EMBEDDED PROCESSOR-BASED SELF-TEST

FRONTIERS IN ELECTRONIC TESTING

Consulting Editor
Vishwani D. Agrawal
Books in the series:
Embedded Processor-Based Self-Test
D. Gizopoulos
ISBN: 1-4020-2785-0
Testing Static Random Access Memories
S. Hamdioui
ISBN: 1-4020-77521
Verification by Error Modeling
K. Radecka and Zilie
ISBN: 1-4020-7652-5
Elements ofSTIL: Principles and Applications ofIEEE Std. 1450
G. Maston, T. Taylor, J. Villar
ISBN: 1-4020-7637-1
Fault Injection Techniques and Tools for Embedded systems Reliability
Evaluation
A. Benso, P. Prinetto
ISBN: 1-4020-7589-8
High Performance Memory Memory Testing
R. Dean Adams
ISBN: 1-4020-7255-4
SOC (System-on-a-Chip) Testing for PIug and Play Test Automation
K. Chakrabarty
ISBN: 1-4020-7205-8
Test Resource Partitioning for System-on-a-Chip
K. Chakrabarty, Iyengar & Chandra
ISBN: 1-4020-7119-1
A Designers' Guide to BuHt-in Self-Test
C. Stroud
ISBN: 1-4020-7050-0
Boundary-Scan Interconnect Diagnosis
J. de Sousa, P.Cheung
ISBN: 0-7923-7314-6
Essentials ofElectronic Testing for Digital, Memory, and Mixed Signal VLSI Circuits
M.L. Bushnell, V.D. Agrawal
ISBN: 0-7923-7991-8
Analog and Mixed-Signal Boundary-Scan: A Guide to the IEEE 1149.4
Test Standard
A. Osseiran
ISBN: 0-7923-8686-8
Design for At-Speed Test, Diagnosis and Measurement
B. Nadeau-Dosti
ISBN: 0-79-8669-8
Delay Fault Testing for VLSI Circuits
A. Krstic, K-T. Cheng
ISBN: 0-7923-8295-1
Research Perspectives and Case Studies in System Test and Diagnosis
J.W. Sheppard, W.R. Simpson
ISBN: 0-7923-8263-3
Formal Equivalence Checking and Design Debugging
S.-Y. Huang, K.-T. Cheng
ISBN: 0-7923-8184-X
Defect Oriented Testing for CMOS Analog and Digital Circuits
M. Sachdev
ISBN: 0-7923-8083-5

EMBEDDED PROCESSOR-BASED
SELF-TEST

by

DIMITRIS GIZOPOULOS

University 0/ Piraeus, Piraeus, Greece

ANTONIS PASCHALlS
University 0/ Athens, Athens, Greece
and

YERVANT ZORIAN
Virage Logic, Fremont, Califomia, U.S.A.

SPRINGER SCIENCE+BUSINESS MEDIA, LLC

A C.I.P. Catalogue record for this book is available from the Library of Congress.

ISBN 978-1-4419-5252-3
ISBN 978-1-4020-2801-4 (eBook)
DOI 10.1007/978-1-4020-2801-4

Printed on acid-free paper

All Rights Reserved

2004 Springer Science+Business Media New York
Originally published by Kluwer Academic Publishers, Boston in 2004
Softcover reprint ofthe hardcover 1st edition 2004
No part of this work may be reproduced, stored in a retrieval system, or transmitted
in any form or by any means, electronic, mechanical, photocopying, microfilming, recording
or otherwise, without written permission from the Publisher, with the exception
of any material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work.

CONTENTS
Contents _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ v
List ofFigures _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ Vlll
List ofTables _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ IX
Preface
Xlll
Acknowledgments
xv
1. INTRODUCTION _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __
1.1
Book Motivation and Objectives
1
1.2 Book Organization
4
2. DESIGN OF PROCESSOR-BASED SOC
7
2.1
Integrated Circuits Technology
7
2.2 Embedded Core-Based System-on-Chip Design
8
2.3
Embedded Processors in SoC Architectures
11
3. TESTING OF PROCESSOR-BASED SOC
21
3.1
Testing and Design for Testability
21
3.2 Hardware-Based Self-Testing
32
3.3
Software-Based Self-Testing
41
3.4 Software-Based Self-Test and Test Resource Partitioning _46
3.5
Why is Embedded Processor Testing Important?
48
3.6 Why is Embedded Processor Testing Challenging?
49
4. PROCESSOR TESTING TECHNIQUES
55
4.1
Processor Testing Techniques Objectives
55
4.1.1
External Testing versus Self-Testing
56
4.1.2
DfT-based Testing versus Non-Intrusive Testing
57
4.1.3
Functional Testing versus Structura1 Testing
58
4.1.4
Combinational Faults versus Sequential Faults Testing _59
4.1.5
Pseudorandom versus Deterministic Testing
60
4.1.6
Testing versus Diagnosis
62
4.1.7
Manufacturing Testing versus On-1inelField Testing _63
4.1.8
Microprocessor versus DSP Testing
63
4.2 Processor Testing Literature
64
4.2.1
Chronological List ofProcessor Testing Research _ _ 64
4.2.2
Industrial Microprocessors Testing
78
4.3
Classification ofthe Processor Testing Methodologies
78
5. SOFTWARE-BASED PROCESSOR SELF-TESTING
81
5.1
Software-based self-testing concept and flow
82
5.2
Software-based self-testing requirements
87
5.2.1
Fault coverage and test quality
88
5.2.2
Test engineering effort for self-test generation
90

