
The comparability graph of an interval graph Gi = (V, Ei ), ∀i ∈ {1, 2} is
the directed graph Gi = (V, Ei ), where: (vk , vl ) ∈ Ei ⇐⇒ vl is "placed
after" vk in 3-th dimension, i.e. (pt (vk ) + tk ≤ pt (vl )).

The relation "place after" defined by the comparatibililty graph is a transi-

tive relation also known as transitive orientation that can be used for orienting
the packing classes.
Temporal Placement 159

Definition 5.9 (Packing Class Orientation) Given a dataflow graph

G = (V, E), a reconfigurable device H and a packing of V into H. The ori-
entation of the packing class corresponding to the packing is define though the
comparability graph of the interval graph in the time dimension.

The orientation of a packing is used to refine information about the place-

ment of components on the time axis. This refinement is the information need
to check if the precedence constraint is fulfilled. Precedence constraints do
not play a role for the placement of component on the x and y directions.
Therefore a large variety of packing classes exist that lead to different 3-D
placements with precedence constraints.

Example 5.10 Figure 5.5 provides a 3-D placement and its characteriza-
tion through interval graphs, complement graphs and the orientation of the
packing class associated to the placement. Shown is only the graph related to
the dimensions x and y.

Figure 5.5. 3-D placement and corresponding interval graphs, complement graphs and or-
einted packing

Now that a good description of a solution is available, we have the possibil-

ity to check if a given arrangement of nodes forms a valid packing with some
precedence constraints fulfilled. We only need a method to build arrangements
160 RECONFIGURABLE COMPUTING

on which we can apply the formulas previously described to check wether or

not the arrangement is a solution. Moreover, if we have a set of available so-
lutions, we also need to extract the optimal one according the the optimality
criteria. A solution or better said an arrangement that can be checked for valid-
ity can be constructed by arbitrarily fixing the edges of the different graphs as
well as their orientations. The set of all possible arrangements that can be built
in this constructive fashion called the solution space. This can be extremely
large and very difficult to explored, if the size of the problem is big.
In [199] [82], a branch and bound strategy for iteratively prune the set of all
possible arrangements is presented. The approach is an incremental that aim
at constructing a tree from the root to the leaf. The root of the tree consists of
the set of available nodes. The tree is constructed from the root to the leaf by
inserting edges to the already existing graph. Two different insertions of edges
on the same node lead to two different branches in the tree. In each step of the
algorithm, a new edge is include into the graphs and the validity of the result-
ing placement is checked. If the resulting placement will not be valid, then all
possible arrangement in which the introduced edges exists are discarded from
the search. Otherwise, a new edge is added to the constructed graph. When-
ever a branch will not lead to valid leaf, the complete subtree resulting from
the introduction of a new branch is discarded. This help to avoid exhaustive
and expensive checks on branches that do not lead to a solution. The formal
characterization of a valid solution previously done, provides at the same time
a means to check if the full or partial arrangement constructed can lead to a
solution. In [199] [82], the authors use the concept of induced cycle to identify
and discard paths that will lead to non valid solutions.

The method presented in this section target problems well specified at compiled-
time with no change on the computational flow at run-time. In many systems,
in particular real-time ones, the set of task as well as their interconnectivity is
not known at compiled-time. Run-time or on-line placement possibility must
be provided to deal with unpredictable event that create tasks, which must be
executed at run-time. The next sections deals with this issue.

2. On-Line Temporal Placement

In section 1, we presented the architecture of a dynamic run-time reconfig-
urable system (figure 3.1). The dynamic management of the device resources
is done by two main components: a scheduler and a placer, each of which can
be implemented as part of an operating system running on a processor. While
the scheduler determines which task should be executed on the RPU, the placer
tries to place the associated bounding box associated to a task on the device,
thus allocating rectangular resource area on the device for the execution of that
task. Whenever the placer receives a request from the scheduler to place a task
Temporal Placement 161

and it is not successful in doing that, the scheduler must decide on what to do
with the task. It can later try again to run the task on the RPU or it can decide
to let the task run with a lower speed on the CPU. In any case, not being able
to satisfy a placement request creates additional delay in the execution of the
program. In order to avoid this penalty, care should be taken on the design of
a the placer. The placer should be able to place as much tasks as possible on
the device in order to avoid delay penalties. This goal can be reached if the
amount of wasted resource is kept small. We first state the on-line placement
problem and then present some solution approaches.
Definition 5.11 (On-line Placement) Given a reconfigurable process-
ing unit H at time t with a given configuration Ct . Find an optimal position
for a new incoming task v such that v does not overlap with any running com-
ponent in the configuration Ct .
As defined in the mathematics, the goal when solving an on-line problem
does not only consists of providing an optimal method at a given time step, but
a method that is optimal for a sequence of computation not fixed in advance.
The method should be developed to cope with unknown parameters that may
arise in the future. A non optimal partial solution can be preferred to an optimal
one at time t, if this partial sub-optimal solution will contributes on a global
optimal in the future. In this section, we limit the on-line definition to the
stepwise optimization of a partial problem in a given time slot, whiteout caring
about the global optimal solution.
We present in the next sections two approaches for on-line placement of
incomming components on the device. While the first approach manages the
free space on the device by keeping track of all empty rectangles from which
one will be selected to place the new module, the second one keeps only track
of the occupied area on the device and try to place a new component in such a
way that no overlap with a running module occurs.

3. Managing the Device’s Free Space with Empty

Rectangles
One possibility to implement the placer is to always keep track of the free
space on the device. On request to place a new component, a free location
which can accomodate the new component is chosen and the module is placed
at the selected location.
Bazargan et al. [18] proposed the used of a list of empty rectangles on the
chip to capture the free locations, an approach that was used in computational
geometry to solve rectangle layout problems as it is the case in glass cutting
and paper cutting industries [116].
The main task to solve is to place a given set of rectangles, so called pieces,
in a given larger rectangular area, the slot sheet, such as to minimize the amount
162 RECONFIGURABLE COMPUTING

of wasted space in the resulting layout. Once again, we start with definitions
and then continue with solution approaches.

Definition 5.12 (Empty Rectangle) An Empty Rectangle (ER) is a

rectangle that does not overlap a placed module on th chip. A maximum empty
rectangle (MER) or Largest Empty Rectangle (LER) is an empty rectangle that
is not included in any other rectangle than the device bounding box.

Example 5.13 An illustration of empty rectangles is provided in figure 5.6.

(E, D), (A,D) and (E,C) are example of empty rectangles. While (E, D) and
(A,D) are maximal, (E,C) is not because it is included in (E,D). (E,F) is not an
empty rectangle bacause it overlaps module v2 .

Figure 5.6. Various types of empty rectangles

Based on the empty rectangle concept, a strategy to solved the on-line tem-
poral placement problem was proposed in [18]. The method, which has the
name "Keeping All Maximum Empty Rectangles (KAMER)", permanently keeps
track of all MERs. Whenever a request for placing a component v arrives, the
list of MERs is searched for a rectangle that can accommodate v. Because it
is possible to have many MERs in which v can fit, strategies like the first-fit
or best-fit are used to select one rectangle from the set of possible free rect-
angles. Once a rectangle is chosen, the candidate points that can be chosen
as reference point for placing the new component are those that do not allow
an overlap with the external part of the rectangle. In [18], the authors use the
bottom left point as the reference one to place the component.
Temporal Placement 163

The great advantage of the KAMER approach is that all the free space on
the device is captured, thus providing a good quality2 . This happens at the cost
of efficiency in the computation time. As the following example shows, the
number of empty rectangle does not grow linear with the number of compo-
nents included. Whenever a new component is placed on large, depending on
the configuration.

Example 5.14 Considere the configuration shwon in figure 5.7(a) with 11

MERs ((D,I), (C,H), (A,P), (A,O), (F,O), (E,N), (Q,G), (Q,H), (E,K), (M,I),
(L,J)) before the placement of a new component of the device. After the place-
ment of module v4 , the number of MERs grows to 14((D,I), (C,H), (N,P), (A,U),
(N,O), (A,S), (F,S), (E,R), (E,K), (L,H), (L,J), (M,S), (V,I), (T,I)) thus increasing
the complexity by a factor more than the number of included modules.

(a) The configuration contains 11 modules (b) The number of modules grows to 14 after
befor the insertion of a new task the insertion of a new task v4

Figure 5.7. Increase of the number of MER through the insertion of a new component.

Also the number of free rectangles can be drastically reduced after the removal
of a module running on the chip. The insertion of components on the chip as
well as their later removal creates large fluctuations on the number of empty
rectangles to be managed, thus increasing the complexity of the alforithm. In
[116], the run-time of the free rectangle placement is shown to be O(n2 )
In order to avoid the quadratic run-time of the KAMER, a simple heuristic
was proposed in [18], with a lower quality and a linear run-time. The strategy
consist of keeping only the non overlapping empty rectangles, thus reducing
the number of empty rectangles to be managed. This allows a linear-run time
at the cost of the quality. The problem with this approach is that several non

2 If a free space exist in which the module can be placed, then there must be a free rectangle that can

accommodate that module and therefore, the MER approach will find a placement
164 RECONFIGURABLE COMPUTING

overlapping rectangles representation of a configuration may exists leading to

multiple possible choices in the representation as illustrated in 5.8. Shown
are two different representations of the free space with different sets of non
overlapping free rectangles.

Figure 5.8. Two different non-overlaping rectangles representations

Because the non overlapping empty rectangles are not necessary maximal,
a module may exists that could fit onto the device, but cannot be placed due to
a a bad non overlapping representation.
As shown in figure 5.9, whenever a new component v1 is placed in a non
overlapping rectangle, two possibilities exist to split that rectangle. An hor-
izontal split that is done using segment Sa and and a vertical split using the
segment Sb .

Figure 5.9. Splitting alternatives after new insertion

Choosing to select one the two split directions may have a negative inpact
on the placement of the next components. Assuming for example that the al-
gorithm keeps selecting the horizontal split, after the placement of module v1
(figure 5.9), the free rectangles left are (A,D) and (C,E). Assumes now that a
new module, whose width is the same as (C,E) but with a height slitghly big-
Temporal Placement 165

ger than that of (C,E) is chosen. The algorithm will not be able to place the
component, although enough free space is available on the chip to accomodate
the new module. Bazargan et al. proposed several strategies for choosing one
split direction with the goal is to favor those splits that create quadratic space
on the chip. In order to increase the quality of the non overlapping free rect-
angle heuristic, a strategy is proposed in [197] that simply consist of delaying
the split decision for a number of step later. This may allow a bad choice that
could have done earlier to be avoid.

While the KAMER algorithm always find a rectangle to place a new com-
ponent, if one exists, the position to place the component within the rectangle
must be selected from a set of points whose number is the area of the rectangle
in worst case. An optimal algorithm must choose one point from the set of
possible points according to some optimization criteria. In [116], the bottom
left position is choosen. Because no relation exists between the rectangle to be
placed and those already placed, an arbitrary position can be choosen for the
new rectangle. In reconfigurable devices where the communication between
pairs of tasks one one side and between a task and the device boundary on the
other side plays an important role, components should be placed in such a way
that the communication will be efficiently realized. If a position in the middle
of the selected rectangle is better adapted for optimizing the gooal seeked, then
the component should be placed there, no mater if the number of empty rectan-
gle increases. An optimal algorithm should considere any single point where
the placement is possible and develop a strategy to choose the best position.
This approach is followed by the next algorithms that we present. Instead of
managing the free space, the strategy consists of managing the occupied space
on the device and use the set of running components for computing the best
position to place the incomming component.

4. Managing the Device’s Occupied Space

The observation that has guided the developmnent of this approach is that
the number of placed modules on the device grows lineary contrary to the
number of free rectangles that has a quadratic grows. Each insertion of a new
component on the chip increases the number of placed components by only
one, the same as any removal of a component decrease the list by only one.
By managing the number of placed modules we do not face large fluctuations
in the list of modules like it is the case with empty rectangles. It is therefore
possible to develop a linear time optimal algorithm for the on-line placement
problem. With this, the temporal placement can be splitted in two subproblems
as first presented in [3].

1 Identify the set of potential sites to place the new component.

166 RECONFIGURABLE COMPUTING

2 Select the best site to place the component according to a set of given cri-
teria.
Assuming that a set of possible placement sites is identified, several criteria
can be used to choose between the feasible positions. Here, we considere the
communication cost betwen the task and its environment as the objective to
be minimized. The connections between two different components as well
as those between modules on the device and the device boundrary are of great
importance. While the first one allow two module to communicate together, the
second one allows a module within the device to communicate with a module
out of the device.
The straightforward way to solve this subproblem is to use a brute force
Algorithm. For each new component c to be placed, the brute force algorithm
solves the first subproblem by scanning all the positions on the device. For
each position p = (xp , yp ), it checks if an overlapping will occur between c
and a placed module, if the component c is placed at location p.
Having solved subproblem through the scanning of all possible positions,
the optimal placement position is computed from the set of feasible positions
by computing the placement cost for each of the locations found in the first
step and then select the best one as the optimal solution.
The Brute force requires O(H × W × n) time to solve subproblem 1. H
being the height, W the width of the reconfigurable device and n is the number
of running tasks on the hardware. This approach is not practicable for large
reconfigurable devices.
Without loss of generality, we will considere that components are placed
relatively to their lower left positions. Later, we will use the middle of the
device as reference point. We next provide some definitions that are important
to understand the placement strategy explained in this section.
Definition 5.15 (IPR relative to placed modules) For a new com-
ponent v to be placed on the device and a placed component v ′ , the Impossible
Placement Region (IPR) Iv′ (v) of v relative to v ′ is the region on the chip,
where v cannot be placed without overlapping with v ′ .

For a set V ′ of placed components, the impossible placement region IV ′ (v)

of v relative to the set V ′ is the region on the chip, where
S v cannot be placed
′
without overlapping with an element of V : IV ′ (v) = v′ ∈V ′ Iv′ (v).
Besides the impossible placement region relative to components of the de-
vice, an impossible placement region relative to the device itself exists and is
defined as follow:
Definition 5.16 (IPR relative to the device) The Impossible Place-
ment Region IH (v) of a module v relative to a reconfigurable device H is the
Temporal Placement 167

region of the device where v cannot be placed without overlapping with the
external area of the device.

Having defined the impossible placement region relative to the device and
that relative to other components on the chip, we can now define the overall
impossible placement region of a component on a device with a running set of
components.

Definition 5.17 (Impossible and Possible Placement Regions)

The impossible placement region IPR I(v) of a new component v on a recon-
figurable device H is the region region on the device, where v is not allowed
to be placed, if the placement has to be valid. I(v) = IV ′ (v) ∪ IH (v), where
V ′ is the set of all the components running on the chip

The possible placement region (PPR) P (v) of a module v is obtained by

subtracting the IPR of the module from device from the device area. If L is the
set of all locations on the device, then P (v) = L − I(v).

As illustrated in figure figure 5.10, for a new component v with height hv

and width wv to be placed on the device and a placed component v ′ , the Im-
possible Placement Region (IPR) Iv′ (v) of v relative to v ′ can be computed by
first defining a left side margin with size wv − 1 and a bottom margin with size
hv − 1 for the component v ′ .

Figure 5.10. IPR of a new module v relative to a placed module v ′

The IPR of v relative to v ′ is the sum of the augmented margin and the area
of v ′ .
The computation of the IPR relative to the device is slightly different from
that of a component. Instead of computing a left side and a bottom margin,
we compute a right side and an upper margin with. The right side margin has
the width wv − 1, while the upper margin has the height hv − 1. Figure 5.11
168 RECONFIGURABLE COMPUTING

Figure 5.11. Impossible and possible placement region of a component v prior to its insertion

show the impossible placement region of a component v on a device with four

running components v1 , v2 , v3 , v4 .
Computing the IRP of a given module can be done in linear time O(n),
where n is the number of components currently running on the chip. For each
component, a constant amount of operations is required to identify the IPR rel-
ative to that component.

Now that we know how to compute the region (set of points) where the
component can be placed, the next sub-problem consists of choosing the best
location for the new component. The position where the new component must
be placed should be a feasible position that minimizes the ommunication cost.
A simple approach consists of scanning all the possible placement positions,
and to compute the communication cost for each position and then selection the
optimal one. This straightforward but inneficient approach requires O(|P P R|∗
n) where n is the number of placed components and |P P R| is the size (number
of point) of the possible placement region. This can be very large according to
the current configuration of the device.
Alternatively, we can first compute the point popt which gives us the optimal
placement cost. If popt is located within the PPR of v, then we have the solu-
tion. Otherwise, the optimal position is not in the PPR and we should choose
the closest point to popt which is located in the PPR and select it as the optimal
placement position.
Temporal Placement 169

The definition of the optimal point is done according to the communication

cost that can be minimized by placing the new component closed to other com-
ponents with which it communicates. If the component to be placed has off-
chip connections, then the boundary of the device that connects to the comp-
ponent is assumed to be a placed module at the corresponding location. The
communication cost that we also called routing cost is therefore formaly de-
fined in terms of weighted distance between modules and device boundary. We
considere that routing cost between two modules is defined by the Euclid dis-
tance. Later we will investigate the Manhatan distance that is better adapted
for the routing cost on a given chip.
Definition 5.18 (Routing Cost) For two modules vi and vj , we define
the routing cost between modules vi and vj as follows:
RoutingCost(Rij ) =
w h
((xj + 2j − xi − w2i )2 + (yj + 2j − yi − h2i )2 ) × wij
In other words, the routing cost between two modules is the weighted Euclid
distance between the two modules. In order to compute the routing cost we use
the center point of the components instead of using the bottom left corner point
as the reference point. For module vi , this is defined by the pair (xi +wi /2, y +
hi /2) where (xi , yi ) define the lower left position of vi , wi the width and hi
the height of vi . If there is no communication between two modules vi , vj ,
then the routing cost between the two modules is zero.
With (n − 1) modules placed on the device and a n-th module to be placed,
the routing cost is the sum of the routing cost between the n-th module and
each of the n − 1 modules running on the chip. The optimal point is therefore
the one, which provides a solution to equation 4.1:

n−1
X wn wi hn hi
min{ (((xn + − xi − )2 + (yn + − yi − )2 ) × win )} (4.1)
2 2 2 2
i=1

In Equation 4.1, xn and yn are variables and other parameters are fixed. Be-
cause xn and yn are positive and independent from each other, we can replace
equation 4.1 by the two equations 4.2 and 4.3.
n−1
X wn wi
min{ ((xn + − xi − )2 × win )} (4.2)
2 2
i=1
n−1
X hn hi
min{ ((yn + − yi − )2 ) × win )} (4.3)
2 2
i=1
The minimums can be computed through the partial derivative of the sums in
equations 4.2 and 4.3. We therefore have for xn
170 RECONFIGURABLE COMPUTING
Pn−1 wn wi 2
∂{ i=1 ((xn + 2 − xi − 2 ) × win )}
=0 (4.4)
∂xn
This leads to the optimal value:
Pn−1
win (xi + wi − wn
2 )
xn = i=1 Pn−1 2 (4.5)
i=1 win

In a similar way, we can compute the optimal value for yn :

Pn−1 hi hn
i=1 win (yi + 2 − 2 )
yn = Pn−1 (4.6)
i=1 win

The computation of this optimal weighted sum of Euclid distance is fast.

However the computed point may fall into the IPR of the component. If this is
case, we need a strategy to quickly move out of the IPR and select the closest
feasible point to optimal. As shown in figure 5.12, there are four nearest possi-
ble points (NPP) associated with a component in whose IPR the optimal point
falls. Four comparisons are necessary for selecting the closest one.

Figure 5.12. Nearest possible feasible point of an optimal location that falls within the IPR

In the worst case, we may face the situation, where several IPRs overlap
such that moving out of the IPR related to a component bring us in the next
IPR of another component. In this case we simple keep moving out of the IPR
untill we get a point in the PPR. However the price of doing such a sucessive
move can be very high. A situation could be constructed such that we have to
go through all IPRs, thus creating a uadratic run-time O(n2 ). Because we need
O(n) time to compute all the IPRs and O(n2 ) in the worst case for moving out
of all IPRs if the computed optimal point falls within one of them, the run-time
of the algorithm is quadratic.
Algorithm 11 provides a pseudo code for the approach presented here.
Temporal Placement 171

Algorithm 11 Pseudo code of the on-line placement algorithm

1: ComputeOptimalPoint(new component)
2: if OptimalPoint is feasible then
3: place the component at the optimal point
4: else
5: Find the 4 near points outside the actual IPR, in which the selected point
is located
6: Insert these points into the optimal points list
7: Select the closest point to the optimal point from the near points list
8: if the selected point is feasible then
9: place the component at the optimal point
10: else
11: remove it from the list, and goto 5
12: end if
13: end if

The implementation of this method can be done using a linked-list of size

O(n) (for n running tasks) to store the placed and running modules on the
reconfigurable device.
A second structure that can be used is a two-dimensional matrix with the
same size as that of the reconfigurable device. This matrix represents the state
of the device. An entry in the matrix defines the value of the point with the
same coordinate on the device. Whenever a point on the device is occupied
by a module, the corresponding element in the matrix will have a pointer to
the related module. Otherwise, the point in the matrix shows its emptiness.
Also for identifying PPR, the effects of extending and deleting margins will
be applied on this matrix. Using the matrix allows to access each element and
obtain its state in just one computing step.
For finding optimal or near optimal point for placing a new module, a linked
list for saving near optimal points can be used. The next element to be removed
from the list is always the one closest to the optimal point. If the removed point
is feasible (this is tested by checking the corresponding entry in the matrix),
then we have the NPP, otherwise the selected point is occupied by a running
module or its margin. In this case, the nearest points, out of the border of that
running module are inserted into the list.

Example 5.19 Figure 5.13 shows an example of recursive move out of the
IPRs. The optimal point falls in an IPR relative to module A. The four nearest
points inserted in the list are 1, 2, 3 and 4. The closet point to the optimal is 4
that is next selected but falls in the IPR of module B.
172 RECONFIGURABLE COMPUTING

Figure 5.13. Moving out of consecutive overlapping IPRs

The points 5, 6 and 7 are inserted into the list. The closest point to the
optimum after the removal of the point 4 is the point 1 that is selected and
maintained as solution, because it is in the PPR.

The approach presented here has the drawback of moving from IPR to IPR
untill a valid placement point is found. In the normal case, we need only a
few number of steps. However the worst case still remain and in the case it
happens, the penalty is a quadratic run-time. A better characterization of the
total set of IPRs can help to improve the efficiency as well as the quality of the
method. This is the topic of the next section.

4.1 Using the Countour of Rectangles

The main drawback of the approach previously presented is the delay that
may arise, when moving a point out of the IPRs. This is due to the fact that
the IPRs of all components are considered separately. If we could merge the
complete IPRs of that component in only one region, then moving out of an
occupied region can be done in constant time, thus improving the run-time of
the algorithm. Several IPRs can be merged in only by computing the contour
of the IPRs.
Like in the previous section, the approach approach presented here is based
on the observation that the occupied space consists of very simple geometric
Temporal Placement 173

objects, namely n placed rectangular modules. The difference is in the use of

the Manhattan Distance metric for capturing the exact routing costs.
The method was first presented in [2] as extension of the work in [3]. The
management of the occupied-space is done through a modification of the well-
known algorithm ContourOfUnionOfRectangles (CUR) [106] [147]
[179], for finding the contour of a union of axis-parallel rectangles. Since the
number of contour segments is linear in n, a running time of O(n log n) can be
achieved. Note that the method does not require the contour to be connected,
it works even on a set of disjoint contours.
in contrast to the previous section and without loss of generality, the impos-
sible placement region of a component v relative to v ′ is now computed with
the reference to the middle point of v, by blowing up the module v ′ by half the
width and half the height of v. The impossible placement region of v relative
to the device H is however computed by shrinking the device’s half the width
and half the height of v.
As defined in the previous section, the new module v is reduced to a point
pv , thus the original problem of finding free space to place a rectangle becomes
the problem of finding a free location to place a point.

Example 5.20 In figure 5.14 the result of the transformation for a new com-
ponent that must be placed is shown. All points out of the impossible placement
regions are feasible placement locations from which the optimal one must be
selected.

Figure 5.14. Expanding existing modules and shrinking chip area and the new
174 RECONFIGURABLE COMPUTING

Among all feasible locations, the points on the contour of the occupied
space, i.e those at the boundary of the free space and the occupied space, form
a set from which the nearest to optimal point will be selected, if the optimal
point falls in an occupied region. This help to avoid the recursive jumping
out of the IPRs of the components as presented in the previous section. The
computation of the contour is therefore at the center of our interest.

4.1.1 Computing the Countour of the Occupied Space

In general finding the contour of a set of n axis-aligned rectangles can be
done in O(n log n + s) by using the CUR algorithm as described in [106]
[147] [179]. Here s is the complexity of the resulting contour. The algorithm
presented in [2] is not simply an implementation of CUR. There are a few sub-
tleties that must be considered. All differences stem from the above mentioned
fact, that the points on the contour are feasible locations. As a consequence
the algorithm has to find free space of height and width 0 as shown on figure
5.15. The CUR returns an exact contour of a set of placed rectangle while
the modified algorithm return the exact set of feasible location. Shown is a
set of rectangles (Left), representing expanded modules (dashed, light). At
the center, we have the contour as returned by CUR, and the resulting feasible
space for placing the center point of the new module (dark, together with the
boundary). The right part of the figure shows the correct feasible positions
for placing the center point of the new module as returned by the modified
CUR (dark, with boundary). The difference is a short segment representing
the boundary of two IPR of the two components.

Figure 5.15. Characterization of IPR of a given component: The set of contours (left), the
contour returned by the modified CUR (middle), the contour returned by the CUR (right)

In the following we will describe CUR and the modification done in [2].
Temporal Placement 175

The building block of CUR is an algorithmic technique from computational

geometry called plane sweep and a data structure called segment tree. For an
in-depth introduction to both, refer to [66].
A plane-sweep algorithm is an algorithm that scans the plane and a set of
objects in it: It consit of moving an axis-parallel line in an orthogonal direction
across the plane and keep track of the structure of the intersection with the set
of objects. The key observation is to notice that updates to this structure only
occur at a discrete set of critical positions called events. By pre-sorting these
events (in time Θ(n log n)), only the updates have to be performed, which can
be done efficiently for all events by using an appropriate data structure. For
our purposes, such a data structure is a segment tree: This is a balanced binary
tree for dynamically storing a set of n intervals. The number of endpoints
of these intervals must be known at construction time. Because it is bounded
by 2n the segment tree can be constructed in O(n). Insertion and deletion of
intervals can be done in O(log n). For more details on segment trees, refer to
[24] [22, 179] [23].
One has to be careful when constructing the segment tree. To find free space
of height and width 0 we have to make sure that two modules starting or ending
on the same coordinate are separated by an elementary interval in the segment
tree. This can be done by disturbing the top left corner of each module by a
sufficiently small value ǫ > 0.
The algorithm uses two plane sweeps: one horizontal sweep that discovers
all the vertical contour segments and one vertical sweep that finds all hori-
zontal segments. As the horizontal and the vertical sweep differ only in the
initialization, only the horizontal one will be described.
For the horizontal sweep we construct a list L of 2n quadruples, denoted
by (pi , ti , bi , ei ): for each module vi′ ∈ V ′ , where V ′ represents the set of
extended modules that form the IPRs of the components on the chip, we add
two elements to L — one for the left side (x′i , Open, yi′ , yi′ + h′i ) and one for
the right side (x′i + wi′ , Close, yi′ , yi′ + h′i ). This list is sorted lexicographically
and we assume that Close < Open.
In the sweep we process all 2n elements of L. In case of an Open event
the corresponding contour points are retrieved from the segment tree and the
segment [yj′ , yj′ + h′j ] is added to the tree. For a Close event the segment
[yi′ , yi′ + h′i ] is removed from the tree and the corresponding contour points are
retrieved. Another vertical plane sweep is necessary to discover all horizontal
segments.

4.2 Routing-Conscious Placement

After describing how to find all feasible locations to place a new module,
we turn to finding an optimal location, which minimizes the weighted commu-
nication between the modules. Instead of using the Euclid distance like in the
176 RECONFIGURABLE COMPUTING

previous section, the Manhattan Distance is used. The Euclid distance between
two components defines the shortest distance between two points. However,
this shortest distance may be a diagonal line that cannot be realized as signal
on the chip. The Manhattan Distance captures the smallest routing distance
between two components on the chip. Such a route is made only upon vertical
and horizontal segments weighted by width of the communication segments.
This can still be achieved in time Θ(n log n), making use of local optimal-
ity properties, the occupied space manager, and another application of plane
sweep techniques.

4.2.1 Coomunication Cost

Given a reconfigurable device H and a compopnent v, the objective of the
placer is to find a point in free space that minimizes communication cost for
the new module v.

1
2
m 2

Figure 5.16. Placement of a component on the chip (left) guided by the communication with
its environement (right).

Like in the previous section, the communication cost of an additional mod-

ule m is the sum of the Manhattan distance between the center point of m and
the center points of the modules that communicate with m.
Also, we consider communication with the chip boundary as indicated in
Figure 5.16. We therefore have to consider the set C = {(x1 , y1 ), . . . , (xk , yk )}
of demand points for communication. We assume that the number k of con-
nections to be established with module m is linear in n, i.e the new component
is connected to a running component by a maximum of one link. Obviously,
the width wij of the communication link between modules i and j must be
considered, thus, we get the objective function
k
X
min{c(xi , yi ) = wij k(xi , yi ) − (xj , yj )k1 }
j=1
Temporal Placement 177

Because we are dealing with the Manhattan metric, this can be reformulated to

c(x, y) = cx (x) + cy (y)

with

k
X
cx (x) = wij |xi − x|,
i=1

and

k
X
cy (y) = wij |yi − y|.
i=1

If we would allow vi to be placed anywhere on the chip this can be reformu-

lated to

li li
min{cx (xi ) : xi ∈ [ , Hl − ]}+
2 2
y hi hi
min{c (yi ) : yj ∈ [ , Hh − ]}.
2 2

As a consequence we can consider two separate minimization problems, one

for each coordinate. If we ignore feasibility, both minima are attained in the
respective weighted medians, so they can be computed in linear time [27]; as
we already sort the coordinates for performing plane sweep, this running time
is not critical, so we may as well use a trivial method. Note that only medians
satisfy unconstrained local optimality, as the gradients for cx and cy are simply
X X
∇cx (x) = wij − wij ,
xi <x xi >x

the sum of the required bandwidths to the left minus the sum of the required
bandwidths to the right and
X X
∇cy (y) = wij − wij ,
yi <y yi >y

the sum of the required buswidths to the bottom minus the sum of the required
buswidths to the top.
178 RECONFIGURABLE COMPUTING

4.2.2 Local Optimality

The optimal point that produces the minimal routing cost with the Manhat-
tan distance is called the median. It is the global optimum that may falls in
the impossible placement region. If it is in the occupied space there are only
two other types of points where the nearest possible point to optimum could be
located.

Median

Figure 5.17. Computation of the union of contours. The point on the boundary represent the
potentially moves from the median out of the IPR.

All points of the first type can be found by intersecting the contour of the
occupied space with the median axes lx = {(xmed , y) : y ∈ [0, Hl ]} and
ly = {(x, ymed ) : x ∈ [0, Hh ]}. In these points one of the gradients ∇cx and
∇cy vanishes. We cannot move in the direction of a better solution because
that way is blocked by either a vertical or a horizontal segment of the contour.
The second type of points are some of the vertices of the contour. These
points are the intersections of horizontal and vertical segments forming an in-
terior angle of π2 pointing in the direction of the median. In these points neither
of the gradients vanishes. Either of the directions indicated by the gradient are
blocked by contour segments.
By simply inspecting all the local optima one finds the one closest to the
global optimum. In the next subsection we describe how this can be done
efficiently.

4.2.3 Algorithm for Global Optimality

The occupied space manager described in the previous section computes the
contour of the occupied space in O(n log n). A simple algorithm that finds the
optimal point to place a new module v would compute the median and check
its feasibility; if the outcome is positive, we have found the optimum. Other-
wise we need to check communication cost for all other possible local minima,
Temporal Placement 179

i.e., for every vertex of the contour and every intersection point of the contour
with one of the median axes. Let L denote this set of points. Computing the
communication cost for a single point takes O(n), so evaluating all objective
values in a brute-force manner would take O(n2 ) time. However, by means of
two more plane sweeps, we can achieve a complexity of O(n log n).
For this purpose, we observed that communication cost for the x- and y-
coordinate of the contour segments can be computed separately, then add the
precomputed values for every point of L. The crucial step is to use the fact that
we only need to compute the communication cost for the leftmost x-coordinate
and for the bottommost y-coordinate; the other values can be obtained by doing
appropriate fast updates during the plane sweep. B

5. Conclusion
This chapter has provided some of the most important approaches for tem-
poral placements. While very solid mathematical backgrounds were developed
and simulated, a physical investigation on reconfigurable chips has failed so
far. This is mainly due to the lack of devices providing partial reconfiguration
facilities, a precondition for the implementation of temporal placement algo-
rithms. One and almost the only device type to provide partial reconfiguration
is very limited and does not allows for evaluating the large set of available al-
gorithms. Besides the difficulty to designed for partial reconfiguration, many
restriction are done on the use of the device resources during and after the
reconfiguration. Some hope were placed on coarse-grained reconfigurable de-
vices. However, they have failed so far to take-off.
Chapter 6

ON-LINE COMMUNICATION

Several methods were presented in the last chapter for the placement of
components at run-time on a reconfigurable device. The last two ones use the
communication costs between a component to be placed and its environments,
i.e the set of modules on the chip and out of the chip that exchange data with
this component, as a mean to select the best position for the new component.
While this communication costs, defined as the average minimum distance of
the new component to its neighbor is useful to guide the placement process,
it does not tell us how data should be exchanged among the different compo-
nents on the chip. In this chapter, we provide some answers to this question by
presenting some approaches to enable the communication at run-time between
modules on the chip. The approaches can be classified on different categories,
depending on the way the communication is realized. We can cite the direct in-
terconnection, the communication over third party, bus-based communication,
the circuit switching approach and network on chip oriented communication.

1. Direct Communication
Direct communication paradigm allows modules placed on the chip to com-
municate using dedicated physical channels, configured at compile-time. The
configuration of the channels remains until the next full reconfiguration of the
device. A configuration defines the set of physical lines to be used, their di-
rection, their bandwidth and speed as well as the terminal, i.e the components
that are connected by the lines. Component must be designed and placed on
the device in such a way that their ports can be connected to the predefined
terminals. Feedthrough channels must also be available in each component to
allow signal used by modules aside the component to cross the components.
182 RECONFIGURABLE COMPUTING

Example 6.1 The configuration of figure 6.1 provides an example of 1-D

communication using direct lines. For this purpose, a set of 10 predefine chan-
nels is fixed and the component must adapt their position and direction to make
use of the channels. The physical channels 1, 5, 6 and 7 are not used. Line 3
is used by module C2 for connection with the device pins on the left and right
sides. It is fed through component C1 which provides the necessary channel
for the signal to cross. Line 4 is used by the two components C1 and C2 for a
direct communication. Additionally, the two components use the same line to
access the device pins. Line 8 and 10 are used by component C5 to access the
device pins on the left and right sides. They cross components C3 and C4. Line
9 is used to connect the components C3 and C5 and runs through component
C5.

Figure 6.1. Direct communication between placed modules on a reconfigurable device

The main disadvantage of this approach is the restriction imposed on the de-
sign of components. For each component, dedicated channels must be foreseen
to allow signals that are not used by this component in its placement location
to cross. This increases the amount of resources needed to implement the com-
ponent. Also, the placement algorithm must deal with additional restrictions
like the availability of signals in a given location. This increases its complexity
and makes the approach only possible an off-line temporal placement, where
all the configurations can be defined and implemented at compile time.

2. Communication Over Third Party

Communication over third party is used for example by Brebner in [37] and
Walder et al. in [211] [196] [212]. A central component exists that behave as
a message reflector. Each message is first sent to the central module, which
forward it to the destination. Brebner in [37] uses this approach to allow the
communication between a reconfigurable module connected to a processor via
On-line Communication 183

a bus and a user programm running on the host processor. The module inputs
and outputs are controlled by registers that are mapped into the address space of
the processor. This approach can be used not only to allow the communication
between a reconfigurable module and a user program, but also between several
reconfigurable modules connected together via a bus. The system is controlled
by an operating system, whose role is to manage the device resources, control
the reconfiguration process and allow the communication to happen between
the components temporally placed on, or removed from the device. All mod-
ules willing to send a message must first copy those messages in their sending
register. Thereafter, the operating system copies the message from those reg-
isters to the input register of the destination module. In [211] [212] the central
module is not an operating system running on a separate processor, but a set of
fixe resources on the device. It provides dedicated channels to access periph-
eral devices and also to connect direct neighbor modules. The communication
between non adjacent modules is done via fixed resources implemented on the
device.

3. Bus-based Communication
The communication between the reconfigurable modules on a given device
can also be done using a common bus. In order to avoid the bus resources
to be destroyed at run-time by components dynamically placed on the device,
predefined slots must be available, where the modules can be placed at run-
time, such that no alteration of the bus is possible. On the predefined slots,
connection ports must be available to dynamically attach the placed component
to the bus. While the predefinition of locations where to place components
allows a simplicity of the placement algorithms, it is not flexible at all. Using
a common bus reduces the amount of resources needed in the system since
only one medium is required for all component. However, the additional delay
increase by the bus arbitration can drastically affect the performance of the
system. The approaches of Walder et al. In [211] [212] as well as that of
Brebner are both based on restricted bus in which no arbitration module is
requirement to manage the bus access.

4. Circuit Switching
Introduced in the 80’s under the name reconfigurable massively parallel
computers [146], [205], circuit switching is the art of dynamically establishing
a connection between two processing elements (PE) at run-time, using a set
of physical lines connected by switches. The system consists of a set of pro-
cessing elements arranged in a mesh. Switches are available at the column and
line intersection to allow a longer connection using the vertical and horizontal
lines at an intersection point. In this way two arbitrary processing elements
184 RECONFIGURABLE COMPUTING

can be dynamically connected at run-time by dynamically setting the switches

on the path from the first processor to the second one. Once the connection is
established, the data is transmitted from the source PE to the destination in just
one clock.
Circuit routing was experienced in several systems like the YUPPIE [146]
in the 80s. It is available today in many systems like the PACT-XPP device
[176]. Also the connection mechanism in FPGA follows the same paradigm.
Although dedicated lines exist in some devices to allow connection between
remode modules on the device, longer connections are normally established
using the switch matrices to connect segmented lines in other to realize the
long communication path.
Despite its application in fine-grained parallel image computing systems,
where it helped to dynamically change the topology of a parallel computer
in order to accommodate the best computation structure of an application un-
der computation, using the circuit switching in reconfigurable devices presents
some drawbacks:
Long communication delay: This will happen, if the connection between
two processing elements has to go through many other processors. The
number of switches on the communication path therefore increases, thus
increasing communication delay and therefore slowing down the clock.
Dynamic computation of routes: In dynamically changing environment,
the communication need between components changes with the placement
of components. Computing new routes at run-time is very time-consuming
and therefore, the performance of the overall computation may be drasti-
cally affected by the run-time routing.
Exclusive use of chip space: Due to the complexity of the synthesis al-
gorithm, components that will be downloaded later onto the reconfigurable
device are implemented at compile time. The synthesis process constrains a
module on a given area of the device, where it uses all the resources, i.e the
processing elements and the interconnects in that area. This has the conse-
quence that the placement of a module at run-time in a region, where a route
is running, will destroy the route, because the connection used by the route
will be assigned to the component. To avoid this, we can place components
only at locations where no route will destroy them, a restriction that will
however increase the chip fragmentation. Consider the placement of figure
6.2 with three routes to connect PE1 to PE2, PE3 to PE4 and PE5 to PE6.
For a new component that needs to use four consecutive processing ele-
ments in a quadratic surface, the region where the routes are implemented
must be prohibited. No region that can accommodate the new component is
therefore left on the device and the component must be rejected, although
enough resources are available on the device.
On-line Communication 185

Figure 6.2. Drawback of circuit switching in temporal placement Placing a component using
4 PEs will not be possible, although enough free resources are available

The previous example has shown how circuit switching might increase the
device fragmentation in devices allowing a 2-D placement. In device allowing
only a 1-D placement like it is the case of Xilinx VirtexII FPGAs, which are
only column wise reconfigurable, circuit switching can be used to connect a
few number of modules, usually, 2 to 8, and allow dynamic communication to
be established between the components running onto the device at run-time.