vi

Contents

5.2.3
Test application time
91
5.2.4
A new self-testing efficiency measure
96
5.2.5
Embedded memory size for self-test execution
97
98
5.2.6
Knowledge ofprocessor architecture
5.2.7
Component based self-test code development
99
5.3
Software-based self-test methodology overview
100
5.4 Processor components c1assification
107
5.4.1
Functional components
108
5.4.2
Control components
111
5.4.3
Hidden components
112
5.5
Processor components test prioritization
113
5.5.1
Component size and contribution to fault coverage _115
5.5.2
Component accessibility and ease oftest
117
5.5.3
Components' testability correlation
119
5.6 Component operations identification and selection
121
5.7 Operand selection
124
5.7.1
Self-test routine development: ATPG
125
5.7.2
Self-test routine development: pseudorandom
133
5.7.3
Self-test routine development: pre-computed tests __ 137
5.7.4
Self-test routine development: style selection
139
141
5.8 Test development for processor components
5.8.1
Test development for functional components
141
5.8.2
Test development for control components
141
5.8.3
Test development for hidden components
143
5.9 Test responses compaction in software-based self-testing _146
5.10
Optimization of self-test routines
148
5.10.1 "Chained" component testing
149
5.10.2 "Parallel" component testing
152
5.11
Software-based self-testing automation
153
6. CASE STUDIES - EXPERIMENTAL RESULTS
157
6.1
Parwan processor core
158
6.1.1
Software-based self-testing of Parwan
159
6.2 PlasmaJMIPS processor core
160
163
6.2.1
Software-based self-testing of Plasma/MIPS
6.3
MeisterlMIPS reconfigurable processor core
168
6.3.1
Software-based self-testing of MeisterlMIPS
170
171
6.4 Jam processor core
6.4.1
Software-based self-testing of Jam
172
173
6.5
oc8051 microcontroller core
6.5.1
Software-based self-testing of oc8051
175
6.6 RISC-MCU microcontroller core
176
6.6.1
Software-based self-testing ofRISC-MCU
177

Contents

vii

6.7 oc54xDSPCore _ _ _ _ _ _ _ _ _ _ _ _ _ _ 178

6.7.1
Software-based self-testing of oc54x
179
6.8 Compaction oftest responses
181
6.9 Summary ofBenchmarks
181
7. PROCESSOR-BASED TESTING OF SOC
185
7.1
The concept
185
7.1.1
Methodology advantages and objectives
188
7.2 Literature review
190
7.3
Research focus in processor-based SoC testing
193
8. CONCLUSIONS
195
References
197
Index
213
About the Authors
217

LIST OF FIGURES
Figure 2-1: Typical System-on-Chip (SoC) architecture.
9
Figure 2-2: Core types of a System-on-Chip.
11
Figure 3-1: ATE-based testing.
28
Figure 3-2: Self-testing of an IC.
34
Figure 3-3: Self-testing with a dedicated memory.
38
Figure 3-4: Self-testing with dedicated hardware.
39
Figure 3-5: Software-based self-testing concept for processor testing. __ 42
Figure 3-6: Software-based se1f-testing concept for testing a SoC core._43
Figure 5-1: Software-based se1f-testing for a processor (manufacturing). _ 82
Figure 5-2: Software-based self-testing for a processor (periodic).
84
Figure 5-3: Application ofsoftware-based self-testing: the three steps. _86
Figure 5-4: Engineering effort (or cost) versus fault coverage.
91
Figure 5-5: Test application time as a function ofthe K!W ratio.
94
Figure 5-6: Test application time as a function ofthe fuP/!;ester ratio.
95
Figure 5-7: Software-based self-testing: overview of the four phases. _102
Figure 5-8: Phase A of software-based self-testing.
103
Figure 5-9: Phase B of software-based self-testing.
104
Figure 5-10: Phase C of software-based self-testing.
105
Figure 5-11: Phase D of software-based self-testing.
107
Figure 5-12: Classes ofprocessor components.
108
Figure 5-13: Prioritized component-Ievel se1f-test pro gram generation._114
Figure 5-14: ALU component ofthe MIPS-like processor.
122
Figure 5-15: ATPG test patterns application from memory.
129
Figure 5-16: ATPG test patterns application with immediate instructions. 131
Figure 5-17: Forwarding logic multiplexers testing.
145
Figure 5-18: Two-step response compaction.
147
Figure 5-19: One-step response compaction.
147
Figure 5-20: "Chained" testing ofprocessor components.
150
153
Figure 5-21: "Parallel" testing of processor components.
Figure 5-22: Software-based self-testing automation.
154
Figure 7-1: Software-based self-testing for SoC.
186

LIST OF TABLES
Table 2-1: Soft, firm and hard IP cores.
10
Table 2-2: Embedded processor cores (1 of3).
15
Table 2-3: Embedded processor cores (2 of3).
16
Table 2-4: Embedded processor cores (3 of 3).
17
Table 4-1: External testing vs. self-testing.
57
Table 4-2: DtT-based vs. non-intrusive testing.
57
Table 4-3: Functional vs. structural testing.
59
Table 4-4: Combinational vs. sequential testing.
60
Table 4-5: Pseudorandom vs. deterministic testing.
62
Table 4-6: Testing vs. diagnosis.
63
Table 4-7: Manufacturing vs. on-line/field testing.
63
Table 4-8: Processor testing methodologies classification.
79
Table 5-1: Operations ofthe MIPS ALU.
124
Table 5-2: ATPG-based self-test routines test application times (case 1). 132
Table 5-3: ATPG-based self-test routines test application times (case 2). 132
Table 5-4: Characteristics of component self-test routines development. _139
Table 6-1: Parwan processor components.
159
Table 6-2: Self-test program statistics for Parwan.
160
Table 6-3: Fault simulation results for Parwan processor.
160
161
Table 6-4: Plasma processor components.
Table 6-5: Plasma processor synthesis for Design I.
162
Table 6-6: Plasma processor synthesis for Design 11.
162
Table 6-7: Plasma processor synthesis for Design III.
163
Table 6-8: Fault simulation results for the Plasma processor Design I. _164
Table 6-9: Self-test routine statistics for Designs 11 and III ofPlasma._164
Table 6-10: Fault simulation results for Designs 11 and III ofPlasma. _165
Table 6-11: Plasma processor synthesis for Design IV.
167
Table 6-12: Comparisons between Designs 11 and IV ofPlasma.
167
Table 6-13: MeisterlMIPS processor components.
168
Table 6-14: MeisterlMIPS processor synthesis.
169
Table 6-15: Self-test routines statistics for MeisterlMIPS processor. __ 170
Table 6-16: Fault simulation results for MeisterlMIPS processor.
170
Table 6-17: Jam processor components.
171
Table 6-18: Jam processor synthesis.
172
173
Table 6-19: Self-test routine statistics for Jam processor.
Table 6-20: Fault simulation results for Jam processor.
173
Table 6-21: oc8051 processor components.
174
Table 6-22: oc8051 processor synthesis.
174

Table 6-23:
Table 6-24:
Table 6-25:
Table 6-26:
Table 6-27:
Table 6-28:
Table 6-29:
Table 6-30:
Table 6-31:
Table 6-32:
Table 6-33:
Table 6-34:
Table 6-35:

List of Tables

Self-test routine statistics for oc8051 processor.

175
Fault simulation results for oc8051 processor.
176
RISC-MCU processor components.
176
177
RISC-MCU processor synthesis.
Self-test routine statistics for RISC-MCU processor.
178
Fault simulation results for RISC-MCU processor.
178
oc54x processor components.
179
oc54x DSP synthesis.
179
Self-test routines statistics for oc54x DSP.
180
Fault simulation results for oc54x DSP.
180
Execution times of self-test routines.
181
Summary ofbenchmark processor cores.
182
Summary ofapplication ofsoftware-based self-testing. __ 183

to Georgia, Dora and Rita

Preface

This book discusses self-testing techniques in embedded processors.