4.1 1-D Circuit Switching: The RMB

In [4] [32] a 1-D circuit switching modell called the Reconfigurable Multi-
ple Bus (RMB), well suited for the column wise reconfigurable Xilinx Virtex
FPGA is designed and implemented. The approach is based on a concept pre-
viously introduced in [204] and [77].
As figure 6.3 shows, the communication structure is a made upon a set of
switches, locally attached to a component using vertical lines. The connec-
tion between the switches is done through a bus, which consists of a set of
segmented horizontal lines. The bandwidth of the horizontal connection may
varies according to the application to be implemented. The function of the
switches is to allow the component on which they are connected to dynami-
cally access the bus and establish communication with other components.
The request for a new connection is done in a wormhole fashion. A sender
components Pk , located in position k and willing to establish a communication
with a receiver component Pt located in position t, must first send a request
for communication to its own switch at location k. Switch k forwards the
request to switch k + 1 in the direction of the receiver, which in turn forward
the request to switch k + 2, until the request arrives at the switch t. Each
switch on the path from the sender to the receiver checks if the request for a
186 RECONFIGURABLE COMPUTING

Figure 6.3. The RMBoC architecture

communication lines can be satisfied, i.e. if a free communication channel is

still available on the switch. If this is not the case, then the request is either
queued or rejected. Otherwise, the switch sets the connection and sends an
acknowledge. The first acknowledge is generated by the last switch on a path
(in this case switch t) and sent back to switch t − 1, which in turn send the
acknowledge to switch t − 2, the process is repeated until the switch at the
sender location. Upon receiving the acknowledge, the sender can start with the
communication. This approach was first presented in [204] and extended later
in [77] with a compaction mechanisms for a quick finding of a free segment.
A redesign as well as an implementation of this approach, called the RMBoC
(RMB on Chip) for the Xilinx FPGA devices was presented in [4], [32].
In of the RMBoC on the Xilinx Virtex FPGAs, the switch controllers are
separated from the processor. In this way, a uniform interface is provided to
designers for connecting their modules on the Bus.
In order to facilitate the partial reconfiguration on Xilinx FPGA devices,
the device is segmented in a set of horizontal slots, each of which can accom-
modate a given module at run-time. The reconfigurable modules can then be
placed only on the foreseen slots. Large modules that cannot fit on one slot can
be implemented using a given amount of consecutive slots. The segmentation
in slots is enforced using dedicated modules, the so called emphbus macros
(figure 6.4), at the boundaries of the slots. A bus macros is a special hard-
ware module that allows established connections between neighbor component
not to be destroyed during the reconfiguration process. The crosspoint locally
manage the connection of the component to the Bus as well as the setting of
switches according the requests and the availability of the channels.
On-line Communication 187

Figure 6.4. RMBoC FPGA implementation

As depicted in figure 6.5, the crosspoint used in the RMBoC consists of

three principal components.
Controller: The controller manages the switch at a local level according to
requests from the left or right crosspoint as well as that of the local module.
Four kinds of command (REQUEST, REPLY, CANCEL and DESTROY)
can be locally processed on a single crosspoint, thus defining the interface
between a processing node and its crosspoint.

 


 

       
   

   

    

   

  

Figure 6.5. Crosspoint architecture

A communication process starts by a REQUEST command from the sender

to its local corresponding crosspoint with the destination address. The com-
mand is successively sent to the next crosspoint in the destination direction
until the receiver location or the earliest crosspoint with no free channel.
188 RECONFIGURABLE COMPUTING

Upon reaching the destination, a connection will be established and a posi-

tive answer (REPLY) is sent back successively until the source. Each pro-
cessing element which receives a REPLY, sets its switch accordingly. If a
processor cannot establish a connection in the destination direction, a neg-
ative answer (CANCEL) is sent back and the connection will not be estab-
lished. Whenever a connection is successfully established, the communica-
tion is done on a point to point manner. At the end of the communication,
the sender sends a DESTROY command to its crosspoint which succes-
sively pass it to the the crosspoints involved in the communication. Each
crosspoint frees the data channel after sending the DESTROY command to
the next crosspoint.
Data network: The function of data network is just to connect corre-
sponding data channels according to the configurations modified by the
controller. The original RMB foreseen that data must be transfered within
one clock cycle from source to destination, once the connection is estab-
lished. However, to avoid long delay and therefore slow clocks, a pipelined
communication is realized in the RMBoC, using registers between the slots.
FIFOs: The purpose of FIFOs is to provide buffer for commands coming
from different side of the crosspoint (left, right, local). The FIFO Selector
sends the command from the FIFOs of each side to the main FIFO. The
policy of arbitration in the FIFO Selector is Round-Robin in the following
order: Left, Right, and Local.

5. Network on Chip
Many well reputed authors [21] [117][60] have predicted that wiring mod-
ules on chip will not be a viable solution in the billion transistor chips in the
future. Instead, they proposed Network on Chip (NoC) as a good solution
to support communication on System on Chip in the future. NoCs encounter
many advantages (performance, structure and modularity)toward global signal
wiring. A chip employing a NoC consists of a set of network clients like DSP,
memory, peripheral controller, custom logic that communicate on a packet base
instead of using direct connection. As shown on figure 6.6, several modules
(network client) placed at fixed locations on the chip, can exchange packets in
the common network. This provides a very high flexibility, since no route has
to be computed before allowing components to start communicating. Compo-
nents just send packets and they don’t care on how the packets are routed in the
network. Network on chip (NoC) is viewed as the ultimate solution to avoid
problems which will arise due to the growing size of the chip. The Quicksilver
chip [89] can be seen as an example of NoC-based architecture.
A generic NoC architecture is characterized by the number of routers, each
of which is attached to processing elements in the array, the bandwidth of
On-line Communication 189

Figure 6.6. A Network on Chip on a 2-D Mesh

the communication channels between the routers, the topology of the network,
and the mechanism used for packet forwarding. The topology of the network is
defined through the arrangement of routers and processor on the device and the
way those processors are connected together. One of the most used topology
is the 2-D mesh network, because it naturally fits the tile-based architecture on
the chip.
Compared to typical macro networks, a network on chip is by far more
resource limited. To minimize the implementation cost, the network should be
implemented with little area overhead. This is especially important for those
architectures composed of tiles with fine level granularity as it is the case in
FPGAs. Thus, instead of having huge memories (e.g. SRAM or DRAM) as
buffer space for the routers like in the macro network, only few registers are
available to be used as buffers for on-chip routers. This leads to a much simpler
model with little overhead, compared to its macro network peer.
We next take a look on the design and implementation of major NoC com-
ponents, which are the routers and the processing elements.

5.1 The Router

The basic building blocks of a NoC are routers that consist of a given set
of components like buffers to temporarily store packets, a controller that deter-
mines how to forward the packet, usually according to the destination address.
The performance and the efficiency of the NoC depend on the underlying com-
190 RECONFIGURABLE COMPUTING

munication infrastructure, which in turn depends on the performance (latency

and throughput) of the on-chip routers. Thus, the design of efficient, high
performance routers represents a critical issue for the success of the NoC ap-
proach. While several possibilities may exist to implement a given router, we
focus on a very simple and intuitive model consisting, as depicted in figure 6.7,
of the following main components:

Figure 6.7. Router Architecture

1 Five Input buffers, usually implemented as FIFOs, to temporally store mes-

sages coming from five directions left, right, north, south and local PE,
before forwarding. Each router willing to send a message in a given direc-
tion copies it into the FIFO of the neighbor router in the direction where the
message will be forwarded. The data are therefore placed on the data lines
and the control signals are use to perform the handshaking between the two
neighbor routers.
2 The controller, which is in charge of managing the dataflow inside the
router. It defines in which direction a message should be sent and sets
the signals in the router to allow this to happen.
3 Five Output Arbiters, each for one direction and the local PE. The output
arbiters manage the assignment of the message to output channels.
We now provide a much detailed description of the router components:
On-line Communication 191

5.1.1 The FIFO

A first-in, first-out (FIFO) is a data storage to model a queue in such a way
that the data that comes in first will be read out first. It consist of a set of
registers for storing the data. The registers are organized in a serial way that
allowed data to be shifted from one register to the next one. Data are pushed
(written) on one side and read from the other side. Figure 6.8 provides the
general architecture of a FIFO.

Figure 6.8. A general FIFO Implementation

It consists of two counters to count the written and read registers. One write
enables signals allow data to be written into the registers and one read enable
signal to notify about the readiness of the data from the registers. Two signals
are also used as flags to inform about the fullness and emptiness of the FIFO. If
full, then no more data can be written to the FIFO. An empty FIFO means that
no data is available in the FIFO registers. If a common clock is used for reading
and writing the data then the FIFO is said to be synchronous. Otherwise, there
will be two different clocks for reading and writing the data. In this case,
we said that the FIFO is emphasynchronous. The FIFO can be parameterized
with the data width (number of bits in a register) and FIFO depth (number of
registers in a FIFO).

5.1.2 Router Control

Each router is identified trough its position in the network. This position
is determined by the (x,y)-coordinate of the processing element on which the
router is attached. Messages are sent in packets, each of which consists of the
address of destination router, control bits and the payload (data). A possible
format for the packet being sent in to the network is shown in figure 6.9
192 RECONFIGURABLE COMPUTING

Figure 6.9. General format of a packet

The first part is the packet address, which is used by the router control to
determine the direction where to send the packet. The controller has an address
decoder that decodes the address into (x,y) coordinate of destination router
or PE. In the simple case of XY-Routing as we will see later, a comparator
is used to compare the (x,y) coordinate of the destination PE to that of the
router in order to compute the direction (LOCAL, EAST, WEST, SOUTH, and
NORTH) where the packet will be sent. The packet is then sent out by writing
in the input FIFO of the corresponding neighbor FIFO if this is not full. If the
FIFO in the neighbor router is full, then the router can decide to take some
action. It can for instance block all incoming packets or send the packet in
another direction to decongest a given data line.

5.1.3 Output Arbiter

In order to increase performance in the router, input FIFOs in all directions
must be read concurrently if they are not empty, i.e. there is at least one packet
in a FIFO. After reading the packets in parallel, the control logic decides in
which directions the packets should be sent. The packets are then written into
the neighbor router in the chosen directions. The case where the control de-
cides to forward many packets in the same direction leads to a contention, be-
cause only one output data line is available and many packets cannot be written
at the same time. For a contention free routing, an arbiter must be provided at
every output port. A simple arbiter may consists of a multiplexer and a Fi-
nite State Machine, which will make the decision in a round-robin fashion. If
the data must be written on the local output line, the the order can be EAST,
WEST, SOUTH, and NORTH as shown in figure 6.10
The incoming packets from the EAST will be written before the one com-
ing from the WEST, which in turn is written before the packet coming from
SOUTH that is written before the one coming from NORTH. Incoming packet
from LOCAL are not considered here, because we assume the a packet coming
from a given direction will not be sent back in that direction. This assumption
is only a design decision that can be made or not. However we can choose to
give the router the freedom to decide where to forward packets. In this case,
a packet can be returned in the same direction from where it was sent. This
capability may lead in some routing algorithm to a ping-pong game between
two routers. However it may also help to decongest some over filled channels.
On-line Communication 193

Figure 6.10. Arbiter to control the write access at output data lines

5.2 The Processing Element

A processing element can be either a processor core, a memory block, an
embedded programmable logic or any custom hardware block. Processing el-
ements are usually connected to the upper right router through an interface in
the network. Each Processing element has a unique address, which is the same
address of router to which it is connected. Thus, any PE can be plugged into
the network if its design fits into an available slot. In order to make this pro-
cess easy, a wrapper is usually used to decouple the network activities from
the computation within a processing element. The wrapper controls all the
transaction on the network and provides to the component a simple interface
to access the network.

5.2.1 Wrapper Design

The PE port is connected through a wrapper to the router. The port width
(or channel width) determines the size of the packet (message) that can be
moved in the network through routers. For consistency, wrappers use the same
port width as the router port width. A wrapper consists of only one FIFO at
input port as shown in figure 6.11. The main functionality of the wrapper is
to provide an interface to the processing element or the module connected to
the network. Its logic includes decoding and encoding the received and sent
packets respectively. Incoming packets to the PE carries the address along with
the payload. This address is removed in the wrapper and only the payload is
passed to the PE. In the same way, wrapper gets the payload from the PE, adds
the address of the destination PE to the payload and formats the packet before
giving it to the connected router. Always an incoming packet is written to the
FIFO in the wrapper and an outgoing packet is written to the FIFO in router to
194 RECONFIGURABLE COMPUTING

which the PE is connected. The wrapper reads the packet from the FIFO only
when there is a packet in that FIFO. In the implementation, the processing
element is instantiated as functional block within the wrapper.

Figure 6.11. A general wrapper architecture

5.3 NoC Design Constraints

During the design of a NoC, important decision must be taken according to
a goal seek by the user. each design decision influences a given parameter in
the final implementation. The area overhead, i.e. the amount of resources that
the final design consumes as well as the latency, i.e. how long in average it
takes to route a packet to destination, are the most important parameters that
guide the design process.

5.3.1 Area Overhead

When considering a NoC architecture, the area used by the routing hard-
ware is important. Costumers will probably not be interested in those design,
where the routing hardware occupies 90% of the device area and only 10%
is left for the implementation of the PEs. The area consumed by the routers
depends on the bandwidth requirements, i.e the packet width, the buffer size
and the complexity of the control algorithm. The packet size determines the
width of connection between routers. This makes it directly proportional to
the amount of internal wire required. The buffer size determines the amount of
memory used for storing the packets within the router before forwarding. The
complexity of the routing algorithm determines how much additional resources
the router consumes.

5.3.2 Latency
The latency consists of the time needed to setup a route and the time need
to transfer the payload to destination. In circuit switching, route setup time
On-line Communication 195

includes request and acknowledgment latency but in packet routing there is no

such set up time. Only the address flit takes initial setup time to reach the des-
tination based on the routing algorithm, thereafter for every cycle the data flit
will be delivered to the destination in a deadlock free network. In deadlock-
free store-and-forward packet switching, each packet spends one cycle to pass
through each router. The total number of routers between source and destina-
tion will determine the latency. For example, if the source and destination are
placed at diagonally opposite corners of a 4x4 mesh, each packet being trans-
ferred will take 8 clock cycles to reach the destination. In destination IP block
perspective, for every clock cycle a packet will be received. The latency com-
pletely depends on the chosen routing algorithm (adaptive or deterministic)
and the buffering method.

5.3.3 Performance Metrics

The performance of a network provides a mean to compare two different
implementations of the network. The two main means used are the latency and
the throughput.
emphLatency: This is the time a message need from its source to its des-
tination. It is the difference between the time where the the last packet of
the message arrives at destination and the time when the first packet of the
message is output from the source.
emphThroughput: Often used in conjunction with latency. This is the max-
imum traffic a network can accept per unit of time, typically measured as
bytes or packets per node per cycle.

5.4 Routing Techniques

Routing is one of the most understood area in computer science. A large
variety of routing algorithms exist in the literature, each with its advantages
and drawback. We provide some of the most used class of routing algorithms
in this section.

5.4.1 Circuit Switching

Circuit switching were presented in section 4. This approach allows a com-
munication path to be created from the source to the destination before trans-
mitting any data. The procedure starts by a routing probe traversing the net-
work and reserving links to transmit the data. This routing probe contains the
source and destination addresses. Once the routing probe reaches the destina-
tion address, an acknowledgment is sent back to the source address then the
data are transferred at the full bandwidth of the hardware. The circuit remains
operational until the end of data to be transmitted. The lock on the links may
be released once all the data have reached the destination by sending back an-
196 RECONFIGURABLE COMPUTING

other acknowledgment through the same route to the source. The disadvantage
of this technique is the time required to establish a dedicated link from source
to destination. It can be advantageous when the time to set up the path is min-
imal, compared to the transfer time of the messages, and when long messages
are present, especially continuous data streams.

5.4.2 Store-and-Forward
At each node, the packets are stored in memory and the routing informa-
tion examined to determine which output channel to direct the packet. This is
why the technique is referred to as store-and-forward (SAF). Additionally, by
transferring an entire packet at a time, the latency for a packet is the number of
routers through which the packet must travel multiplied by the time to transfer
the packet between the routers.

5.4.3 Virtual Cut-Through

Virtual cut-through (VCT) was conceptualized to address the deficiency in
SAF-based packet routing, which results from the buffering of messages at
each node along the communication path. Since the routing information is
carried in the header of a packet, the packet should not be stored in the current
node’s memory if an output buffer is available. The packet simply cuts through
the router of the node to an available output channel. This alleviates the need
for an excessive amount of memory along the path of a message, but enough
memory has to be allocated if an output channel is not available. Of course
at high volumes of messages on the network, VCT will behave similarly to
SAF packet switching. The unit of message flow control for VCT is the packet
[184].

5.4.4 Wormhole Routing

Wormhole routing was conceived to address the deficiency in VCT, that is,
if an output channel is not available the packet must be stored in the current
node’s memory. Wormhole routing divides a message into smaller flow-control
digits than packets, which are called flits. Each message contains one header
flit which carries the routing and control information and the remaining data
for the message is stored in data flits. The header flit always goes first to
allocate a path for the data flits. Thus, smaller memory requirements exist
for each node on the network (buffers flits instead of packets). If an output
channel is available, the header flit is routed and the remaining data flits follow
in a pipeline style fashion. During any instance of a message traversing a
network, the flits of a message will be located in multiple routers. This routing
technique got its name in that it looks like a worm traversing the network.
While this method benefits from very low latency and buffer space, blocking
On-line Communication 197

and deadlock problems can occur. Therefore, other techniques, such as virtual
channels, are needed. An advantage of virtual channels is the ability to share a
single physical channel.

5.4.5 Deadlock and Livelock

Deadlock is a situation that occurs when a packet is waiting for an event
that can never happen due to a circular dependence on resources. Livelock, on
the other hand, is a configuration of the network in which packets continue to
move, but never reach their destination.
A routing algorithm should determine the optimal routing path in order to
transport packets from source to destination through the network. The optimal-
ity is usually defined according to the following parameters: high performance,
low overhead, Deadlock and livelock freeness, fault-tolerance and flexibility.
In general, routing can be done in two ways: emphdeterministic routing and
emphadaptive routing.

5.5 Deterministic Routing

A deterministic routing algorithm provides a unique path from a source to
destination. XY-routing also called dimension ordering routing is a simple
deterministic routing algorithm, where messages are transmitted fully in each
dimension, beginning with the lowest dimension available. In a 2-D mesh
network, XY-routing first routes packets along the X-axis. Once it reaches
the destination’s column, the packet is then routed along the Y-axis until the
destination’s line. Therefore, any packet moving in the Y-direction can never
return to the X-direction. The XY-routing algorithm routes the packets based
on the destination address, irrespective of the traffic pattern on the link and
the link delay. The router compares its own address (Xrouter, Yrouter) to the
destination address of a packet. Packets are forwarded based on destination
address (Xdest, Ydest) as follows.

1 If Xrouter < Xdest, the packet ist forwarded in the east direction

2 If Xrouter > Xdest, the packet ist forwarded in the west direction

3 If Xrouter = Xdest and Yrouter > Ydest, the packet is sent to the south of
the current router

4 If Xrouter = Xdest and Yrouter < Ydest, the packet is sent to the north of
the current router

5 If Xrouter = Xdest and Yrouter = Ydest, the packet is sent to the local PE
198 RECONFIGURABLE COMPUTING

5.6 Adaptive Routing

In adaptive routing algorithms, the direction where to send an incoming
packet is not fixed a priori. The routing algorithm may decides to use more
complex approaches to decide on the further direction of a packet. Adaptive
algorithms are usually used to improve the performance in the presence of lo-
calized traffic or to provide fault-tolerance in the network. Packets are not al-
ways routed along the shortest path. A longest routing path may be preferred to
a shortest one, if the longest path provides the best better fault tolerance result
or if deadlock and livelock problems can be avoided. An example of adaptive
routing is the Q-routing, an adaptive routing algorithm based on Q-learning,
which is a form of reinforcement learning [36]. It routes packets based on the
learnt routing information from its neighbors. Initially this algorithm builds
a routing table based on the delivery times (Q values) of the packets to every
router in the network. These delivery times are updated every time a router
forwards a packet for a particular destination, which changes depending on the
traffic at a given time to the destination.
Based on this information, a router can choose an alternative route when
the queues are congested in the intermediate routers, thus resulting in faster
delivery compared to the XY-routing algorithm.
With Reinforcement learning in general and Q-learning in particular, each
router is an autonomous structure which learns with the time the efficient route
to all possible destinations. On communication request, the best router, i.e. the
router through which the fastest communication is possible (from the point of
view of that router) is selected and the packet is sent there. With this method all
the routers need to learn the new structure of the network whenever a change
is done.
With a frequently changeable network, the router will spend most of its
time for learning the new network structure, thus decreasing the network per-
formance.
Adaptive routing algorithms are more complex than deterministic routing
techniques, which means that the amount of resources consumed by the router
that implements adaptive routing is much higher than the amount of resources
consumed by router implementing deterministic routing. Therefore, the com-
plexity of those algorithms does not qualify them to be used on a chip. The
XY-algorithm usually performs well in the practice and routes packets accord-
ing to the lower Manhattan distance.

Network on a chip presents a viable communication infrastructure for com-

munication among task dynamically placed on a reconfigurable device. How-
ever it is still too inflexible to dynamically support the communication among
modules in a changing network. When placing a component on a reconfig-
urable device, we face two possibilities: First, the components fits on one PE,
On-line Communication 199

i.e. the resources needed by the component is no more than that available on
a PE. In this case, we place the component on one PE and attach it to the cor-
responding router to allow a communication with other components. In the
second case, the component does not fit on one PE, i.e. more than the amount
of resources offered by a PE is necessary to implement the component. In this
case, the component will be splitted in pieces, each of which can fit on a PE.
The communication between the pieces placed on different PE will used the
router network for communication. In other words, the communication within
a component boundary must be done using packets that are sent and routed in
the network. This increase the complexity of the module and more resources
are wasted than needed. This situation is best illustrated in the figure 6.12.

A
PE PE PE PE
2

3
1

PE PE PE PE
6

PE PE PE PE

Figure 6.12. Implementation of a large reconfigurable module on a Network on Chip

There, a large module needs twelve PEs to be implemented. It is splitted

in parts, each of which is implemented on one PE. The connection among the
parts is done using routers. A better implementation would connect all the
PE using direct wiring. This has no disadvantage, because the PEs are close
to each other and therefore long connections are avoided. In figure 6.12, the
routers inside the area of the components (1,2,3,4,5,6) become redundant.
In order to overcome this handicap, the concept of dynamic network on chip
that we next presented, was developed. In the dynamic network on chip, the
redundant routers can be used as additional resource to implement an even
bigger module.

6. The Dynamic Network on Chip (DyNoC)

A Dynamic Network on Chip (DyNoC) [35] [32] is a Network on Chip
whose structure can be dynamically changed at run-time. In a DyNoC, a
routers is a programmable element basically configure as router, but that can
be configured at run-time to implement any function that can fit on it. This
200 RECONFIGURABLE COMPUTING

concept provides a mean to overcome the defficieny of network on chip as

communication paradigm in reconfigurable computing (figure 6.12).
Whenever a module is placed at a given location where it uses the resources
of several PEs, the redundant routers within the module boundary can be used
as additional resources for the module. The placed module only needs one
router to access the network. With this the routers in the area of a placed
component are no more accessible by other components in the network. Upon
completion, the module is removed and the network must be reactivated in the
area where the component where previously placed. This can be done quickly,
because the router are programmable components that can be quickly reset to
their basic configuration, i.e. the routers.
With this, we have a network in which some parts may be deactivated at a
given period of time and reactivated in the future. In order for such a network
to efficiently operate, some prerequisites on the communication infrastructure
of the network must be fulfilled. We explain these next.

6.1 Communication infrastructure

The goal is to have a communication infrastructure in which the reachabil-
ity of packets is insured, independent on the changing topology which occurs
when components are placed and removed on the chip. In its basic state, the
communication infrastructure is a normal NoC. Processing elements access the
network via corresponding routers. Additionally, direct communication lines
exist between neighbor PEs. In this way, the routers are only used for com-
munication between non neighbor PEs. With this modification, component
splitted accross several PEs with router inbetween do not need the router any-
more for communication, since direct lines already exist to connect the neigh-
bor PEs. The direct lines can also be modified to connect two PEs not directly
neighbor and allow more flexibility. The cost of realizing such an hybrid archi-
tecture with a network on one side and direct routing on the other side will be
more than that of a normal NoC, however the gain in flexibility is enormous.
Routers in those area where a component is placed become useless and
therefore they can be used to provide more resources (logic and memory) to
the component. The realization of such routers on the current programmable
logic device is too costly, thus making the realization of a DyNoC communi-
cation infrastructure only possible if the router are directly available as hard
macros on the device, as it is the case with the switch matrices, the embedded
multipliers or the block memories on FPGA devices. Whenever a component
is placed in a given region, only one router is necessary for this element to ac-
cess the network. We use without loss of generality the router attached to the
upper right PE of the module (figure 6.12).
Figure 6.13 shows a communication infrastructure as described below on
a reconfigurable device. Note that the arrangement of the router is done in
On-line Communication 201

Figure 6.13. The communication infrastructure on a DyNoC

such a way that the device is surrounded by a ring of routers. This arrange-
ment increases the flexibility and, as we will see later, is a prerequisite for a
accessibility of placed components.
A further requirements on the structure of the network is that routers should
be able to notify their neighbors about their activity. This can be done using
an additional emphactivation line that is set to one if the router is active, i.e.
no component is placed on top of that router and zero if a component is using
that router in its internal implementation. In this case, the activation line is
controlled by the component, which notifies the routers around on the states of
the routers that it internally uses.

6.2 Network access

As state earlier we would like to have a network in which all component
are reachable, despite the dynamic modification of the network. Let first recall
the way component are developed for partial reconfiguration. Each task that
can be executed at run-time is implemented as a component, represented by a
rectangular box and stored in a database. The bounding box delimits the area of
resources used by that component. In fact the component is available as partial
bitstream that will be downloaded onto the device at run-time, whenever the
corresponding task needs to be executed.
Those components are hardware modules that required to be synthesize
from a given description. Since synthesis is a time consuming task, it cannot
202 RECONFIGURABLE COMPUTING

be done on-line. Therefore the synthesis of components is done at compiled

time. The result of the synthesis of a component is box that encapsulates the
circuit implemented with the resources in a given area (routers logic and PEs).
After the placement of a new component on the device at run-time, its coor-
dinate is set to that of its corresponding router. As the placement destroys some
part of the network, we must insure that packets sent to other components will
still reach their destination.

6.2.1 Reachability of Components and Pins

In this section, we define some conditions that must be fulfilled by a tempo-
ral placement of components on the device in order to enforce the reachability
of all the components on the device an those of the pins around the device.
Definition 6.2 (reachability of component and pin) Given a re-
configurable device and a configuration, i.e. a set of components actually run-
ning on the device, we said that a component (pin) on reconfigurable device
at a given time is reachable iff each message sent to this component (pin) can
reach the component (pin).
Because the communications between components are established at run-
time and since the configuration of the chip is not known in advance, we must
insure that all components and pins on the device are reachable at any time
during the temporal placement. This condition is fulfilled if at any time the set
of components and pins on the device is strongly connected.
Definition 6.3 (strongly connected configuration) Given a re-
configurable device, we said that a configuration of the device is strongly con-
nected iff for each pair of components A and B, a path of active routers that
connects the two components, exists on the device.
One way to enforce the strongly connectedness in a DyNoC is to require that
each components placed on the chip must always be surrounded by a ring of
active routers. This can be reached either by synthesizing components in such
a way that when placed on the device, they are always surrounded by a ring of
routers. The second possibility is to let the job to a temporal placer. This will
considerably increase the complexity of the placer. Besides the computation
of free space to place a new component it must insure that all placements are
strongly connected.
Theorem 6.4 If each component is synthesized in such a way that it is in-
ternally surrounded only by processing elements, then all placements on the
reconfigurable device are strongly connected.

Proof Assume that a set of components developed as require in theorem 6.4

and placed on the device is not strongly connected. In this case, at least one
On-line Communication 203

pair of components abuts or a component abut the device boundary. Let con-
sider the first case. The second one can be handle in a similar way. Either
the two components overlap or at least one component uses some routers on
its internal boundary (this is illustrated in figure 6.14) . The first case is im-
possible because only overlapping free placements are valid. The second case
contradicts our requirement of the theorem, thus completing the proof.

Example 6.5 Consider the placement of figure 6.14 with two abutting com-
ponent. The first component attached to router A is not implemented according
to the guideline of theorem 6.4, since router are used in its internal boundary.
The second component attached to router B is developed according to the rule
of theorem 6.4. No route is available between those two component, because
the only routers available are consumed by the first component.

Figure 6.14. A impossible placement scenario

The configuration of figure 6.15 allows the reachability of all components

and pins. All the components are implemented according to the guideline of
theorem 6.4. Therefore, no matter where a component is placed on the device,
it will always be surrounded by a ring of routers.
204 RECONFIGURABLE COMPUTING

Figure 6.15. A stronly connected configuration on a DyNoC

While in a static NoC, each router always has four active neighbor routers1 ,
this is not always the case in the DyNoC presented here. Whenever a compo-
nent is placed on the device, it covers the routers in its area. Since those routers
cannot be used, they are deactivate. The component therefore set the activation
signal to the neighbor routers to notify them not to send packets in its direction.
Upon completing its execution, the deactivated routers are set to their default
state. A routing algorithm used for common NoC cannot work on the DyNoC,
without modification. An improvement or an extension of the existing routing
algorithms should be done in such a way that packets may be able to surround
the components on their way to the destinations.

6.3 Routing Packets

Because the dynamic placement and removal of modules on the chip creates
unpredictable obstacles in a DyNoC, adaptive routing algorithms must be used
to handle the dynamic changing configuration of the network. The routing
algorithm must be fully local-decisive2 , deadlock free and live lock free. We
must insure that each packet sent in the network will reach its destination after
a finite number of steps. While very complex and intelligent algorithm may
be developed, a very simple strategy was proposed in [35] [32] to allow the
packets to surround the obstacle, whenever they encounter one.

1 Routers around the chip does not have four neighbour. The package pins can be considered as further

neighbors where the routers can sent a message. This leads to the availability of the four neighbor routers
everywhere on the chip.
2 The decision where to send a packet is taken locally.
On-line Communication 205

The strategy is based on the XY-routing, because of its simplicity, its effi-
ciency and its deadlock freeness, the XY-routing algorithm presented in section
5.5 was adapted for the DyNoC. Recall that in a full mesh, the XY-routing is a
deadlock free, shortest path routing algorithm that first routes packets in X di-
rection to the correct X-coordinate and then in the Y direction until the correct
Y-coordinate. The adapted XY-routing, the emphS-XY-Routing (Surrounding
XY-routing) operates in three different modes:
emphThe N-XY (Normal XY) mode. In this mode the router behave as a
normal XY router.
emphThe SH-XY (Surround horizontal XY) mode. The router enter this
mode when its left neighbor or its right neighbor is deactivated.
emphThe SV-XY (surround vertical XY) mode. The router enter this mode
when its upper neighbor or its lower neighbor is deactivated.
The normal mode is available when all the neighbors of a routers are active.
In this case, it is the XY-routing which is used.

6.3.1 The SH-XY mode: Surrounding Obstacles in the X-direction

Assume without loss of generality that a packet moving from right to left
is blocked by an obstacle. As shown in figure 6.16 there exist two alternative
paths for the packet to reach its destination.
According to the XY-routing, which is the base of this algorithm, the first
path is chosen if the Y-coordinate of destination of the packet is greater or equal
than that of the router and the packet is sent upward. Otherwise, the second
path is chosen and the packet is sent downward. One problem occurs when a
packet with destination Ydest is sent for example upward and riches a router
r with coordinate Yr > Ydest . According to the previous defined scheme the
packet will be sent downwards to the router with coordinate Yr − 1 which will
send it upward, thus producing a "ping-pong" game. Figure 6.16 illustrates
this situation. In this case, the ping pong game happens between routers 1 and
router 2.
To avoid this, a strategy consists of stamping the packet by setting a "stamp-
bit" to 1 in order to notify the next routers that the packet is willing to surround
the component and it should not be sent back. In our example, the packet is
stamped by the router 1.
Upon reaching the router upper right of the module to be surrounded, the
stamp is removed and the packet is sent left, until its destination column or
until another obstacle is found. Because each component is always surrounded
by a ring of routers, this algorithm for surrounding a component in the X-
direction will always work and a packet will never be blocked. In the example
of figure 6.16 path 1 will be chosen.
206 RECONFIGURABLE COMPUTING

Figure 6.16. Obstacle avoidance in the horizontal direction

6.3.2 Surrounding Obstacles in the Y-direction

The situation is similar, to the one where a packet moving in the X-direction
is blocked. Assume without loss of generality that a packet moving from top
to bottom is blocked by a placed component. Dealing with this case as the
previous one, the packet will be sent left or right. No preference is done here,
because the packet is already in its right column. Lets assume that the packet is
sent to next router right to the current one. Because the basic routing algorithm
is the XY-routing, the next router will first compare the X position of the packet
with its own position. With the packet X-destination being smaller, it will send
the packet back to the router from which it was received. The two routers will
keep sending the same packet to each other, thus creating a deadlock. Figure
6.17 in which a "ping-pong" game will then happen between router 2 and router
4 illustrates this situation.
Like in the previous case, the packet is stamped to notify all the routers
above the obstacle that the packet is willing to surround the component. In our
example the packet will then be sent right until the last router (5) above the
component. There, the router removes the stamp and send the packet down-
ward. From there on we have the same situation as defined in the previous
section (Surrounding Obstacles in the X-direction).
On-line Communication 207

Ping-Pong

1
Game

5
6
Obstacle
Component
Routing
Path1 Routing
Path 2

Destination
Component

Figure 6.17. Obstacle avoidance in the verticalal direction

In the two cases, a path will be always available from the source to destina-
tion. With the XY-routing used as underlying routing technique, we conclude
that each packet sent will reach its destination in a finite number of steps.
Theorem 6.6 With very high probability, the S-XY algorithm presented here
is deadlock free and livelock free.

Proof We need to first prove that there is always a path from packet to desti-
nation. With this, the only possibility to have a livelock is that the packet move
permanently through the network and never reach its destination, despite the
availability of a path.
Second we must prove that each packet will reach its destination after a
finite number of steps. The first requirement is guaranteed through theorem
6.4.
We now assume that a packet never reach its destination. This will happen
only if the packet is blocked or if the packet is looping in a given region. Be-
cause a path always exists from one active router to all other active routers,
no packet can be blocked in the network, i.e. a packet is looping. Since this
situation is not possible in the normal XY, it can only arise in the surrounding
phase. When a packet is blocked in a given direction, it takes the perpendicular
direction. This is done untill the last router on the component boundary which
208 RECONFIGURABLE COMPUTING

is at one corner of the module to be surrounded. From there on, the normal
XY-routing resumes. A looping of a packet around a component is therefore
not possible.
It may be possible to construct a placement sequence of components that
blocks a single packet whenever that packet moves in a given direction, forcing
it to take the perpendicular direction. Such a sequence may leads to a packet
that will loop in a given region. However, doing this will only block a single
packet and other packets will move to their destination as previously described.
This leas to a very low live locking because only one packet out of an infinite
number is live locking. However such placement sequences are very unlikely
to happen in the reality.
In the S-XY routing, the direction where to send a packet whenever an ob-
stacle is encountered is fixed for all routers a priori. In this particular case,
packets are always sent to the right, whenever they are blocked in the vertical
direction. This may leads to extreme long routing paths like that of figure fig-
ure 6.18. Those extreme paths are not usual, but may occur as result of some
placements.

Figure 6.18. Placement that cause an extreme long routing path

On-line Communication 209

6.3.3 Router Guiding

A simple and efficient strategy to avoid such path consists of instructing
the router around a component on the direction where a packet must be sent,
whenever the incoming packet is blocked by the component in a given direc-
tion. Only one line is necessary to realize this concept.
Figure 6.19 illustrates this approach that we call the emphrouter guiding.
Besides the activation line, which provides information on the state of a router(1=activate,
0=deactivated), an additional line is used for instructing the router on the di-
rection where they should sent packet to, in order for them to surround the
component. A value of 0 (=east or south) will mean that a packet blocked in
the vertical direction must be forwarded east and a packet blocked in the hori-
zontal direction must be forwarded south. A value of 1(=west or north) means
that a packet blocked in the vertical direction must be forwarded west, while a
packet blocked in the horizontal direction must be forwarded north.
The best direction where to send packets whenever they are blocked by a
component can easily be determine by the component shape, a process that
can be managed at compile time.
This approach considerably limits the complexity of the routers and there
is no need for stamping anymore. We call this modification the emphrouter
guiding because the routers are guided by the components.

01 01 00 00

01 01

01 01

01 01

00 00

00 00

01 01 00 00

Figure 6.19. Router guiding in a DyNoC

210 RECONFIGURABLE COMPUTING

6.4 Analysis of Efficiency

We define the communication latency of a packet as the number of routers
on a path from its source to its destination. On a reconfigurable device with
width W and height H, the maximum latency is given by W + H. However,
the unpredictable structure of the network which changes with the placement
of components creates unpredictable delays. The worst case delay Dwc (p) of
a packet going trough k obstacles from its source to its destination is given by
equation.
k
X
i
Dwc (p) = H + W + (Dwc (p)) (6.1)
i=0
Dwci (p) is the worst case additional delay caused by the obstacle i. If w
i
is the weight of the components i and hi its height, then we have Dwc i (p) =

max(wi , hi ). The half perimeter of the component is not added here, since it
is already in the sum H + W . Only the maximum between the components
height and its width is considered. This additional delay is unpredictable and
depends on the temporal placement of components on the device.
In a static NoC where each processing element is assigned a fixed place, the
delay is also fixed. Packet communication can be seen as pipelined compu-
tation in which the length of the pipeline is the packet delay from source to
destination. Once the pipeline if filled, one packet is received at the destination
on each clock.
In the DyNoC, we have the same situation, however delays created by newly
placed components create new situations for which the path must first be filled
with packets. It is even possible that some packets sent later reach their desti-
nation before other, which were sent earlier, but which had to go through some
obstacles. This makes the problem difficult for a formal analysis.

6.5 DyNoC Implementation on FPGAs

DyNoCs with different size and width were prototyped on two FPGAs [35]
[32], the VirtexII-1000 and VirtexII-6000 from Xilinx. While the prototype on
the VirtexII-6000 were mostly for statistically (area, latency) purpose, the im-
plementation on the VirtexII-1000 were done on the RC200 FPGA-board from
Celoxica [43]. The result is given in table 6.1 in term of area (A) occupation
for different bit-widths, memory (M) usage and speed (S) in MHz.
Two video applications with a VGA controller running at 25Mhz for normal
640x480 VGA were implemented. In the first one, a color generator module
(CG) communicates with the VGA controller (VC). The color generator gets
the X and Y coordinates of the current pixel position from the VGA module,
computes the color to be placed at that position and sent it back to the VGA
module which displays the color at the corresponding position. The color gen-
On-line Communication 211

Table 6.1. Router Statistics

VirtexII-1000 VirtexII-6000
A/M/S(8 bit) 8% /4% / 77.2 1% /0% / 77.2
A/M/S(16 bit) 12% 7% / 75.4 2% /1% / 75.4
A/M/S(32 bit) 21% / 12% / 77.3 3% /2% / 74.9
A/M/S(64 bit) 46% / 28% / 70.1 7% /4% / 73.7

Table 6.2. TLC and CG Statistics

VirtexII-1000 VirtexII-6000
A/M/S(TLC) 53% /32% / 100.2 8% / 5% / 100.2
A/M/S(CG) 47% 33% / 91.2 20% /11% / 121.9

erator application is a nice method to detect changes in the communication,

since this will directly have a visual effect on the screen. The X and Y posi-
tions are each 12 bits wide and the color is 24 bits wide. Therefore packets
with 32 bits width in each direction were used. Implemented on the RC200
board with the DyNoC, no change was detected in the displayed pattern, event
with a full network traffic due to the communication among remaining routers.
The second application is the implementation of a traffic light controller
(TLC) containing three modules. A VGA controller (VGA), a Traffic light vi-
sual module (LV) and a traffic control module (TC) to capture the pedestrians
wishes. As in the first case, the VGA module is used to display the state of a
traffic intersection on which the light and the button used by the pedestrians
can be seen. The traffic visual module is in charge of building the traffic light
infrastructure which is then displayed by the VGA module. The VGA sends
the X and Y pixel scan positions to the traffic visual module and receive a color
to be displayed. According to the pixel positions, the traffic light visual com-
putes the pattern to be place at that position. This generates the traffic light
infrastructure. The last module is a FSM which monitor the pedestrian inputs
(two push buttons on the board) and sent a message for the transition of state of
the traffic infrastructure to the traffic light visual, which in turn generates the
corresponding color to be seen. The traffic light controller were successfully
implemented on 3x3 DyNoC. The router at position (2,2) were disable to en-
force a surrounding. All the remaining routers keep communicating with each
other to keep the traffic high in the network. Also here the application runs as
it use to do without any interruption and without any malfunction. Table 6.2
provides some statistics on the implementation on VirtexII-1000 and VirtexII-
6000. The implementation of the TLC on a 3x3 DyNoC is shown in figure
6.20
212 RECONFIGURABLE COMPUTING

(2,1) (3,1)
(1,1) (3,1)
(1,1) (2,1)

LV LV
TC TC

(2,2)
(3,2)
(2,2) (3,2) (1,2)
(1,2)

(2,3)
(2,3) (3,3)
(1,3)
(1,3) (3,3) VGA
VGA

Figure 6.20. DyNoC implementation of a traffic light controller on a VirtexII-1000

As we can see, the size of the routers are very large, making the used of a
NoCs very difficult on FPGAs. This is somehow due to the fact that the routing
resources in the router are built on top the programmable resources available
(LUT and routing matrices) in an FPGA. A direct access to the available re-
sources in the FPGA will reduce the amount of the resources need to a min-
imum. However, this can be done only if the vendors provides the necessary
knowledge to control the resources on a very low level.