These techniques are based on the execution of test pro grams aiming to
lower the cost oftesting for processors and surrounding blocks.
Manufacturing test cost is already a dominant factor in the overall
development cost of Integrated Circuits (IC). Consequently, cost effective
methodologies are continuously seeked for test cost reduction. Self-test, the
ability of a circuit to test itself is a widely adopted Design for Test (Dff)
methodology. It does not only contribute to the test cost reduction but also
improves the quality of test because it allows a test to be performed at the
actual speed of the device, to detect defect mechanisms that manifest
themselves as delay malfunctions. Furthermore, self-test is a re-usable test
solution. It can be activated several times throughout the device's life-cycle.
The self-testing infrastructure of a chip can be used to detect latent defects
that do not exist at manufacturing phases, but they appear during the chip' s
operating life
The application of self-testing, as weIl as, other testing methods, face
serious challenges when the circuit under test is a processor. This is due to
the fact that processor architectures are particularly sensitive to performance
degradation due to extensive design changes for testability improvement.
Dff modifications of a circuit, including those that implement self-testing,
usually lead to area, performance and power consumption overheads that
may not be affordable in a processor design. Processor testing and selftesting is a particularly challenging problem due to sophisticated complex
processor structure, but it is also a very important problem that needs special
attention because of the central role that processors play in every electronic
system.
In this book, an emerging self-test methodology that recently captured
the interest of test technologists is studied. Software-based self-testing, also
called processor-based self-testing, takes advantage of the programm ability
of processors and allows them to test themselves with the effective execution
of embedded self-test programs. Moreover, software-based self-testing takes
advantage of the accessibility that processors have to all other surrounding

xiv

Preface

blocks of complex designs to test these blocks as weIl with such self-test
programs. The already established System-on-Chip design paradigm that is
based on pre-designed and pre-verified embedded cores employs one or
more embedded processors of different architectures. Software-based selftesting is a very suitable methodology for manufacturing and in-field testing
of embedded processors and surrounding blocks.
In this book, software-based self-testing is described, as a practical, lowcost, easy-to-apply self-testing solution for processors and SoC designs. It
relaxes the tight relation of manufacturing testing with high-performance,
expensive IC test equipment and hence results in test cost reduction. If
appropriately applied, software-based self-testing can reach a very high test
quality (high fault coverage) with reasonable test engineering effort, small
test development cost and short test application time.
Also, this book sets a basis for comparisons among different softwarebased self-testing techniques. This is achieved by: describing the basic
requirements of this test methodology; focusing on the basic parameters that
have to be optimized; and applying it to a set of publicly available
benchmark processors with different architectures and instruction sets.

Dimitris Gizopoulos Piraeus, Greece

Antonis Paschalis
Athens, Greece
Yervant Zorian
Fremont, CA, USA

Acknowledgments
The authors would like to acknowledge the support and encouragement
by Dr. Vishwani D. Agrawal, the Frontiers in Electronic Testing book series
consulting editor. Special thanks are also due to earl HaITis and
Mark de Jongh of Kluwer Academic Publishers for the excellent
collaboration in the production ofthis book.
The authors would like to acknowledge the help and support of several
individuals at the University of Piraeus, the University of Athens and Virage
Logic and in particular the help ofNektarios Kranitis and George Xenoulis.

Chapter

1
Introduction

1.1

Book Motivation and Objectives

Electronic products are used today in the majority of our daily activities.
Thus, they enabled efficiency, productivity, enjoyment and safety.
The Integrated Circuits (ICs) realized today consist of multiple millions
of logic gates and even more memory cells. They are implemented in, very
deep sub-micron (VDSM) process technologies and often consist of
multiple, pre-designed entities called Intellectual Property (IP) cores. This
design methodology that allowed the integration of embedded IP cores is
known as Embedded Core-Based System-on-Chip (SoC) design
methodology. SoC design flow supported by appropriate Computer Aided
Design (CAD) tools has dramatically improved design productivity and has
opened up new horizons for successful implementation of sophisticated
chips.
An important role in the architecture of complex SoC is played by
embedded processors. Embedded processors and other cores buBt around
them constitute the basic functional elements of today's SoCs in embedded
systems. Embedded processors have optimized design (in terms of silicon
area, performance, power consumption, etc), and provide the means for the
integration of sophisticated, flexible, upgradeable and re-configurable
functionality of a complex Soc. In many cases, more than one embedded
Embedded Processor-Based Self-Test
D.Gizopoulos, APaschalis, Y.Zorian
Kluwer Academic Publishers, 2004

Chapter 1 - Introduction

processors exist in a SoC, each of which takes over different tasks of the
system and sharing the processing workload.
Issues such as the quality of the final SoC, the reliability of the
manufactured ICs, and the reduced possibility of delivering malfunctioning
chips to the end users, are rapidly getting more importance today with the
increasing criticality of most of electronic systems applications.
In the context of these quality and reliability requirements, complex SoC
designs, realized in dense manufacturing technologies face serious problems
that need special consideration. Manufacturing test of complex chips based
on external Automatic Test Equipment (ATE),as a method to guarantee that
the delivered chips are correct1y operating, is becoming less feasible and
more expensive than ever. The volume of test data that must be applied to
each manufactured chip is becoming very large, the test application time is
increasing and the overall manufacturing test cost is becoming the dominant
part ofthe total chip development cost.
Under these circumstances, which are expected to get worse as circuits
size shrinks and density increases, the effective migration of the
manufacturing test resources from outside of the chip (ATE) to on-chip,
built-in resources and thus the effective replacement of external based
testing with internally executed self-testing is, today the test technology of
choice for all SoCs in practice. Self-testing allows at-speed testing, i.e. test
execution at the actual operating speed of the chip. Thus all physical faults
that cause either timing miss-behavior or an incorrect binary value can be
detected. Also, self-testing drastically reduces test data storage requirements
and test application time, both of which explode when external, ATE-based
testing is used. Therefore, the extensive use of self-testing has a direct
impact on the reduction of the overall chip test cost.
Testing of processors or microprocessors, even when they are not deeply
embedded in a complex system, is known to be achallenging task itself.
Classical testing approaches used in other digital circuits are not adequate to
the carefully optimized processor designs, because they can't reach the same
efficiency as in other types of digital circuits. Also, self-test approaches,
successfully used to improve the testability of digital circuits, are not very
suitable for processor testing because such techniques usually add overheads
in the processor's performance, silicon area, pin count and power
consumption. These overheads are often not acceptable for processors which
have been specifically optimized to satisfy very strict area, speed and power
consumption requirements.
This book primarily discusses the special problem of testing and selftesting of embedded processors in SoC architectures, as well as the problem
of testing and se1f-testing other cores of the SoC using the embedded
processor as test infrastructure.

Embedded Processor-Based Seit-Test

First, the general problem of testing complex SoC architectures is

discussed and the particular problem of processor testing and self-testing is
analyzed. The difficulties are revealed and the requirements for successful
solutions to the problem are discussed. Then, a comprehensive review of
different approaches is given and the work done so far is classified into
different categories depending on the targets of each methodology or
application. This part of the book serves as a comprehensive guide to the
readers that want to identifY particular topics of processor test and apply the
most suitable approach to their problem.
After this review, the processor testing and self-testing problems are
discussed considering reduction of the test cost for the processor and the
overall SoC. In the case of modem, cost-effective embedded processors, the
extensive application of DIT techniques is limited and, in many cases,
prohibited since such processors can't afford performance degradation, high
hardware overhead and increased power consumption. F or these reasons, the
inherent processing ability of embedded processors can be taken advantage
of, for successful testing of all the processor's internal modules. Several
approaches have been proposed that use the processor instruction set
architecture to develop efficient self-test programs which when executed
perform the testing task ofthe processor. This technique, known as softwarebased self-testing (SBST) adds minimal or zero overheads to the normal
operation of the processor and SoC in terms of extra circuit, performance
penalty and power consumption. It seems that software-based self-testing, if
appropriately applied, can be considered as the ultimate solution to low-cost
testing of embedded processors. The requirements of software-based selftesting and different self-test styles are presented and optimization
alternatives are discussed in this book. Software-based self-testing of
processors is presented as an effective test approach for cost-sensitive
products. A set of experimental results on several publicly available
processors illustrates the practicality of software-based self-testing.
Embedded processors used in SoC designs have excellent access to other
cores of the complex chip, independent of the architecture of the SoC and
the interconnect style between the embedded cores. Therefore, embedded
software routines, developed for software-based self-testing, can be
successfully used for testing other embedded cores of the SoC. Going even
further, the existence of powerful embedded processors in a design, can be
used for a wide variety of tasks apart from manufacturing test alone. These
tasks include: on-line testing in the field of operation, debugging, diagnosis,
etc. The last part of the book briefly discusses the use of embedded
processors for self-testing of SoC blocks, again from the low-cost test point
ofview.

Chapter 1 - Introduction

The book provides a guide to processor testing and self-testing and an

analysis of low-cost, software-based self-testing of processors and processorbased SoC architectures. It also sets the framework for comparisons among
different approaches of software-based self-testing focusing on the main
requirements of the technique. It also reveals the practicality of the method
with the experimental results presented.