7. Conclusion
In this chapter, we have addressed the communication paradigm needed to
allow components dynamically placed at run-time on a reconfigurable devices
to communicate with their environment. The most promising approach is the
Network on Chip, which has attracted a lot of attention in the research com-
munity in the last decade. A large amount of work is available, each with its
own level of complexity. The DyNoC is an extension of the NoC paradigm.
In our opinion, it provides the best NoC-based approached to solve the dy-
namic need of communication in temporal placement. We have also presented
a circuit switching approach and show why it is not a good solution in tem-
poral placement. However, the 1-D RMB circuit switching approached can be
used to connect a small number of components in predefined location on a 1-D
reconfigurable device.
Chapter 7

PARTIAL RECONFIGURATION DESIGN ∗

Partial reconfiguration is one the prerequisite in reconfigurable computing,

since it allows for swapping modules into and out of the devices without having
to reset the complete device for a total reconfiguration. As we will see later in
chapter applications, the possibilities offered by the partial reconfigurable de-
vices in the implementation of system adaptivity are enormous. Unfortunately
very few devices on the market have the capability to be partially reconfigured
in order to replace some modules, while the rest of the design keeps working.
One of the few devices falling in the category of partial reconfigurable chips
are the FPGAs from Xilinx, in particular those from the Virtex class.
Designing for partial reconfiguration on Xilinx device is a complex exer-
cise that requires to be well prepared, technically, but also psychologically.
This chapter is much like a tutorial for designing partial reconfigurable appli-
cations on the Xilinx Virtex FPGA-devices. Our goal in this chapter is to guide
the designer in the design process for partial reconfiguration by providing the
essential steps and materials in order for him to avoid long search and long
introduction phases, which sometimes leads to a possible discouragement.
We therefore explained the three design flows used for the partial reconfig-
uration on Xilinx devices: The JBits approach, the Modular Design Flow and
the Early Access Design Flow. While the JBits approach apply for very few
old Virtex FPGA, the modular design and the early access design flows sup-
port newer devices. We therefore explain those two design flows on the basis
of working examples, whose sources can be downloaded from the book home
page.

∗ This chapter and the correponding Appendix was prepared by Christophe Bobda and Dominik Murr
214 RECONFIGURABLE COMPUTING

The development of partial reconfigurable applications on the Xilinx plat-

form is done according to the Modular Design Flow provided by Xilinx in
[227]. The modular design flow provides a guided flow on how to generate the
constraints required by the placement and configuration tools for the genera-
tion of full and partial bitstreams representing modules to be downloaded at
run-time into the device for a partial of for a full configuration.
Several tutorials for partial reconfiguration on the Xilinx devices exist to be
downloaded from the internet. Despite their usefulness, almost all of them tar-
gets only VHDL and Verilog designs. With the increasing interest in software-
like hardware design language like SystemC and Handel-C, the need for incor-
porating those languages in partial reconfiguration design flows is high. We
therefore described in this chapter how to incorporate Handel-C designs in a
partial reconfiguration design flow.
Finally, in order to efficiently use the capability of partial reconfigurable
chips, the platform that host the chip, i.e. the board, must be designed in such
a way that the use of the flexibility provided by the reconfigurable chip can be
maximized. This issue is addressed at the end of the chapter where we provide
some guidelines in the design of partial reconfigurable platforms. As example
of such platforms, we present the Erlangen Slot Machine, which was designed
to allow the maximal use of partial reconfiguration.

1. Partial Reconfiguration on Virtex Devices

Virtex FPGAs like, other FPGAs fro Xilinx are organized as a two dimen-
sional array of CLBs containing a certain amount of logic. They are configured
with configuration data called a bitstream, which can be downloaded into the
device via the configuration port.
The idea behind partial reconfiguration is to realize reconfiguration by only
making the changes needed to bring the device in the desired state. Fragments
of the complete bitstream, also call packets, are sent to the device in order to
reconfigure the needed part of the device. A copy of the last configuration is
maintained in a dedicated part of the processor memory, called the configura-
tion memory. Partial reconfiguration is done by synchronization between the
configuration memory and the device. Changes made between the last config-
uration and the present one is marked as dirty packets which are then sent to
the device for partial reconfiguration. A packet is a sequence of (command,
data) pairs of varying length which are used to read or write internal registers
and configuration state data.
Virtex devices are organized in frames or column-slices which are the small-
est unit of reconfiguration. Partial reconfiguration can then be performed by
just downloading the configuration data needed to modify the part of the de-
vice needed. Up to the Virtex II Pro, a frame use to span a complete column,
with the consequence that the reconfiguration of one component affects all the
Partial reconfiguration design 1 215

components, which share a common column with that component. From the
Virtex 4 and upwards, the frames does not spans a complete column any more,
but only a complete tile. Therefore, the reconfiguration of a component affects
only those components in the same blocks, which share a common column
with the reconfigurable module.
It is then the matter of the partial reconfiguration design flow to produce the
partial bitstream, i.e. the set of data required to configure the needed frames
and therefore to allow the replacement of a component on the chip. The ex-
tracted partial bitstreams are then used to move the device from one configura-
tion state to the next one without reconfiguring the whole device.
Two possibilities exist to extract the partial bitstream representing a given
module. A constructive approach which allows us to first implement any single
component separately using common development tools. The so developed
components are then constrained to be place at a given location using bounding
box and position constraints. The complete bitstream is finally built as the sum
of all partial bitstreams. This approach is the one followed by the modular
design and the early access design flows. The second possibility consist of
first implementing the complete bitstreams separately. The fix parts as well
as the reconfigurable parts are implemented with components constrained at
the same location in all the bitstreams, which differs from each other only on
the reconfigurable parts. The difference of two bitstreams is then computed to
obtain the partial bitstream needed to move from one configuration to the next
one.

Example 7.1 Consider for instance the two designs of figure 7.1, where the
full designs are named Top1 and Top2. The fix part consists of the module F-
Module 1, F-Module 2, and F-Module 3, which are placed at the same location
in the two designs. In Top1 , the reconfigurable module is R-Module 1, which
is placed at the same location, with the same bounding box constraints as the
reconfigurable module R-Module 2 in Top2.
The addition of R-Module 2 to the fix parts of the two bitstreams generate
the full bitstream Top2. The same is true for R-Module 2 and Top1. Also,
as shown on the figure, adding the two bitstreams Top1 and R-Module 2 will
produce the bitstream Top2, because R-Module 2 will replace R-Module 1 in
Top1. The substraction of Top2 from Top1 produce the partial bitstream for
R-Module 1.

2. Bitstream Manipulation with JBits

JBits [102] is an application interface developed by Xilinx to allow the end
user to set the content of the LUT and make connections inside the FPGA
without the need for other CAD tools. JBits does permit the changes to be
made directly to the device, but to the configuration data.
216 RECONFIGURABLE COMPUTING

Figure 7.1. Generation of bitstreams for partial reconfiguration

The JBits API is constructed from a set of java classes, methods and tools
which can be used to set the LUT-values as well as the interconnections. This
is all what is required to implement a function in an FPGA. JBits also provides
functions to read back the content of a FPGA currently in use.
The content of a LUT can be set in JBits using the set function:

set(row, col, SliceN umber T ype, value)

which means that the content of the LUT Type (Type is either the LUT F or
the LUT G of the corresponding CLB) in the slice SliceNumber (SliceNumber
can be 0 or 1) of the CLB in position row and col should be set to the value of
the array value. The Virtex LUTs have four inputs and 16 entries representing
the 16 possible values a 4-inputs function can take.
A connection is defined using the function connect:

connect(outpin, inpin)
This function uses the JBits routing capabilities to connect the pin outpin to
the pin inpin anywhere inside the FPGA. outpin and inpin should be one of the
CLB terminals.
Connecting the output of a LUT to a CLB terminal can be done with the
function.
Partial reconfiguration design 2 217

set(row , col , terminal control , LUT output)

where terminal control sets the correct terminal to be connected to the out-
put LUT output of the corresponding LUT. JBits provides hundreds of prede-
fine cores like adders, subtractors, multipliers, CORDIC Processor, encoder
and decoders, network modules, etc... which can directly be used in designs.
They can also be combined to generate more complex cores.
Although any function can be implemented entirely in JBits using the basic
primitives previously described as well as the predefined libraries modules, it
is usually not the best approach, due to the high complexity in building com-
ponents from the scratch. Instead, the bitstream substraction described in the
previous section is use to extract the part of the bitstream needed at run-time
for partial reconfiguration.
The approaches used to extract partial bitstream of the component that will
be replaced at run-time, consist of implementing as many full bitstream, also
called top-level, as required and perform the substraction among them to ex-
tract the differences. Those are the parts that will be used to configure the
device later. Implementing the substraction using JBits is straightforward. The
two bitstreams need to be scanned and compared for each element. Only the
difference is copied in the resulting bitstream.

3. The Modular Design Flow

One of the main drawback of JBits is the difficulty to provide fixed routing
channels for a direct connection between two modules. This is important be-
cause one must insure that signals will not be routed on the wrong paths after
the reconfiguration. Consider for instance the two full designs of figure 7.2 in
which the fixed as well as the reconfigurable components are place at the same
locations. The connection F-Module 1 ↔ R-Module 1 is runing on a different
path than the connection F-Module 1 ↔ R-Module 2. On reconfiguration the
design will probably not work correctly.
One of the main contribution in the modular design flow [128] is the use of
the so-called Bus Macro primitive to guarantee fixed communication channels
among components that will be reconfigured at run-time and the fixed part of
the design.
The modular design flow however was not initially developed to support
partial reconfiguration. It is a an approach that allows several Engineers to
cooperatively work on the same project. The project leader identifies the com-
ponents of the whole project, estimates the amount of resources that will be
consumed by each component, defines the placement locations for the com-
ponents on the device, and let the single parts be developed by the different
engineers. Finally, the single developed parts are integrated into the whole de-
sign to build a workable circuit. The modular design flow is a 4-steps method
218 RECONFIGURABLE COMPUTING

Figure 7.2. Routing tools can route the same signals on different paths

(design entry and synthesis, initial budgeting, active implementation, and as-
sembly) for which a viable directory structure must be use.
We recommend to use the ISE version 6.3, which is on our opinion, the
most stable version for implementing partial reconfiguration with the modular
design flow.

3.1 Directory Structure

Xilinx recommends a structure of design directories as shown in figure 7.3:

The HDL Design directory (HDL): This directory is used to store the HDL
descriptions of the modules and that of the top level designs. The HDL
description can be in VHDL, Verilog or any other hardware description
language.

Synthesis directory (synth): In this directory the team leader synthesizes

the top-level design, which includes the appropriate HDL file, the project
file, and project directories. In the same way, each team member will also
have a local synthesis directory for his or her synthesized module.

Modules directory (Modules): In this directory, the active implementation

of individual modules is done in the respective module’s subdirectory.

Implementation directory (Top): It includes two subdirectories, one for the

initial budgeting phase and one for the final assembly phase. In the initial
directory the team leader performs the initial budgeting operations for the
corresponding top-level design. In final assembly directory, the team leader
assembles the top-level design from the individual modules implemented
and published in PIMs directory.
Partial reconfiguration design 3 219

PIMs (Physically Implemented Modules) directory (Pims): The active im-

plementation of individual module creates appropriate module directory in
PIMs directory where implemented module files are copied.

Figure 7.3. The recommended directory structure for the modular design flow

We next focus on the single steps of the modular design flow

3.2 Design Entry and Synthesis

This step must be carried out before going to the modular design imple-
mentation step. In this step, the members of the team develop modules using
a hardware description language (HDL), and synthesize them using synthesis
tools. The team leader completes top-level design entry and synthesis before
moving to the initial budgeting phase. Team members complete individual
module design and synthesis before moving to the active module implemen-
tation phase for that module. However, the team members can work on their
modules while the team leader is working on initial budgeting phase of modu-
lar design implementation. The restriction that must be done on the top-level
design are the following:
All global logic, including I/Os must be included in the top-level descrip-
tion.
All design modules are instantiated as black boxes in the top-level design,
with only ports and port directions specified.
220 RECONFIGURABLE COMPUTING

Signals that connect modules to each other and to I/O pins must be defined
in the top-level design.

Top-level design and individual modules are synthesized as described in

the documentation of the synthesis tool selected. In synthesis properties, the
insert I/O pads settings must be enable for top-level design and disable for each
submodule. Separate unique netlists will be created for each of the modules
and for the top-level design. For each instantiated module, exactly the same
module name must be used as its file name. Some synthesis tools like the
Xilinx xst, will not be able to match the module names specified in the top-
level netlist to the module netlists.

3.3 Initial Budgetting

In this stage, the team leader assigns top-level constraints (pin locations,
area constraints for each modules, timing constraints, etc...) to the top-level
design, which results in topX.ucf file (user constrain file). The same topX.ucf
file is used for the active module implementation of individual modules.
Only the netlist for the top-level design file (topX.ngc, if compiled with
Xilinx XST) and the constraint file (topX.ucf) are copied to the initial directory
in implementation directory of the corresponding top-level design.
The same top-level design name should be used for both files. The following
command is then first executed

ngdbuild -modular initial design name

In this case design name is top.ngc. In this step, the compiler ngdbuild pro-
duces two files, design name.ngd and design name.ngo. The design name.ngd
file is used during subsequent modular design steps, while design name.ngo is
not. The constraint editor that is used to edit design constraints such as clock
periods, in the top-level design, can be invoked with the following command:

constraints editor design name.ngd

The following command invokes Floorplaner.

floorplanner design name.ngd

Using Floorplanner, modules sizes, module locations, and information on
module ports can be created. The modifications are copied in the top-level
constraint file topX.ucf. This file can also be directly edited with the constraints
as specified in the user manual. Constraints that must be set in the .ucf file,
either by hand or by the floorplanner are the following:

1 Module area constraints:

Partial reconfiguration design 4 221

each module must have a minimum width of four-slices.

a module’s width is always a multiple of four slices (e.g., 4, 8, 12, ..)
the module bounding box must span the full height of the device.
each module must be placed on an even four-slice boundary.

besides those constraints, area group must be defined for each component
and specific properties for the partial reconfiguration design flow must be
added to the to the area group of components.

2 IOB constraints: The Input Output Block (IOB) must be constrained within
the columnar space of their associated reconfigurable modules. This condi-
tion is important to avoid the connection between a module and a pin to be
destroyed during the reconfiguration of another component. Also, all IOBs
must be locked down to exact sites.

3 Global logic constraint: All global logic like clocks, power and ground
lines must be constrained in the top level. No global logic should be locked
in the module topX.ucf file.

4 Insertion of Bus macros

Partial reconfiguration is the process of transiting from one full configura-
tion to a new full configuration using a partial configuration bitstream. This
partial bitstream represents a component or a set of components present
in the final bitstream, but not in the first one. The two bitstreams have a
common fix part, which does not change. The reconfigurable parts are con-
nected to the fix part by defining signal and connecting them to the ports.
The path followed by the signals in two designs may be different due to
the routing tools. On partial reconfiguration, the resulting application may
not work anymore, if the signals connecting the fix and dynamic part are
not using the same paths in the two designs. The role of the bus macros
in the Xilinx FPGAs is to provide fix communication channels, which can
be used by reconfigurable module. Those are tri-state lines which help to
keep the signal integrity of modules being reconfigured. Those channels
are not affected by the reconfiguration and the signals keep their integrity,
thus insuring a correct operation of the application.
The approach consists of connecting the two components at the two ends
of a dedicated long line segment within the FPGA. The two component use
tri-state buffers to define the direction of the communication. A bus macro
is a construct that allow such connection to be specified in the design tool.
Each bus macro occupies a 1-row by 8-column section of tri-buffer site
space.
222 RECONFIGURABLE COMPUTING

Besides the constraints specified, global-level timing constraints can be cre-

ated for the overall design. The output of the initial budgeting phase is a
topX.ucf file containing all placements and timing constraints. Each module is
implemented using this topX.ucf file, extended with the module-specific con-
straint requirements.
To finalize the initial budgeting phase, the following command is run in the
initial subdirectory of the corresponding top-level design, after copying the
files topX.ucf file and topX.ngc file to the the same directory directory.

ngdbuild -p device type -modular initial topX.ngc

3.4 Implementing Active Modules

During this phase, the team members implement the top-level design with
only the active module expanded using the top-level topX.ucf file. Team mem-
bers can implement their modules in parallel. Once each de design has been
synthesized, floorplanned, and constrained, the implementation (mapping, place
and route) of all modules (both fixed and partially reconfigurable) for the de-
sign can start. Each module will be implemented separately, but always in the
context of the top-level logic and constraints. Bitstreams will be generated
for all reconfigurable modules. This section describes the overview on how to
independently implement each module.
The following instructions should be performed for each of the module in
the design.

1 Copy the module implementation files to the to the active implementation

directories for each module (active/module name)under the module direc-
tory in which active module will be implemented:

Synthesized module netlist file module name.ngc

The file topX.ucf, which the team leader created in the initial budgeting
phase must be copied in the module local directory and renamed to
module name.ucf.

2 Build the module in active module mode as follows:

ngdbuild -uc module name.ngc -modular module -active topX.ucf \\

top level design directory path/topX.ngo

The output NGD file is named after the top-level design and contains im-
plementation information for both the top-level design and the individual
module. The -uc option ensures that the constraints from the local topX.ucf
file are annotated.
Partial reconfiguration design 5 223

3 Map the module to the library elements using the following command.

map topX.ngd

In this step, the logic of the design only with expanded active module is
mapped.
4 The place and route of the library elements assigned to the module is done
by invoking the placer and router with:

par -w module name.ncd top routed.ncd

The -w option ensures that any previous versions of the produced files de-
sign name routed.ncd are overwritten. If the area specified for the module
cannot contain the physical logic for the module because it is sized incor-
rectly, then resizing must be done again and the .ucf file generated during
the initial budgeting phase must be regenerated with the changes made.
The entire initial budgeting phase should be performed again. This would
then imply that new .ucf files would need to be copied to each module and
that each module needs to be reimplemented. If everything went fine, then
the module is now placed and routed at the correct location in the given
bounding box.
5 Run trace on the implemented design to check the timing report in order to
verify that the timing constraints are met.

trce top routed.ncd

If necessary, the Floorplanner is used to reposition logic, if the module

implementation is unsatisfactory, for example, if it does not meet timing
constraints. In this case the Map, place, and route steps must be rerun again
as described in the previous steps.
6 The last step in the active implementation is the publishing of the imple-
mented module file to the central located PIMs directory, which was set up
by the team leader. This step is performed using the command:

pimcreate pim directory path -ncd top routed.ncd

This command creates the appropriate module directory inside the PIMs
directory. It then copies the local, implemented module files, including the
.ngo, .ngm and .ncd files, to the module directory inside the PIMs direc-
tory and renames the .ngd and .ngm files to module name.ncd and mod-
ule name.ngm. The -ncd option specifies the fully routed NCD file that
should be published.
224 RECONFIGURABLE COMPUTING

3.5 Module Assembling

Module assembling is the final phase of modular design. Here the team
leader assembles the previously implemented modules into one top-level de-
sign. In this phase previously implemented modules, which were published to
the PIMs directory, and the top-level design file are used. The resulting full de-
sign can be used to generate a bitstream. As before, we provide the necessary
steps.

1 In order to incorporate all the logic for each module into the top-level de-
sign, run ngdbuild as follows:

ngdbuild -modular assemble -pimpath pim directory path topX.ngo

Ngdbuild generates an topX.ngd file from the top-level topX.ucf file, the
top-level .ngo file, and each PIM’s topX.ngo file.
2 The logic for the full design can then be mapped as follows:

map topX.ngd

MAP uses the topX.ncd and topX.ngm files from each of the module direc-
tories inside the PIMs directory to accurately recreate the module logic.
3 The full design can then be placed and routed using the following com-
mand:

par -w topX.ncd topX routed.ncd

The place and route command, par, uses the topX.ncd file from each of the
module directories inside the PIMs directory to accurately reimplement the
module logic.
4 In order to verify if timing constraints were met, the trace command must
be invoked as follows:

trce top routed.ncd

Partial reconfiguration with the modular design flow can be performed as

batch job using one or more scripts that contain the large amount of commands
that must be invoked. An example of such scripts and design examples are
available for download on the book’s website. Also tools like Part-Y [71]
provide a conformable graphical user interface on top of the modular design.
Partial reconfiguration design 6 225

4. The Early Access Design Flow

With the introduction of the Early Access Design Flow for creating partially
reconfigurable designs, Xilinx updated and enhanced the existing Modular De-
sign Flow [227]. The package consists of a number of scripts that modify the
user’s ISE installation. Up to now the update is only viable with the exact ISE
version 8.1i with service pack 1. It can be obtained by registered users from
[226].
The main changes compared to the Modular Design Flow concern usability
and the relaxation of some rigid constraints for partial reconfiguration designs.
Also new bus macros improve constraint setting and allow more possibilities
for reconfiguration.
First of all Virtex-4 platform support is added to the PR tools. Besides the
known Spartan 3, Virtex-II and -II Pro this enables users to deploy the current
top product of Xilinx’s linecard.
Partially reconfigurable areas may now also be defined block-wise and do
not have to span over the full chip height. This provides the possibility, for
Virtex-4 devices and later to reconfigure the tiles rather than whole columns.
The Virtex-II and Spartan FPGAs, in contrast, physically lack this ability.
When a small region is reconfigured, the whole column is written, but the
configuration controller on the FPGA checks whether the reconfiguration would
actually change the content of the configurable logic block, and perform re-
configuration only where changes are necessary. The reconfiguration will not
touch blocks that would remain unchanged at all as depicted in figure 7.4.
With this, applications may also be realized that require a bus that is placed
through the whole width of the chip and won’t be destroy when reconfigura-
tion occurs somewhere in between. Figure 7.5 shows an example of design for
a video streaming application that uses several stages of filters and a bus that
transports the pictures from one module to the next. The single filters in the
modules are partially reconfigurable and may be swapped without interrupt-
ing and stalling the bus. In this case, we will not want to alter the bus while
configuring a module on the chain.
Another relaxation is that signals crossing partially reconfigurable blocks
without interacting, won’t have to be passed through bus macros anymore.
The routing algorithm is now capable of determining those signals and using
longer range communication lines that will not be touched by a partial recon-
figuration. This greatly alleviates timing constraints to be met and simplify
the designing process as a whole. Before, the developer had to build up the
design, have the signals be routed and check for signals that had been guided
through the partially reconfigurable area. Also resources that use to lie within
this block but have to be used in another fixed module, would have had to be
fed signals through some bus macros. Both cases, given the signal will not ex-
change information with the PR area, are now fortunately handled by the new
226 RECONFIGURABLE COMPUTING

Figure 7.4. Limitation of a PR area to a block (dark) and the actual dimensions (light)

Figure 7.5. Scheme of a PR application with a traversing bus that may not be interrupted

design flow. Now, traversing signals must not be guided through a bus macro,
a tedious handwork.
Partial reconfiguration design 7 227

4.1 New directory structure

The Early Access Tools now support a modified, more sensible directory
structure for PR projects as depicted in figure 7.6. There, the folders’ contents
are presented in the order of usage in the design flow.

Figure 7.6. Improved directory structure for the Early Access Design Flow

The non pr folder should contain a working version of every combination

of modules that can be later use for partial reconfiguration. Although not
required it is strongly recommended to verify the functionality of the dif-
ferent variants of the design before trying to do any partial reconfiguration!
synth holds the HDL-definitions of the modules. There should be at least
one subfolder for every partially reconfigurable module, a folder base for
all the fixed modules and top for the top-level design. The modules are then
synthesized herein. synth thus replaces the folders hdl and synth from the
Modular Design Flow.
The initial budgeting phase on the top-level module is done in top using the
synthesized netlist from synth/top, the UCF and the Bus Macro definition
files.
228 RECONFIGURABLE COMPUTING

Initial budgeting for the static modules is accomplished in static

The activation phase takes place in the reconfigmodules’ subdirectories for

each partially reconfigurable module

Finally the design is assembled in merges and full and partial bitstreams are
generated.

4.2 Design Entry and Synthesis

Except the slightly changed locations because of the modified directory
structure the general implementation work stays the same as with the Mod-
ular Design. Thus top-level modules contain:

all global logic, including I/Os,

all design modules, instantiated as black boxes with only ports and port
directions specified and

signals that connect modules to each other and to I/O pins.

Again, top-level modules include no logic except clock generation and dis-
tribution circuits.

4.3 Bus Macros

Xilinx ships a new set of Bus Macros to go with the Early Access package.
They are now distinguished between device type, direction, synchronicity and
width. They are now directed and not deployable in any direction and either
clocked to run synchronously with the data or unclocked as the old bus macros.
The naming convention yields to terms like busmacro device direction synchronicity width.nmc
where
Partial reconfiguration design 8 229

device can be
xc2vp - for a Virtex-II Pro,
xc2v - for a Virtex-II or
xc4v - for a Virtex-4 FPGA.
direction is either one of
r2l - right-to-left
l2r - left-to-right
b2t - bottom-to-top (Virtex-4 only)
t2b - top-to-bottom (Virtex-4 only).
synchronicity may be
sync - synchronous or
async - asynchronous
width is either
wide - spanning four CLBs or
narrow - reaching over two CLBs.
A sample name is busmacro xc2vp r2l async narrow.nmc.
The direction of the macro is to be seen geographically: right-to-left means
data flowing from the right side of a border to the left side. If the right side
is inside a partially reconfigurable area, the bus macro is an output bus macro,
otherwise it is an input bus macro. The same applies to the other cases with
left-to-right and the bottom-to-top and top-to-bottom bus macros. The usage is
shown in picture 7.7.

Figure 7.7. Usage of the new EA bus macros

Figure 7.8 shows the difference between a wide and a narrow type bus
macro. The classification concerns only the range of CLBs that one of these
bus macros bridges. Both kinds offer eight signal lines to cross a boundary.
230 RECONFIGURABLE COMPUTING

The unoccupied CLBs between wide bus macros can be used by user logic
or other bus macros. They can also be nested up to three times in the same y-
row as depicted in figure 7.9. With the old bus macros, only four signals could
traverse the boundary at this position whereas the new macros put through up
to 24 lines. The three staggered bus macros do not have to be of the same type
or direction.

(a)

(b)

Figure 7.8. Narrow (7.8(a)) and wide (7.8(b)) bus macro spanning two or four CLBs

Especially during the reconfiguration signals from partially reconfigurable

modules might toggle unpredictably. The new bus macros now provide ver-
sions that may therefore be disabled when appropriate to output a stable zero
instead of the fluctuant signals.

4.4 Initial Budgeting

For the initial budgeting as well as for the activation phase similar steps
have to be taken as with the Modular Design. The UCF generation will have
a different outcome since area constraints may now be set more freely and bus
macros have to be deployed more carefully concerning their type and direction.
The placement rules are generally defined as follow:
Partial reconfiguration design 9 231

Figure 7.9. Three nested bus macros

the x-coordinate of the CLBs on which the bus macros is locked to, has to
be divisible by two
the y-coordinate has also to be divisible by two
Virtex-4 devices require coordinates divisible by four due to the location of
block RAMs and DSPs
bus macros have to traverse a region boundary. Each of the endpoints has
to completely lie in a different area
Further actions for the Initial Budgeting like setting the timing constraints
follow the same guidelines as the Modular Design. One must keep in mind
that signals that will not interfere with a module can cross that module without
having to run through a bus macro.
To close the Initial Budgeting Phase, the top-level design has to be imple-
mented. In directory pr project/top, the command
ngdbuild -uc system.ucf -modular initial system.ngc
has to be invoked. pr project/static now budgets the static portion of the de-
sign. The .nmc files for the used bus macros have to be placed in this directory.
The necessary commands are:
ngdbuild -uc ../top/system.ucf -modular initial ../top/system.ngc
map system.ngd
par -w system.ncd sytem base routed.ncd
The routed outputs are placed in pr project/top respectively pr project/static.
The process yields another file, static.used, that contains a list of routes within
the partially reconfigurable regions that are already occupied with static de-
sign connections. This is used as an input for the placement and routing of the
partially reconfigurable modules later on. The file allows for guiding signals
through PRMs without using bus macros.
232 RECONFIGURABLE COMPUTING

4.5 Activation
In the Activation Phase the partially reconfigurable modules are implemented.
First the previously generated static.used is copied to each of the modules di-
rectories (e.g. pr project/ reconfigmodules/ prm a1) and renamed to arcs.exclude.
Each module is then run through the known sequence of ngdbuild, map and
par as follows for the prm a1. In directory pr project/ reconfigmodules/ prm a1,
do the following:
ngdbuild -uc ../../top/system.ucf -modular module -active \\
prm component name ../../top/system.ngc
map system.ngd
par -w system.ncd prm des routed.ncd
Note that prm component name is the component name not the instance
name of the module. par automatically looks for and incorporates the arcs.exclude
file.

4.6 Final Assembly

The last step is to assemble the static and reconfigurable modules into the
top-level design. This work is greatly facilitated by two new scripts: pr verifydesign
and pr assemble. For this, a routed top-level design, a set of the static modules
and one PRM design has to be copied to a separate subdirectory for each PRM
under pr project/merges. For every PRM (e.g. in pr project/merges/prm a1),
the following has to be executed:
pr verifiydesign sytem base routed.ncd prm a1 routed.ncd
pr assemble sytem base routed.ncd prm a1 routed.ncd
It should be checked if the final netlists (e.g. pr project/merges/prm a1/
static routed.ncd) meet timing constraints. The timing analysis is performed
with the .pcf file for the static design.
The partial and full bitstreams to reconfigure the targeted FPGA are then
found in merges. Before using them the developer has to make sure that the
assembly process produced proper files by examining the .summary files gen-
erated.

4.7 Merging Multiple Partially Reconfigurable Areas

To merge more than one partially reconfigurable region in one design an
iterative invocation of pr verifydesign and pr assemble is needed. With
each iteration a new partially reconfigurable module is merged into the existing
overall design.
Partial reconfiguration design 10 233

The corresponding actions are

1 Merge each PRM of one reconfigurable area with the static design
pr verifydesign and pr assemble generate intermediate designs with
just the one PRM contained and partial bitstreams that can be used to re-
configure the FPGA at this particular region afterwards. The necessary
commands are

merges/prmA/pr verifiydesign sytem base routed.ncd \\

prm a1 routed.ncd

merges/prmA/pr assemble sytem base routed.ncd prm a1 routed.ncd

merges/prmA/pr verifiydesign sytem base routed.ncd \\

prm a2 routed.ncd

merges/prmA/pr assemble sytem base routed.ncd prm a2 routed.ncd

2 Merge additional modules for other partially reconfigurable areas

For every additional region, the output of the previous proceeding is taken
as input: the partially reconfigurable module b1 is to be merged with the
combined a1 and base design.

merges/prmB1withA1/pr verifydesign \\
../../prmA/prm a1/base routed full.ncd \\
../../reconfigmodules/prm b1/prm b1 routed.ncd

merges/prmB1withA1/pr assemble \\
../../prmA/prm a1/base routed full.ncd \\
../../reconfigmodules/prm b1/prm b1 routed.ncd

Every combination of partially reconfigurable modules have to be merged

together. This requires for two partially reconfigurable regions accordingly:

merges/prmB1withA2/pr verifydesign \\
../../prmA/prm a2/base routed full.ncd \\
../../reconfigmodules/prm b1/prm b1 routed.ncd

merges/prmB1withA2/pr assemble \\
../../prmA/prm a2/base routed full.ncd \\
../../reconfigmodules/prm b1/prm b1 routed.ncd
234 RECONFIGURABLE COMPUTING

merges/prmB2withA1/pr verifydesign \\
../../prmA/prm a1/base routed full.ncd \\
../../reconfigmodules/prm b2/prm b2 routed.ncd

merges/prmB2withA1/pr assemble \\
../../prmA/prm a1/base routed full.ncd \\
../../reconfigmodules/prm b2/prm b2 routed.ncd

merges/prmB2withA2/pr verifydesign \\
../../prmA/prm a2/base routed full.ncd \\
../../reconfigmodules/prm b2/prm b2 routed.ncd

merges/prmB2withA2/pr assemble \\
../../prmA/prm a2/base routed full.ncd \\
../../reconfigmodules/prm b2/prm b2 routed.ncd

With the iterations several files that will be created are equivalent to each
other. In this case for partially reconfigurable area B

prmB1withA1/prm b1 blank.bit,
prmB1withA2/prm b1 blank.bit,
prmB2withA1/prm b1 blank.bit and
prmB2withA2/prm b1 blank.bit
are equal. Equivalent bitstreams for module b1 are

prmB1withA1/prm b1 routed partial.bit and

prmB1withA2/prm b1 routed partial.bit

and for module b2

prmB2withA1/prm b2 routed partial.bit and

prmB2withA2/prm b2 routed partial.bit.

Four variations of the full bitstream are generated representing the four
combinations of the two partially reconfigurable modules in this example:

prmB1withA1/base routed full.bit,

prmB1withA2/base routed full.bit,
prmB2withA1/base routed full.bit and
prmB2withA2/base routed full.bit
Partial reconfiguration design 12 235

In practice only one full bitstream is needed to load the FPGA. The recon-
figuration can then be conducted with the partial bitstreams. Which full
bitstream to take is incumbent on the user.

5. Creating Partially Reconfigurable Designs

This chapter gives a detailed description of the steps to take when recre-
ating a design to be partially reconfigurable. The next part summarizes two
sample designs to explain the single tasks to accomplish. The understanding
of this guide depends on a profound knowledge of VDHL, ISE and EDK. To
get a quick glance of VHDL see [208], for a in depth introduction consult, e.g.
[131]. Introductions to ISE and EDK are provided in [229] and [230]. For the
Early Access Reconfiguration Flow software to work, it is crucial to deploy the
correct version of ISE. The current issue of the update package requires ISE
8.1.01i (Service Pack 1) and updates it to ISE 8.1.01i PR8 resp. PR12. The
used EDK system has version 8.1.02i.

5.1 Animated Patterns

The sample design Animated Patterns is a partially reconfigurable pattern
generator for the XUP development board [228]. It is base on a project taken
from a previously published tutorial [40]. The design creates a simple mov-
ing pattern and directs it out to a VGA-compatible screen. Depending on the
current pixel position that is drawn on the monitor the reconfigurable module
Produce Color evaluates the color to be set for the next pixel. The system
with its two areas, reconfigurable and fixed, is depicted in figure 7.10.

Figure 7.10. Scheme of Animated Patterns

236 RECONFIGURABLE COMPUTING

Partially reconfiguring the device at run-time here means exchanging Produce

Color while the output to the monitor is continuing uninterruptedly. Control
for the user is provided through a simple c-program accessible over a serial
terminal. If a new partial bitstream is chosen the program loads the necessary
partial bitstream from a Compact-Flash card connected to the development
board and transmits the data to the ICAP, the Internal Configuration Access
Port, to reconfigure the device. The ICAP allows certain families of FPGAs11
to be reconfigured from within the FPGA by a processor, in this case the Pow-
erPC sitting in the Virtex II Pro on the XUP development board itself. The
design can be found on the book’s web page.

Figure 7.11. Two patterns that can be created with modules of Animated Patterns. Left: the
middle beam is moving. Right: the diagonal stripes are moving.

5.2 Video8
The second sample design is called Video8 and is used for the code and
command execution examples in the following guide. It deploys the Digilent
VDEC1 extension card [70] for the Virtex-II Pro XUP Development Board
[228] to digitize an analog video stream from a camera. The extension card
is connected to the XUP via a proprietary slot and is technically connected
through an I2 C Bus.
As shown in figure 7.12, the data stream is taken into the system by video in.
The module converts to the RGB color space and hands over to a filter, employ-
ing e.g. a Sobel-implementation for edge detection. This filter module will be
reconfigurable at the end of this tutorial. The filtered data is then processed in
vga out to be send to a VGA compliant screen.
The embedded hardcore processor on the FPGA is used to configure the
video digitizer card and the ICAP on start up. It is then used to run a simple
program to choose which partial bitstream to load next. The in- and output is
brought to the user’s console on a remote PC through the serial port.

11 currently the Xilinx Virtex-II, -II Pro, Virtex-4 and -5

Partial reconfiguration design 13 237

Figure 7.12. Scheme of Video8

5.3 Measures to take for PRM creation

Initially a completely working project should mark the starting point for the
creation of a partially reconfigurable design. The initial project Video8 non pr
marks the starting point for the instructions to follow.

1. Move partially reconfigurable parts to the top-level module

Partially reconfigurable modules may not be submodules to other than the
top-level module. To move its instatiation to the top-level generally two
approaches are possible:

reconstruct the initial design to place the reconfigurable parts and also
therewith connected module instantiations in the top-level. On the on
hand the resulting code is much easier to understand and modules can
be developed by different engineers according to the modular design
flow straight forward. On the other hand the facilitation that comes for
example from using EDK for the basic design might be outweighed
by the lot of work involved when extracting and reconnecting all the
necessary modules. This militates all the more regarding the synthesis
process. The placement tool places modules according to their connec-
tivity with each other and the constraints given in the UCF not accord-
ing to their actual position in the hierarchy. Thus reconstructing the
design might be useless since the placement is done independently.
In the Video8 non pr example video in and vga out are directly affili-
ated to the partially reconfigurable module to-be. As suggested in fig-
ure 7.13 these two modules at least would have to be instantiated in
238 RECONFIGURABLE COMPUTING

the top-level module. Additional difficulties arise from the use of the
OPB to connect the filter module. Either the developer has to adapt the
filter module and provide other means of connection to the rest of the
system or the OPB has to be brought up to the top-level as well. This
again affords work and other modules to be changed accordingly.

Figure 7.13. Reconstructing example Video8 to place the partially reconfigurable part and
connected modules in the top-level.

The other solution is to only replace the old instantiations of the par-
tially reconfigurable modules with instantiations on the top-level and
amend the entities that formerly hosted the modules: Ports have to be
added to redirect the input data to the partially reconfigurable modules
up the module hierarchy to its new location at the top-level module and
the output data back down to be fed into the system again. Figure 7.14
depicts the new structures. The big advantage with this approach is the
minimal adaption work to be done. An initial EDK design may be used
as is and can easily be changed.

The aimed at module for partial reconfiguration, rgbfilter, is initially

instatiated in entity system, the top-level module of the initial EDK de-
sign. To place rgbfilter in the top-level module according to the second
approach, video in has to be expanded as follows:

entity video_in is {
...

//ports going out to the rgbfilter

LLC_CLOCK_to_filter : out std_logic;
R_out_to_filter : out std_logic_vector(0 to 9);
G_out_to_filter : out std_logic_vector(0 to 9);
B_out_to_filter : out std_logic_vector(0 to 9);
h_counter_out_to_filter : out std_logic_vector(0 to 9);
v_counter_out_to_filter : out std_logic_vector(0 to 9);
Partial reconfiguration design 14 239

Figure 7.14. Moving a partially reconfigurable module to the top-level design on the example
of Video8

valid_out_to_filter : out std_logic;

//ports coming back from rgbfilter

R_in_from_filter : in std_logic_vector(0 to 9);
G_in_from_filter : in std_logic_vector(0 to 9);
B_in_from_filter : in std_logic_vector(0 to 9);
h_counter_in_from_filter : in std_logic_vector(0 to 9);
v_counter_in_from_filter : in std_logic_vector(0 to 9);
valid_in_from_filter : in std_logic
...
};

Since the ports of a submodule cannot be connected directly to those of its

hosting module, intermediate signals have to be introduced in the hosting
module. The entity vide in from the sample design has to be appended
accordingly after defining the signals themselves:

...
-- signals out to the filter:
R_out_to_filter <= R1;
G_out_to_filter <= G1;
B_out_to_filter <= B1;
h_counter_out_to_filter <= h_counter1;
v_counter_out_to_filter <= v_counter1;
valid_out_to_filter <= valid1;
-- processed signals in from the filter:
240 RECONFIGURABLE COMPUTING

R2 <= R_in_from_filter;
G2 <= G_in_from_filter;
B2 <= B_in_from_filter;
h_counter2 <= h_counter_in_from_filter;
v_counter2 <= v_counter_in_from_filter;
valid2 <= valid_in_from_filter;
...

The top-level entity must contain a similar construction.