1.2

Book Organization

The remaining ofthis book is organized in the following Chapters:

Chapter 2 discusses the trends in modem SoC design and

manufacturing. The central role of embedded processors in SoC
architectures is remarked. Emphasis is given to the importance
that classic processors are getting today when used as embedded
processors in SoC designs.
Chapter 3 deals with the challenges of processors and processorbased SoC testing and self-testing and focuses on the particular
difficulties of processor testing and the importance of the
problem. Software-based self-testing is introduced and its main
benefits are presented.
Chapter 4 consists of two parts. The first part discusses several
different ways of classifying processor testing approaches, since
each one focuses on specific aspects of the problem. The second
part presents a comprehensive, chronological list of processor
testing related research work of the recent years, giving
information on the focus of each work and the results obtained.
Finally, each of the presented works is linked to the one or more
classifications it belongs to.
Chapter 5 discusses the concept and details of Software-Based
Self-Testing. The basic objectives and requirements of the
methodology from the low-cost test point of view are analyzed.
Self-test code generation styles are discussed and all steps of the
methodology are detailed. Optimization of self-test pro grams is
also discussed.
Chapter 6 is a complement to Chapter 5, as it presents a set of
experimental results showing the application of software-based
self-testing to several embedded processors. Target processors
are of different complexities, architectures and instruction sets
and the methodology is evaluated in several special cases where
its pros and cons are discussed. Efficiency of software-based
self-testing is, in all cases, very high.

Embedded Processor-Based Self-Test

Chapter 7 briefly discusses the extension of software-based selftesting to SoC architectures. An embedded processor can be used
for the effective testing of other cores in the Soc. The details of
the approach are discussed and a list of recent, related works
from the literature is given.
Chapter 8 concludes the book, gives a quick summary of what
has been discussed in it and outlines the directions in the topic
that are expected to gain importance in the near future.

Chapter

2
Design 0/ Processor-Based SoC

2.1

Integrated Circuits Technology

Integrated circuit (IC) manufacturing technologies have reached today a

maturity level which is the driving force for the development of
sophisticated, multi-functional and high-performance electronic systems.
According to the prediction of the 2003 International Technology Roadmap
for Semiconductors (ITRS) [76], by year 2018 the half pitch 1 of dynamic
random access memories (DRAM), microprocessors and application-specific
Ies (ASIC) will drop to 18 nm and the microprocessor physical gate length
will drop to 7 nm. The implementation of correctly operating electronic
circuits in such sm all geometries - usually referred to as Very Deep
Submicron (VDSM) manufacturing technologies - was believed, just a few
years ago, to be extremely difficult, if at all possible, due to major hurdles
imposed by the fundamental laws of microelectronics physics when circuit
elements are manufactured in such small dimensions and with such sm all
distances separating them. Despite these skeptical views, VDSM

The half pitch of the first level interconnect is a measure of the technology level,
calculated as Y2 of the pitch which is the sum of widths of the metal interconnect and the
width ofthe space between two adjacent wires.

Embedded Processor-Based Self-Test

D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic Publishers, 2004

Chapter 2 - Design of Processor-Based SoC

technologies are successfully used today to produce high performance

circuits and continue providing evidence that Moore's law is still valid.
The increasing density of ICs realized in VDSM technologies allows the
integration of a very large number of transistors, either in the form of logic
gates or in the form of memory ceIls, in a single chip. In most cases, both
types of circuits (logic and memory) are combined in the same chip.
Furthermore, today, a single chip can contain digital circuits as weIl as
analog or mixed-signal ones. Multi-million transistor ICs are used nowadays
not only in high-end systems and performance-demanding applications, but,
practicaIly, in almost all electronic systems that people use daily at work, at
home, while traveling, etc, just to mention only some examples. More
system functionality can be put in electronic devices because hardware costs
per transistor are now several orders of magnitude lower than in the past and
the integration of more functionality to serve the final user of the system is
offered at lower costs.
Successful design and implementation of circuits of such complexity and
high density have become a reality not only because microelectronics
manufacturing technologies matured, but also because sophisticated tools for
Electronic Design Automation (EDA) or Computer-Aided Design (CAD)
emerged to cope with the design complexity of such systems. An important
step forward was the development and standardization of Hardware
Description Languages (HDL) [51], such as VHDL [75] and Verilog [74].
HDLs, supporting simulation and synthesis EDA tools arrive together,
matured together and are continuously being optimized to help designers
simulate their designs at several levels of abstraction and high simulation
speeds, as weIl as, to quickly synthesize high level (behavioral or Register
Transfer Level - RTL) descriptions into working gate-level netlists that
perform the desired functionality. Such a design flow based on HDLs,
simulation and synthesis increases the design productivity, allows quick
prototyping and early design verification at the behavioral level. This way,
design errors can be identified at early stages and effectively corrected.
Therefore, the possibility for a first-time-correct design is much larger.
The synergy of VDSM technologies, EDA tools and HDLs is also
supported by the System-on-Chip design paradigm, discussed in section 2.2,
and altogether boost design productivity.