2. Base design rebuilding

Many designs for the Xilinx FPGAs take advantage of the use of EDK and
its advanced wizards. Complete functioning systems on chip including the
software controlling can be created with a simple sequence of point-and-
clicks. In many cases the static base system, including the microprocessor,
the communication over a serial lines, RAM-controller and so forth, is cre-
ated with EDK. The controlling software that runs on the microprocessor,
may also be part of the system, leading to the creation of a bistream file
in the .elf format. This is a file supported by the EDK, which describe
the hardware configuration and the memory mapping with the software that
will be executed on the processor.
Up to now EDK does not support partial reconfiguration directly. The nec-
essary phases for the design flow still have to be done manually. Systems
implemented with EDK have to be exported to an ISE-format project, syn-
thesized as required and then be plugged into the top-level module as a
submodule.
In the case of Video8 non pr a basic system consisting of the infrastructure
for the PowerPC on the Virtex-II Pro, a RS-232 controller, a Video-in and
a VGA-out module. The design is created so that the partial bitstreams to
reconfigure the FPGA are loaded from a Compact Flash Card attached to
the FPGA.
Since the reconfigurable module in Video8 non pr has been included as a
core (Pcore) in the EDK project, the additional port definitions previously
described have to be applied to the system entity generated by the EDK
as well. This can be easily accomplished by adding External Ports to the
project. These changes will be written to the project’s MHS file similar to
the following sample lines:

PORT LLC_CLOCK_OUT = VIDEO_IN_PR_0_LLC_CLOCK_to_filter, DIR = O

PORT R_in = VIDEO_IN_PR_0_R_out_to_filter, DIR = O, VEC = [0:9]
PORT G_in = VIDEO_IN_PR_0_G_out_to_filter, DIR = O, VEC = [0:9]
PORT B_in = VIDEO_IN_PR_0_B_out_to_filter, DIR = O, VEC = [0:9]
Partial reconfiguration design 15 241

PORT h_counter_in = VIDEO_IN_PR_0_h_counter_out_to_filter, \\

DIR = O, VEC = [0:9]
PORT v_counter_in = VIDEO_IN_PR_0_v_counter_out_to_filter, \\
DIR = O, VEC = [0:9]
PORT valid_in = VIDEO_IN_PR_0_valid_out_to_filter, DIR = O
PORT R_out = RGBFILTER_0_R_out, DIR = I, VEC = [0:9]
PORT G_out = RGBFILTER_0_G_out, DIR = I, VEC = [0:9]
PORT B_out = RGBFILTER_0_B_out, DIR = I, VEC = [0:9]
PORT h_counter_out = RGBFILTER_0_h_counter_out, \\
DIR = I, VEC = [0:9]
PORT v_counter_out = RGBFILTER_0_v_counter_out, \\
DIR = I, VEC = [0:9]
PORT valid_out = RGBFILTER_0_valid_out, DIR = I

2.1. Additional Modules

To make the best use of the EDK, some of the needed components can
be added and connected in the EDK. A JTAG-PPC controller can help
to debug components connected to the PowerPC core. The opb sysace
module is the necessary interface to the SystemACE device that is de-
ployed for the management of the a Compact Flash Card attached to the
FPGA. opb timer is used here to measure the time the reconfiguration
takes. Finally the opb hwicap functions as a interface to the ICAP for
self-reconfiguration.
2.2. Bus Connections
The needed bus connections for the additional modules can also be
set up with EDK using the graphical user interface. This prevents a
lot of errors originating from misspelling of variables and names. It
also eases the task for less experienced users. Thus opb hwicap and
opb timer have to be connected to the same opb on which the Sys-
temACE module is connected to.
2.3. Setting Address Ranges and Software Locations
To ensure the correct function with the parameters set in the control-
ling software some of the project’s modules have to be set to a certain
address as stated in table 7.1.

Module Name Start Address Size

opb sysace 0x41800000 64KB
opb timer 0x41c00000 64KB
opb hwicap 0x40200000 64KB
docm ctlr 0xe8800000 32KB
iocm ctrl 0xffff8000 32KB
Table 7.1. New address ranges to be set in EDK
242 RECONFIGURABLE COMPUTING

docm ctlr and iocm ctrl are the controllers for data and instructions
to be stored in on-chip block RAMs. The program that controls the
reconfiguration of the FPGA in this case is to large to fit in a smaller
ram area. It can generally be put into the DDR-RAM as well but in
this special design, the framebuffer for the VGA output is placed in
the DDR-RAM, consuming a great share of the bandwidth of the bus
it is connected to. Thus placing the program in the DDR-RAM as well
generates unwanted distortions in the VGA output.
2.4. Rewriting the Control Software
The program deployed to control the non-pr design has to be extended
by the initialization of the ICAP module and a mean to choose which
partial bitstream to load next.
For the sample design, the control program main.c contains just the
initialization of the video decoder card. The initialization routine for
the ICAP and the SystemACE has to be added as well as a menu that
prompts the user for the next partial bitstream.
2.5. Export to ISE
The generated design has to be synthesized to be a submodule to the
top-level design. For this EDK supports the export of a project to
the ISE where detailed options for the synthesis can be set. Alterna-
tively, the module can be synthesized using xst. To export the project
in Project|Project-Options|Hierarchy and Flow the follow-
ing settings have to be made as follows:
Design is a sub-module with name system i
use project navigator implementation flow (ISE)
don’t add modules to an existing ISE-project

3. Creating a Top-Level Module

In order to integrate the different modules, a top-level module has to be
created. It should contain just module instantiations and clocks. Here, I/O-
buffers are put up and bus macros are instantiated. Input and output ports
are simply directed through the appropriate buffer instance. Bidirectional
inout-signals are split up in in-, out- and a tristate control signals, using
three corresponding buffers.
Instances have to be named, especially those, that will be static and recon-
figurable since their definition is placed in the UCF and the connection to
the design is established over the instance name.
The partially reconfigurable module rgbfilter requires bus macros for
signals R,G and B, v and h counter and valid from and to the rgbfilter.
Tools like PlanAhead or FPGA-Designer can facilitate this work. For de-
Partial reconfiguration design 16 243

tailed information on how and why to use bus macros see sections 3.3 and
4.3. An examples of the resulting UCF can be found on the book web page.
4. Synthesis
The synthesis of the partially reconfigurable project takes place in the sub-
folder synth of the project base folder. Here, corresponding subfolders for
static and reconfigurable modules and the top-level module can be found.
It should be mentioned again, that any submodules, including the partially
reconfigurable ones, have to be synthesized without automatically adding
I/O-buffers. Thus the option -IOBUF has be set to no. I/O-buffers only have
to be included in the top-level module. The exported project contains the
system as a submodule system i of type system and should already in-
clude the needed ports for the data flow to and from the reconfigurable mod-
ule. system i is ordered under a generated super module system stub
and can therefore directly be included in a top-level module. These ports
should be created by EDK when entering and wiring external ports as de-
scribed above. If not so the generated HDL files describing and instantiat-
ing entity system resp. system i (system.vhd resp. system stub.vhd)
have to be extended with the necessary ports.
For the sample design folders, the reconfigurable modules (mod *) as well
as a top and an edk directory must be created. There is no separate static
module, except the EDK system. Therefore, no static folder is required.
The ports to add to entity system are the same as when changing the former
super module of the one that should be partially reconfigurable.
5. Appending the UCF Taking a UCF built for example from EDK, several
amendments have to be made to let the partial reconfiguration work. As
stated before the definition of the area to place the partial and static blocks
and the location of the bus macros have to be declared in the UCF.
To set a region to be partially reconfigurable the UCF is extended by lines
like

INST "reconfig_module" AREA_GROUP = "pblock_recon_module";

AREA_GROUP "pblock_recon_module" RANGE=SLICE_X30Y64:SLICE_X45Y95;
AREA_GROUP "pblock_recon_module" MODE=RECONFIG;

With this the reconfigurable module reconfig module is associated with

the area group pblock recon module whose bounding box is defined by
the rectangle between slices X30Y64 and X45Y95. It is then marked recon-
figurable. A fixed part like a basic system created with EDK, can be set
with:

INST "system_i" AREA_GROUP = "pblock_fixed_area"; AREA_GROUP

244 RECONFIGURABLE COMPUTING

"pblock_fixed_area" RANGE=SLICE_X46Y16:SLICE_X87Y149;

Omitting MODE=RECONFIG tells the parser that the block will be fixed. Al-
though the UCF is a simple text file and can be edited manually, tools like
PlanAhead can be a big help with their architecture dependant graphical
user interfaces.
6. The Build Process
The build process starts with the synthesis of all the involved modules. The
submodules should not include I/O-buffers, which must all be included only
in the top-level module. The initial budgeting is partly done already, and
must be complete now. All other phases, from the activation to the final
assembly follow the steps described in section 4.

6. Partial Reconfiguration using Handel-C Designs

The structure of the directory previously described is the same, if another
HDL like Handel-C or even SystemC have to be used. Using Handel-C, the
goal is to provide the implementation of top-levels designs and separately the
implementation of each module in a much comfortable way. The Handel-C
must code for the entire system must be divided in code for modules, each
with its own interface for connecting other modules. The integration is then
done by connecting the modules together using signals in the top-level design
as shown in figure 7.1. For each module to be reconfigured later, a separate
top-level must be produce. The transition from one design to the next one will
then be done later using either full reconfiguration or partial reconfiguration
with modules representing the difference from one top-level to the next one.
We explain this modular design for Handel-C by mean of one example pro-
vided below.
We consider a system containing an adder in its first configuration. The
system can be partially reconfigured to act implement a subtractor in place
of the adder. Both modules (adder and subtractor) have to be implemented
separately. The implementation of the adder is provided in the following code
segment in Handel-C.

void main() {
unsigned 32 res;

// Interface definition
interface port_in(unsigned 32 var_1
with {busformat="B<I>"}) Invar_a();

interface port_in(unsigned 32 var_2

with {busformat="B<I>"}) Invar_b();
Partial reconfiguration design 17 245

interface port_out() Outvar(unsigned 32 Res = res

with {busformat="B<I>"});

// Addition
res = Invar_a.var_1 + Invar_b.var_2; }
The first part of the code is the definition of the interfaces for communication
with other modules and the second part realizes the addition of the input values
coming from the input interface, and the output values are sent to the output
interface. We need not to provide the implementation of the subtractor, since
it is the same like that of the adder, but instead of adding we subtract. Having
implemented the two modules separately, each module will be inserted in a
separated to-level. Up to the use of the adder or subtractor, the two top-levels
are the same. The design can now be reconfigure later to change the adder
against the subtractor. Connecting the module in a top-level design is done as
shown in the following code segment.
unsigned 32 operand1, operand2;

unsigned 32 result;

interface adder(unsigned 32 Res)

my_adder(unsigned 32 var_1 = operand1,

unsigned 32 var_2 = operand2)

with {busformat="B<I>"};

void main() {
operand1 = produceoperand( 0);

operand2 = produceoperand( 3);

result = my_adder.Res;

dosethingwithresult(result);
}
According to the top-level design being implemented, the adder will be re-
placed by a subtracter. The next question is how to keep the signal integrity
of the interfaces. Since there is no bus macro available in Handel-C, bus-
macros provided by Xilinx must be used as VHDL-component in a Handel-C
246 RECONFIGURABLE COMPUTING

design. For instruction on integrating a VHDL code in Handel-C, consult the

Celoxica Handel-C reference manuals [125]. Bus Macros are provided with
the Xilinx application note on partial reconfiguration [128] [227]. Before set-
ting constraints on a Handel-C design, the design have to be compiled first.
Afterwards, the resulting EDIF-files must be opened with any text-editor and
the longest pattern that contains the name of the module as declared in the
top-level module is selected. This is useful because the Handel-C compiler
generate EDIF-code and automatically adds some characters to the module
name used in the original design. If the original module name is used it will
not be recognized in the following steps of the Modular Design Flow. With the
EDIF for the modules and that of the top-level designs we now have all what
we need to run the modular design flow explained in section 3.

7. Platform design
One of the main factor that have hindered the implementation of on-line
placement algorithms developed in the past and presented in chapter 5, is the
development process of reconfigurable modules for the Xilinx platforms. As
we have seen in the previous section, partial reconfiguration design is done
with many restrictions, which make a systematic development process for par-
tial reconfiguration very difficult. Each module placed at a given location on
the device is implicitly assigned all the resources in that area. This includes
the device pins, clock managers and other hard macros components like the
embedded multipliers and the embedded BlockRAMs.

Figure 7.15. Modules using resources (pins) not available in their placement area (the com-
plete column) must use feed-through signals to access those resources

As illustrated in figure 7.15, a module using resource outside its placement

area must use connection running trough other modules in order to access its
resources. We call those signal used by a given module and crossing other
Partial reconfiguration design 18 247

modules feed-trough signals. Using feed-through lines to access resources

have two negative consequences:

Cumbersome design automation: Each module must be implemented with

all possible feed-trough channels needed by other modules to reach their
resources. This problem was already identified in section 1, where dedi-
cated channels are preconfigured on the device to be used by modules at
run-time. Because we only know at run-time, which module needs to feed
the signal through, many channels reserved for a possible feed-through be-
come redundant.

Non relocation of modules: Modules accessing resources, like pins n a

given area of the chip, are no more relocatable, because they are compiled
for fixed locations where a direct access to their pins must be available.

Most of the FPGA platforms available on the market do not provide solu-
tion to the problems previously mentioned. Many systems on the market offer
various interfaces for audio and video capturing and rendering, for communi-
cation and so forth. However each interface is connected to the FPGA using
dedicated pins in a fix location. Modules willing to access a given interface
like the VGA must be placed in the area of the chip where the VGA signals
available such that they can be assigned those signals among other resources.
This makes a relocation of module at run-time impossible. Relocation pro-
vides the possibility to have only one bitstream for a module representing the
implementation of that module for a given location. At run-time, the coordi-
nate of the module is modified to match the location assigned to the module
by the on-line placer. Without relocation, each module must be compile for
each possible position on the device where it can be later placed. A copy of
the component bitstream must therefore be available for each location on the
chip. The amount of storage needed for such a solution does not make such an
approach attractive.
A research platform, the Erlangen Slot Machine (ESM) that we next present,
was developed at the university of Erlangen to overcome the drawbacks of
existing platform, and therefore allow unrestricted on-line placement on an
FPGA platform. Through the presentation of the ESM, we also hope to high-
light some of the requirements in the platform design for reconfigurable com-
puting.
The goals that the ESM-designers had in mind while developing a new plat-
form was to overcome the deficiency of existing FPGA-Platforms by provid-
ing:

A new highly flexible FPGA Platform in which each component must not
be fixed all the time at a given chip location.
248 RECONFIGURABLE COMPUTING

A suitable tool support which goal is to ease the development process of

modules for run-time reconfiguration and communication and to make an
efficient use of the architecture.

While the tooling aspect is important, it is not part of the basic requirements
for designing the platform itself. We therefore just focus on the architectural
aspect of the platform design in the next sections.

7.0.1 Drawback of existing systems

The practical realization and use of partial and dynamic reconfiguration on
the current existing FPGA-based reconfigurable platforms is highly restricted,
due to the following factors:

1. Limitation of reconfiguration on actual FPGAs: Very few FPGAs al-

lowing partial reconfiguration exist on the market. Those few FPGAs like
the Virtex series, allow only a restricted partial reconfiguration. As stated
in section 1, up to the Virtex II Pro, the reconfiguration had to be done col-
umn wise meaning that loading a module in a given area will affect all the
modules on top and below the module. From the Virtex 4 and upwards,
the reconfiguration of a module affects only those components in the same
blocks, which share a common column with the reconfigurable module.
The block does not span the complete device column, but only a few num-
ber of lines.

2. I/O-Pin-Problematic: The design of existing FPGA platforms makes it

more difficult to design for partial reconfiguration. Most of the existing
platforms have the peripheral components like Video, RAMs, Audio, ADC
and DAC connected at fixed location on the device. This has the conse-
quence that a module must be stuck on a given region where it will have
access to his pins, and therefore making a relocation very difficult. An-
other problem related to the pin connection is that the pins belonging to
a given logical group like video, audio, etc... are not physically grouped
together. On many platform they are spread around the device. A module
willing to use a device will have to feed many lines through many different
components.
This situation is illustrated on figure 7.16 for the Celoxica RC200 platform.
To see are two modules among which a VGA module, that are implemented
on the FPGA. The VGA module uses a large number of pins at the bottom
part of the device and also on the right side. It is possible to implement
a module without feed-through only on the two first columns on the left
side. Most of the reconfigurable module will need more than the two avail-
able column of resources to be implemented. Feed-through lines that run
Partial reconfiguration design 19 249

Figure 7.16. Illustration of th pin problematique on the RC200-Board

through the reconfigurable module must be used, in order for the VGA
module to access the pins cover by the reconfigurable module.
Similar situation are not only present on the Celoxica boards. The XESS
boards [56], the Nallatech boards [165], the Alpha boards [8] face the same
limitations.
On the XF-Board [177] [211] from the ETH in Zurich, the connection to
the peripherals are done on the side of the device. Each module access
its peripheral through an operating system (OS) layer implemented on the
left and right part of the device and between the component swapped in
and out. This approach provides only a restricted solution to the problem,
since all modules must be implemented with a given number of feed-trough
lines and interfaces to access the OS layer for communication with the pe-
ripheral. The intermodule communication as well as the communication
between a module and its peripheral is done via buffers provided by the
OS. This indirect communication will affect the performance of the sys-
tem. Many other existing platforms like the RAPTOR-Board [132], Celox-
ica RC1000 and RC2000 [43] are PCI systems, which require a workstation
for operation. The use in stand-alone systems as it is the case in embedded
is not possible.

3. Intermodule communication: One of the biggest problem not considered

in the design of reconfigurable platform is the intermodule communication.
Modules placed at run-time on the device may need to exchange some data
among each other. This request of communication is a dynamic task due to
the dynamism in on-line placement. For module placed far from each other
it will not be possible to feed communication line through several module
in order to establish a connection. New paths are then necessary in order to
allow a dynamic communication to take place at run-time.
250 RECONFIGURABLE COMPUTING

The previous limitations was the primary motivation in the design of the
Erlangen Slot Machine (ESM), whose concept we next present.

7.0.2 The Erlangen Slot Machine

The Erlangen Slot machine (ESM) is made upon a BabyBoard mounted on
a MotherBoard. The separation of the system in two boards allows the Baby-
Board which contains the reconfigurable module to be used on a wide variety
of systems. For the integration of the ESM in a new platform, no redesign of
the complete system must be done. Only a new MotherBoard must be pro-
vided according to the computational requirement in the new environment. A
multimedia system for example will provide a MotherBoard with multimedia
devices like video and audio. In an automotive system, the MotherBoard will
mostly contain sensor and actuators. The BabyBoard can be mounted on a
PCI MotherBoard in a computer system or on a stand alone MotherBoard in
an embedded system. Actually, only one variety of MotherBoard is available,
that is mainly targeted for applications in multimedia and signal processing. In
the next, we provide a detailed description of the boards functionality.

7.0.3 The BabyBoard: Computation and Reconfigurable Engine

The reconfigurable engine of the ESM platform is a BabyBoard, which fea-
tures an FPGA Virtex II-6000 FPGA from Xilinx, several SRAMs and a con-
figuration circuit. Due to the restriction20 in the reconfiguration of the Virtex
FPGAs, the architecture was adapted to match the following:
Free relocation of modules: On-line placement of modules on a recon-
figurable device, in this case the FPGA, is done by downloading a partial
bitstream implementing a given task in the FPGA. The full and partial bit-
streams represent circuits implemented on a given region of the device at
compile time. In other to place a module in another location other than
the location for which it was compiled, a relocation of the module must be
done. The relocation of a given module therefore modifies the coordinates
of all the resources used by the module in its previous location (the location
in which the module was constrained at compile time) with the coordinates
of the module in the new location. Relocation can be done only if all the
resources are available and structured in the same way in the previous lo-
cation and in the new location. This include the device pins used by the
module. If the connection of a module with its peripheral is done via some
pins fixed at a given chip location, the module must be constrained at that
location. A relocation will be possible, but the module must feed signals
through all the other modules between the pins and the new location.

20 The reconfiguration can be done only column wise

Partial reconfiguration design 21 251

This problem is solved on the ESM by avoiding a direct connection of pe-

ripheral on the FPGA. As shown in figure 7.17, all the bottom pins from
the FPGA are connected to an interface controller (crossbar). The Interface
devices are connected to the interface controller as well. This makes it pos-
sible to establish any connection from one module to its peripheral via the
crossbar. No mater where the module resides in the device, the crossbar
can be set at run-time to connect the module to its peripheral.

Figure 7.17. Architecture of the ESM-Baby board

Uniform repartition of resource: The ESM like any other reconfigurable

platform is primarily a parallel computing engine in which each module
carries its own computation and communicate with other modules. There-
fore a certain amount of resources must be allocate to each module for its
operation, at least for a given period of time. Memory is very important
in application like video streaming in which a given module must exclu-
sively access a picture at time for computation. However, the capacity of
the available BlockRAMs in FPGA is limited. External memory (SRAM
or DRAM) must therefore be added to allow the storage of large amount of
data by each module. To allow a module to exclusively accesses its mem-
ory, the SRAM are connected at the top part of the FPGA. In this way, a
module will have its connection to its peripheral from the bottom part and
the top part will be used to temporally store the computation data. No mod-
ule will need to feed its lines through the other modules for accessing its
resources. According to the number of memory banks which can be con-
nected on the top level, the device is then divided in a set of exchangeable
252 RECONFIGURABLE COMPUTING

slots, each of which have access to the SRAM on the top part and to the
crossbar at the bottom.
The name Erlangen Slot Machine was choosen, because it was developed
at the University of Erlangen-Nuremberg in Germany, but mainly due to
this organization in slots. This modular organization of the device simpli-
fies the relocation, primary condition for a viable reconfigurable computing
system. Each module moved from one slot to another one will find the same
resources there. The architecture is illustrated in figure 7.17. The Baby-
Board is logically divided into columns of 2 CLBs called micro slots. The
micro slots are the smallest allocatable units in the system. Each module
must therefore be implemented in a given number of micro slots. Due to
the number of pins needed to access one the external SRAM, each module
willing to used an SRAM module must be implemented in a multiple of 3
micro slots.

7.0.4 The configuration manager

A part from the main FPGA, the BabyBoard also contains the configura-
tion circuitry. This consists of a CPLD, a configuration FPGA, that is a small
Spartan II-FPGA, and a Flash.

The CPLD is used to download the Spartan II’s configuration from the flash
on power-up. It also contains board initialization routines for the on board
PLL and the Flash.

The configuration program for the main FPGA is implemented in the Spar-
tan II. Due to its small size, this reconfiguration program could be directly
implemented in the CPLD, which could configure the main FPGA at power
on. But a much larger device was chosen to increase the degree of freedom
in the reconfiguration management. As state before, the relocation is an
important operation of the ESM. This can be done in different ways: The
first approach is to keep a bitstream for each possible module and each pos-
sible position in memory. However, the size of the Flash cannot allow us to
implement this solution. The second approach is the on-line modification
of the coordinate of the resources used in the module’s bitstream to match
the new position. This modification can be done for example through a
manipulation of the bitstream file with the Java program JBits [102] previ-
ously presented. However file manipulation is a time consuming operation
which will increase the reconfiguration time. The last and most efficient
solution if to compute the new module’s position while the bitstream is be-
ing downloaded, using a dedicated circuitry. This requires a much larger
hardware code than the simple reconfiguration. This is why the Spartan II
was chosen.
Partial reconfiguration design 22 253

The Flash provides a capacity of 64 MByte, thus enabling the storage of up

to 32 Full configurations and few hundreds partial bitstreams for the FPGA
Virtex II 6000 that was used.

7.0.5 Memory
6 SRAM banks with 2 MBytes each are vertically attached to the board on
the top side of the device, thus providing enough memory space to six different
slots for temporal data storage. The SRAMs can also be used for shared mem-
ory communication between neighbor modules e.g for streaming applications.
They are connected to the FPGA in such a way that the reconfiguration of a
given module will not affect the access to other modules.

7.0.6 Debugs lines

Debugging capabilities are offered through general purpose I/O provided in
regular distance between the basic slots. A JTAG port provides debug capabil-
ities for the main FPGA, the CPLD and the Spartan II.

7.0.7 The MotherBoard

The MotherBoard provides programmable links from the FPGA to all pe-
ripherals. The physical connections are established at run-time through a pro-
grammable crossbar implemented in an a Spartan FPGA on the MotherBoard.
This crossbar functionality basically solves the I/O pin dilemma of many ex-
isting FPGA platforms, thus allowing free relocation of modules requiring I/O
pin connectivity. Besides the run-time programmable crossbar, many peripher-
als for multimedia and communication are available on the board (figure 7.18).
Video capture and rendering interfaces as well as high speed communication
links are available on the MotherBoard. This interfaces allows the platform to
be used for instance in autonomous system, where a robot may collect pictures
from the environment and send them after a preprocessing step to a central
station for evaluation.

7.1 Intermodule communication

One of the central point in dynamic reconfiguration is the communication.
Each module that is placed on the device must collects its data from a given
source and send its results to a given destination. This problem is given an im-
portance only in few systems like the XF-Board of Zurich [177] [211]. How-
ever the XF-Board implements only a communication through a third party
which is slow, due to the long path that messages have to go.
In the ESM, the communication among different modules (figure 7.19) can
be realized in three different ways: The first one is a direct communication
using bus macros between adjacent modules. The second one is the shared
254 RECONFIGURABLE COMPUTING

Figure 7.18. Architecture of the ESM MotherBoard

memory using the external SRAMs or the internal BlockRAMs. However,

only neighbor modules can use those two communication modes. For modules
placed in non-adjacent slots, we provide a dynamic signal switching communi-
cation, the reconfigurable multiple bus (RMB) presented in section 4.1. Also,
the crossbar switch can be use as well for communication between modules.

7.1.1 Adjacent Communication

Adjacent communication is done between two modules, which abut each
other. Bus macros must be used for a direct communication between neigh-
boring modules. The number of Bus-macros needed is limited by the amount
of signal that can go through the Bus-macros and the number of signals to be
connected.

7.1.2 Communication via shared memory

The communication between two neighboring modules can be done in two
different ways using shared memory:
Communication using BlockRAM: The BlockRAMs are dual ported RAMs
used in the FPGA mostly to implement FIFOs and other memory. It is a
nice possibility to implement communication among two neighbor mod-
ules, working in two different clock domains. The sender will just write on
Partial reconfiguration design 23 255

SRAM SRAM SRAM

FPGA

FPGA
M1 M2 M3 M1 M2 M3
FPGA

FPGA
M1 M2 M3 M1 M2 M3

Crossbar

Figure 7.19. Intermodule communication possibilities on the ESM

one side with its own frequency, while the receiver will read the data at the
other end with its own frequency.

Communication using external RAM: The communication through the ex-

ternal RAM is particular useful in applications in which each module must
process a large amount of data and then send the processed data to the next
module. This is mostly the case in video streaming application in which a
first module captures a stream, image by image. The image are alternately
store in two different memory which are accessed by the next module for
reading. On the ESM, each SRAM can be accessed by three modules as
shown in figure 7.20.
Since a SRAM can only be accessed by one module at a time, a controller
is then used to manage the concurrent SRAM access between the three
modules. Depending on the application, the user may set the priority of
accessing the SRAM for the three modules.
256 RECONFIGURABLE COMPUTING

Figure 7.20. SRAM-based intermodule communication on the ESM

7.1.3 Communication via RMB

The communication between non adjacent modules happens through the 1-
D circuit switching paradigm presented in section 4.1.

7.1.4 Communication via the crossbar

The last possibility to establish a communication among modules is to use
the crossbar switch available on the mother board. Because all modules are
connected to the crossbar via the pins at the bottom of the FPGA, the commu-
nication among the modules can be set in the crossbar as well. This possibility
should be use only as emergency solution, due to the high latency caused by
the of-chip routing.

8. Enhancement in the Platform Design

The design of the Erlangen Slot Machine was strongly influenced by the ar-
chitectural limitations of the Xilinx Virtex II FPGA, in particular their colum-
nwise reconfiguration. This limitation was the primary motivation for the im-
plementation of the crossbar switch on a separate FPGA. All the connections
between running modules and their peripherals could be implemented in the
crossbar switch and be safe from the reconfiguration of the module. Substan-
tial efforts were therefore place in the design of the two boards previously
described.
With the introduction of the Virtex 4, Xilinx has introduced a new paradigm
in which, complete column must not be replaced on reconfiguration. This new
Partial reconfiguration design 24 257

development can lead to a very great enhancement on the platform design in

the two following aspects:

First, the crossbar switch can now be implemented in the FPGA rather than
on a separated device. In this case, we need to attach all the peripheral on
one side of the block in which the crossbar is implemented. Modules can
then be connected on the other side of the block. We can even think of a
system in which all components are attached around the chip. The crossbar
can the be implemented as a ring, distributed module around the chip.
The distribution of the resources, like the external RAM must not be done in
a column wise manner anymore. Resources should now be homogeneous
spread around the device, in order to allow different modules, which are
placed on different blocks to access their own resources.

Figure illustrate the enhancements on the Erlangen Slot Machine, using the
Virtex 4 and Virtex 5 FPGAs as previously explained.

Figure 7.21. Possible enhancement of the Erlangen Slot Machine on the Xilinx Virtex 4 and
Virtex 5 FPGAs

Despite the advantages provided by the new class of Virtex devices, a great
disadvantage that results from this architectural enhancement is the communi-
cation: for the columnwise reconfigurable devices, the 1-D circuit was a sim-
ple and efficient possibility to enforce communication among different mod-
ules running on the FPGA. In a two dimensional reconfigurable device as it
258 RECONFIGURABLE COMPUTING

is the case with the Virtex 4 and Virtex 5, the difficulty of implementing the
communication increases. In 2-D we have a quadratic growth in the amount
of resources needed to implement the communication infrastructure. I.e., the
amount of resources needed to implement a 2-D communication infrastruc-
ture is not only twice the amount needed to implement a 1-D communication
infrastructure but four times.

9. Conclusion
We have presented in this chapter the different possibilities to design for
partial reconfiguration on Xilinx devices. Our goal was not to rewrite a com-
plete manual on partial reconfiguration, since several description on this exist
[128] [227]. The Xillinx manuals as well as the work of Sedcole [190] provide
very good descriptions on using the modular design flow and the early access.
our primary motivation was to provide a king of tutorial, based on our experi-
ence and for which a workable design, not too complex, but also not so easy,
exists. The designs as well as all the scripts need for compiling are available
to download from the book’s web page. The difficulty in designing for partial
reconfiguration can be reduced, if the target platform is well designed. One of
such platform is the Erlangen Slot Machine that was presented, with the goal
to emphasize the challenges in designing such a platform. The ESM is how-
ever strongly influenced by the column wise reconfigurable Virtex. The price
we pay for flexibility is very high. With the advent of new Virtex 4 and virtex
5, enhancements can be made in the design in order to increase the flexibility,
while reducing the costs.
One of the main problem remain the communication between the module
at run-time. While this problem is somehow better solved in a 1-D reconfig-
urable device through the use of circuit switching and dedicated channels on
the module, its extension on a 2-D reconfigurable device is not feasible due
to the amount of resources needed. Viable communications approaches like
the DyNoC was presented in chapter 6 however, with the amount of resources
needed by those approaches, their feasibility is only possible if manufacturers
provide coarse-grained communication element in their devices.
Chapter 8

SYSTEM ON A PROGRAMMABLE CHIP

Developments in the field of FPGA have been very amazing in the last two
decades. FPGAs have move from tiny devices, with few thousands of gates,
only able to implement some finite state machines and glue-logic to very com-
plex devices with millions of gates as well as coarse-grained cores. In year
2003, a growth rate of 200% was observed in the capacity of the Xilinx FPGA
in less than 10 years, while in the meantime, a 50% reduction rate in the power
consumption could be reached with the prices also having the same decrease
rate. Other FPGA vendors have face similar development and this trend is
likely to continue for a while. This development, together with the progress
in design tools, has boosted the acceptance of FPGAs in different computation
fields. With the coarse-grained elements like CPU, memory, arithmetic units,
available in recent FPGAs, it is now possible to build a complete system con-
sisting of one or more processors, embedded memory, peripherals and custom
hardware blocks in a single FPGA. This opportunity limits the amount of com-
ponents that must be soldered on a printed circuit board in order to get a FPGA
system working.
In this chapter, we present the different possibilities that exist to build those
systems consisting of one or more processors, peripherals and custom hard-
ware component.
We start with an introduction in system on programmable chip, and then
we present some of the various elements usually needed to build those sys-
tems. We then present at the end of the chapter a design approach for adaptive
multiprocessor systems on chip.

1. Introduction to SoPC
A system on chip (SoC) is usually defined as a chip that integrates the major
functional elements of a complete end product. Figure 8.1 presents the achieve-
260 RECONFIGURABLE COMPUTING

ment of a system on chip, namely the replacement of a complete printed circuit

board by a single chip.

Figure 8.1. Integration of PCB modules into a single chip: from system on PCB to SoC

A system on chip usually consists of a processor, memory, peripheral mod-

ules and custom hardware modules. Contrary to a system on chip that is man-
ufactured on an integrated circuit and cannot be modified again, a system on
programmable chip (SoPC), or programmable system on chip, is built on pro-
grammable logic. Its structure can be modified by the end user, either at com-
pile time or at run-time by means of full or partial reconfiguration.
In the next sections, we take a brief look on the main components of sys-
tem on programmable chip and present some key implementations of those
components by manufacturers.

1.1 Processor
Processors are the central processing and coordination units in system on
programmable chips. Besides the coordination of the complete system, they
are in charge of collecting the data from different peripheral modules or from
the memory, process those data and store them into the memory or sent them
to the peripheral modules. Also the processor initializes the peripherals as
well as the dedicated hardware modules on the chip and manages the memory.
The most widely used processors are those used on Xilinx and Altera devices,
since those two companies control the largest part of the programmable device
market. Besides the processors offered by those two companies, other platform
independent implementations exist. We list some of the available processors
next.
System on a Programmable Chip 261

1.1.1 The MicroBlaze Processor Core

The MicroBlaze [127] is a soft 32-bit RISC processor core designed and
commercialized by Xilinx. The processor is not directly available on a chip. It
exists only as reference design, optimized for the Xilinx FPGA. With a clock
frequency of up to 200 MHz and a to possibility to include a 32-bit single
precision floating point unit (FPU) in the IEEE-754 format into the datapath,
the MicroBlaze is one of the fastest available soft core processor. It consists of
a set of fixed components and a set of optional features that can be configured
at design time by the user. The main fixed features of the MicroBlaze are:

An Arithmetic and Logic (ALU) Block unit featuring a shifter

thirty two 32-bit general purpose registers

Harvard architecture with separate 32-bit address bus and 32-bit data bus

a 5 stage pipeline in the version 5.1a

seven Fast Simplpex Links (FSL) that can be used to connect the processors
to custom hardware

an interface for connecting the processor to an On-chip Peripheral Bus

(OPB)

Besides these fixed features, the MicroBlaze processor can be parameterized

by the users at compile time to include additional functionalities. The latest
version of the MicroBlaze, the version v5.1a, can be configured to support the
following additional features:

a barrel shifter, a multiplier and a divider can be added to the ALU-Block

a 32-bit single precision Floating Point Unit (FPU) in the IEEE-754 format

A data cache and an instruction cache

an interface to the Local Memory Bus (LMB), a high-speed synchronous

bus used primarily to access on-chip block RAM.

A CacheLink for accessing external memory

debug logic

The MicroBlaze soft processor core is available as part of the Xilinx Embed-
ded Development Kit (EDK), which includes a comprehensive set of system
tools to design an embedded application in a Xilinx FPGA.
262 RECONFIGURABLE COMPUTING

1.1.2 The PowerPC 405 Processor

Contrary to the MicroBlaze that exists only as reference design compiled to a
netlist, the PowerPC 405 processor [126] is immersed into some Xilinx Virtex
FPGAs (the Virtex II Pro, the Virtex 4 and the Virtex 5). The integration of the
PowerPC directly into the chip at fabrication allows for a great performance
increase compare configurable processor cores.
The PowerPC 405 processor core used in the Xilinx devices is a 32-bit im-
plementation of thePowerPC embedded environment architecture that can be
clocked at up to 450 MHZ. It provides a dual mode operation, which allows
programs to run either in privileged mode or in user mode. The main compo-
nents of the PowerPC 405 are:

the Central Processing Unit (CPU) with the following features:

– a five stage (fetch, decode, execute, write-back) pipelined datapath. A
fetch queue consisting of two prefetch and a decode buffer is used to
queued instructions in case of stalls. Instructions flow directly to the
decode buffer if the prefetch buffers are empty.
– a general purpose register file (GPR) consisting of thirty two 32-bit reg-
isters accessible via three reads and two write ports. The five ports on
the GPR allow the processor to execute load/store operations in parallel
with ALU or MAC operations.
– a issue execution unit consisting of ALU and a single cycle throughput
Multiply Accumulate Unit (MAC).
an exception handling module for servicing a total of 19 critical and non
critical exceptions that can be caused by error conditions, the internal timers,
debug events, and the external interrupt controller (EIC) interface.
a memory management unit, which provides address translation, protection
functions, and storage-attribute control for a total of 4 GB non segmented
address space.
a 16 KB instruction-cache and a 16 KB data-cache, which can be accessed
respectively through the data cache and instruction cache units.
additional resources like timers and debug units.
a Processor Local Bus (PLB) interface providing a 32-bit address and three
64-bit data buses attached to the instruction-cache and data-cache units.
a Device Control Register (DCR) interface for the attachment of on-chip
device control, clock and power management interfaces for clock distribu-
tion and power management.
System on a Programmable Chip 263

a JTAG debugging port, on-chip interrupt controller interface and on-chip

memory controller interface for attaching additional memory.
an Auxiliary Processor Unit (APU) interface and a APU controller that are
used for extending the native PowerPC 405 instruction set with custom in-
structions, implemented as coprocessor on the FPGA Fabric. With the APU
a tighter integration of application-specific function in to the processor is
possible. The APU can be used for processor interconnection as well. The
PowerPC 405 does not have a floating point unit, however Xilinx provides
in its module library a dedicated floating point unit that can be attached to
the processor through the APU interface.

1.1.3 The Nios Processor Core

Like the MicroBlaze, Altera’s Nios II [124] processor core is available as
reference design to be synthesized and to be downloaded into FPGAs. The
architecture supports operation in user mode as well as in supervisor mode,
thus allowing for a protection of the control registers. The Nios II core can
be customized by the designer by adding optional functionalities to the basic
system. It main features are:
a register file that consists of thirty two 32-bit general purpose integer reg-
isters, and six 32-bit control registers. Floating point registers can be add
to the register file by the customization of the processor.
an Arithmetic Logic Unit supporting the main arithmetic and logic opera-
tions as well additional shift and rotate operations.
a single precision floating point as specified by the IEEE Std 754-1985, that
can be added to the core by customization.
optional, a NIOS core may include one instruction cache and/or one date
cache, depending on the user application. Caches can be avoided in order
to keep the design small.
one data bus selector module and one instruction-bus selector interfaces,
each of which can be used to provide memory and I/O access. Each con-
troller provides an Avalon master port for connecting the memory via the
Avalon switch fabric, and an interface to fast memory outside the Nios II
core.
various interface for debugging and interrupt handling.
a custom instruction port that can be used to extend the native instruction
set. A set of instructions implemented in a hardware module can be acces-
sible by connecting that hardware module to this port.
264 RECONFIGURABLE COMPUTING

1.1.4 The LEON Procesor Core

Developed by the company Gaisler Research, the LEON processor core [1]
is an open source 32-bit processor core based on the SPARC V8 architecture.
Several versions of the LEON exists, however we present only the LEON3 in
this section. The LEON 3 is available as a synthesizable VHDL model, which
means that it can be implemented on any FPGA platform with the correspond-
ing vendor tool. The complete source code is open and available under the
GNU GPL license. The main features of the LEON 3 processor, which can be
clocked at up to 125 MHz, are:

Advanced 7-stage pipelined datapath

Hardware multiply, divide and MAC units

a fully pipelined floating point unit in the IEEE-754 format

Separated instruction and data cache, whose size, associativity, and replace-
ment policy, can be configured

a memory management unit with configurable TLB

an interface to the AMBA-AHB bus

support for Symmetric Multiprocessor (SMP)

various interfaces for debugging, clocking and power-down mode

1.2 Memory
A system on programmable chip needs memory for storing instructions and
data. Memory elements are available in small amount on almost all modern
FPGAs. This is usually used to build the caches directly on the chip. For
applications that require more memory, external SRAM or DRAM must be
used. On-chip memory can be built from the memory elements, the block
RAMs, or from the LUTs. The use of the LUT in the second case to built
memory has two drawbacks: First the amount of available computing resources
decreases, thus reducing the complexity of the logic that can be implemented
on the chip. Second, the memory built with LUTs is distributed across the
chip, meaning that chip interconnection must be used to connect the spread
LUTs. This leads to decrease in performance in the memory access. On-chip
block RAMs are usually dual-ported. They therefore provide a nice possibility
to integrate a custom hardware module, in particular those working in two
different clock domains. The memory is used in this case to synchronize the
module with each other.
System on a Programmable Chip 265

1.3 Peripheral
Peripherals are used in the system for communication with the external
world and for debugging purpose. The peripheral components are usually pro-
vided by the board manufacturers as ready to the module that can be inserted
in the SoPC-design. Because of their lower frequency, compare to the one of
processors, peripherals should be connected to the low-speed bus, which in
turn can be connected to the processor bus through a bridge. The communi-
cation between the processor and the available peripheral can be done either
through the I/O mapping paradigm or through the memory map paradigm. In
the first case, the processor addresses each peripheral using special instructions
that directly address the peripheral on the bus. In the second the peripheral is
assigned an address space in the memory and the communication happens in
this space with the processor writing in control register that can be read by the
peripheral. The peripheral responds by writing in status registers that can be
read by the processor.