2.2

Embedded Core-Based System-oo-Chip Desigo

Further improvements in design productivity have been obtained recently

with the wide adoption of a new systematic design methodology, weIl
known as the System-on-Chip (SoC) design paradigm [7], [48], [60], [79],
[88], [171]. SoC architectures are designed with the use of several,

Embedded Processor-Based Self-Test

embedded Intellectual Property (IP) cores coming from various ongms

(collectively called IP core providers). IP cores are pre-designed, preoptimized, pre-verified design modules ready to be plugged in the SoC (in a
single silicon chip) and interconnected with the surrounding IP cores, to
implement the system functionality. Each IP core in a SoC architecture may
deliver a different form of functionality to the system: digital cores, analog
cores, mixed-signal cores, memory cores, etc. IC design efficiency is
expected to speed up significantly because of the emergence of the SoC
design tlow as weIl as because ofthe improvements in EDA technology. The
2003 ITRS [76] predicts that the SoC design cyc1e will drop from today's
typical 12-months cyc1e to a 9-months cyc1e in year 2018.

Memory
DSP subsystem

Figure 2-1: Typical System-on-Chip (SoC) architecture.

A typical SoC architecture containing representative types of cores is

shown in Figure 2-1. A SoC typically consists of a number of embedded
processor cores, one of which may have a central role in the architecture and
others may have specialized tasks to accomplish (like the Digital Signal
Processor - DSP - core ofFigure 2-1, which takes over the functionality and
control of the DSP subsystem of the SoC). Several memory cores (dynamic
or static RAMs, ROMs, etc) are also employed in a SoC architecture, as
shown in Figure 2-1, each one dedicated to a different task storing
instructions, storing data or a combination of both. Other types of embedded
cores are used to implement the interface between the SoC and systems out
of it in aserial or parallel fashion and also other types of cores are used to
interface with the analog world converting analog to digital and vice versa.

Chapter 2 - Design of Processor-Based SoC

10

IP cores are released from IP core providers either as soft cores, firm
cores or hard cores depending on the level of changes that the SoC designer
(also called IP cores user) can make to them, and the level of transparency
they come with when delivered to the final SoC integrator [60]. A soft core
consists of a synthesizable HDL description that can be synthesized into
different semiconductor processes and design libraries. Afirm core contains
more structural information, usually a gate-level netlist that is ready for
placement and routing. A hard core includes layout and technologydependent timing information and is ready to be dropped into a system but
no changes are allowed to it [60].
Hard cores have usually smaller cost as they are final plug-and-play
designs implemented in a specific technology library and no changes are
allowed in them. At the opposite end, soft cores are available at HDL format
and the designer can use them very flexibly, synthesizing the description
using virtually any tool and any design library and thus the cost of soft cores
is usually much higher than that of hard cores. Adescription level in
between the hard and soft cores both in terms of cost and design flexibility is
the firm cores case where the final SoC designer is supplied with a gate-level
netlist of a design which can be altered in terms of technology library and
placement/routing but not in such a flexible way as a soft core. Table 2-1
summarizes the characteristics of these three description levels of IP cores
used in SoC designs.
Core category
softcore
firm core
hardcore

Changes
Many
Some
No

Cost
High
Medium
Low

Description
HDL
Netlist
Layout

Table 2-1: Soft, firm and hard IP COfes.

A tremendous variety of IP cores of all types and functionalities is

available to SoC designers. Therefore, designers are given the great
advantage to select from a rich pool of well-designed and carefully verified
cores and integrate them, in a plug-and-play fashion, in the system they are
designing. An idea ofthe variety oftypes ofIP cores that a SoC may contain
is also given in Figure 2-2.

Embedded Processor-Based Seit-Test

Memory
Gores

Processor
Gores

DSP
Gores

Memory
Interface

iH' Peripherals f-+

iH'

f-+

User
Defined
Logic

DSP

Peripheral
Gores

11

User
Defined
Logic

I-.

Peripheral
Interface

Analog and
Analog and
Mixed-Signal ~ Mixed-Signal ~
Gores
Interface

tt t

Test
Interface

Pin I Port
Mapping
~

Figure 2-2: Core types of a System-on-Chip.