1.4 Custom hardware

The final types of component in a system on programmable chip are the
custom hardware. Those are hardware components that implement special
functions for which one intend to speed-up the computation. Custom hard-
ware modules are usually streaming-based, meaning that the computation is
performed on a stream of data that flow through the hardware module. Care
should be taken while integrating custom hardware modules in a system on
programmable chip. Connecting a custom hardware module on a peripheral
bus may result on a performance decrease, if the data rate of the computation
is much higher than that of the bus, which is usually the case. Direct connec-
tion and communication using dual-ported RAM can help in this case to allow
the custom module to work with its maximal frequency.

1.5 Interconnection
The interconnection provides the mechanism for integrating integrate all the
components previously described in a workable system on programmable chip.
The interconnection mechanism provides the communication channels, the in-
terface specification, the communication protocols and the arbitration policy
on the channels. The design of the interconnection infrastructure depends on
the target application. In traditional system on chip, where the infrastructure
is fixed at chip production, the communication infrastructure must be as gen-
eral as possible to serve all class of applications that can be implemented on
the chip. In programmable devices, the flexibility can be used to modify the
interconnection infrastructure according to the type of application to be imple-
mented. In this case, each application can first be analyzed in order to derive
266 RECONFIGURABLE COMPUTING

and implement the best communication infrastructure for its computation. De-
spite the great interest in network on chip in the last couple of years, intercon-
nection on chip is dominated by the SoC communication paradigm which is in
most of the case bus-based. Leading existing solutions were previously devel-
oped for SoCs before adapted to SoPCs. The general connection mechanism
is provided in figure 8.2
It usually consists of two different buses. A high performance bus that is
used by the processor to access the memory and a slow bus used to connect
the peripherals. High performance dedicated hardware modules are connected
to the high performance bus, while low performance custom hardware compo-
nents are connected to the low performance bus. A bridge is used to allow for
communication to happen between two modules attached on the two different
busses. Besides those two buses, several possibilities exists to directly connect
the component. Dedicated lines or crossbar switches can be used, even in co-
habitation with the two previously described buses. The two well established
bus systems in the world of programmable system on chip are the CoreConnect
from IBM and the ARM AMBA.

1.5.1 The IBM CoreConnect

As shown on figure 8.2 the IBM CoreConnect communication infrastructure
consists of three different buses:

the Processor Local Bus (PLB), which is a high-performance bus, used to

connect high-bandwidth devices such as high-performance processor cores,
external memory interfaces and DMA controllers.

the On-Chip Peripheral Bus (OPB), which is a secondary bus that can be
used to decoupled the peripherals from the PLB in order to avoid a lost of
system performance. Peripherals like serial ports, parallel ports, UARTs,
GPIO, timers and other low-bandwidth devices should be attached to the
OPB. Access to the peripherals on the OPB bus by PLB masters is done
through a bridge, which is used as a slave device on the PLB and as master
on the OPB. The bridge performs dynamic bus sizing, in order to allow
devices with different data widths to communicate.
Figure 8.3 illustrates the implementation of the OPB. Note that no tri-state
is required. The address and date buses are instead implemented using
a distributed multiplexer. This is a common technique for implementing
buses in programmable logic devices. All the master inputs to the bus are
ORed and the result is provided to all the slaves. The arbitration module
defines which master is granted the bus. All other masters must then place
a zero signal on their output. The ORed is then used to write the value of
the bus master on the bus.
System on a Programmable Chip 267

Figure 8.2. Example of system ingration with CoreConnect buses

the Device Control Register (DCR) Bus to allow lower performance status
and configuration registers to be read and written. It is a fully synchronous
bus that provides a maximum throughput of one read or write transfer every
two cycles. The DCR bus removes configuration registers from the memory
address map, reduces loading and improves bandwidth of the processor
local bus.

1.5.2 The Amba

The Advanced Microcontroller Bus Architecture (AMBA) from ARM shares
many similarities with the IBM CoreConnect. Both architectures support data
bus widths of 32-bits and higher, they utilize separate read and write data chan-
268 RECONFIGURABLE COMPUTING

Figure 8.3. Implementation of the OPB for two maters and two slaves

nels, they allow multiple masters, and support split transactions and burst trans-
fers. The AMBA 2 interconnection infrastructure consists of two main buses:

the Advance High-Speed Bus (AHB) that is used as high-performance sys-

tem interconnect to allow communication between high-performance mod-
ules and the processor. The AHB plays the same role in the AMBA that the
PLB plays in the CoreConnect. A bridge from AHB or ASP is required to
interface to APB.

the Advance Peripheral Bus (APB), used to connect the slower peripherals.
A bridge is used to interface between the AHB and APB buses and allow
low peripheral to be accessed from components hung on the AHB.

2. Adaptive Multiprocessing on Chip

The continuous performance improvement of the last decades in micropro-
cessor is not likely to hold in the future. This positive trend observed in past,
and also known Moore’s law was due to two main factors: high-speed clocks
and improvement in instruction level parallelism.
While the speed was increased by constantly growing clock frequency, the
capacity has always been boosted by the reduction of transistor size. Sequen-
tial programs could therefore be speeded-up through instruction level paral-
lelisms (ILP) in a transparent way for the user. The increase in speed is getting
more and more difficult to maintain. Enhancement on ILP-design for data-
path isn’t coupled anymore with a performance increase of the system [169]
[213]. Moreover the power is becoming one of the main limitations. In order
to solve this problem, the research community and the industry have started
System on a Programmable Chip 269

using the large amount of available chip area to implement several processors,
thus creating the so called chip multi processing systems.

Basically, two paradigms exist for parallel computation on multiproces-

sors. The first one is the Message Passing Interface (MPI), usually available
where communication among processors happens in a network. The second
paradigm, the Shared Memory or Symmetrical Multi Processing (SMP) is ap-
plied in bus-based systems where communication happens through a share
memory. Multiprocessor on chip has been studied according to the second
paradigm. In most of the existing works, the purpose is to have an efficient
mapping of a set of threads onto the processors. Two possibilities have been
presented that address this purpose: Simultaneous Multithreading (SMT) [75,
120, 103] and Chip Multiprocessor (CMP) [169, 109, 139, 94, 16]. An SMT-
Chip is based on a superscalar processor with a set of units for a parallel ex-
ecution of instructions. Given a set of threads, the resources are dynamically
allocated by extracting several instructions to be executed in parallel. Threads
that need long memory access are preempted in order to avoid idle states of
processors. Base on an Alpha 21164 processor with 10 function units and 8
instructions per clock, an SMT-system were shown to provide a speedup of
up to 4 compared to common superscalar implementations [75]. In Hirata et
al [120], ray-tracing-application on SMT-architecture could be simulated with
considerable performance improvement. Gulati and Bagherzadeh [103] could
also provide simulation results of SMT with performance improvements.

Instead of using only one superscalar architecture to execute instructions in

parallel, Chip Multiprocessors use many processor cores to compute threads
in parallel. Each processor has a small first-level local instruction and data
cache. A second level cache is available for all the processors on the chip.
CMPs target applications consisting of a set of independent computations that
can be easily mapped to threads. This is usually the case in database or web-
server where transactions do not rely on each other. In CMPs, threads having
long memory access are preempted to allow others to use the processor. Bar-
roso et al [16] have presented a scalable CMP-architecture called Piranha in
which the processors -in this case alpha processors- use a crossbar switch for
communication. Despite the good simulation results no physical chip could be
manufactured. A most concrete approach is the work of Lammon and Oluku-
tun that leads to the design of the Hydra Chip [109], whose technology is in
used in the Niagara chip [139]. In this architecture, a hardware support for
speculation is available, besides the parallel execution of threads. The commu-
nication among the processors is done using a two-level on-chip cache. The
simulation provides almost a linear performance increase in the number of pro-
cessors. Other developments in multiprocessor on chip include the IBM Cell-
Chip, consisting of a 64-Bit Power processor and eight so called synergistic
270 RECONFIGURABLE COMPUTING

processors. The Intel Pentium D and the AMD Dual Core Athlon 64 Processor
are other examples of multiprocessor on chip.
The work in [167] proposed an adaptive chip-multiprocessor architecture,
where the number of active processors is dynamically adjusted to the current
workload needed in order to save energy while preserving performance. The
proposed architecture is a shared memory based, with a fix number of embed-
ded processor cores. The adaptivity results in the possibility of switching the
single processors on and off.
Programmable logic devices in general and FPGA in particular have experi-
enced a continuous growth in their capacity, a constant decrease in their power
consumption and permanent reduction of their price. This trend, which has
increased interest in using FPGAs as flexible hardware accelerators, is likely
to continue for a while.
After an unsuccessful attempt in the 90’s to use FPGA as co-processor in
computing system, the field FPGA in high-performance computing is experi-
encing a renaissance with manufacturers like Cray who now produce systems
made upon high speed processors couple to FPGAs that act as hardware accel-
erators [58]. However, actual FPGA-solutions for high-performance comput-
ing still use fast microprocessors and the results are systems with a very high
power consumption and power dissipation.
A combination of the chip multiprocessor paradigm and flexible hardware
accelerators can be used to increase the computation speed of applications. In
this case, FPGAs can be used as target devices in which a set of processors
and a set of custom hardware accelerators are implemented and communicate
together in a network.
FPGA manufacturers like Xilinx have been very active in this field by pro-
viding ready to use components (soft of hard-cores processors, bus-based in-
tercommunication facilities, and peripherals) on the base of which, multipro-
cessor on chip can be built. The Virtex II Pro FPGA for example provides up to
four PowerPC processors and several features like memory and DSP-modules.
The adaptivity of the whole system can be reached by modifying the com-
putation infrastructure. This can be done at compile-time using full reconfig-
uration or at run-time by means of partial device reconfiguration. This results
in a multiprocessor on-chip, whose structure can be adapted at run-time to the
computation paradigm of a given application.
Unfortunately, the capabilities of those devices are not exploited. FPGA
devices like the Xilinx Virtex Pro are usually used only to implement a small
hardware component. The two available Power PC processors remains most
the time unused. This is in part due to the lack of support for the multiprocessor
design.
In this section, we investigate the use of Adaptive Multiprocessor on Chip
(AMoC) and present a design approach for such systems. We also present an
System on a Programmable Chip 271

environment developed at the University of Kaiserslautern in Germany, for a

seamless generation and configuration of the hardware infrastructure in order
to produce a complete system.
One of the main applications of such infrastructure is in high-performance
computing in embedded systems. We next present our understanding of the
system architecture that must be first generated onto the device. Thereafter, we
present a two steps design approach and the design automation for AMoCs.

2.1 Hardware Infrastructure

Many possibilities exist to build a multiprocessor hardware infrastructure on
a chip. In this section we however focus on the general computing infrastruc-
ture shown on figure 8.4. For the sake of modularity, the infrastructure consists
of a set of processor blocks, each having a custom hardware accelerator and
memory directly attached the processor, either through the available fast pro-
cessor bus, or through a dedicated communication link. The processor blocks
are connected in a network and are accessible from off-chip components via
peripherals. Additional external RAM might be available.

Figure 8.4. General adaptive multiprocessor hardware infrastructure

One of the main tasks in the design and implementation of a system like
that of figure 8.4, is the interconnection network. Contrary to other modules
(processors, peripherals, memory controllers and some custom hardware) that
272 RECONFIGURABLE COMPUTING

are available in the vendor libraries, communication facilities, in particular

network on chip modules are not available at all. Also, most of the existing
tools provide only a very limited support for multiprocessing. We therefore
focus on the communication aspect next.

2.2 Communication Infrastructure

The communication between the processor blocks can basically be done in
two different ways: Shared-memory and message passing: in the first case, the
processor blocks are attached to a common bus on which a memory is also
attached. The exchanged of message is done via the common accessible mem-
ory. Although such a system can be built with the current vendor tools, the
amount of handwork required to have a system working is quite high. Further-
more, some modules needed to build a workable system, for instance a module
to handle the cache coherence amount the processors is not provided and must
be designed and implemented by the user. In the second case, where message
passing is used, the processor blocks are accessible via a network on which
they hung. The network is in charge of routing messages from source to des-
tination. Although message passing can be implemented with shared-memory,
we focus in this section on those implementations where a set of independent
processors communicate with messages that are sent through a network.
The design of the communication network should be carefully done, in order
to avoid a high consumption of chip resources, while maximizing the speed and
the throughput. A network on chip (NoC) normally consists of set of routers
each connected to a processor, and direct connections between the routing ele-
ments. Each router provides a processor an interface to the network for sending
and receiving messages in form of packets. Also, the routers are in charge of
placing incoming packets on the best tracks and insure a smooth routing to
destination.
While the routers have the advantage of flexibility, thus allowing any topol-
ogy to be built, their resource consumption has been shown to be quite high in
FPGAs, due to the large amount of multiplexers used [32]. In order to avoid
wasting too much chip resources, a limited amount of routers may be used.
This can be reached by using a ring architecture. Like a Bus, a ring provides
a great advantage in the resource consumption. The disadvantages of the bus,
like the exclusive bus access by masters and the lower fault tolerance can be
can be compensated by having several masters on different ring segments. In
[108], a ring architecture was designed for the interconnection of a set of pro-
cessors blocks on a FPGA.
As figure 8.5 illustrates, the processors are primarily connected using a ring.
However, because a ring allows only one kind of topology to be built, routers
are available for building cross connections over different rings. The routers
therefore provide a means to build other topologies than the ring. A router
System on a Programmable Chip 273

is able to connect up to four rings. Building a mesh for example is done by

connecting two "perpendicular and disconnected rings" via the router. In figure
8.5, two rings connected to a router are shown.

Figure 8.5. Structure of the on chip network

A ring has the advantage that components connected to it do not have to

deal with routing and therefore, the amount of multiplexers (one of the most
resource consuming element) needed to implement the complete infrastructure
decreases drastically. Components only need transceivers to access the ring
for sending and receiving messages (figure 8.5). Each transceiver consists of
a sender (S) and a receiver (R) that operate independent from each other. The
receivers behave on the ring as slave and on the local bus as master, while the
sender behave on the local bus as slave and on the ring as master. The ring
slave in the receiver collects data from the ring and passes them to local bus
master in the same receiver that. The message can then be copied into the local
memory via the local bus. The local bus slave in the sender collects data from
the local bus and send them to ring master in the sender, which then place the
message onto the ring.
Within a transceiver, each slave and each master is basically a finite state
machine(FSM), which is in charge of initiating and handle the requires trans-
274 RECONFIGURABLE COMPUTING

Figure 8.6. Implementation of the transceiver

actions. This modularization helps to decouple the complex bus operations

from the cumbersome ring transactions.
The transceivers operate in two different modes: active copying and DMA
(direct memory access). The active copy mode is used to send small number of
packets while DMA deals with the transmission of large data segments. With
the DMA capability on transceivers, a remote DMA between two processors
is implemented, in order to quickly transfer large amount of data from one
processor to another one, without loss of performance. The communication
protocol used in this case is next explained.

2.3 Data Transfer and Communication Protocol

In active copy mode, the processors pass each single byte to be send to the
transceiver while in DMA mode, the processor just specifies the destination
address and provides only the start address of the data block to be sent. The
transceiver can then access the memory and transmits the data over the net-
work, while the processor continues its computation.
A sending operation is initiated by a processor by copying the destination
address together with some control data to the control register of the FSM
acting as local bus slave. While this first FSM collects data to be sent from
the local bus, the second FSM, acting as master on the ring, builds packets
and send them via the ring. In case of an active copy, the data is place on
the bus by the processor, while in DMA, the slave FSM accesses the memory
itself to collect data, thus freeing the processor. This second possibility is used
to implement a remote DMA (RDMA) which allows a processor to directly
copy a chunk of data in the memory of another processor on the network. The
destination-processor must therefore dynamically allocate enough space in its
local memory for incoming data from the sender-processors. A processor Ps ,
willing to send data to another processor Pr , first sends a request of RDMA
data transfer to processor Pr . The request specifies the length (in bytes) of
data to transfer. Pr then tries to allocate enough memory for the data and
System on a Programmable Chip 275

allows Ps to transfer data, if the reservation is successful. The transfer is done

by the sender and receiver modules of the corresponding transceivers while the
two processors might perform other tasks.
In order to implement the handshaking between the two processors, a com-
munication area is foreseen on each processor. This is a table with two entries
for each processor. The first one specifies the request/acknowledge of data
transfer and the second one specifies the amount of bytes to transfer. Each pro-
cessor performs a request/acknowledge through a remote writing in the table
of the sender/destination processor at the corresponding location.

Figure 8.7. The communication protocoll

A possible communication scenario is shown in figure 8.7; The network

consists of n + 1 processors also called clients. client0 and clientn consec-
utively request a RDMA-transfer of data to client1 (step 1 and step 2). Ac-
cording to a first-request-first-grant mechanisms, client1 first grants RDMA
access to client0 by writing the granted-code at the right location in the hand-
shaking table of client0 (step 3). From this point on, the sender transceiver
and the receiver transceiver are used to realize the data transfer, while the two
processors keep computing (step 4). Upon completion, client0 removes its
request (step 5), and client1 removes its grant. clientn is now granted the
RDMA-access (step 6) and the data transfer is done according to the scheme
previously described (step 7 and step 8).
276 RECONFIGURABLE COMPUTING

2.4 Design Automation

The goal in design automation is to ease the implementation of an adaptive
multiprocessor system described in figure 8.4, on an FPFA.
The generation of such a system is a complex task whose support through
the current vendor tools is very limited. A substantial amount of hand work
most be done in order to have many processors working in an FPGA. The only
paradigm supported by the current tools, albeit very limited, is the SMP-like
computation paradigm. For the design of MPI-like systems the user must do
everything from scratch. The network components must be provided, also sup-
port for cache coherence in the case of SMP-like computation. This requires
a great amount of hand work and strong experience in hardware design. In
order to make multiprocessing implementation on chip easier, in particular for
software designers, but also for those hardware designers not willing to spend
too much time in the hardware part of the system, a seamless approach must
be available to generate the hardware infrastructure. The designer can then fo-
cus only on parallelizing he’s code by defining the part to be executed by each
processor and the communication points in the code.
The work in [130] [183] has contributed, besides the development of the
communication infrastructure to the design and implementation of a vendor
independent framework in which the user, no matter if beginner or experienced,
can specify the complete requirements of he’s system and let the tool produce
the necessary configuration and start-up code.
The approach that followed is a two-steps one, consisting of first providing
an abstract specification of the system. The abstract specification will then
be refined and mapped to a concrete description according to the component
library available in the second step.

2.5 Automatic Generation of the Hardware Infrastructure

In the first step, the hardware is automatically generated from an abstract
specification. This specification may be provided by the user or it can be the
result of an analysis process that generates the best multiprocessor infrastruc-
ture as well as the best topology to compute a given application.
As shown in figure 8.8, the description at the highest level is very abstract
and specifies only the amount of the processors, their interconnection paradigm
and the characteristics of the computing blocks used in the system.
The description is done in XML in conjunction with a document type defi-
nition file (DTD). The document type definition describes the syntax as well as
valid components and their attributes. At this level, no information about spe-
cialized and platform-dependent system components has to be supplied. Valid
modules that can be described at this stage are for instance CPU, Memory,
CommMedium, Periphery and HWAccelerator. CommMedium is an identifier
System on a Programmable Chip 277

for the whole class of communication media. It comprises buses, networks and
point-to-point connections. Periphery and HWAccelerator differ only in their
type of connections. While a HWAccelerator is used for internal components,
Periphery is used to describe modules with external connections.
Having specified the system, the description may be functionally simulated
to observe the system behaviour, and afterwards be validated.

Figure 8.8. Automatic hardware generation flow

In the last step, the final system can be created by supplying a concrete
system description for the target platform.
When transforming an abstract system description into a concrete one, trace-
ability must be maintained. Therefore, in each of both steps, the respective
XML input file is validated against its document type definition to prevent in-
valid component specifications.
Subsequently, the given specification is processed by a Java application con-
sisting of three consecutive phases. In phase one, an XML parser extracts re-
278 RECONFIGURABLE COMPUTING

quired information from the XML sources and an internal representation of

the resulting XML tree is built up for further use. For each component, an ap-
propriate class is instantiated, which knows how to process its own parameters.
Obviously, the application can be easily extended to support new user-designed
IP-cores by just adding a corresponding class.
Both parsers -the abstract one as well as the concrete one- share this ap-
proach. Whereas the abstract parser has finished its work at this point, the
concrete XML parser can now go on processing the gathered information to
create the platform specific hardware information files. Individual mappers are
created in the second phase, one for each component and each target platform.
In the third and final step, a mapper is invoked to create the desired platform-
dependent hardware description files. Again, it has to be emphasized that ex-
tending the application to support a new platform can be simply accomplished
by adding another mapper class.
Mapper classes for system components are derived from an abstract platform-
dependent super class that contains a generic algorithm which enables derived
classes to inherit mapper functions. The current system component which has
to be mapped is passed from the corresponding class itself to its mapper class
via constructor. The resulting hardware description files mentioned above are
then passed to the vendor’s tool chain, for example to Xilinx EDK to generate
the bitstream for the configuration of the FPGA.
The system does not only enable the user to create hardware information
files from a given specification. Existing design can also be loaded from files
in which they were saved. Also, system components can be added to or remove
from the system in a very comfortable way, and the changes cane be rewritten
back to the XML file.
A tool called PinHat (Platform-independent hardware generation tool) was
developed with the goal to integrate all the functionalities and features de-
scribed above into one single, easy-to-use application. PinHat provides a
graphical user interface (figure 8.9) that allows the user to comfortably edit
the desired system settings and guides him through the process of deriving the
concrete system description from the abstract one.
It integrates seamlessly with vendor supplied toolchains and allows the gen-
eration of a bitstream to configure any supported FPGA based on the created
concrete specification.
After having adapted the abstract specification to the end-user’s needs, Pin-
Hat offers a transformation wizard which supports the procedure of refining
abstract node elements to concrete ones. First of all, the root element has to
be refined by specifying the target toolchain and the target board. This piece
of information is used to identify correct specified components when refining
all the other abstract nodes. Selecting for instance Xilinx as target toolchain
and ML 310 as target board would allow an abstract CPU node to be refined
System on a Programmable Chip 279

Figure 8.9. The Platform-independent Hardware generation tool (PinHat)

to either a Microblaze or alternatively a PowerPC 405 concrete node element.

Subsequently, the refinement of all lower-level nodes can be done. PinHat
has integrated parsers to read out component and board description files where
information about peripherals, FPGA pin connections, board architecture, in-
ternal component parameters and much more, can be stored and thus does not
need to be entered manually. This increases the comfort by avoiding the users
to enter hundreds of parameters for a single component. Furthermore, this
procedure guarantees that no non-existing parameters can be specified. Param-
eters which are not specified are initialized with standard values which are also
part of the parsed description files. The role of PinHat in the design flow is
illustrated in figure 8.10, where the abstract specification is mapped later to the
Xilinx Platform.

2.6 Automatic configuration of the hardware

infrastructure
In contrast to the downloading of the bitstream into the FPGA, we under-
stand under automatic configuration the generation of the binary code that will
280 RECONFIGURABLE COMPUTING

[ht]

Figure 8.10. The PinHaT framework

be run on each processor, the configuration of the operating system and the
generation of the files needed for system start-up.
Like in the Automatic generation of the hardware infrastructure, we have
a SW-specification, which may be -as the HW-specification- provided by the
user or it can be the result of an analysis and code generation process. The SW-
specification describes whether a given processor is to be run with a standalone-
application or if an operating system is foreseen for that processor. For each
standalone-application additional information is given about the task it exe-
cutes; the task allocation results from splitting the application to be computed
System on a Programmable Chip 281

into a separated set of independent and communicating tasks. The distribution

of the jobs and the mapping to the processors is stored in XML, complementing
the abstract component input described in section 2.5. We are not focussing on
the partitioning process here. We assume that this part has been done and that
the segments of the overall code that will be implemented on each processor
are available.
As the abstract system description is transformed into a concrete one, the
mapping "task → abstract processor" is refined into a mapping "task → con-
crete processor".
Thereafter, parameters of the software for each processor must be specified:
this includes information about either the application (source-code, libraries,
etc.) or about the operating system(s)(type, configuration, require function-
alities and modules, ...). Once those information are provided, scripts and
compiling files for building the standalone software and the operating systems
are generated. Building the standalone software is straightforward because it
requires only compiling and linking the files, using vendor tools on the basis
of the hardware generated in the previous step. The building of the operating
system is however a more complex issue. Depending on the OS chosen, differ-
ent steps are necessary; for some operating systems only the kernel-parameters
may be configured explicitly, while the file-system and therewith the available
programs is pre-build. Other operating systems allow for the complete configu-
ration of the kernel, file-system and the available programs. The configuration
of the operating systems is done by generated scripts.
The result of the configuration step is one executable file for each processor
in the system. This file can be the same for a pair of processors in the system,
if the parameters provided to build the file are the same.
Using vendor tools, the FPGA-configuration bitstream, resulting from the
generation of the hardware infrastructure is complemented with the generated
executable files. These results in a configuration file containing hardware and
software configuration of the whole system for the chosen FPGA.
According to the operation paradigm chosen, we may have the operating
system running on all the nodes or on just one node as illustrates in (figure
8.12). The last approach makes more sense. In this case, the node attached to
the peripherals uses the operating system for a better control of those periph-
erals. This node will then behave as a master in the system, collecting the data
and dispatching them to all the processors involved in the parallel computa-
tion of a given application, collect the partial results from the single processors
and send them trough the peripherals to off-chip components. The operating
system can also be made available on many processors, if those processors
manage some peripherals separately.
As case study, several implementations were done on FPGA, using the Pin-
Hat tool. A three processor system consisting of a PowerPC, and two MicroB-
282 RECONFIGURABLE COMPUTING

Figure 8.11. software configuration flow

laze processors on the Xilinx ML310 board, featuring a VirtexII Pro 30 FPGA,
were built. The primary goal was to test the bandwidth performance of a sys-
tem with multiple rings and a router for the ring interconnection. The design,
in a first non optimized implementation, reveals an average real bandwidth of
36 Mbytes/s. This includes the complete transaction need to code, send and
decode a message.
In order to have the system running with a real life application, the singu-
lar value decomposition (SVD) on the ML310-Board is implemented, with the
number of processors varying from one to eight, and with matrices varying
from (4 × 200) to (128 × 200). The performance increase due to the use of
multiprocessors was shown to be almost linear in the number of micro proces-
sors.
System on a Programmable Chip 283

Figure 8.12. 4-Processor infrastructure for the SVD on the ML310-Board

Since the bulk of the computations in the SVD is the computation of the dot-
products of the column pairs, which is a good candidate for hardware imple-
mentation. Custom multiply accumulate (MAC) modules for computing those
column-pair multiplications were designed and directly attached connected to
the MicroBlaze processors via the fast simplex link (FSL) ports (figure 8.12, in
order to increase the performance of the whole system.

2.7 Adaptivity
As state at the beginning of the chapter, the adaptivity can be used for several
purposes. According to the application, a physical topology consisting of a
given arrangement of routers and processors may be built to match the best
computation paradigm of that application. The topology can be modified, even
at run-time to better adapt to the changes observed in the application.
On the other hand, the adaptivity of the system can be reached by using
reconfigurable hardware accelerators, provide that the target device supports
partial reconfiguration.
Designing partially reconfigurable systems with unlimited capabilities is a
complex engineering task that requires a lot of hand work. In order to avoid this
and keep the thing very simple, a template for partial reconfiguration can first
be created for the target device. In this template, fixed locations must be se-
lected, where the reconfiguration is allows to take place. It acts as place holder
in order to accommodate reconfigurable modules at run-time. Any hardware
accelerator can be plugged on the hardware block at run-time by means of par-
284 RECONFIGURABLE COMPUTING

tial reconfiguration. This step can even be performed from within the device
using the appropriated port like the ICAP in the case of Xilinx devices.

3. Conclusion
In this chapter, we have addressed the system on programmable chip paradigm
and provided some advantages of using those architectures. The main compo-
nents needed to build such a system were presented, with a focus on the leading
manufacturers. The goal was not a comparative study on the different system
on programmable chip infrastructure. We rather wanted to provide the user a
brief view on the existing solution.
In the second part of the chapter, we have presented a design approach for
multiprocessor systems on FPGAs. Using the reconfiguration, it is possible to
modify the hardware infrastructure to adapt it to the best computation paradigm
of the application being computed.
We have presented a framework to specify and implement a complete sys-
tem without knowledge of the vendor tools. Obviously, all what is feasible
with the PinHat tool is also feasible with current vendor tools. However, the
complexity and the difficulty of using those tools do not allow any new comer
to generate design with it. The goal here was to hide the complexity in the de-
sign of multiprocessor, which is very limited in the current tool to the user and
allow him to focus on the efficient analysis and partitioning of its application.
We believe that multiprocessing on FPGA has a great potential to speed-up
computation in embedded system. In order to facilitate their use, systems like
the PinHat are welcome to hide the complexity of the hardware generation and
let the designer focus on the analysis of the application and the definition of
the best adapted structure for its computation.
The adaptivity of the whole system is provided through the use of partial
reconfiguration in order to exchange running hardware modules at run-time.
However in order to overcome the difficulty of designing for partial reconfigu-
ration with the current tool, a template base approach is recommended, which
foresee predefined location for placing components at run-time.
Chapter 9

APPLICATIONS

Up to now, we have focussed on the technology behind reconfigurable de-

vices, their integration in different systems and their programmability at dif-
ferent level of abstraction. The natural questions which arises it the one to
know, in which area does reconfigurable computing could be helpful? if ar-
eas of application exist, then we would also like to know what the gain of
using reconfigurable technology is, compared to other processing devices used
before? The gain of using a solution rather another one is usually expressed
as the computation speed, or as the improvement in power consumption, or
as the price, or even as the easiness in the programmability. In [57], Craven
and Athanas provided a performace/price comparative study between FPGA-
based high-performance computing machines and traditional supercomputers.
Based on the gain in accelerating floating-point computations, they came to
the conclusion that the price to pay was too high compare to the marginal gain
obtained, when using FPGAs in supercomputers. Other researchers have re-
ported amazing gains with FPGA-based computing machines in several fields
of computation and several companies that have been offering FPGA-based
computing solutions a decade ago are still operating and even growing. Our
goal in this chapter is to present some applications that benefit somehow from
an implementation on reconfigurable devices. Almost all the experiments re-
ported were done on FPGA platforms. We will not report all the experiment in
all details. We rather focus on the coarse-grained structure of the applications
and point out where the reconfiguration can be helpful. We are primary con-
cerned with the speed improvement and the use of flexibility of reconfigurable
computing systems.
Despite the very large number of applications that could benefit from an
implementation in FPGAs, we have selected only 6 domains of application the
286 RECONFIGURABLE COMPUTING

we present here: pattern matching, video streaming, signal processing using

distributed arithmetic, control, and super computing.

1. Pattern Matching
Pattern matching can be defined as the process of checking if a character
string is part of a longer sequence of characters. Pattern matching is used in
a large variety of fields in computer science. In text processing programs like
Microsoft Word, pattern matching is used in the search function. The pur-
pose is to match the keyword being searched against a sequence of characters
that build the complete text. In database information retrieval, the content of
different fields of a given database entry are matched against the sequence of
characters that build the user request. Searching in genetical database also use
pattern matching to match the content of character sequence from a database
entries with the sequence of characters that build a given query. In this case
the alphabet is built upon a set of a genomic characters. Speech recognition
and other pattern recognition tools also use pattern matching as basic opera-
tions on top of which complex intelligent functions may be built, in order to
better classify the audio sequences. Other application using pattern matching
are: dictionary implementation, spam avoidance, network intrusion detection
and content surveillance.
Because text mining, whose primary role is the categorization of documents,
makes a heavy use of pattern matching, we choose in this section to present
the use of pattern matching in text mining and to point out the possible use of
reconfiguration.
Documents categorization is the process of assigning a given set of docu-
ments from a database to a given set of categories. The categories can either
be manually defined by a user, or it can be computed automatically by a soft-
ware tool. In the first case, the categorization is supervised and in the second
cased we have an unsupervised categorization. In categorization, the first step
usually consist of indexing the collection of available documents. This is usu-
ally done through the so called vector space representation. In this model a
document is represented as a vector of key words or term present in the doc-
ument. A complete collection of n documents over a list of m keywords is
then represented by a term by documents matrix A ∈ Rm × Rn . An entry aij
in A represents the frequency of the word i in the document j. The term by
documents matrix is then used for indexing purpose or statistical analysis like
LSI (Latent Semantic Indexing) [67]. Building a terms by documents matrix
is done by scanning the documents of the given collection in order to find the
appearance of key words and return the corresponding entry in the matrix for
each document. Pattern matching is therefore used for this purpose.
The first advantage of using the reconfigurable device here is the inherent
fine-grained parallelism that characterizes the search. Many different words
Applications 287

can be searched for in parallel by matching the input text against a set of words
on different paths. The second advantage is the possibility of quickly exchange
the list of searched words by means or reconfiguration.
Pattern matching was investigated in different work using different approaches
on FPGAs [52] [85] [104] [151] [101] [180] [34], each with a different over-
head in compiling the set of words down to the hardware and different capac-
ities. We define the capacity of a search engine as the number of words that
can de searched for in parallel. A large capacity also means a high complexity
of the function to be implemented in hardware, which in turn means a large
amount of resources. The goal in implementing a search engine in hardware is
to have a maximal hardware utilization, i.e. as many words as possible that can
be searched for in parallel. We present in this section various hardware imple-
mentation of the pattern matching, each with its advantages and drawbacks.

1.1 The Sliding Windows Approach

One approach for text searching is the so called sliding window (SL)[85]
[151]. In the 1-keyword version, the target word is stored in one register, each
character being stored in one register field consisting of a byte. The length of
the register is equal to the length of the word it contains. The text is streamed
through a separate shift register, whose length is the same as that of the key-
word. For each character of the given keyword stored as a byte in one register
field, an 8-bit comparator is used to compare this character with the corre-
sponding character of the text, which streams through the shift register. A hit
occurs when all the comparators return the value true.
The sliding window can be extended to check a match of multiple patterns
in parallel. Each target word will then be stored in one register and will have
as many comparators as required. The length of the window (the number of
characters in the shift register) is defined by the segment of the text to be pro-
cessed. Words with length bigger than the length of the window can not be
handled. To overcome this problem one can define the length of text to be
considered as the maximum length over all the target words. In this case, all
the words with a length smaller than the maximum length should be filled with
don‘t care characters to be handled correctly. The example of figure 9.1 shows
the structure of a sliding windows with three key words.
The main advantage of this approach resides in the reconfiguration. Be-
cause each keyword is stored in an independent register, the reconfiguration
can happen without affecting the other words, thus providing the possibility to
gradually and quickly modify the dictionary.1

1 With dictionary, we mean here the set of keyword compiled into the reconfigurable device
288 RECONFIGURABLE COMPUTING

Figure 9.1. Sliding windows for the search of three words in parallel

This method however has a main drawback, which is the redundancy in the
amount of register fields used to stored the words as well as the number of
comparators. Redundancy in the sliding windows approach reduce the capac-
ity of such implementation, thus making its use not competitive. In the case of
the Xilinx Virtex FPGA XCV300 with a maximum of 6144 flip flops, a sliding
windows of length 10 needs 10 × 8 = 80 flip flops to store one target word in
the device. Consider that the device placement and routing permits an utiliza-
tion of 80%. The number of words which can be folded in such an FPGA will
theoretically be in the range of 60, which is small compared to the number of
words which can be folded in the same divide with a more efficient implemen-
tation. A nice approach would be to just perform the comparison on a given
alphabet consisting of the set of common character and to use the result in the
evaluation of each single word. This would lead to a reduce amount of register
fields and a reduce amount of comparators.

1.2 Hashed Table-Based Text Searching

The hashed table-based text searching approach was implemented on the
SPLASH II Platform [180]. The text to be processed is streamed in a pair of
consecutive characters called superbyte. An incoming superbyte is mapped on
one character of the target alphabet. Each non alphabet character is mapped
to a delimiter. To determine whether or not a word is in the target list, a hash-
Applications 289

ing technique is used. A hash table in memory is used to store the value of a
presence bit for a particular target word. An entry zero in this table means that
the word is not present and a value one means that the word will probably be
present. A hash register (with a length of 22 bit in the SPLASH implementa-
tion), which is initially set to zero, is incrementally modified by the incoming
superbytes. At the end of a word marked by a delimiter, the hash register is
set to zero to allow the next word to be processed. The hash register is used to
address the hash table of 222 pseudo-random mappings of words. The modifi-
cation of the hash register happens as follows: When a non delimiter superbyte
is encountered the contents of the 22-bit hash register is updated by first XOR-
ing the upper 16-bit of the register with the incoming superbyte and a value
of a hash function. The modified value of the hash register is then circularly
shifted by a fixed number of bits. Upon reaching the end of a word, the content
of the register is used to address the hash table and determine the value of the
presence bit. To reduce the likelihood of false hits a number of independent
hash functions is calculated for each word of the text, with the stipulation that
each hash function lookup of a word must result in a hit for that word to be
counted as a target word.
Estimations on performance of the SPLASH text searching have been done
mainly on the basis of the critical path length returned by the place and route
tool. Communication between the host computer and the boards like memory
latency have not been taken in consideration.

1.3 Automaton-Based text searching

It is well known that any regular grammar can be recognized by a deter-
ministic finite state machine (FSM). In an automaton-based search algorithm,
a finite state machine is built on the basis of the target words. The target words
define a regular grammar which is compiled in an automaton acting as a recog-
nizer for that grammar. When scanning a text, the automaton changes its state
with the appearance of characters. Upon reaching an end state, a hit occurs and
the corresponding word is set to be found. One advantage of the FSM-based
search machine is the elimination of the preprocessing step done in many other
methods to remove stop words (like ’the’, ’to’, ’for’ etc., which does not affect
the meaning of statements) from documents. In [52], the implementation of
the hardware based retrieval machine is done using a memory to store the tran-
sition table. Two registers hold the current state and the incoming character.
The logic is only used to decode the address of the next state in memory and
to set the correct values in the registers.
The performance (speed) of such an implementation is governed by the path
(RAM output → state register → RAM address → RAM output) which can
considerably slow down the search process, because of the multiple memory
accesses. To increase the speed of an FSM implementation, memory accesses
290 RECONFIGURABLE COMPUTING

can be suppressed by compiling the transition table in hardware and having a

mechanism to execute the state machine. Folding big transition tables in FP-
GAs can be difficult because of the limited capacity of those devices. Transi-
tion tables for word recognizers are often full of backtracking edges and cross-
ing edges (edge which connects two nodes on different paths in the FSM-tree).
With the sequential structure of the memory based recognizer, it is almost im-
possible to eliminate the backtracking edges, which constitute in most cases
the biggest part of the transition table.
The SPACE (Scalable Parallel Architecture for Concurrency Experiments)
machine [104] [105] makes use of the parallel structure of FPGAs to eliminate
backtracking edge from the FSM. This search machine is a FSM like imple-
mentation for text searching capable of handling a search over a list of almost
100 target words. The search is performed simultaneously on a board with 16
CAL1024 FPGAs without using a memory to store the transition table. Each
key word is compiled in a separate FSM which is mapped to the hardware
device. The text is then streamed into the device, characters by charactes in
ASCII form. For the case insensitivity purpose the 8-bit characters are mapped
by the preprocessor to 5-bit characters. With the incoming characters, which
clock the flip flops, all the FSM move simultaneously to their next state. Upon
reaching an end state, the corresponding accumulator is incremented and the
result is returned. The backtracking edges appearing in the formal represen-
tation of the FSM become redundant and will not be taken in consideration.
Figure 9.2a) shows a transition graph of the FSM for the word "conte". The
transition table requires 5 × 6 memory locations to be stored (figure9.2b)). For
the same FSM the hardware representation will need only 4 flip flops, 4 AND
gates, and 5 comparators (figure 9.2c)).
While having the advantage of replacing a word already folded in the FPGA
with a new target word of the same length, the SPACE implementation presents
two major inconveniences: First, it does not take into consideration words
which share a common prefix, thus increasing the amount of resources used to
stored the character states. Second, it uses one comparator for each character
of a target word, which increase the resource needed to implement the com-
parators. Those two factors lead to redundancy in flip flop and look-up table
utilization as it is the case in the sliding window recognizer. On the SPACE
board, with a total of 8192 flip flops it has been possible to implement a list of
only 96 target words with average length 12.
When taking in consideration the common prefix of words, it is possible to
save a considerable amount of flip flops. For this purpose, the search machine
could be implemented as an automaton recognizer, common to all the target
words. This approach was first presented in [34]. As shown in figure 9.3, the
resulting structure is a tree in which a path from the root to a leaf determines
the appearance of a corresponding key word in the streamed text. Words which
Applications 291

Figure 9.2. FSM recognizers for the word "conte": a) sate diagram, b) transition table, c) basis
structure the hardware implementation: 4 flip flops will be need to code a 5 × 6 transition table

share a common prefix use a common path from the root corresponding to the
length of their common prefix. A split occurs at the node where the common
prefix ends. In the hardware implementation of a group of words with a com-
mon prefix, common flip flops will be used to implement the common path
(figure 9.3a)).

Figure 9.3. a) Use of the common prefix to reduce the number of flip flops of the common
word detector for "partir", "paris", "avale","avant". b) implementation without use of common
prefix and common comparator set

For each character in the target alphabet only one comparator is needed.
The comparison occurs in this case, once at a particular location in the device.
292 RECONFIGURABLE COMPUTING

Incoming characters can be directly sent to a set of all possible comparators.

Upon matching a particular one, a corresponding signal is sent to all the flip
flops, which need the result of that comparison. This method will reduce the
number of comparators needed almost by the sum of the length of all target
words.
The overall structure of the search engine previously explained is given in
figure 9.4. It consists of of an array of comparators to decode the characters
of the FSM alphabet, a state decoder which moves the state machine in the
next states and a preprocessor to map incoming 8-bit characters to 5-bit char-
acters thus mapping all the characters to lower cases. Since case insensitivity
is considered in most application in information retrieval, the preprocessor is
designed to map upper and lower case characters to the binary code of 1 to 26
and the rest of character to 0 [104]. Characters are streamed to the device in
ASCII notation. An incoming 8-bit character triggers the clock and is mapped
to a 5-bit character which is sent to the comparator array. All the 5-bit signals
are distributed to all the comparators which operate in parallel to check if a
match of the incoming character with an alphabet character occurs. If a match
occurs for the comparator k, the output signal chark is set to one and all the
other are set to zero. If no match occurs, all the output signal are set to be zero.
Figure 9.5 illustrates this concept on the base of the word "tictic".

Figure 9.4. Basis structure of a FSM-based words recognizer that exploits the common prefix
and a common set of characters

The one-hot approach (figure 9.2c)) is probably the best implementation

for the kind of automaton realized in the state decoder. First it takes a major
advantage toward the regular structure of the FPGA and it is the fastest way to
implement FSM in hardware [142]. Moreover, there is no need to incorporate
backtracking edges in a suitable hardware representation. To implement a set
of different words, the resulting tree structure of the common FSM is mapped
in hardware. All the leafs corresponding to target words are connected to the
hit port. The width of the hit port is determined by the number of positions
needed to code the number of target words. Upon matching a particular word,
Applications 293

one leaf of the automaton is reached and the hit output is set as to the index of
the matched target word.

Figure 9.5. Processing steps of the FSM for the word "tictic"

1.4 Automaton generation

A tool was provided in [34] to allow for an automatic generation of a term by
document matrix from a document collection with a given list of target words.
This tool automatically generates a VHDL description of the search machine,
and the synthesis can be done with available vendors or third party tools. The
tool is based on the method described in [14] to generate the corresponding
tree structure which is then mapped in hardware. The resulting state machine
is optimized (in terms of number of states) and is able to solve the Mississippi
problem2 for a set of target words. With the union, concatenation, Kleene star
(Kleene closure) operations, all the basic FSMs (FSM related to one word) are
used to build one FSM capable of recognizing all the target words. A set of
final states is linked to the appearance of target words in a streaming text. The
resulting transition table is then translated into VHDL code in such a way that

2 Can the state machine recognize the word ’issip’ in a file containing the word Mississippi?[52]
294 RECONFIGURABLE COMPUTING

no backtracking and crossing edges appear in the configuration. The resulting

VHDL code is then synthesized, mapped, placed, routed and configured to be
downloaded into the FPGA with the help of the generated scripts. The doc-
ument collection is then scanned and a column vector representing the weigh
of target words in this document is returned for each document. The perfor-
mance of this approach was evaluated using the Spyder Virtex board from the
FZI Karlsruhe [214], a PCI board featuring a Xilinx Virtex XCV300. A list of
330 key words with an average of 10 characters each could then be compiled
into the FPGA, leading to an automaton with 2370 states. The utilization of
the FPGA was 59% (1817 CLB slices out of 3072 slices), including bus inter-
face. The stream rate of the characters through the FPGA was by 3 million3
characters per second.

2. Video Streaming
Video streaming is the process of performing computations on video data,
which are streamed through a given system, picture after picture.
Implementation of video streaming on FPGA have attracted several researchers
in the past, resulting in the building of various platforms [39] [115] [198] [72]
[207] [140] [201] for various computations. Two main reasons are always
state as motivation for implementation video streaming in FPGAs: the perfor-
mance, which results from using the exploitation of the inherent parallelism in
the target applications in order to build a tailored hardware architecture and the
adaptivity, which can be used to adjust the overall system by exchanging parts
of the computing blocks with better adapted ones.
In [193], Richard G.S discusses the idea of parameterized program gener-
ation of convolution filters in an FPGA. A 2-D filter is assembled from a set
of multipliers and adders, which are in turn generated from a canonical serial-
parallel multiplier stage. Atmel application notes [13] discuss 3x3 convolu-
tion implementations with run-time reconfigurable vector multiplier in Atmel
FPGAs. To overcome the difficulties of programming devices with classic
Hardware Description Languages(HDL) like VHDL and Verliog, Celoxica has
develop the Handel-C language. Handel-C is essentially an extended subset
of the standard ANSI-C language, specifically designed for use in a hardware
environment. The Celoxica Handel-C compiler, the DK1 development envi-
ronment includes a tool, the PixelStream, for an easy generation of video pro-
cessing functions for FPGA implementations. PixelStream offers a set of ba-
sic modules and functions that can be (graphically) combined to build a more
complex datapath in FPGA. The resulting design can be mapped to one of the
Celoxica FPGA development boards.

34 million if less than 255 words are compiled into the word recognizer.
Applications 295

We will not focus in this section on the details of video or image process-
ing. A comprehensive description of video processing algorithms and their
hardware implementation on FPGA is provided in [153].
The Sonic architecture [115] is a configurable computing platform for accel-
eration and real-time video image processing. The platform consists of plug-
in processing elements (PIPEs), which are FPGA daughter card that can be
mounted on a PCI-board plugged on a PCI-slot of a workstation. A PIPEflow
bus exists to connect adjacent PIPES, while a the available PIPE provide global
connection to the PIPES. Sonic’s architecture exploits the reconfiguration ca-
pabilities of the PIPES to adapt part of the computation flow at run-time.
The Splash [97] architecture was also used in video processing. Its systolic
array structure makes it well suited to image processing.
The ESM platform introduced in section 7.0.2 presents an optimal pipelined
architecture for the modular implementation of video streams. Its organization
in exchangeable slots, each of which can be reconfigured at run-time to per-
form a different computation, makes it a viable platform in video streaming.
Moreover, the communication infrastructure available on the platform provides
an unlimited access of modules to their data, no matter on which slot they are
placed, thus increasing the degree of flexibility in the system.
The computation on video frames is usually performed in different steps,
while the pictures stream through the datapath. It therefore presents a nice
structure for a pipelined computation. This has led to a chained architecture
on the base of which most video streaming systems are built. The chain con-
sist of a set of modules, each of which is specialized for a given computation.
The first module on the chain is in charge of capturing the video frames, while
the last module output the processed frames. Output can be rendered on a
monitor or can be sent as compressed data over a network. Between the first
and the last modules, several computational modules can be used according
to the algorithm implemented. A module can be implemented in hardware or
in software according to the goals. While software provides a great degree of
flexibility, it is usually not fast enough to carry the challenging computations
required in video applications. ASICs can be used to implement the computa-
tional demanding parts, however ASIC does not provide the flexibility needed
in many systems. On a reconfigurable platform, the flexibility of a processor
can be combined with the efficiency of ASICs in order to build a high perfor-
mance flexible system. The partial reconfiguration capability of reconfigurable
devices provides the possibility to replace a given module on the chain at run-
time.
Most of the video capture modules provide the frames to the system on a
pixel by pixel manner, leading to a serial computation on the incoming pixels.
Since many algorithms need the neighborhood of a pixel to compute its new
value, a complete block must be stored and processed for each pixel. Cap-
296 RECONFIGURABLE COMPUTING

turing the neighborhood of a pixel is often done using a sliding window data
structure with varying size. The size of the sliding window corresponds to that
of neighbor region of a pixel needed for the computation of the new value. As
shown in figure 9.6, a sliding window is a data structure used to sequentially
capture the neighborhood of pixels in a given image.

FIFO1 W11 W12 W13 W14 W15 Disposed

FIFO2 W21 W22 W23 W24 W25

FIFO3 W31 W32 W33 W34 W35

FIFO4 W41 W42 W43 W44 W45

RAM W51 W52 W53 W54 W55

Figure 9.6. Implementation of a 5x5 sliding windows

A given set of buffers (FIFO) is used to update the windows. The number
of FIFOs vary according to the size of the window. In each step, a pixel is read
from the memory and placed in the lower left cell of the window. Up to the
upper right pixel which is disposed, i.e. outputted, all the pixels in the right
part of the window are placed at the queue of the FIFOs one level higher. The
processing part of the video is a normal image processing algorithm combining
some of the basic operators like:

Median filtering
Basic Morphological operations
Convolution
Edge detection

In the field of video compression, the processing is usually done in a block

by block manner, different from the sliding window. However the overall struc-
ture is almost the same. The neighborhood must always be saved in order to
provided the necessary data for the computation of a new pixel value.
As explained earlier in this section and as figure 9.7 shows, the architecture
for a video streaming system is usually built on a modular and chained basis.
The first module deals with the image captured from an image source. This
can be a camera or a network module which collects the picture through a net-
work channel, or any other source. The frames are alternately written to the
RAM1 and RAM2 by the capture module. The second module collects the pic-
tures from the RAM1 or RAM2, if this is not in use by the first module, builds
Applications 297

Figure 9.7. A modular architecture for video streaming on the ESM

the sliding windows and passes it to the third module, which processes the
pixel and saves it in its own memory or directly passes it to the next module.
This architecture presents a pipelined computation in which the computational
blocks are the modules that process the data frames. RAMs are used to tempo-
rally store frames between two modules, thus allowing a frame to stream from
RAM to RAM and the processed pictures to the output.

2.1 Enabling Adaptivity through Reconfiguration

An adaptive video streaming system is characterized by its ability to opti-
mize the computation according to changing environmental condition. In most
cases, only one module on the computation chain must be changed while the
rest keep running. The video capture module for example can be changed if we
want to optimize the conversion of pixel to match the current brightness or the
current landscape. It is also possible to change the video source from camera
to a new one with different characteristics or the source can be for instance
a network channel collecting the frames in a compressed form. In an adap-
tive system, the functionality of a module on the computation path should be
changed very fast, and without affecting the rest of the system. Two possibili-
ties exist for this purpose. The first one consists of providing some parameters
to the module to instruct it to switch from one algorithm to the next one. How-
ever, the structure of the basic algorithms are not always the same. A Sobel
filter cannot be changed in a Laplace filter by just changing the parameters.
298 RECONFIGURABLE COMPUTING

This is also true for a median operator which cannot be replaced by a Gauss
operator by just changing the parameters. The network capture module and
the camera capture module require two different algorithms for collecting the
pixels and bringing them in the system. The second possibility consists of re-
placing the complete module at run-time with a module of the same size, but
different in its structure, while the rest of the system keeps running. Recon-
figurable devices in general and FPGAs in particular fulfill this requirements.
Many available FPGAs can be partly configured, while the rest of the system
is running. Moreover many FPGAs provide small size on-chip memories, able
to hold part of a frame (the so called region of interest). It is therefore possible
to perform many computations in parallel on different regions of interest, thus
increasing the performance and flexibility of video applications.

3. Distributed Arithmetic
Distributed arithmetic (DA) is certainly one of the most powerful tool for
the computation of the product of two vector products, one of which is con-
stant, i.e it consists of constant values. DA exploits the nature of LUT-based
computation provided by the FPGAs by storing in a LUT, all the possible result
for a set of variable combinations. Computation at run-time only consists of
retrieving the results from the LUT, where they were previously stored. The el-
ements of the variable vector are used to address the look-up table and retrieve
partial sums in a bit-serial (1-BAAT, 1 Bit At A Time) manner. Investigations
in distributed arithmetic have been done in [215] [158] [223] [224] [33] [95]
[63] [62] [31].
One of the notorious DA-contribution is the work of White [215]. He pro-
posed the use of ROMs to store the precomputed values. The surrounding logic
to access the ROM and retrieve the partial sums had to be implemented on a
separate chip. Because of this moribund architecture, distributed arithmetic
could not be successful. With the appearance of SRAM based FPGAs, DA
became an interesting alternative to implement signal processing applications
in FPGA [158] [223] [224]. Because of the availability of SRAMs in those
FPGAs, the precomputed values could now be stored in the same chip as the
surrounding logic.
In [33], the design automation of DA architecture was investigated and a
tool was designed to help the user in the code generation of a complete DA
design, and to perform investigations on the various tradeoffs. Also, an initial
attempt to implement DA for floating-point numbers numbers in order to in-
crease the accuracy, was presented. However, the amount of memory required
to implement such a solution makes it applicable only on small examples.
The idea behind the distributed arithmetic is to distribute the bits of one
operand across the equation to be computed. This operation results in a new
Applications 299

one, which can then be computed in a more efficient way. The product of two
vectors X and A is given by the following equation:
n
X
Z =X ×A= (Xi × Ai ) (3.1)
i=0

We assume that A = (A1 , · · · , An ) is a constant vector of dimension n and

X = (X1 , · · · XP
n ) is a variable vector of the same dimension. With the binary
representation, w−1 j
j=0 Xij 2 , of Xi , where w is the width of the variables and
Xij ∈ {0, 1} is the j-th bit of Xi , equation (3.1) can be rewritten as:
n
X w−1
X w−1
X n
X
Z= Ai × Xij 2j = 2j Xij Ai (3.2)
i=0 j=0 j=0 i=0

Spreading equation 3.2 for a better understanding leads to equation 3.3

Z = (X00 × A0 + X10 × A1 + ... + Xn0 × An )20

+ (X01 × A0 + X11 × A1 + ... + Xn1 × An )21
(3.3)
...
+ (X0(w−1) × A0 + X1(w−1) × A1 + ... + Xn(w−1) × An )2(w−1)

Equation 3.2 represents the general form of a distributed arithmentic. It

shows how the variable operands are distributed in the P equation. Each bit of
each variable operand contributes only once P to the sum ni=0 Xij Ai . Because
Xij ∈ {0, 1}, the number of possible values ni=0 Xij Ai can take is limited to
2n . Therefore, they can be precomputed and stored in a look-up table, provide
that enough place is available in the LUT. Lets call such a look-up table a
distributed arithmetic look-up table (DALUT). The DALUT has the size w×2n
bits. At run-time, the n-tupel (X1j , X P2j , ..., Xnj ) will be used to address the
DALUT and retrieve the partial sums ni=0 Xij Ai . the n-tupel (0, 0, ..., 0) will
be used to address the location 0, (0, 0, ..., 1) is used to address the location 1,
and (1, 1, ..., 1) is used to address the location 2n − 1.
The complete dot-product Z requires w steps to be computed in a 1-BAAT
manner. At step j, the j-thPnbit of each variable are used to address the DALUT
and retrieved the value i=0 Xij Ai . This value is then left shifted by a factor
j (which corresponds to a multiplication by 2j ) and accumulated. After w ac-
cumulations, the values of the dot-product can be collected. The DA datapath
for this computation is obvious as figure 9.8 shows.
The DALUT address consists of a set of bits, each of which belongs to a
component of the input vector. A barrel shifter is used to shift the number
retrieved from the DALUT at the j-th step on j positions. The shifted value is
300 RECONFIGURABLE COMPUTING

Figure 9.8. Architecture of a distributed arithmetic datapath

then add to the value in the accumulator. After w steps, the results is collected
from the accumulator.
Many enhancements can be done on a DA implementation. The size of the
DALUT for example can be halved if only positive values are stored. In this
case, the sign bit, which is the first bit of a number, will be used to decide if the
retrieved value should be added to or subtracted from the accumulated sum.
On the other hand, it is obvious that all the bit operations, i.e. the retrievals of
a value from the DALUT, are independent from each other. The computation
can therefore be performed in parallel. The degree of parallelism in a given
implementation depends on the available memory to implement the DALUTs.
In the case where w DALUTs and datapaths can be instantiated in parallel,
the retrieval of all partial sums can be done in only one step, meaning that the
complete computation can be done in only one step insted of w as in the serial
case.
In general, if k DALUTs are instantiated in parallel, i.e. for a computation
performed on a k-BAAT basis, then w/k steps are required to retrieved all the
partial sums, which can be directly accumulated, thus leading to a run-time of
w/k steps. Figure 9.9 shows a datapath for the computation of the DA.
The input values are segmented in different fields, each of which is assigned
to a datapath for the retrieval of the partial sums from the corresponding DA-
LUT. The retrieved values from the k DALUTs are shifted by the correspond-
ing values and added to the accumulated sum. After w/k steps, the result can
be collected from the accumulator.
Applications 301

Figure 9.9. k-parallel distributed arithmetic datapath

3.1 Distributed Arithmetic in High Dimension

In many computational areas like in mechanical control, computations are
not only limited to dot-product. Matrix operations are usually required, whose
general form can be defined by equation 3.4.
   
z1   x1
z2  a11 ... a1r  
 = ...  x2  (3.4)
 ...   ... 
a1s ... asr
zs xr
This equation can be implemented using s DALUTs.PThe i-th (i ∈ {1, .., s})
DALUT is used to compute the vector product zi = rj=0 xj × aij with the
constants ai1 to air . If there is enough space on the chip to hold all the DA-
LUTs, then equation 3.4 can be computed in few clocks. As we will see later,
partial reconfiguration can also be used to implement many dot-products, if
there is not enough memory on the chip to hold all the DALUTs.

3.2 Distributed Arithmetic and Real Numbers

The straightforward approach to handle real numbers in distributed arith-
metic is the use of fixed-point. The representation of a real number using the
fixed-point approach is done by defining an integer part and a fractional part
302 RECONFIGURABLE COMPUTING

for each number. The integer part is separated from the fractional part using
a point. Because the point is only an imaginary and not physically available
in the number representation, operations on fixed-point numbers are not differ-
ent than those on normal integer numbers. The datapath previously presented
can be used without modifications for for real numbers represented as fixed-
point. Representing real numbers as fixed-point is advantageous in the compu-
tation speed and the amount of resources required to implement the datapath.
However, the ranges of fixed-point representations as well as their precisions
is small compared to those of a floating-point representations. Therefore we
would like to handle real numbers as floating-point.
In this section we present a concept first investigated in [33] for handling
real numbers, represented as floating-point in the IEEE 754 format, using dis-
tributed arithmetic. In IEEE 754 format, a number X is represented as follows:

X = (−1)SX 2eX × 1.mX (3.5)

In this equation, eX is the exponent (we consider that the substraction with
the bias is done and the result is eX ), mX is the mantissa and SX is the sign
of X. Without loss of generality, we will consider the sign to be part of the
mantissa, thus the new representation is X = 2eX × mX . With A, X and
Z being all floating-point numbers, the floating-point counterpart of equation
(3.1) is given by:
n
X n
X
Z =X ×A= (Xi × Ai ) = (2eAi × mAi ) × (2eXi × mXi )
i=0 i=0
n
X n
X
= (2eAi × 2eXi ) × (mAi × mXi ) = (2eAi +eXi ) × (mAi × mXi )
i=0 i=0
(3.6)
The goal here is to compute and provide Z as floating-point number, us-
ing the same principle of the partial sums previously described. At each step
of the computation, a floating-point value Fi must be read from the floating-
point DALUT and added to an accumulated sum, which is also a floating-point
number. Since the adder used at this stage is a floating-point adder, we as-
sume that issues like rounding and normalization have been integrated in its
implementation. From the last part of equation 3.6, we see that the value
(2eAi +eXi ) × (mAi × mXi ) represents a floating-point number with expo-
nent eAi + eXi and mantissa mAi × mXi . By setting eFi = eAi + eXi and
mFi = mAi × mXi , we have the requested values at each computation step.
Instead of computing the exponential part (2eAi +eXi ) of Fi as well as its man-
tissa (mAi × mXi j ) on line, the idea here consists of using two floating-point
DALUTS for each constant Ai . The values eAi + eXi will then be precom-
puted and saved in the first DALUT and mAi × mXi in the second one. Lets
Applications 303

call the first DALUT which stores the exponents the EDALUT and the sec-
ond DALUT which stores the mantissas the MDALUT. The size of EDALUT,
size(EDALU T ) as well as that of MDALUT, size(M DALU T ) are given in
equations (3.7).

size(EDALU T ) = n × 2|E| × |E| bits (3.7)

size(M DALU T ) = n × 2|M | × |M | bits (3.8)

|E| is the exponent width and |M | is the mantissa width of the floating-point
representation.
The main argument against this approach is the size of the DALUTs used
for this floating-point DA implementation, which can be so large that an im-
plementation cannot be possible. However, if we consider a DA implementa-
tion involving five variables and five coefficients represented in the IEEE 754
floating-point format with 8 bits exponent and 10 bits mantissa, the total mem-
ory requirement for the EDALUTs and the MDALUTs is: ((5 × 28 × 8) + (5 ×
210 × 10))/1024 = 60 Kbits. Therefore the EDALUTS and the MDALUTs
will easily fit into a very small low cost FPGA like the Xilinx Spartan III 50
with a total of 72 Kbits Block RAM [225].
For large vectors with numbers represented on large mantissas and expo-
nents, this approach will require a large amount of memory, certainly not di-
rectly available on the chip. An external memory must therefore be used.
Having the EDALUTs and the MDALUTs for each coefficient, the datapath
will not be implemented as in the fixed-point or integer case. The variables are
no more input in a bit-serial way. At step i the variable Xi is used to address
the two i-th EDALUT and MDALUT. The bits of eXi are used to access the
EDALUT, while the bits of mXi are used to access the MDALUT in parallel.
The values collected from the EDALUT and the MDALUT are used to build
the floating point number Fi . After n steps the floating-point dot-product is
computed. Figure 9.10 shows the datapath for the floating-point DA. Since
all the DALUTs for the n coefficients are available on the device, they can be
accessed in parallel to compute the dot-product in only one step.
Signal processing applications are usually based on the computation of the
dot-product of one vector of variables with a vector of constants. They are
ideal candidates for a DA implementation in reconfigurable hardware. As state
earlier in this section, in order to facilitate the design of DA applications a
tool was implemented in [33] for the design automation. Also, the possibil-
ity is given to investigate the different tradeoffs for a given application. For a
given device and a given set of constants, different DA implementations can
be generated, from the full parallel implementation (w-BAAT), which can be
computed in only one step, to the full sequential DA (1-BAAT). For each possi-
304 RECONFIGURABLE COMPUTING

e xn m xn

e x1 m x1

e x0 m x0

eA1 mA1
eA1 + 1 mA1 * 2
Control
LUTs 1

eA1 + 2|E| mA1 * 2|M|

eA2 mA2

LUTs 2 e Fi m Fi
eA2 + 2|E| mA2 * 2|M|

+
eAn mAn
e Zi m Zi

LUTs n
eAn + 2|E| mAn * 2|M|

Figure 9.10. Datapath of the distributed arithmetic computation for floating popin numbers

bility, the user is provided the area and speed of the design. Real numbers can
be handled either as fixed-point or as floating-point in the IEEE 754 format
with the technique previously defined. The width of the mantissa as well as
that of the exponent has to be provided. For the final design, the tool generates
a description in the Handel-C language.

3.3 Use of Reconfiguration

The modification of design implemented as distributed arithmetic can be
done at run-time through the replacement of the DALUTs. The remaining part
of the design must not be changed, because it is the same for all designs. In
this case, the designer must not have to deal with all the cumbersome aspects
of partial reconfiguration design. However, if the datawidth of the operators
changes, then the computing structure (shifter, adder, accumulator) must be
replaced as well. In this case the partial reconfiguration of the device cannot
be avoided. In section 4, we present the use of reconfiguration in DA designs,
using the concept of adaptive controller concept, which makes use of DA and
partial reconfiguration.
Applications 305

3.4 Applications
In this section, we present one application, the implementation of recur-
sive convolution algorithm for time domain simulation of optical multimode
intrasystem interconnects that was substantially speeded-up through the use of
distributed arithmetic. Another application that has benefit from the imple-
mentation as DA is the adaptive controller. Because section 4 is devoted to
adaptive controller, we explain only the first application here.

3.4.1 Recursive Convolution Algorithm of Time Domain Simulation of

Optical Multimode Intrasystem Interconnects
In general, an optical intrasystem interconnect contains several receivers
with optical inputs driven by transmitters with optical outputs. The intercon-
nections of transmitter and receivers are made by a passive optical waveguide
which are represented as a multiport (figure 9.11) using ray tracing approach.
y 1 1

x 1
y 1 e 1

y 2 1
x 2
y 2 e 2

x e y a 1

y a e a

Figure 9.11. An optical multimode waveguide is represented by a multiport with several trans-
fer paths.

The transfer of an optical signal along the waveguide can be computed by a

multiple convolution process. Frequency domain simulation methods are not
applicable regarding to the high number of frequency. Pure time domain sim-
ulation methods are more efficient if the pulse responses can be represented by
exponential functions. The application of a recursive method for the convolu-
tion of the optical stimulus signals at the input ports with the corresponding
pulse responses enables a time efficient computation of the optical response
signals at the belonging output ports. The recursive formula to be implemented
in three different intervals is given by equation 3.9.

y(tn ) = f0 · y(tn−1 ) + (3.9)

f4 · x0 − f5 · x1 + f24 · x2 + f53 · x3 .
Because the values f0 , f4 , f4 , f24 and f53 are constants, while tn−1 , x0 , x1 ,
x2 , x3 are variables, equation 3.9 is best adapted for a DA-implementation.
306 RECONFIGURABLE COMPUTING

For this equation, different tradeoffs were investigated in the framework pre-
sented in [33], for generation and evaluation of DA-trade-off implementation.
A Handel-C code were generated and the complete design was implemented
on a system made upon the Celoxica RC1000-PP board equipped with a Xilinx
Virtex 2000E FPGA and plugged into a workstation.

Workstation 1 interval 3 intervals

Sun Ultra 10 73.8 ms 354.2 ms
Athlon (1.53 GHZ) 558 ms 1967.4 ms
FPGA (time) 1 interval 3 intervals
Custom dot-product design 25.6 ms 76.8 ms
Sequential DA 19.4 ms 19.4 ms
3-parallel DA 6.4 ms 6.4 ms
FPGA (area) 1 interval 3 intervals
Custom dot-product design could not fit could not fit
Sequential DA yes (7 % ) yes (14 % )
3-parallel DA yes (14 % ) yes (42 % )
Table 9.1. Results of the recursive convolution equation on different platforms

Table 9.1 provides the implementation results on different platforms. The

workstation is used for sending the variable to the FPGA and collecting the re-
sult of the computation. The implementation of equation 3.9 in three intervals
occupies about 14 % of the FPGA area while running at 65 MHZ. In order to
obtain a maximal parallelism in the computation, 6 DALUT could be imple-
mented in parallel leading to a 6 times speed-up. As can be seen on figure 9.12,
the 6-parallel DALUT and the corresponding adder three occupy about 76 %
of the FPGA area. Enough space is left for the routing.
The same design was implemented without the use of DA. It could not fit
into the same FPGA and the run-time was much bigger that that of the DA-
implementation. The performance as well as the area consumption of our the
DA-implementation is more efficient than that of all the other architectures.

4. Adaptive Controller
In this section, we investigate the next field of application of reconfigurable
devices, namely the implementation of adaptive controllers, also identified here
as multi-controller, using partial reconfiguration.
We will first introduce the multi-controller concept. Thereafter, we inves-
tigate its implementation using the distributed arithmetic. Finally, the imple-
mentation of the whole design using partial reconfiguration is shown.
Applications 307

Figure 9.12. Screenshot of the 6-parallel DA implementation of the the recursive convolution
equation on the Celoxica RC100-PP platform

4.1 Adaptive Control

The task of a controller is to influence the dynamic behavior of a system
referred to as plant. A control feedback is available, if the input values for the
plant are calculated on basis of the plant’s outputs.
Adaptive controllers deal with the control of a plant, for instance a com-
plex mechatronic system, in a changing environment. In this case, the control
is modeled as a physical process that operates in a limited set of operating
regimes. Depending on the environmental conditions, the operation regime of
the plant must be changed to adapt to new conditions. With conventional meth-
ods it might be possible to design one robust controller that optimally controls
the plant in all operating regimes. According to the conditions, the control
regime may be changed using a set of parameters. Besides the possible slow
response of such a controller [166], the size can be too big due the large amount
of possible cases to be considered. To overcome this handicap, the controller
modules can be implemented as separate modules and stored in a database.
They will be use at run-time to replace the active controller, whenever a new
control algorithm must be implemented. The price to pay for this solution is
the amount of memory to store all possible module implementations as well as
the time needed to switch from one module to the next one.
The general architecture of an adaptive controller as described in [161] is
shown in figure 9.13. It consists of a set of controller modules (CM), each of
which is optimized for an operating regime of the plant. A supervisor com-
ponent is used to switch from module to module an set the active one.4 The
decision to switch from one CM to the next one is made on basis of mea-

4 At a given time, the active controller is the one which controls the plant.
308 RECONFIGURABLE COMPUTING

surements of physical values of the plant. The strategy of the supervisor can
vary from simple boolean combinations of the input values to very complex
reasoning techniques [206].
Multi−Controller

Supervisor

i CM 1

CM i
Plant MUX

CM n

Figure 9.13. Adaptive controller architecture

The computation of the plant inputs is normally done in a time discrete

manner, using a set of discrete sample points tk . . . tk+1 . The time between two
sampling points is constant and defines the sampling period T . The controller
application can therefore be seen as a periodic real-time task, which must be
performed at every sampling point and must be completed within one period
T.
While in a Von Neumann processor based implementation of the system,
each module is implemented as sequence of instructions, in a reconfigurable
device, the tasks are implemented as hardware modules to be downloaded onto
the device at run-time. Lets take a look on the behavior of the single modules
and their implementations.

4.2 Controller-Module Specification

A common approach is to model the plant as a linear time-invariant system.
Based on this model and the requirements of the desired system behavior, a
linear controller is systematically derived using formal design methods.
Lets assume that every controller module CMi of figure 9.13 is designed
as a linear time-invariant system, optimized for a given operating regime of
the plant. After a time discretization has been performed, the behavior can be
captured in equation 4.1.
The input vector of the controller module, representing the measurements
from sensors of the plant, is represented by u, y is the output vector of the
controller, which is used to drive the actuators of the plant, and x is the inner
Applications 309

state vector of the controller. The matrices A, B, C and D are used for the
calculation of the outputs based on the inputs.

x(k + 1) = Ax(k) + Bu(k)

(4.1)
y(k) = Cx(k) + Du(k)
p = dim(u), s = dim(x), q = dim(y)
The task of the digital controller is to calculate equation 4.1 during one
sampling interval. The new state xk+1 and the output yk must be computed
before the next sampling point k + 1.
The state space equations of a digital linear time invariant controller (equa-
tion 4.1) can be written as a product of a matrix of constant coefficients and a
vector of variables (equations 4.2 and 4.3).

z(s+q,1) = M(s+q,s+p) v(s+p,1) (4.2)

    
x1 (k + 1) a11 ... a1s b11 ... b1p x1 (k)
 .   . .. . . .. .  . 
 .   . . . . . .  . 
 .   . . . .  . 
    
 s + 1)
x (k  as1 ... ass bs1 ... bsp x (k)
 y (k)
 1



=  c
 11 ... c1n d11 ... d1p



 s
 u (k) 
 1


(4.3)
    
 .   . .. . . .. .  . 
 .   .. . . . . .  . 
. . . . .
yq (k) cq1 ... cqs dq1 ... dqp uq (k)
| {z } | {z }| {z }
z M v

Equations 4.2 and 4.3 define a classical distributed arithmetic equation in

higher dimension. Each of the s + q entries of the resulting vector v defines
a separate DA equation. We therefore need to compute s + q different DA-
equations, all independent from each other. The computation of the s + q
equations can then be performed in parallel, provide that enough resources
are available on the chip to store all the DALUTs. If enough memory is not
available for storing all the DALUTs as well as the corresponding logic for the
datapath, then the different equations must be implemented sequentially. In
general, compromises must be found between the amount of parallelism that
one need to have and the speed of the design.

4.3 Use of Reconfiguration

As we explained earlier, the controller modules can be implemented as hard-
ware modules and store in a database from which they will be downloaded at
run-time by the supervisor module.
The straightforward approach to realize this is to build a structure consisting
of a fixed part in which the supervisor resides. A reconfigurable slot is then
foreseen as place holder for the controller module that will be downloaded at
run-time (figure 9.14 left). Each controller module is downloaded at run-time
310 RECONFIGURABLE COMPUTING

Figure 9.14. Adaptive controller architecture. Left the one slot implementation, and right the
two slot implemenation

into the reconfigurable slot, which has predefined interfaces to the supervisor
and to the plant.
The problem with this approach is the vacuum, which arises on switching
from one module to the next one. Because only one slot is available, the re-
configuration of this slot will place the plant in a "floating state" on reconfig-
uration, where it is not really controlled. In order to avoid this situation, the
approach developed in [64] was the use of two slots. One active slot an a pas-
sive one (figure 9.14 right). The active slot is in charge of controlling the plant,
while the reconfiguration takes place in the passive slot. Whenever the super-
visor decides to replace a controller module, the reconfiguration will first take
place on the passive slot. Thereafter, the control of the plant will be given to
the configured module, which becomes the active one, while the former active
becomes passive.

5. Adaptive Cryptographic Systems

Sensitive data exchange is always coupled with security issues. Cryptogra-
phy provides a means to allow two communicating entities to exchanged data
in such a way that a third party that intercept the data will not be able to under-
stand their meaning. A protocol is therefore required, which allow the sender
to crypt the data in such a way that only the receiver will be able decrypt them.
Considerable advances have been done in communication in the past and
this trend is likely to continue. The development of the internet has also pushed
the development in several branches where a network processing could not be
Applications 311

tough off few years ago. Applications like e-commerce, e-government, virtual
private network, on-line banking must provide a high degree of security.
Over years, a large variety of standards like Triple-DES, Advanced Encryp-
tion Standard AES), Data Encryption Standard (DES), RSA, OpenPGP, Ci-
pherSaber, IPsec, Blowfish, ANSI X9.59, CAST, RC4 and RC6, have been
developed to provide high security. With this large variety of standards and the
customized implementation possibilities for each standard, cryptography can
be seen as one of the most versatile application domains of computer science.
Depending on criteria like speed, degree of flexibility, and degree of security,
single implementations of cryptography application were developed either as
software or as intellectual property component.
In [221] the advantages of using reconfigurable hardware in cryptography
are listed. The author focus on the traditional advantage of flexibility and
performance. The flexibility is so far important because it offers the possi-
bility to used the same hardware to switch from one algorithm to the next
one at run-time, according to factors like the degree of security, the compu-
tational speed, the power consumption. Also, according to some parameters,
a given algorithm can be tuned. Moreover, algorithms that has been broken
and where the security is no more insured can be changed by means of re-
configuration. The system can easily be upgraded to include new standards,
developed while the system was already deployed. The corresponding algo-
rithm can therefore be compiled and included in the library of bitstreams for
the device configuration. The second advantage provided by reconfigurable
hardware, namely the performance can be used to efficiently implement the
components, by using the inherent parallelism and building efficient operators
for computing boolean operation on a very large amount of data. This results
on a large throughput and a cost efficiency. Experiments reported a throughput
of 12 GBit/s for an implementation of the block cipher AES on an FPGA Virtex
1000, using 12,600 slices and 80 RAMs [90], while an ASIC implementation,
the Amphion CS5240TK [150] clocked at 200 MHz, could provided twice
the throughput of the FPGA solution. The same algorithm, implemented on a
DSP TI TMS320C6201 provided a throughput of 112.3 Mbits/s [222], while a
throughput of 718.4 Mbit/s could be reached on a counterpart implementation
on a on a Pentium III [10].
The general architecture of an adaptive cryptographic engine proposed by
Prasanna and Dandalis [61] [178], basically consists of a database to hold the
different configuration that can be downloaded at run-time onto the device,
like an FPGA for instance, to perform the computation and a configuration
controller to perform the reconfiguration, i.e. downloading the corresponding
bitstream form the database into the FPGA. Each bitstream represent a given
algorithm implementing a given standard and tuned with some parameters ac-
cording to the current user’s need.
312 RECONFIGURABLE COMPUTING

Figure 9.15. Architecture of an adaptive cryptographic system

With the recent development in FPGA, it is possible to have a complete sys-

tem on programmable chip (processor, peripherals, custom hardware compo-
nents, interconnection) implemented on an FPGA. The configuration controller
therefore must no more resides off chip. It can be implemented as custom on-
chip hardware module or as software running on am embedded processor. Also
the possibility to reconfigure only part of the chip opens new possibilities. In
the architecture presented in [61] [178], the whole device must be reset on re-
configuration, thus increasing the power consumption because of the amount
of the data that must be downloaded on a chip. Power consumption is usually
a big problem in mobile environments and the implementation must consider
such issues. Besides this power saving, partial reconfiguration also provides
the possibility of keeping a skeleton structure into the device and perform only
few modifications at run-time on the basic structure to move from one algo-
rithm to the next one. In [42] a cryptographic application is implemented as
exchangeable module of a partial reconfigurable platform. The system, which
is based on the AES algorithm algorithm consumes only 954 Virtex slices and
3 block RAMS. The cryptographic algorithm is used in this case just as a block
to test the partial reconfigurable platform, instead of using the partial reconfig-
uration to enhance the flexibility of the cryptographic algorithm.
Figures 9.15 present a possible architecture of an adaptive cryptography
system, using the previous mentioned advantages of partial reconfigurable de-
vices.
The main difference with the architecture presented in [61] is the used of
partial reconfiguration, which allows for an integration of all components on a
Applications 313

single chip. Also, in contrast to the adaptive architecture for a control system
presented in figure 9.14, the loader module resides into the device. Depend-
ing on the implementation chosen, the loader can reside inside or outside the
device. If a Xillinx Virtex chip is used and the configuration takes place via
the ICAP port, then the loader, which is the ICAP module in this case, is au-
tomatically available inside the device. However, if the configuration happens
through the normal SelectMap port, then we need an external loader module
for collecting configuration data from the database and copy them on the con-
figuration port.
In figure 9.15, the architecture is logically divided in two main blocks. A
fix one, which remains continuously on the chip. It consist of the parts, which
are common to all the cryptographic algorithms in general or common to the
algorithms in a given class only. On the figure, we show only one reconfig-
urable slots, however, it can be implemented as set of configurable blocks,
each of which can be changed by mean of reconfiguration to realize a given
customized standard.
The last point concerns the development of building blocks that will be com-
piled in bitstreams to be downloaded into the device at run-time. A designer
is no more required to focus on the hardware implementation of the crypto-
graphic algorithm. A lot of work was done in this direction and the results are
available. We need mostly to focus on the architecture of the overall system
and find out how a viable partitioning can be done, according to the reconfigu-
ration scheme.
Most of the work have focussed in various implementations of a given ap-
proach like the RSA [185] in [155] [88] [154] [232] [50] or the implementa-
tions described in [78] [19] [20] [47] [145] [152] [15], mostly parameterizable
and based on the Elliptic Curve approach [138] [157]. Generators for pro-
ducing a customized description in a hardware description language have been
developed for example in [47]. This can be used to generate the variety of
configurations that will be used at run-time to move from one implementation
to the next one.

6. Software Defined Radio

Software defined radio (SDR) was defined in 1991 by Joe Mitola [129] as
follow: "A software radio is a radio whose channel modulation waveforms
are defined in software. That is, waveforms are generated as sampled digital
signals, converted from digital to analog via a wideband DAC and then pos-
sibly upconverted from IF to RF. The receiver, similarly, employs a wideband
Analog to Digital Converter (ADC) that captures all of the channels of the soft-
ware radio node. The receiver then extracts, downconverts and demodulates
the channel waveform using software on a general purpose processor."
314 RECONFIGURABLE COMPUTING

As it is the case in many new technologies, software defined radio (SDR)

was initiated by the military in an attempt to alleviate numerous problems as-
sociated with traditional radio systems [162]. In the project SPEAKeasy [162],
which can be seen as precursor of the SDR technology, several military entities
(DARPA, AIR FORCE/AFRL, ARMY/CECOM, NAVY/NRaD/SPAWAR, NSA)
joined efforts in the development of "programmable" radios in order to demon-
strate the feasibility of a Multiband, Multimode Radio (MBMMR) and to de-
velop technologies that facilitate programmability and implementation of MB-
MMR. The achievement was a system with two programmable channels, the
Texas Instrument quad-TMS320C40 multi-chip module for digital signal pro-
cessing, and a SUN Sparc 10 workstation for interfacing with human.
The main goal in software defined radio is to move the analog-digital sig-
nal conversion as closer as possible to the antenna and to process the digital
signals in software rather than using a set of dedicated hardware components.
The expected benefits of using software defined radio have drained great atten-
tion and interest from governmental entities, from the academia and from the
industry.
SDR is the next evolutionary stage of wireless technology. Many believe
that it is just a question of time up to the the adoption of SDR in the broad
range of communication facilities, and several initiatives like [121] [189] were
born with the goal of accelerating the proliferation of SDR.
The advantages of SDR can be categorized as follows:
Interoperability in order to provide support of multiple standards through
multimode, multiband radio capabilities
Flexibility for an efficient shift of technology and resources
Adaptivity to provide a faster migration towards new standards and tech-
nologies through programmability and reconfiguration
Sustainability for increasing resource utilization through generic hardware
platforms
Reduced ownership costs because of the reduced infrastructure, the lower
maintenance, and the easier deployment
The general architecture of a SDR is presented in figure 9.16.
At least one digital to analog (DAC) and one analog to digital converter
(ADC) must ba available in a SDR system. While the ADC performs conver-
sion of the incoming analog signals to digital signals, the DAC performs the
inverse operation.
The signal processing part collects digital signals from the ADC or from
another source like a base station input, processed them and send them to the
DAC.
Applications 315

Figure 9.16. Architecture of a software defined radio system

The ideal realization of a software defined radio would push the ADC and
DAC as close as possible to the antenna. The practical realization is however,
as explained in [217], very challenging.

6.1 use of reconfiguration

Reconfigurable devices provide the best prerequisites to fulfill the expecta-
tions (interoperability, flexibility, adaptability, sustainability, and reduced own-
ership costs) placed on software defined radio systems. As explained in the
previous section, it is possible to realize a complete system consisting of a pro-
cessor, peripherals and reconfigurable custom hardware modules as system on
programmable chip. In such a system, the maximal flexibility will be insured
by the processor, while lower degree of flexibility can be realized by mean
of reconfiguration. For many algorithm in signal processing, software alone
will not be able to provide the required performance. A migration of functions
in hardware is then required. The simple approach as realized in [9] [122]
[48] is to provided a single coprocessor for computation intensive tasks, like
the CORDIC [209] implementation, directly attached to the processor. This
solution provides the possibility to configure the system as whole, either to im-
plement a new functions at run-time, or event to upgrade the system structure
according to the availabilities of new standards. The use of partial reconfig-
uration [69] will be welcomed, to allow the upgrading to take place without
interrupting the system, while saving a considerable amount of power.

7. High Performance Computing

One of the first and most investigated fields of application of FPGAs in
the 90s was high-performance computing. The goal of many platform built at
that time was to provide acceleration of application like searching in genomic
databases, signal and image processing, matrix factorization. Several architec-
tures were developed, mostly consisting of a workstation on which one or more
FPGA were attached. By reconfiguration of the FPGAs, custom instructions
could be added to the computing system, sometimes even at run-time. While
remarkable speed-ups were reported for some applications, the resulting cus-
tom computing machines (CCMs) could not be widely adopted as viable com-
puting platforms. This was mainly due to architectural limitations and lack of
316 RECONFIGURABLE COMPUTING

high-level programming tools. The capacity of FPGAs was too small to allow
real migration of software components in hardware. Furthermore, the perfor-
mance bottleneck posed by the slow communication between host processor
and FPGAs often busted speed-ups. Finally, there were no useful software-like
description languages that could have encourage software designers to write
programs for such machines and implement efficient libraries for reusable data
stream management. Programming CCMs meant tedious low-level hardware
design, which was not attractive at all.
The last couple of years brought a pivotal change. FPGA’s logic densities
increased dramatically, now allowing for massive bit-level parallelism. Avail-
able high-level languages like Handel-C [125] and ImpulseC [175] as well
as progress in compilation technology makes the logic resources available to
high-level application programmers.
Boosted by the recent engagement of companies such as Cray [58], SRC
Computers [59], and ClearSpeed [51], Nallatech [165], and SGI [191], CCMs
are experiencing a renaissance. Despite the use of new high-speed interfaces
and efficient datastream management protocols between the components in the
new systems, the new architectures are built on the same models like the old
ones in the 90s. It usually consist of a set of FPGAs, on the same board with
one or more processors. The processors control the whole systems and con-
figured the FPGAs at run-time with dedicated function to be accelerated in
hardware.
While considerable progress have been done, the rentability of the ma-
chines provided by the firms previously mentioned still have to be proven.
The rentability cannot be evaluated anymore only on the base of speed-ups ob-
served in some class of applications. It is not sure that a comparison made in
terms of performance/$, performance/J, and performance/sqm will favor fur-
ther deployment of those systems. Craven and Athanas recently provided in
[57] a study on the viability of the FPGA in supercomputers. Their study is
based on the cost of realizing floating-point into FPGAs. The cost include not
only the performance gain, but also the price to pay. The authors came to the
conclusion that the price to pay for a FPGA-based supercomputer is to high for
the marginal speed-ups.
High-precision Floating-point arithmetic is usually the bulk of computa-
tion in high-performance computing. Despite the progress done in develop-
ing floating-point units in hardware, Floating-point computation and FPGA is
still a difficult union. The advantage provided by FPGA in customized im-
plementation of floating-point as described Shirazi [192] is unlikely to help
here because the datapath must provide customization that consume the largest
amount of resources. Coarse-grained elements like the embedded multipliers
in FPGAs or coarse-grained reconfigurable device provide the best prerequi-
sites for the use of reconfigurable device in supercomputers.
Applications 317

8. Conclusion
We have presented in this chapter a couple of applications that can take
advantage of the flexibility as well as performance of reconfigurable devices.
For each application targeted, we placed the focus mostly on the architectural
organization because the goal was not to necessary present the detailed im-
plementation of an application. However, we have presented a comprehensive
description of the pattern matching application, while keeping the other pre-
sentation short. This is due to the low attention paid on pattern matching in
hardware in available textbooks. With the large amount of literature in other
presented topics like image processing and control, we choose not to focus in
details of those applications.
A lot of work have been done in the last two decades and a large number
of applications was implemented in FPGA, which is the main representative
of reconfigurable devices. We could not present nor cite all available imple-
mentations here. We have rather focussed on few ones where we could show
the usefulness of reconfigurability. Finally, we would like to emphasize that
despite two decades of research, the existence of "killer applications" could
not be shown for reconfigurable device, thus limiting the acceptance of recon-
figuration in the industry. The existence of a killer application, will certainly
boost the development of reconfigurable devices, leading to new class of de-
vices, tools and programmers. Despite this missing step, research keep going
and the support of the industry is needed more than it have ever be.
318 RECONFIGURABLE COMPUTING
References

[1] G. R. AB, MicroBlaze Processor Reference Guide: Embedded Development Kit EDK
8.2i, 2005, http://www.gaisler.com. [Online]. Available: http://www.gaisler.com

[2] A. Ahmadinia, C. Bobda, S. Fekete, J. Teich, and J. van der Veen, “Optimal routing-
conscious dynamic placement for reconfigurable computing,” in 14th International
Conference on Field-Programmable Logic and Application, ser. Lecture Notes in
Computer Science, vol. 3203. Springer-Verlag, 2004, pp. 847–851, available at
http://arxiv.org/abs/cs.DS/0406035.

[3] A. Ahmadinia, C. Bobda, M. Bednara, and J. Teich, “A new approach for on-line place-
ment on reconfigurable devices,” in Proceedings of the 18th International Parallel and
Distributed Processing Symposium (IPDPS) / Reconfigurable Architectures Workshop
(RAW), 2004.

[4] A. Ahmadinia, C. Bobda, J. Ding, M. Majer, J. Teich, S. P. Fekete, and J. C. van der
Veen, “A practical approach for circuit routing on dynamic reconfigurable devices,”
in RSP ’05: Proceedings of the 16th IEEE International Workshop on Rapid System
Prototyping (RSP’05). Washington, DC, USA: IEEE Computer Society, 2005, pp.
84–90.

[5] C. Alpert and A. Kahng, “Geometric embeddings for faster and better multi-way netlist
partitioning,” 1993.

[6] C. J. Alpert and A. B. Kahng, “Multi-way partitioning via spacefilling curves and dy-
namic programming,” in Design Automation Conference, 1994, pp. 652–657.

[7] C. J. Alpert and S.-Z. Yao, “Spectral partitioning: The more eigenvectors, the better,” in
Design Automation Conference, 1995, pp. 195–200.

[8] ADM-XRC-II Xilinx Virtex-II PMC, Alpha Data Ltd., 2002, http://www.alpha-
data.com/adm-xrc-ii.html.

[9] Altera, “Software defined radio,” www.altera.com/literature/cp/fpga-cores-for-sdr.pdf.

[Online]. Available: www.altera.com/literature/cp/fpga-cores-for-sdr.pdf

[10] K. Aoki and H. Lipmaa, “Fast Implementations of AES Candidates,” in The

Third Advanced Encryption Standard Candidate Conference. New York, NY,
320 RECONFIGURABLE COMPUTING

USA: National Institute of Standards and Technology, 13–14 Apr. 2000,

pp. 106–120, entire proceedings available from the conference homepage
http://csrc.nist.gov/encryption/aes/round2/conf3/aes3conf.htm. [On-
line]. Available: citeseer.ist.psu.edu/aoki00fast.html

[11] P. Athanas, J. Becker, G. Brebner, and H. ElGindy, “Dynamically reconfigurable archi-

tectures,” in Dagstuhl Semiar 03301, 2003.

[12] P. M. Athanas and H. F. Silverman, “Processor reconfiguration through

instruction-set metamorphosis,” IEEE Computer, vol. 26, no. 3, pp.
11–18, 1993, citeseer.nj.nec.com/athanas93processor.html. [Online]. Available:
citeseer.nj.nec.com/athanas93processor.html

[13] Atmel, AT6000 FPGA configuration guide. Atmel Inc.

[14] R. Baeza-Tates, Handbook of Algorithms and Data Structures. Addison-Wessley,

1991.

[15] S. Bajracharya, C. Shu, K. Gaj, and T. El-Ghazawi, “Implementation of elliptic curve

cryptosystems over gf(2n) in optimal normal basis on a reconfigurable computer,” in
FPGA ’04: Proceedings of the 2004 ACM/SIGDA 12th international symposium on
Field programmable gate arrays. New York, NY, USA: ACM Press, 2004, pp. 259–
259.

[16] L. Barroso, K. Gharachorloo, R. McNamara, A. Nowatzyk, S. Qadeer, B. Sano,

S. Smith, R. Stets, and B. Verghese, “Piranha: A scalable architecture based on
single-chip multiprocessing,” 2000. [Online]. Available: http://citeseer.ist.psu.edu/
barroso00piranha.html

[17] V. Baumgarte, G. Ehlers, F. May, A. Nueckel, M. Vorbach, and M. Weinhardt, “PACT

XPP A self-reconfigurable data processing architecture,” J. Supercomput., vol. 26, no. 2,
pp. 167–184, 2003.

[18] K. Bazargan, R. Kastner, and M. Sarrafzadeh, “Fast template placement for reconfig-
urable computing systems,” In IEEE Design and Test - Special Issue on Reconfigurable
Computing, vol. January-March, pp. 68–83, 2000.

[19] M. Bednara, M. Daldrup, J. Shokrollahi, J. Teich, and J. von zur Gathen, “Tradeoff anal-
ysis of fpga based elliptic curve cryptography,” in Proc. of the IEEE International Sym-
posium on Circuits and Systems (ISCAS-02), Scottsdale, Arizona, U.S.A, May 2002.

[20] M. Bednara, M. Daldrup, J. von zur Gathen, J. Shokrollahi, and J. Teich, “Reconfig-
urable implementation of elliptic curve crypto algorithms.” in IPDPS, 2002.

[21] L. Benini and G. Micheli, “Network on chips: A new soc paradigm,” IEEE Computer,
January 2001.

[22] J. L. Bentley, “Multidimensional binary search trees used for associative searching,”
Commun. ACM, vol. 18, no. 9, pp. 509–517, 1975.

[23] J. L. Bentley, “Solutions to Klee’s rectangle problems,” Carnegie-Mellon Univ., Pitts-

burgh, PA, Technical Report, 1977.
REFERENCES 321

[24] J. L. Bentley and D. Wood, “An optimal worst case algorithm for reporting intersections
of rectangles,” IEEE Trans. Comput., vol. C-29, pp. 571–577, 1980.

[25] P. Bertin, D. Roncin, and J. Vuillemin, “Introduction to programmable ac-

tive memories,” in Systolic Array Processors, J. McCanny, J. McWhirter,
and E. Swartzlander, Eds. Prentice Hall, 1989. [Online]. Available:
citeseer.ist.psu.edu/bertin89introduction.html

[26] N. B. Bhat and D. D. Hill, “Routable technologie mapping for lut fpgas,” in ICCD ’92:
Proceedings of the 1991 IEEE International Conference on Computer Design on VLSI
in Computer & Processors. Washington, DC, USA: IEEE Computer Society, 1992,
pp. 95–98.

[27] M. Blum, R. W. Floyd, V. Pratt, R. L. Rivest, and R. E. Tarjan, “Linear time bounds for
median computations,” in Proc. Fourth Annual ACM Symposium on Theory of Comput-
ing, 1972, pp. 119–124.

[28] C. Bobda, “Temporal partitioning and sequencing of dataflow graphs on reconfigurable

systems,” in International IFIP TC10 Stream on Distributed and Parallel Embedded
Systems (DIPES 2002). Montreal, Canada: IFIP, 2002, pp. 185–194.

[29] C. Bobda and N. Steenbock, “A rapid prototyping environment for distributed recon-
figurable systems,” in 13th IEEE International Workshop On Rapid System Prototyp-
ing(RSP’02), Darmstadt Germany. IEEE Computer Society, 2002.

[30] C. Bobda, “Synthesis of dataflow graphs for reconfigurable systems using temporal par-
titioning and temporal placement,” Dissertation, Universität Paderborn, Heinz Nixdorf
Institut, Entwurf Paralleler Systeme, 2003, ? 35,-, ISBN 3-935433-37-9.

[31] C. Bobda, “Coremap: a rapid prototyping environment for distributed reconfigurable

systems,” International Journal of Embedded Systems, vol. 2, no. 3-4, pp. 274 – 290,
2006.

[32] C. Bobda and A. Ahmadinia, “Dynamic interconnection of reconfigurable modules on

reconfigurable devices.” IEEE Design & Test of Computers, vol. 22, no. 5, pp. 443–451,
2005.

[33] C. Bobda, A. Ahmadinia, and J. Teich, “Generation of distributed arithmetic designs

for reconfigurable application.” in ARCS Workshops, 2004, pp. 205–214.

[34] C. Bobda and T. Lehmann, “Efficient building of word recongnizer in fpgas for term-
document matrices construction,” in Field Programmable Logic and Aplications FPL
2000, R. Hartenstein and H. Grünbacher, Eds. Villach, Austria: Springer, 2000, pp.
759–768.

[35] C. Bobda, M. Majer, D. Koch, A. Ahmadinia, and J. Teich, “A dynamic noc approach
for communication in reconfigurable devices,” in Proceedings of International Con-
ference on Field-Programmable Logic and Applications (FPL), ser. Lecture Notes in
Computer Science (LNCS), vol. 3203. Antwerp, Belgium: Springer, Aug. 2004, pp.
1032–1036.

[36] J. A. Boyan and M. L. Littman, “Packet routing in dynamically changing

networks: A reinforcement learning approach,” in Advances in Neural Information
322 RECONFIGURABLE COMPUTING

Processing Systems, J. D. Cowan, G. Tesauro, and J. Alspector, Eds., vol. 6.

Morgan Kaufmann Publishers, Inc., 1994, pp. 671–678. [Online]. Available:
citeseer.ist.psu.edu/boyan94packet.html

[37] G. J. Brebner, “A virtual hardware operating system for the xilinx xc6200,” in FPL ’96:
Proceedings of the 6th International Workshop on Field-Programmable Logic, Smart
Applications, New Paradigms and Compilers. London, UK: Springer-Verlag, 1996,
pp. 327–336.

[38] R. P. Brent and F. T. Luk, “The solution of singular-value and eigen-value problems on
multiprocessor arrays,” SIAM J. Sci. Stat. Comput., vol. 6, no. 1, pp. 69–84, 1985.

[39] D. A. Buel, J. M. Arnold, and W. J. Kleinfelder, Splash 2 FPGAs in a Custom Comput-

ing Machine. IEEE Computer Society Press, Los Alamitos, California, 1996.

[40] C. Bobda, M. Huebner, A. Niyonkuru, B. Blodget, M. Majer, A. Ahmadinia, “Design-

ing partial and dynamically reconfigurable applications on xilinx virtex-ii fpgas using
handelc,” in International Workshop on Field Programmable Logic and Applications
(FPL), Antwerp, Belgium.

[41] J. M. P. Cardoso and H. C. Neto, “An enhance static-list scheduling algorithm for tem-
poral partitioning onto rpus,” in IFIP TC10 WG10.5 10 Int. Conf. on Very Large Scale
Integration(VLSI’99). Lisboa, Portugal: IFIP, 1999, pp. 485 – 496.

[42] J. Castillo, P. Huerta, V. López, and J. I. Martínez, “A secure self-reconfiguring archi-

tecture based on open-source hardware,” reconfig, vol. 0, p. 10, 2005.

[43] RC2000 Development Board, Celoxica Ltd., 2004,

http://www.celoxica.com/products/boards/rc2000.asp.

[44] P. K. Chan, M. D. F. Schlag, and J. Y. Zien, “Spectral-based multi-way fpga partition-

ing,” in FPGA ’95: Proceedings of the 1995 ACM third international symposium on
Field-programmable gate arrays. New York, NY, USA: ACM Press, 1995, pp. 133–
139.

[45] D. Chang and M. Marek-Sadowska, “Partitioning sequential circuits on dynamicaly

reconfigurable fpgas,” in International Symposium on Field Programmable Gate Ar-
rays(FPGA 98). Monterey, California: ACM/SIGDA, 1998, pp. 161 – 167.

[46] K.-C. Chen, J. Cong, Y. Ding, A. B. Kahng, and P. Trajmar, “Dag-map: Graph-based
fpga technology mapping for delay optimization,” IEEE Design and Test of Computers,
vol. 09, no. 3, pp. 7–20, 1992.

[47] R. C. C. Cheung, N. J. Telle, W. Luk, and P. Y. K. Cheung, “Customizable elliptic curve

cryptosystems.” IEEE Trans. VLSI Syst., vol. 13, no. 9, pp. 1048–1059, 2005.

[48] F. Christensen, “A scalable software-defined radio development system,” Xcell Journal,

Oct. 2004.

[49] K. S. Christos H. Papadimitrio, Combinatorial Optimization. Englewood Cliffs, NY:

Prentice Hall, 1982.

[50] M. Ciet, M. Neve, E. Peeters, and J.-J. Quisquater, “Parallel fpga implementation of
rsa with residue number systems - can side-channel threats be avoided? - extended
REFERENCES 323

version,” Cryptology ePrint Archive, Report 2004/187, 2004. [Online]. Available:

http://eprint.iacr.org

[51] p. ClearSpeed Technology, http://www.clearspeed.com/.

[52] W. Cockshott and P. Foulk, “A low-cost text retrieval machine,” IEEE PROCEEDINGS,
vol. 136, no. 4, pp. 271–276, July 1989.

[53] J. Cong and Y. Ding, “Flowmap: an optimal technology mapping algorithm for delay
optimization in lookup-table based fpga designs.” IEEE Trans. on CAD of Integrated
Circuits and Systems, vol. 13, no. 1, pp. 1–12, 1994.

[54] J. Cong and Y. Ding, “Combinational logic synthesis for lut based field programmable
gate arrays,” ACM Trans. Des. Autom. Electron. Syst., vol. 1, no. 2, pp. 145–204, 1996.

[55] T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson, Introduction to Algorithms.

McGraw-Hill Higher Education, 2001.

[56] X. Corp, http://www.xess.com.

[57] S. Craven and P. Athanas, “Examining the viability of fpga supercomputing,” EURASIP
Journal on Embedded Systems, vol. 2007, pp. Article ID 93 652, 8 pages, 2007,
doi:10.1155/2007/93652.

[58] Cray, “Cray xd1 supercomputer,” GNU’s home page. [Online]. Available: http:
//www.cray.com/products/xd1/

[59] I. CRC Computers, “General purpose reconfigurable computing systemsr.” [Online].

Available: http://www.srccomp.com/

[60] W. J. Dally and B. Towles, “Route packets, not wires: on-chip interconnection net-
works,” in Proceedings of the Design Automation Conference, Las Vegas, NV, Jun.
2001, pp. 684–689.

[61] A. Dandalis and V. K. Prasanna, “An adaptive cryptographic engine for internet protocol
security architectures,” ACM Trans. Des. Autom. Electron. Syst., vol. 9, no. 3, pp. 333–
353, 2004.

[62] K. Danne, “Distributed arithmetic FPGA design with online scalable size and perfor-
mance,” in Proceedings of 17th SYMPOSIUM ON INTEGRATED CIRCUITS AND SYS-
TEMS DESIGN (SBCCI04). ACM Press, New York, NY, USA, 7 - 11 Sep. 2004, pp.
135–140.

[63] K. Danne and C. Bobda, “Dynamic reconfiguration of distributed arithmetic controllers:

Design space exploration and trade-off analysis,” ipdps, vol. 04, p. 140a, 2004.

[64] K. Danne, C. Bobda, and H. Kalte, “Run-time exchange of mechatronic controllers

using partial hardware reconfiguration,” in Proc. of the International Conference on
Field Programmable Logic and Applications (FPL2003), Lisbon, Portugal, Sep. 2003.

[65] S. Davidson, D. Landskov, B. Shriver, and P. W. Mallett, “Some experiments in local

microcode compaction for horizontal machines.” IEEE Trans. Computers, vol. 30, no. 7,
pp. 460–477, 1981.
324 RECONFIGURABLE COMPUTING

[66] M. de Berg, M. van Kreveld, M. Overmars, and O. Schwarzkopf, Computational Geom-

etry: Algorithms and Applications, 2nd ed. Berlin, Germany: Springer-Verlag, 2000.

[67] S. Deerwester, S. Dumai, G. Furnas, T. Landauer, and R. Harshmann, “Indexing by

latent semantic analysis,” Journal of American Society for Information Science, vol. 41,
no. 6, pp. 391–407, 1990.

[68] A. DeHon, “DPGA-coupled microprocessors: Commodity ICs for the early 21st
century,” in IEEE Workshop on FPGAs for Custom Computing Machines, D. A. Buell
and K. L. Pocek, Eds. Los Alamitos, CA: IEEE Computer Society Press, 1994, pp.
31–39. [Online]. Available: citeseer.ist.psu.edu/dehon94dpgacoupled.html

[69] J. P. Delahaye, G. Gogniat, C. Roland, and P. Bomel, “Software radio and

dynamic reconfiguration on a dsp/fpga platform,” 2004. [Online]. Available:
http://hal.ccsd.cnrs.fr/ccsd-00089395/en/

[70] Digilent Inc., “The vdec1 video decoder board.” [Online]. Available: http:
//www.digilentinc.com/Products/Detail.cfm?Prod=VDEC1

[71] F. Dittmann, A. Rettberg, and F. Schulte, “A y-chart based tool for reconfigurable sys-
tem design,” in Workshop on Dynamically Reconfigurable Systems (DRS). Innsbruck,
Austria: VDE Verlag, 17 Mar. 2005, pp. 67–73.

[72] P. S. B. do Nascimento, M. E. de Lima, S. M. da Silva, and J. L. Seixas, “Mapping

of image processing systems to fpga computer based on temporal partitioning and de-
sign space exploration,” in SBCCI ’06: Proceedings of the 19th annual symposium on
Integrated circuits and systems design. New York, NY, USA: ACM Press, 2006, pp.
50–55.

[73] W. E. Donath, “Hierarchical placement method,” IBM Technical disclosure Bulletin,

vol. 17, no. 10, pp. 3121–3125, 1975.

[74] W. E. Donath and A. J. Hoffman, “Algorithms for partitioning of graphs and computer
logic based on eigenvectors of connection matrices,” IBM Technical disclosure Bulletin,
vol. 15, no. 3, pp. 938–944, 1972.

[75] S. J. Eggers, J. S. Emer, H. M. Levy, J. L. Lo, R. L. Stamm, and D. M.

Tullsen, “Simultaneous multithreading: A platform for next-generation processors,”
IEEE Micro, vol. 17, no. 5, 1997. [Online]. Available: http://citeseer.ist.psu.edu/
eggers97simultaneous.html

[76] Ejnioui and N. Ranganathan, “Circuit scheduling on time-multiplexed fpgas.” [Online].

Available: citeseer.nj.nec.com/17858.html

[77] H. A. ElGindy, A. K. Somani, H. Schroeder, H. Schmeck, and A. Spray, “RMB - A Re-

configurable Multiple Bus Network,” in Proceedings of the Second International Sym-
posium on High-Performance Computer Architecture (HPCA-2), San Jose, California,
Feb. 1996, pp. 108–117.

[78] M. Ernst, M. Jung, F. Madlener, S. Huss, and R. Bl&#252;mel, “A reconfigurable sys-

tem on chip implementation for elliptic curve cryptography over gf(2n),” in CHES ’02:
Revised Papers from the 4th International Workshop on Cryptographic Hardware and
Embedded Systems. London, UK: Springer-Verlag, 2003, pp. 381–399.
REFERENCES 325

[79] G. Estrin and C. R. Viswanathan, “Organisation of a "fixed-plus-variable" structure

computer for computation of eigenvalues and eigenvectors of real symetric matrices,”
Journal of the ACM, vol. 9, pp. 41–60, 1962.

[80] G. Estrin, B. Bussell, R. Turn, and J. Bibb, “Parallel processing in a restructurable

computer system,” IEEE Transactions on Electronic Computers, vol. 12, no. 5, pp. 747–
755, 1963.

[81] G. Estrin and R. Turn, “Automatic assignment of computations in a variable structure

computer system,” IEEE Transactions on Electronic Computers, vol. 12, no. 5, pp. 755–
773, 1963.

[82] S. P. Fekete, E. Köhler, and J. Teich, “Optimal fpga module placement with temporal
precedence constraints,” Technische Universität Berlin, Tech. Rep. 696.2000, 2000.

[83] C. M. Fiduccia and R. Mattheyses, “A linear-time heuristic for improving network par-
titions,” in Proceedings of the 19th Design Automation Conference, 1982, pp. 175–181.

[84] L. Ford and D. Fulkerson, Flows in Networks. Princeton University Press, 1962.

[85] P. Foulk, “Data-folding in SRAM configurable FPGA,” in IEEE Workshop on FPGAs

for Custom Computing Machines. IEEE, 1993, pp. 163–171.

[86] R. FRANCIS, “Technology mapping for lookup-table based fieldprogrammable ga

te arrays,” Ph.D. dissertation, University of Toronto, 1992. [Online]. Available:
citeseer.ist.psu.edu/francis93technology.html

[87] R. J. Francis, J. Rose, and K. Chung, “Chortle: a technology mapping program for
lookup table-based field programmable gate arrays,” in DAC ’90: Proceedings of the
27th ACM/IEEE conference on Design automation. New York, NY, USA: ACM Press,
1990, pp. 613–619.

[88] J. Fry and M. Langhammer, “Rsa and public key cryptography in fpgas.” [Online].
Available: www.altera.com/literature/cp/rsa-public-key-cryptography.pdf

[89] F. C. Furtek, E. Hogenauer, and J. Scheuermann, “Interconnecting heterogeneous nodes

in an adaptive computing machine.” in Proceedings of International Conference on
Field-P rogrammable Logic and Applications (FPL), ser. Lecture Notes in Computer
Science (LNCS), vol. 3203. Antwep, Belgium: Springer, Aug. 2004, pp. 125–134.

[90] K. Gaj and P. Chodowiec, “Comparison of the hardware performance of the aes candi-
dates using reconfigurable hardware.” in AES Candidate Conference, 2000, pp. 40–54.

[91] D. D. Gajski and L. Ramachandran, “Introduction to high-level synthesis,” IEEE Des.

Test, vol. 11, no. 4, pp. 44–54, 1994.

[92] D. D. Gajski and M. Reshadio, “Application and advantages,” CECS, UC Irvin, Tech-
nical Report 04-12, 2004.

[93] C. H. Gebotys, “An optimal methodology of synthesis of dsp multichip architectures,”

Journal of VLSI Signal Processing, vol. 11, pp. 9–19, 1995.

[94] L. Geppert, “Sun’s big splash,” IEEE Spectrum Magazine, pp. 21–29, M, January 2005
2005.
326 RECONFIGURABLE COMPUTING

[95] J. Gerling, K. Danne, C. Bobda, and J. Schrage, “Distributed arithmetics for recursive
convolution of optical intercannects,” in EOS Topical Meeting, Optics in Computing
(OIC), Engelberg (Switzerland), Apr. 2004, pp. 65–66.

[96] P. B. Gibbons and S. S. Muchnick, “Efficient instruction scheduling for a pipelined ar-
chitecture,” in SIGPLAN ’86: Proceedings of the 1986 SIGPLAN symposium on Com-
piler construction. New York, NY, USA: ACM Press, 1986, pp. 11–16.

[97] M. Gokhale, W. Holmes, A. Kopser, D. Kunze, D. P. Lopresti, S. Lucas, R. Minnich,

and P. Olsen, “Splash: A reconfigurable linear logic array.” in ICPP (1), 1990, pp. 526–
532.

[98] S. C. Goldstein, H. Schmit, M. Moe, M. Budiu, S. Cadambi, R. R. Taylor, and

R. Laufer, “Piperench: A coprocessor for streaming multimedia acceleration,” in ISCA,
1999, pp. 28–39. [Online]. Available: citeseer.ist.psu.edu/goldstein99piperench.html

[99] G. H. Golub and C. F. V. Loan, Matrix Computations. North Oxford Academic Pub-
lisching, 1983.

[100] C. E. D. C. Green and P. Franklin, “RaPiD – reconfigurable pipelined datapath,”

in Field-Programmable Logic: Smart Applications, New Paradigms, and Compilers.
6th International Workshop on Field-Programmable Logic and Applications, R. W.
Hartenstein and M. Glesner, Eds. Darmstadt, Germany: Springer-Verlag, 1996, pp.
126–135. [Online]. Available: citeseer.ist.psu.edu/ebeling96rapid.html

[101] S. Guccione, “Programming fine-grained reconfigurable architechture,” Ph.D. disserta-

tion, The University of Texas at Austin, May 1995.

[102] S. Guccione, D. Levi, and P. Sundararajan, “Jbits: A java-based interface for reconfig-
urable computing,” 1999.

[103] M. Gulati and N. Bagherzadeh, “Performance study of a multithreaded superscalar mi-

croprocessor,” in Proc. Intl. Symp. High-Performance Computer Architecture, ACM,
1996, pp. 291–301.

[104] B. Gunther and G. Milne, “Accessing document relevance with run-time reconfigurable
machines,” in IEEE Workshop on FPGAs for Custom Computing Machines. Napa
California: IEEE, 1996, pp. 9–16.

[105] B. Gunther and G. Milne, “Hardware-based solution for message filtering,” school of
computer and information science, Tech. Rep., 1996.

[106] R. H. Güting, “An optimal contour algorithm for iso-oriented rectangles,” J. Algorithms,
vol. 5, pp. 303–326, 1984.

[107] K. Hall, “An r-dimensional quadratic dimensional quadratic placement algorithm,”

Journal of Management Science, vol. 17, no. 3, pp. 219–229, 1970.

[108] T. Haller, “Adaptive multiprocessing on reconfigurable chip,” 2006, master Thesis.

[109] L. Hammond, B. A. Nayfeh, and K. Olukotun, “The hydra chip,” IEEE

MICRO Magazine, pp. 71–83, March-April 2000. [Online]. Available: http:
//citeseer.ist.psu.edu/287939.html
REFERENCES 327

[110] R. Hartenstein, “A decade of reconfigurable computing: A visionary Retrospective,” in

Proceedings of the International Conference on Design, Automation and Test in Europe
(DATE). IEEE Computer Society, March 2001, pp. 290–295.

[111] R. Hartenstein, Morphware and Configware, A. Y. Zomaya, Ed. New York: Springer-
Verlag, 2006.

[112] R. Hartenstein, Basics of Reconfigurable Computing, S. P. J. Henkel, Ed. New York:

Springer-Verlag, 2007.

[113] R. Hartenstein, A. Hirschbiel, and M.Weber, “Xputers - an open family of non von neu-
mann architectures,” in 11th ITG/GI Conference on Architektur von Rechensystemen.
VDE-Verlag, 1990.

[114] J. R. Hauser and J. Wawrzynek, “Garp: A MIPS processor with a reconfigurable

coprocessor,” in IEEE Symposium on FPGAs for Custom Computing Machines, K. L.
Pocek and J. Arnold, Eds. Los Alamitos, CA: IEEE Computer Society Press, 1997,
pp. 12–21. [Online]. Available: citeseer.nj.nec.com/hauser97garp.html

[115] S. D. Haynes, J. Stone, P. Y. K. Cheung, and W. Luk, “Video image processing with the
sonic architecture,” Computer, vol. 33, no. 4, pp. 50–57, 2000.

[116] P. Healy and M. Creavin, “An Optimal Algorithm for Rectangle Placement,” Dept. of
Computer Science and Information Systems, University of Limerick, Limerick, Ireland,
Tech. Rep. UL-CSIS-97-1, Aug. 1997.

[117] A. Hemani, A. Jantsch, S. Kumar, A. Postula, J. Oberg, M. Millberg, and D. Lindqvist,

“Network on chip: An architecture for billion transistor era,” in In Proceeding of the
IEEE NorChip Conference, November 2000., 2000.

[118] B. Hendrickson and R. Leland, “An improved spectral graph partitioning algorithm for
mapping parallel computations,” SIAM Journal on Scientific Computing, vol. 16, no. 2,
pp. 452–469, 1995. [Online]. Available: citeseer.nj.nec.com/hendrickson95improved.
html

[119] M. R. Hestenes, “Inversion of matrices by biorthogonalization and related results,” J.

Soc. Indust. Appl. Math., vol. 6, no. 1, pp. 51–90, 1958.

[120] Hirata et al., “An elementary processor architecture with simultaneous instruction is-
suing from multiple threads,” in Proc. Intl. Symp. Computer Architecture, Assoc. of
Computing Machinery, 1992, pp. 136–145.

[121] G. home page, “Gnu radio - the gnu software radio,”

http://www.gnu.org/software/gnuradio/. [Online]. Available: http://www.gnu.org/
software/gnuradio/

[122] J. Huie, P. DŠAntonio, R. Pelt, and B. Jentz, “Synthesizing fpga cores for software-
defined radio.” [Online]. Available: www.altera.com/literature/cp/fpga-cores-for-sdr.
pdf

[123] S. Iman, M. Pedram, C. Fabian, and J. Cong, “Finding uni-directional cuts based on
physical partitioning and logic restructuring,” in Fourth International Workshop on
Physical Design. IEEE, 1993.
328 RECONFIGURABLE COMPUTING

[124] A. Inc, Nios II Processor Reference Handbook, November, 2006. [Online]. Available:
http://www.altera.com/literature/lit-nio2.jsp

[125] C. inc, Handel-C Reference Manual, 2000, www.celoxica.com/techlib/files/CEL-

W0410251JJ4-60.pdf. [Online]. Available: www.celoxica.com/techlib/files/
CEL-W0410251JJ4-60.pdf

[126] X. Inc, PowerPC 405 Processor Block Reference Guide: Embedded Development Kit,
July 20, 2005. [Online]. Available: www.xilinx.com/bvdocs/userguides/ug018.pdf

[127] X. Inc, MicroBlaze Processor Reference Guide: Embedded Development Kit EDK 8.2i,
June 1, 2006. [Online]. Available: www.xilinx.com/ise/embedded/mb ref guide.pdf

[128] X. Inc, XAPP290: Two Flows for Partial Reconfiguration: Module Based or Difference
Based, September 9, 2004. [Online]. Available: www.xilinx.com/bvdocs/appnotes/
xapp290.pdf

[129] J. M. Joseph Mitola, Software Radio Architecture: Object-Oriented Approaches to

Wireless Systems Engineering. Wiley, John & Sons, Incorporated, 2001.

[130] S. Jung, “Entwurf eines verfahrens und einer umgebung zur durchgängigen konfigura-
tion adativer on-chip multiprozessor,” 2006, master Thesis.

[131] Jürgen Reichardt, Bernd Schwarz, VHDL-Synthese, 3rd ed. Oldenbourg-Verlag, Dec
2003. [Online]. Available: http://users.etech.fh-hamburg.de/users/reichardt/buch.html

[132] H. Kalte, M. Pormann, and U. Rueckert, “Rapid prototyping system f"ur dynamisch
rekonfigurierbarer hardware struckturen,” in AES 2000, 2000, pp. 149–157.

[133] K. Karplus, “Xmap: A technology mapper for table-lookup field-programmable gate

arrays,” in DAC ’91: Proceedings of the 28th conference on ACM/IEEE design automa-
tion. New York, NY, USA: ACM Press, 1991, pp. 240–243.

[134] R. Kastner, A. Kaplan, and M. Sarrafzadeh, Synthesis Techniques and Optimizations

for Reconfigurable Systems. Amsterdam: Kluwer Academic Publishers, 2003.

[135] M. Kaul and R. Vemuri, “Optimal temporal partitioning and synthesis for reconfig-
urable architectures,” 1998.

[136] M. Kaul, R. Vemuri, S. Govindarajan, and I. Ouaiss, “An automated temporal

partitioning tool for a class of dsp applications,” 1998. [Online]. Available:
citeseer.nj.nec.com/160625.html

[137] B. W. Kernighan and S. Lin, “An efficient heuristic procedure for partitioning graphs,”
The Bell System Technical Journal, vol. 49, no. 2, pp. 291–307, 1970.

[138] N. KOBLITZ, “Elliptic curve cryptosystems,” Mathematics of Computation, vol. 48,

pp. 203–209, 1987.

[139] P. Kongetira, K. Aingaran, and K. Olukotun, “The hydra chip,” IEEE MICRO
Magazine, pp. 21–29, March-April 2005. [Online]. Available: http://citeseer.ist.psu.
edu/287939.html
REFERENCES 329

[140] H. Kropp, C. Reuter, M. Wiege, T.-T. Do, and P. Pirsch, “An fpga-based prototyping
system for real-time verification of video processing schemes.” in FPL, 1999, pp. 333–
338.

[141] K. Kucukcakar and A. Parker, “Chop: A constraint-driven-system-level partitioner,” in

Design Automation Conference, 1991, pp. 514–519.

[142] P. Kurup and T. Abbasi, Logic Synthesis Using Synopsys. Kluwer Akademic publisher,
1997.

[143] D. Landskov, S. Davidson, B. Shriver, and P. W. Mallett, “Local microcode compaction

techniques,” ACM Comput. Surv., vol. 12, no. 3, pp. 261–294, 1980.

[144] T. Lengauer, Combinatorial Algorithm for Integrated Circuit Layout. Teubner

Stuttgart, 1990.

[145] K. H. Leung, K. W. Ma, W. K. Wong, and P. H. W. Leong, “Fpga implementation

of a microcoded elliptic curve cryptographic processor,” in FCCM ’00: Proceedings
of the 2000 IEEE Symposium on Field-Programmable Custom Computing Machines.
Washington, DC, USA: IEEE Computer Society, 2000, p. 68.

[146] H. Li and Q. Stout, Reconfigurable Massively Parallel computers, H. Li and Q. Stout,

Eds. Glasgow, UK: Prentice-Hall, 1991.

[147] W. Lipski, Jr. and F. P. Preparata, “Finding the contour of a union of iso-oriented rect-
angles,” J. Algorithms, vol. 1, pp. 235–246, 1980, errata in 2(1981), 105; corrigendum
in 3(1982), 301–302.

[148] H. Liu and D. F. Wong, “Network flow-based circuit partitioning for time-multiplexed
FPGAs,” in IEEE/ACM International Conference on Computer-Aided Design, 1998,
pp. 497–504.

[149] H. Liu and D. F. Wong, “Circuit partitioning for dynamicaly reconfigurable fpgas,” in
International Symposium on Field Programmable Gate Arrays(FPGA 98). Monterey,
California: ACM/SIGDA, 1999, pp. 187 – 194.

[150] A. S. Ltd., High Performance AES Encryption Cores. [Online]. Available:

http://http://www.jat.co.kr/newsletter/2001 11 24/ds5210-40.pdf

[151] S.-M. Ludwig, “Hades-fast hardware synthesis tools and a reconfigurable coprocessor,”
Ph.D. dissertation, Swiss Federal Institute of Technologie, Zürich, 1997.

[152] J. Lutz and A. Hasan, “High performance fpga based elliptic curve cryptographic co-
processor,” in ITCC ’04: Proceedings of the International Conference on Information
Technology: Coding and Computing (ITCC’04) Volume 2. Washington, DC, USA:
IEEE Computer Society, 2004, p. 486.

[153] P. S. G. Maya Gokhale, Reconfigurable Computing: Accelerating Computation with

Field-programmable Gate Arrays. Berlin: Springer, 2005.

[154] A. Mazzeo, L. Romano, G. P. Saggese, and N. Mazzocca, “Fpga-based implementation

of a serial rsa processor,” in DATE ’03: Proceedings of the conference on Design, Au-
tomation and Test in Europe. Washington, DC, USA: IEEE Computer Society, 2003,
p. 10582.
330 RECONFIGURABLE COMPUTING

[155] A. Michalski and D. Buell, “A scalable architecture for rsa cryptography on large
fpgas,” in FCCM ’06: Proceedings of the 14th Annual IEEE Symposium on Field-
Programmable Custom Computing Machines (FCCM’06). Washington, DC, USA:
IEEE Computer Society, 2006, pp. 331–332.

[156] G. D. Micheli, Synthesis and Optimization of Digital Circuits. McGraw-Hill Higher

Education, 1994.

[157] V. S. Miller, “Use of elliptic curves in cryptography,” in Lecture notes in computer sci-
ences; 218 on Advances in cryptology—CRYPTO 85. New York, NY, USA: Springer-
Verlag New York, Inc., 1986, pp. 417–426.

[158] L. Minzer, “Programmable silicon for embedde signal processing,” Embedded Systems
Programming, pp. 110–133, March 2000.

[159] E. Mirsky and A. DeHon, “MATRIX: A reconfigurable computing architecture with

configurable instruction distribution and deployable resources,” in IEEE Symposium
on FPGAs for Custom Computing Machines, K. L. Pocek and J. Arnold, Eds. Los
Alamitos, CA: IEEE Computer Society Press, 1996, pp. 157–166. [Online]. Available:
citeseer.ist.psu.edu/mirsky96matrix.html

[160] T. Miyamori and K. Olukotun, “REMARC: Reconfigurable multimedia array

coprocessor (abstract),” in FPGA, 1998, p. 261. [Online]. Available: citeseer.ist.psu.
edu/miyamori98remarc.html

[161] A. Morse, “Control using logic-based switching,” Trends in Control, Springer, London,
1995.

[162] M. C.-G. H. Mr. Adam Harrington and D. S. S. Jones, “Software-defined radio: The
revolution of wireless communication,” Annual Review of Communications, vol. 58,
2005.

[163] R. Murgai, Y. Nishizaki, N. Shenoy, R. K. Brayton, and A. Sangiovanni-Vincentelli,

“Logic synthesis for programmable gate arrays,” in DAC ’90: Proceedings of the 27th
ACM/IEEE conference on Design automation. New York, NY, USA: ACM Press,
1990, pp. 620–625.

[164] R. Murgai, N. V. Shenoy, R. K. Brayton, and A. L. Sangiovanni-Vincentelli, “Perfor-

mance directed synthesis for table look up programmable gate arrays.” in ICCAD, 1991,
pp. 572–575.

[165] I. Nallatech, http://www.nallatech.com.

[166] Narendra and Balakrishnan, “Adaptive control using multiple models: Switching and
tuning,” Yale Workshop on Adaptive and Learning Systems, 1994.

[167] M. Nikitovic and M. Brorsson, “An adaptive chip-multiprocessor architecture for future
mobile terminals.” in CASES, 2002, pp. 43–49.

[168] A. Nohl, G. Braun, O. Schliebusch, R. Leupers, H. Meyr, and A. Hoffmann, “A uni-

versal technique for fast and flexible instruction-set architecture simulation,” in DAC
’02: Proceedings of the 39th conference on Design automation. New York, NY, USA:
ACM Press, 2002, pp. 22–27.
REFERENCES 331

[169] K. Olukotun and L. Hammond, “The future of microprocessors,” ACM Queue, vol. 3,
no. 7, pp. 26–29, September 2005. [Online]. Available: http://doi.acm.org/10.1145/
1095408.1095418

[170] OSCI, SystemC Reference Manual, 2003. [Online]. Available: www.systemc.org

[171] I. Ouaiss, S. Govindarajan, V. Srinivasan, M. Kaul, and R. Vemuri, “An integrated

partitioning and synthesis system for dynamically reconfigurable multi-FPGA
architectures,” in IPPS/SPDP Workshops, 1998, pp. 31–36. [Online]. Available:
citeseer.nj.nec.com/ouaiss98integrated.html

[172] Y. Pan and M. Hamdi, “Singular value decomposition on processors arrays with a
pipelined bus system,” Journal of Network and Computer Applications, vol. 19, pp.
235–248, 1996.

[173] A. Pandey and R. Vemuri, “Combined temporal partitioning and scheduling for recon-
figurable architectures,” in Reconfigurable Technology: FPGAs for Computing and Ap-
plications, Proc. SPIE 3844, J. Schewel, P. M. Athanas, S. A. Guccione, S. Ludwig, and
J. T. McHenry, Eds. Bellingham, WA: SPIE – The International Society for Optical
Engineering, 1999, pp. 93–103.

[174] P. G. Paulin and J. P. Knight, “Force-directed scheduling for the behavioral synthesis of
asic’s.” IEEE Transactions on CAD, vol. 6, no. 8, pp. 661–679, 1989.

[175] D. Pellerin and S. Thibault, Practical FPGA Programming in C. Prentice Hall„ April
2005.

[176] M. Petrov, T. Murgan, F. May, M. Vorbach, P. Zipf, and M. Glesner, “The XPP ar-
chitecture and its co-simulation within the simulink environment.” in Proceedings of
International Conference on Field-P rogrammable Logic and Applications (FPL), ser.
Lecture Notes in Computer Science (LNCS), vol. 3203. Antwep, Belgium: Springer,
Aug. 2004, pp. 761–770.

[177] M. Platzner and L. Thiele, “XFORCES - executives for reconfigurable embedded

systems.” [Online]. Available: http://www.ee.ethz.ch/\∼platzner

[178] V. K. Prasanna and A. Dandalis, “Fpga-based cryptography for internet security.”

[Online]. Available: halcyon.usc.edu/∼pk/prasannawebsite/papers/dandalisOSEE00.
pdf

[179] F. P. Preparata and M. I. Shamos, Computational Geometry: An Introduction. New

York, NY: Springer-Verlag, 1985.

[180] D. Pryor, M.R.Thisle, and N.shirazi, “Text searching on splash 2,” in IEEE Workshop
on FPGAs for Custom Computing Machines. IEEE, 1993, pp. 172–177.

[181] K. M. G. Purna and D. Bhatia, “Temporal partitioning and scheduling data flow graphs
for reconfigurable computers,” IEEE Transactions on Computers, vol. 48, no. 6, pp.
579–590, 1999.

[182] F. J. Rammig, “A concept for the editing of hardware resulting in an automatic

hardware-editor,” in Proceedings of 14th Design Automation Conference, New Orleans,
1977, pp. 187–193.
332 RECONFIGURABLE COMPUTING

[183] D. Rech, “Automatische generierung und konfiguration von adaptiven on-chip symmet-
rical multiprocessing (smp) systemen,” 2006, master Thesis.

[184] E. Rijpkema, K. G. W. Goossens, A. Radulescu, J. Dielissen, J. van Meerbergen,

P. Wielage, and E. Waterlander, “Trade offs in the design of a router with both guar-
anteed and best-effort services for networks on chip,” in DATE ’03: Proceedings of the
conference on Design, Automation and Test in Europe. Washington, DC, USA: IEEE
Computer Society, 2003, p. 10350.

[185] R. L. Rivest, A. Shamir, and L. Adleman, “A method for obtaining digital signatures and
public-key cryptosystems,” Communications of the ACM, vol. 21, no. 2, pp. 120–126,
1978.

[186] R. L. Rivest and C. E. Leiserson, Introduction to Algorithms. New York, NY, USA:
McGraw-Hill, Inc., 1990.

[187] C. Rowen and S. Leibson, “Flexible architectures for engineering successful socs.” in
DAC, 2004, pp. 692–697.

[188] S. M. Scalera and J. R. Vázquez, “The design and implementation of a context switch-
ing fpga,” in IEEE Symposium on FPGAs for Custom Computing Machines. Napa
Valey, CA: IEEE Computer Society Press, April 1998, pp. 78–85.

[189] SDR, “The sdr forum’s website.” [Online]. Available: http://www.gnu.org/software/

gnuradio/

[190] P. Sedcole, “Reconfigurable platform-based design in fpgas for video image process-
ing,” Ph.D. dissertation, Imperial College, London, January 2006.

[191] SGI, http://www.sgi.com/products/rasc/.

[192] N. Shirazi, A. Walters, and P. M. Athanas, “Quantitative analysis of floating point arith-
metic on fpga based custom computing machines.” in FCCM, 1995, pp. 155–163.

[193] R. G. Shoup, “Parameterized convolution filtering in an FPGA,” in Proceedings of In-

ternational Conference on Field Programmable Logic and Arrays, 1993, pp. 274–280.

[194] H. Singh, M.-H. Lee, G. Lu, F. J. Kurdahi, N. Bagherzadeh, and E. M. C. Filho,

“Morphosys: An integrated reconfigurable system for data-parallel and computation-
intensive applications,” IEEE Transactions on Computers, vol. 49, no. 5, pp. 465–481,
2000.

[195] M. J. S. Smith, Application-specific integrated circuits. Addison-Wesley Longman

Publishing Co., Inc., 1997.

[196] C. Steiger, “Operating systems for reconfigurable embedded platforms: Online schedul-
ing of real-time tasks,” IEEE Trans. Comput., vol. 53, no. 11, pp. 1393–1407, 2004,
member-Herbert Walder and Member-Marco Platzner.

[197] C. Steiger, H. Walder, M. Platzner, and L. Thiele, “Online Scheduling and Placement
of Real-time Tasks to Partially Reconfigurable Devices,” in Proceedings of the 24th
International Real-Time Systems Symposium (RTSS’03), December 2003.
REFERENCES 333

[198] M. A. Tahir, A. Bouridane, and F. Kurugollu, “An fpga based coprocessor for glcm and
haralick texture features and their application in prostate cancer classification,” Analog
Integr. Circuits Signal Process., vol. 43, no. 2, pp. 205–215, 2005.

[199] J. Teich, S. P. Fekete, and J. Schepers, “Optimizing dynamic hardware reconfiguration,”

Angewante Mathematic Und Informatik Universität zu Köln, Tech. Rep. 97.228, 1998.

[200] G. R. G. S. J. Thomas, “An optimal parallel jacobi-like solution method for the singular
value decomposition,” in Proc. Int. Conf. on Parallel Processing, January 1988.

[201] C. Torres-Huitzil and M. Arias-Estrada, “Fpga-based configurable systolic architecture

for window-based image processing,” EURASIP Journal on Applied Signal Processing,
vol. 2005, no. 7, pp. 1024–1034, 2005, doi:10.1155/ASP.2005.1024.

[202] S. Trimberger, “Scheduling designs into a time-multiplexed fpga,” in International

Symposium on Field Programmable Gate Arrays(FPGA 98). Monterey, California:
ACM/SIGDA, 1998, pp. 153 – 160.

[203] F. Vahid and T. Givargis, Embedded System Design: A Unified Hardware/Software In-
troduction. New York, NY, USA: John Wiley & Sons, Inc., 2001.

[204] R. Vaidyanathan and J. L. Trahan, Dynamic Reconfiguration: Architectures and Algo-

rithms. IEEE Computer Society, 2003.

[205] R. Vaidyanathan and J. L. Trahan, Dynamic Reconfiguration: Architectures and Algo-

rithms (Series in Computer Science (Kluwer Academic/Plenum Publishers).). Plenum
Publishing Co., 2004.

[206] A. van Breemen and T. de Vries, “An agent-based framework for designing multi-
controller systems,” Proc. of the Fifth International Conference on The Practical Ap-
plications of Intelligent Agents and Multi-Agent Technology, pp. 219-235, Manchester,
U.K, Apr. 2000.

[207] C. Veciana-Nogu&#233;s and J. Domingo-Pascual, “Adaptive video on demand service

on rsvp capable network,” in ECMAST ’99: Proceedings of the 4th European Confer-
ence on Multimedia Applications, Services and Techniques. London, UK: Springer-
Verlag, 1999, pp. 212–228.

[208] “Vhdl online.” [Online]. Available: http://www.vhdl-online.de/

[209] J. E. Volder, “The birth of cordic,” J. VLSI Signal Process. Syst., vol. 25, no. 2, pp.
101–105, 2000.

[210] E. Waingold, M. Taylor, D. Srikrishna, V. Sarkar, W. Lee, V. Lee, J. Kim, M. Frank,

P. Finch, R. Barua, J. Babb, S. Amarasinghe, and A. Agarwal, “Baring it all to
software: Raw machines,” Computer, vol. 30, no. 9, pp. 86–93, 1997. [Online].
Available: citeseer.ist.psu.edu/waingold97baring.html

[211] H. Walder, S. Nobs, and M. Platzner, “Xf-board: A prototyping platform for reconfig-
urable hardware operating systems.” in ERSA, 2004, p. 306.

[212] H. Walder and M. Platzner, “A runtime environment for reconfigurable hardware oper-
ating systems.” in FPL, 2004, pp. 831–835.
334 RECONFIGURABLE COMPUTING

[213] D. W. Wall, “Limits of instruction-level parallelism,” in Proceedings of the 4th

International Conference on Architectural Support for Programming Languages and
Operating System (ASPLOS), vol. 26, no. 4. New York, NY: ACM Press, 1991, pp.
176–189. [Online]. Available: http://citeseer.ist.psu.edu/wall90limits.html

[214] K. Weiss, T. Steckstor, C. Ötker, I. Katchan, C. Nitsch, and J. Philipp, Spyder - Virtex -
X2 User’s Manual, 1999.

[215] S. A. White, “Application of distributed arithmetic to digital signal processing: A tuto-

rial review,” IEEE ASSP Magazine, pp. 4–19, July 1989.

[216] G. Wigley and D. Kearney, “The development of an operating system for reconfigurable
computing,” in Proceedings of the 9th IEEE Symposium Field-Programmable Custom
Computing Machines(FCCM’01). IEEE-CS Press, April 2001.

[217] S. E. G. William Lehr, Fuencisla Merino, “Software radio: Implications for wireless
services, industry structure, and public policy,” Massachusetts Institute of Technology,
Program on Internet and Telecoms Convergence, Tech. Rep., Aug. 2002.

[218] J. W. Williams and N. Bergmann, “Embedded linux as a platform for dynamically

self-reconfiguring systems-on-chip,” in Proceedings of the International Conference
on Engineering of Reconfigurable Systems and Algorithms, June 2004, pp. 163–169.
[Online]. Available: http://espace.library.uq.edu.au/view.php?pid=UQ:410

[219] M. Wirthlin and B. Hutchings, “A dynamic instruction set computer,” in IEEE

Symposium on FPGAs for Custom Computing Machines, P. Athanas and K. L. Pocek,
Eds. Los Alamitos, CA: IEEE Computer Society Press, 1995, pp. 99–107. [Online].
Available: citeseer.nj.nec.com/wirthlin95dynamic.html

[220] P. L. E. Wolfgang Rosenstiel, New Algorithms. Architectures and Applications for Re-
configurable Computing. Berlin: Springer, 2005.

[221] T. Wollinger, J. Guajardo, and C. Paar, “Security on fpgas: State-of-the-art implementa-

tions and attacks,” Trans. on Embedded Computing Systems, vol. 3, no. 3, pp. 534–574,
2004.

[222] T. J. Wollinger, M. Wang, J. Guajardo, and C. Paar, “How well are high-end dsps suited
for the aes algorithms? aes algorithms on the tms320c6x dsp.” in AES Candidate Con-
ference, 2000, pp. 94–105.

[223] Xilinx, “A guide to using field programmable gate arrays (fpgas) for application-specific
digital signal processing performance,” http://www.xilinx.com, 1995.

[224] Xilinx, “The role of distributed arithmetic design in fpga-based signal processing,”
http://www.xilinx.com, 2000.

[225] Xilinx, “Spartan-3 fpgas,” http://www.xilinx.com, 2000.

[226] Xilinx Inc., “The early access partial reconfiguration lounge,” (registration required).
[Online]. Available: http://www.xilinx.com/support/prealounge/protected/index.htm

[227] Xilinx Inc., “Early access partial reconfiguration user guide,” xilinx User Guide
UG208, Version 1.1, March 6, 2006. [Online]. Available: http://www.xilinx.com/
bvdocs/appnotes/xapp290.pdf
REFERENCES 335

[228] Xilinx Inc., “Virtex-II Pro XUP Development Board,”

http://www.xilinx.com/univ/xupv2p.html.

[229] Xilinx Inc., “Xilinx ISE 8 Software Manuals and Help - PDF Collection,” 2005.
[Online]. Available: http://toolbox.xilinx.com/docsan/xilinx8/books/manuals.pdf

[230] Xilinx Inc., “Edk platform studio documentation,” 2007. [Online]. Available:
http://www.xilinx.com/ise/embedded/edk docs.htm

[231] H. Yang and D. F. Wong, “Efficient network flow based min-cut balanced partitioning,”
in International Conference on Computer-Aided Design, 1994.

[232] Y. Yang, Z. Abid, and W. Wang, “Two-prime rsa immune cryptosystem and its fpga im-
plementation,” in GLSVSLI ’05: Proceedings of the 15th ACM Great Lakes symposium
on VLSI. New York, NY, USA: ACM Press, 2005, pp. 164–167.
Appendix A
Hints to Labs

This chapter gives a step-by-step guide in note form to create a partially reconfigurable
system. It has been developed while reconstructing the Video8 example and is intended to
give hints to the reader on how to create own designs. A more detailed description of the
demonstration project can be found in 5.2 on page 236. The sources of the guide can be found
on the book’s web page. Following the directory and entity names refer to the Video8 project.
A profound knowledge of VHDL, ISE and EDK will be needed to comprehend the instructions.
338 RECONFIGURABLE COMPUTING

Tutorial
Creation of partially reconfigurable designs
On the Example of Video8
Version 1.0

1. Prerequisites
Following basic conditions should be present when accomplishing this guide. In case of
deviations the given approach might not be feasible.
ISE 8.1.01i PR8 or PR12 (Early Access Partial Reconfiguration patch for ISE)
EDK 8.1.02i

Abbreviations:
PR partially reconfigurable
PRM partially reconfigurable module
TL top-level
TLM top-level module
DIR directory
<base> root directory of the project

2. Reorganization of the Project Video8 non pr

1. elevate PRM (rgbfilter) to the TLM (top.vhd): initially rgbfilter is instantiated in
entity video in. But PRM may not be sub-modules of a static part. This change now is
only the the first step. rgbfilter will have to be brought up to completely to the top-level
module.

1.1. In video in.vhd: additional ports have to be adjoined to the entity declaration of
video in so that rgbfilter can be instantiated outside this entity but the rest may
stay the same as before.

entity video_in is {
...
LLC_CLOCK_to_filter : out std_logic;
R_out_to_filter : out std_logic_vector(0 to 9);
G_out_to_filter : out std_logic_vector(0 to 9);
B_out_to_filter : out std_logic_vector(0 to 9);
h_counter_out_to_filter : out std_logic_vector(0 to 9);
v_counter_out_to_filter : out std_logic_vector(0 to 9);
valid_out_to_filter : out std_logic;
R_in_from_filter : in std_logic_vector(0 to 9);
G_in_from_filter : in std_logic_vector(0 to 9);
B_in_from_filter : in std_logic_vector(0 to 9);
h_counter_in_from_filter : in std_logic_vector(0 to 9);
v_counter_in_from_filter : in std_logic_vector(0 to 9);
valid_in_from_filter : in std_logic
...
Appendix A: Hints to Labs 339

};

1.2. In video in.vhd: modules that formerly accessed the rgbfilter instance have to
redirected to outward ports. E.g. simply by rewriting corresponding signals:

-- signals out to the filter:

R_out_to_filter <= R1;
G_out_to_filter <= G1;
B_out_to_filter <= B1;
h_counter_out_to_filter <= h_counter1;
v_counter_out_to_filter <= v_counter1;
valid_out_to_filter <= valid1;
-- processed signals in from the filter:
R2 <= R_in_from_filter;
G2 <= G_in_from_filter;
B2 <= B_in_from_filter;
h_counter2 <= h_counter_in_from_filter;
v_counter2 <= v_counter_in_from_filter;
valid2 <= valid_in_from_filter;

2. making the PRM available to EDK (not necessary; good for testing purposes)

2.1. adapt .mpd and .pao (PCORE-definition file und -synthesizing directives) for the
video in-module:
create .mpd from .vhd with:
psfutil -hdl2mpd video_in_pr.vhd -bus opb m \\
-lang vhdl -o VIDEO_IN_PR_v2_1_0.mpd

In .pao: delete rgbfilter since it’s a PRM. It’ll be synthesized seperately.

lib VIDEO_IN_v1_00_a rgbfilter

löschen.
rename video in to video in pr
2.2. Create an own PCORE for rgbfilter:
copy video in
fit .pao:
lib RGBFILTER_v1_00_a rgbfilter

use only rgbfilter.vhd and create .mpd from it

2.3. rename VIDEO IN v1 00 a to VIDEO IN PR v1 00 a
2.4. include Video in pr and rgbfilter in the EDK project. The custom PCOREs
can be found under ’IP Catalog’, ’Project Repository’. Erase the PCORE video in
from the EDK project. Now wire the ports: rgbfilter receives several signals from
video in. video in takes in the results from rbgfilter.
2.5. Attention has to be paid to the correct address ranges of opb2plb-bridge :
0x0 - 0x0fffffff
The modules accessible over these addresses in this block will be supplied to the bus
340 RECONFIGURABLE COMPUTING

participants on every side of the bridge. This is important because the data flows from
video in to the frame buffer situated in the RAM. (→ Memory-Mapped I/O)
2.6. Synthesis in EDK (for testing). If errors occur like
ERROR:HDLParsers:3317 - "D:/murr/edkProjects/ \\
Video8_81_PR/hdl/video_in_pr_0_wrapper.vhd" Line 10.
Library VIDEO_IN_PR_v1_00_a cannot be found.

ERROR:HDLParsers:3014 - "D:/murr/edkProjects/ \\
Video8_81_PR/hdl/video_in_pr_0_wrapper.vhd" Line 11.
Library unit VIDEO_IN_PR_v1_00_a \\
is not available in library work.
then the .pao files have not been altered correctly as stated above (e.g. not all oc-
curences of video in have been changed to video in pr)
2.7. add additional filters if desired (e.g. a mean-value filter)
2.8. place the address of the frame buffer at the end of the DDR-RAMs (at address
0x00001111111000000000000000000000) to allow the Linux System to start from
0x0
→ adapt files, since Video8 unfortunately received hard-coded addresses from its au-
thor:
In rgbstream2framebuffer.vhd:
next_wb_counter <= ram_out(24 to 33) + \\
(ram_out(34 to 43) * "00000000001010000000") ;

changes to:
next_wb_counter <= ram_out(24 to 33) + \\
(ram_out(34 to 43) * "00000000001010000000") + \\
"000011111110000000000000000000" ;

adapt simple timer unit.vhd:

reader_addr_plb_unsynced <= (others => ’0’);

changes twice to:

reader_addr_plb_unsynced <= \\
"00001111111000000000000000000000"

3. prepare the EDK-project for PR:

3.1. supply additional IP-Cores:
JTAG-PPC controller (debugging)
opb sysace
opb timer (to measure the time the reconfiguration process takes)
opb hwicap
3.2. Bus Connections
connect opb timer with the opb the sysace is already placed at
Appendix A: Hints to Labs 341

connect opb hwicap with the same opb

3.3. fill in addresses:
sysace: 0x41800000 - 64K
opb timer: 0x41c00000 - 64K
opb hwicap: 0x40200000 - 64K
docm ctlr: 0xE8800000 - 32K → so that the program fits into the BRAMs. If
put in the RAM unwanted deviations occur in the resulting picture.
iocm ctrl: 0xFFFF8000 - 32K /
3.4. extend the .ucf with sysace-pinouts. Entries of the form:
Net fpga_0_SysACE_CompactFlash_SysACE_CLK_....
→ see appendix B
3.5. connect ports:
opb sysace:
OPB_Clk - sys_clk_s
SysACE_MPA -> make external
SysACE_CLK -> make external
SysACE_MPD -> make external
SysACE_MPD_I -> no connection
SysACE_MPD_O -> no connection
SysACE_MPD_T -> no connection
SysACE_CEN -> make external
SysACE_OEN -> make external
SysACE_WEN -> make external
SysACE_MPIRQ -> no connection
→ not to forget to adapt the pins in the .ucf
opb timer:
OPB_clk -> sys_clk_s

also fit the external ports to the ones in the .ucf

3.6. Extend the C-programm for PR with ICAP:
copy xhwicap* files from John William’s ICAP driver [218]
adapt main.c: first configure the video decoder card, then present the PR menu
can e.g. sysace stdio.h not be found add the xilfatfs-library under Software|Software
Platform Settings|Platform Settings
create a linker script for the application (if possible put all the program parts to
iocm- resp. docm-memory (→ RAM very often causes problems ⇒ if necessary
extend the address space of the corresponding controller)
3.7. Testing: completely compile with EDK ad see if the system works so far
3.8. enter ports for the PRM in the port declaration of entity system as external ports:
close the EDK project
redirect ports to and from rgbfilter from the local instance to the corresponding
external ports:
→ add to system.mhs at the external port (at the beginning of the file):
342 RECONFIGURABLE COMPUTING

PORT LLC_CLOCK_OUT = VIDEO_IN_PR_0_LLC_CLOCK_to_filter, DIR = O

PORT R_in = VIDEO_IN_PR_0_R_out_to_filter, DIR = O, VEC = [0:9]
PORT G_in = VIDEO_IN_PR_0_G_out_to_filter, DIR = O, VEC = [0:9]
PORT B_in = VIDEO_IN_PR_0_B_out_to_filter, DIR = O, VEC = [0:9]
PORT h_counter_in = VIDEO_IN_PR_0_h_counter_out_to_filter, \\
DIR = O, VEC = [0:9]
PORT v_counter_in = VIDEO_IN_PR_0_v_counter_out_to_filter, \\
DIR = O, VEC = [0:9]
PORT valid_in = VIDEO_IN_PR_0_valid_out_to_filter, DIR = O
PORT R_out = RGBFILTER_0_R_out, DIR = I, VEC = [0:9]
PORT G_out = RGBFILTER_0_G_out, DIR = I, VEC = [0:9]
PORT B_out = RGBFILTER_0_B_out, DIR = I, VEC = [0:9]
PORT h_counter_out = RGBFILTER_0_h_counter_out, \\
DIR = I, VEC = [0:9]
PORT v_counter_out = RGBFILTER_0_v_counter_out, \\
DIR = I, VEC = [0:9]
PORT valid_out = RGBFILTER_0_valid_out, DIR = I
reopen EDK project
erase instance rgbfilter in System Assembly|Bus Interface
3.9. export to ISE:
→ in Project|Project-Options|Hierarchy and Flow:
Design is a sub-module: name: system i
do not synthesize non-Xilinx IPs
use project navigator implementation flow (ISE)
don’t add modules to an existing ise
4. Synthetizing the project with ISE
4.1. copy the content of the EDK-directory to <base>/synth/edk
4.2. Verify the ISE project:
→ open system.ise in base/synth/edk/projnav with ISE:
→ verify if the design is structured as follows:
system stub.vhd (system stub)
system.vhd (system i)
ppc405 0 wrapper (ppc405 0)
ppc405 1 wrapper (ppc405 1)
jtagppc 0 wrapper (jtagppc 0)
reset block wrapper (reset block)
..
.
system stub.bmm
→ if the structure is not like the above, check for the correct export settings in EDK.
Mind the "don’t add modules to an existing ISE project"-setting!
4.3. mark and rightclick system stub, select ’Synthesize - XST’-> properties: Xilinx Spe-
cific Options-> don’t add I/O Buffers (since the system is to be a sub-module)
4.4. create netlists for the EDK-module with doubleclicking ’Synthesize - XST’
4.5. the suiting Block-RAM memory mapping file (system stub bd.bmm) can be created
with doubleclicking ’Generate Programming File’ (every step necessary will be done,
including synthetization)
Appendix A: Hints to Labs 343

→ if that doesn’t work out there’s probably something wrong (a workaround might be com-
piling with I/O-buffers to get the system stub bd.bmm and then again without to generate
the appropriate .ngc-files)
5. Changes to the skript framework: in directory synth
The example Video8 comes with a bunch of scripts to support up the PRM generation. This
part gives hints on how to change those appropriately.
Directories:
mod * directories according to the modules (sobel, sharpen, mean-value, unfiltered)
ändern
top-directory remains
static - not needed (in this case contained in edk
edk - copy previously generated design here
Files:
adapt doit.cmd → descend in new DIR and execute xst
in these DIR: adapt .prj and .xst (are the same in all the DIR)
6. Changes to the project files
6.1. top.vhd (takes the longest time to accomplish):
adapt the entity-definition (ports) that are connected to the I/Os and further to the
ucf
not to forget the I, O and I/O buffers → IBUFs resp. OBUFs remain simple,
IOBUFs are extended to I, O and T (tristate)-line (e.g. for Sda decoder: one pin
to the outside, but * I, * O and * T towards the entity system i → very time and
work consuming!
add/erase components that are (not) needed:
→ adapt component system ⇒ comes up to the system from the EDK project →
can be copied from <base>/synth/edk/hdl/system.vhd (system stub.vhd
here is only a dummy to be able to instantiate system.vhd as a sub-module in
the ISE project)
→ guide the PRM signals as ports out in entity system
add/remove corresponding instances
care for intermediate signals if signals have to run directly from component in-
stances to I/Os
insert bus macros for signals coming from/going to the PRM (here: R,G u. B to
and from rbgfilter, v and h counter to and from rgbfilter and valid to
and from rgbfilter
→ again, add intermediate signals
additional instances, that are to be static or reconfigurable later on have to be
named → the definition will be done in the ucf (e.g. recon module: rgbfilter
portmap( ....)
6.2. top.ucf:
place bus macros (PlanAhead or FPGA-Designer can pose a big help here)
assign areas for static and reconfigurable modules (→ in PlanAhead this can be
done with drag-and-drop manner!)
344 RECONFIGURABLE COMPUTING

assign which instance from top.vhd is member of which group (static ↔ recon-
figurable)
7. The Build Process:
7.1. synthesize all files:
if not done yet: synthesize the ISE project in
<base>/synth/edk/projnav
(system stub bd.bmm) may not be omitted!)
execute <base>/synth/synthesize all.cmd
7.2. begin with the Early Access Partial Reconfiguration Design Flow:
→ execute buildit.cmd ⇒ the full and partial bitstreams should be built if the script
has been adapted properly.
Appendix B
Example of a User Constraints File

The following chapter gives an excerpt of a User Constraints File (UCF). It has been
created for the partial reconfiguration of the Video8 sample design. The complete file can be
found on the book’s web site.

# area group for reconfig module

INST "recon_module" AREA_GROUP = "pblock_recon_module";
AREA_GROUP "pblock_recon_module" RANGE=MULT18X18_X3Y8:MULT18X18_X3Y11;
AREA_GROUP "pblock_recon_module" RANGE=RAMB16_X3Y8:RAMB16_X3Y11;
AREA_GROUP "pblock_recon_module" RANGE=SLICE_X30Y64:SLICE_X45Y95;
AREA_GROUP "pblock_recon_module" RANGE=TBUF_X30Y64:TBUF_X44Y95;
AREA_GROUP "pblock_recon_module" MODE=RECONFIG;

# INST "fixed_area" AREA_GROUP = "pblock_fixed_area";

INST "system_i" AREA_GROUP = "pblock_fixed_area";
INST "alibi" AREA_GROUP = "pblock_fixed_area";
INST "dcm_25_MHZ" AREA_GROUP = "pblock_fixed_area";

AREA_GROUP "pblock_fixed_area" RANGE=SLICE_X46Y16:SLICE_X87Y149;

# Clock and reset constraints

Net sys_clk_pin LOC=AJ15;
Net sys_clk_pin PERIOD = 10000 ps;

Net reset LOC=AG5 | IOSTANDARD = LVTTL; # Center

Net reset TIG;

Net sys_rst_pin LOC=AH5;

Net sys_rst_pin IOSTANDARD = LVTTL;
346 RECONFIGURABLE COMPUTING

INST "dcm_25_MHZ" loc="DCM_X2Y0";

# leds
net LED_0 loc = AC4 | IOSTANDARD = LVTTL;
net LED_1 loc = AC3 | IOSTANDARD = LVTTL;
net LED_2 loc = AA6 | IOSTANDARD = LVTTL;
net LED_3 loc = AA5 | IOSTANDARD = LVTTL;

# push buttons
net button_start loc = AH2 | IOSTANDARD = LVTTL; # Right
net button_stop loc = AH1 | IOSTANDARD = LVTTL; # Left

#### Diligent general purpose expansion port (where the VDEC1 is connected)
NET "YCrCb_in* LOC = "AA8" | IOSTANDARD = LVTTL ;
.
.
.

NET "LLC_CLOCK" LOC = "B16" | IOSTANDARD = LVCMOS25 ;

NET "RESET_VDEC1_Z" LOC = "AF6" | IOSTANDARD = LVTTL | DRIVE = 8 ;
NET "VDEC1_PWRDN_Z" LOC = "G1" | IOSTANDARD = LVTTL | DRIVE = 8 ;
NET "VDEC1_OE_Z" LOC = "AF3" | IOSTANDARD = LVTTL | DRIVE = 8 ;
NET "Sda_decoder_pin" LOC = "AE5" | IOSTANDARD = LVTTL | DRIVE = 8 | PULLUP = TRUE;
NET "Scl_decoder_pin" LOC = "AB8" | IOSTANDARD = LVTTL | DRIVE = 8 | PULLUP = TRUE;

#### VGA
NET "BLANK_Z" LOC = A8 | DRIVE = 12 | SLEW = SLOW | IOSTANDARD = LVTTL;
NET "COMP_SYNC" LOC = G12 | DRIVE = 12 | SLEW = SLOW | IOSTANDARD = LVTTL;
NET "H_SYNC_Z" LOC = B8 | DRIVE = 12 | SLEW = SLOW | IOSTANDARD = LVTTL;
NET "V_SYNC_Z" LOC = D11 | DRIVE = 12 | SLEW = SLOW | IOSTANDARD = LVTTL;
NET "PIXEL_CLOCK" LOC = H12 | DRIVE = 12 | SLEW = SLOW | IOSTANDARD = LVTTL;

INST "red_out_DAC*
INST "green_out_DAC*
INST "blue_out_DAC*

#### Bus Macros

INST "BM_1" LOC = SLICE_X44Y94;
INST "BM_3" LOC = SLICE_X44Y92;
INST "BM_2" LOC = SLICE_X44Y90;
INST "BM_4" LOC = SLICE_X44Y88;
INST "BM_5" LOC = SLICE_X44Y86;
INST "BM_6" LOC = SLICE_X44Y84;
Appendix B: Example of a User Constraints File 347
INST "BM_7" LOC = SLICE_X44Y82;
INST "BM_8" LOC = SLICE_X44Y80;
INST "BM_9" LOC = SLICE_X44Y78;
INST "BM_10" LOC = SLICE_X44Y76;
INST "BM_11" LOC = SLICE_X44Y74;
INST "BM_12" LOC = SLICE_X44Y72;
INST "BM_13" LOC = SLICE_X44Y70;
INST "BM_14" LOC = SLICE_X44Y68;
INST "BM_1" AREA_GROUP = "AG_BM_1";
INST "BM_2" AREA_GROUP = "AG_BM_2";
INST "BM_3" AREA_GROUP = "AG_BM_3";
INST "BM_4" AREA_GROUP = "AG_BM_4";
INST "BM_5" AREA_GROUP = "AG_BM_5";
INST "BM_6" AREA_GROUP = "AG_BM_6";
INST "BM_7" AREA_GROUP = "AG_BM_7";
INST "BM_8" AREA_GROUP = "AG_BM_8";
INST "BM_9" AREA_GROUP = "AG_BM_9";
INST "BM_10" AREA_GROUP = "AG_BM_10";
INST "BM_11" AREA_GROUP = "AG_BM_11";
INST "BM_12" AREA_GROUP = "AG_BM_12";
INST "BM_13" AREA_GROUP = "AG_BM_13";
INST "BM_14" AREA_GROUP = "AG_BM_14";

#### Module RS232_Uart_1 constraints

Net fpga_0_RS232_Uart_1_RX_pin LOC=AJ8;
Net fpga_0_RS232_Uart_1_TX_pin LOC=AE7;

#### Module SysACE_CompactFlash constraints

Net fpga_0_SysACE_CompactFlash_SysACE_*
.
.
.

#### Module DDR_256MB_32MX64_rank1_row13_col10_cl2_5 constraints

Net fpga_0_DDR_256MB_32MX64_rank1_row13_col10_cl2_5_DDR_*
.
.
.
Appendix C
Quick Part-Y Tutorial

This quick tutorial gives an introduction to the basic features of the Part-Y[71] tool devel-
oped by Florian Dittmann from the "Heinz Nixdorf Institute" at the University of Paderborn.

It is intended to guide a novice to Part-Y and partial reconfiguration through a first example
of a PR design. The targeted system was the Xilinx XUP Development Board with a Virtex-II
Pro FPGA.

Prerequisites:
– running version of Part-Y (this tutorial was created with version 1.2)
– the tutorial sources.zip file
– ISE 6.1 or higher installed (tutorial is known to work on 7.1.04i)
Notes:
– always hit OK before changing the tab in a view - changes won’t be saved otherwise
– StdOut will be directed to MiscellaneouskStandard Output
– you might have extend the Part-Y Project by the proper definition for your FPGA.
These can be found under
de.upb.cs.roichen.party.device.concrete
– encountering problems you might find Part-Y FAQ.txt useful
Proceeding:
1. Open the level configuration window and ensure that the following items from the
different views are checked:
– BehavioralModuleSelectionController
– BehavioralTopLevelController
– BehavioralDownloadController
– StructureTopAssemblyController
– StructureTreeLevelController
– GeoSystemLevelController
350 RECONFIGURABLE COMPUTING

– MiscStorageController
– MiscGenerateBitstreamsController
– MiscOutputController
2. In Module Selection import the modules rt1.vhd, rt2.vhd and fixed.vhd from tuto-
rial sources.zip. Add them to current modules and mark each of them as reconfig-
urable. Hit OK!
3. In Top import global.vhd and global2.vhd. Add global.ucf to each of the Top-Level-
Designs marking one at a time and adding the .ucf separately. OK!
4. Top Assembly: mark Top0 global and Top1 global2 each at a time and add every
module to both of them. You should see the added modules at the ’modules instances
of selected top’-window.
5. If everything worked out up to now, you’ll see an hierarchy like the one exposed in
figure [C.1] on page 351. in Structural View|Tree.
6. In Geometrical View|System: set the correct target platform and a proper bus macro
file
7. Miscellaneous|Storage: set a name for the Part-Y project and a project path. Clicking
OK will create a file hierarchy at the given spot, INTEGRATE will copy the vhdl files
and others to the correct places. Be sure not to use spaces in path names since Part-Y
as well as ISE cannot cope with them properly. If everything went right there should
be a bunch of subfolders in your specified directory containing copies of the vhdl files
you declared and a couple of additional files.
8. In Behavioral View|Synthesis all of the VHDL-files have to be synthesized.
9. With Miscellaneous|Bitstream Generation you can conduct the initialization phase, the
activation as well as the final assembly phase by clicking on the run buttons. For some
reason the won’t look pressed when they are - might be a bug in the GUI. You can
check if the action is taking place in the ’standard output’ window.
10. After everything has been accomplished, the bitstreams can be downloaded to the de-
vice with Behavioral|Download using iMPACT
Appendix C: Quick Part-Y Tutorial 351

Figure C.1. Tree View after Top Assembly