With the adoption of the SoC design paradigm embedded core-based ICs
can be designed in a more productive way than ever and first-time-correct
design is much more likely. Electronic systems designed this way have much
shorter time-to-market than before and better chances for market success.
We should never forget the importance of time-to-market reduction in
today's high competition electronic systems market. A product is successful
if it is released to its potential users when they really need it, under the
condition, of course, that it operates acceptably as expected. A "perfect"
system on which hundreds of person-months have been invested may
potentially fail to obtain a significant market share if it is not released on
time to the target users. Therefore, successful practices at all stages of an
electronic system design flow (and a SoC design flow in particular) that
accomplish their mission in a quick and effective manner are always looked
for. The methodology discussed in this book aims to improve one of the
stages ofthe design flow.
We continue our discussion on the SoC design paradigm emphasizing on
the key role of embedded processors in it.

2.3

Embedded Processors in SoC Architectures

Essential parts of the functionality of every SoC architecture are assigned

to one or more embedded processors which are usually incorporated in a
design to accomplish at least the following two tasks.

Realization of a large portion of the system's functionality in the

form of embedded code routines to be executed by the processor(s).

12

Chapter 2 - Design of Processor-Based SoC

Control and synchronization of the exchange of data among the

different IP cores ofthe Soc.

The first task offers high flexibility to the SoC designers because they
can use processor's inherent programmability to efficiently update, improve
and revise the system's functionality just by adding or modirying existing
embedded software (code and data) stored in embedded memory cores.
ActuaIly, in many situations, an updated or new product version is only a
new or revised embedded software module which runs on the embedded
processor of the SoC and offers new functionality to the end user of the
system.
The second task that embedded processors are assigned to, offers
excellent accessibility and communication from the processor to all internal
cores of the SoC and, therefore, the processor can be used for several
reliability-related functions of the system, the most important of them being
manufacturing testing and field testing. This strong connection that an
embedded processor has with all other cores of the SoC makes it an excellent
existing infrastructure for the access of all SoC internal nodes, controlling
their logic states and observing them at the SoC boundaries. As we will see,
embedded processors can be used as an effective vehicle for low-cost selftesting oftheir internal components as weIl as other cores ofthe SoC.
In the majority of modem SoC architectures, more than one embedded
processors exist together; the most common situation is to have two
embedded processors in a SoC. For example, an embedded microcontroller
(/-lC) or embedded RISC (Reduced Instruction Set Computer) or other
processor can be used for the main processing parts of the system, while a
Digital Signal Processor (DSP) can take over part of functionality of the
system which is related to heavier data processing for specialized signal
processing algorithms (see Figure 2-1 and Figure 2-2). In architectures
where the SoC communicates with different external data channels, a
separate embedded processor associated with its dedicated memory
subsystem may deal with each of the communication channels while another
processor can be used to co ordinate the flow of data in the entire SoC.
The extensive use of embedded processors in SoC architectures of
different complexities and application domains has given new life to classic
processor architectures, with a word length as small as 8-bits, which were
widely used in the past. Successful architectures of microcontrollers,
microprocessors and DSPs were used for many years in the past, in a big
variety of applications as individually packaged devices (commercial-of-theshelf components - COTS). These classical processors are now used as
embedded processor cores in complex SoC architectures and can actually

Embedded Processor-Based Self-Test

13

boost up system's performance, while taking over simple or more complex

tasks ofthe system's functionality.
A wide range of architectures of classical processors are now used as
embedded cores in SoC designs: accumulator-based processors, stack-based
processors, RISC processors, DSPs, with word sizes as sm all as 8 bits and as
large as 64 bits. One of the most common formats that these classical
processors designs appear is that of a synthesizable HDL description (VHDL
or Verilog). The SoC designer (integrator) can obtain such a synthesizable
processor model, synthesize it and integrate the processor in the SoC design.
The SoC may be realized either using an ASIC standard cells library or an
FPGA family device. Depending on the implementation technology, these
processor architectures (that were considered "old" or even obsolete) can
give to the SoC a processing power that is considered more than sufficient
for a very large set of applications.
Furthermore, these new forms of classical processor architectures have
been seriously improved in terms of performance because of the flexibility
offered by the use of synthesizable HDL and today's standard cells libraries.
Processors as COTS components used in boards could not be re-designed;
neither could they be re-targeted in a faster technology library. On the other
hand, embedded processors available in HDL format can be re-designed to
meet particular needs of a SoC design as well as be re-targeted to a new
technology library to obtain better performance. Even the instruction set of
the processor can be altered and extended to meet the application
requirements.
The reasons for the re-birth and extensive re-use of classical processor
architectures in the form of embedded processor cores in SoC designs, as
wen as the corresponding benefits derived from this re-use are discussed in
the following paragraphs.

Classical processors have very wen designed architectures, and have

been extensively used, programmed, verified and tested in the past
decades. The most successful of them have been able to penetrate
generations of electronic products. Many algorithms have been
successfully implemented as machine/assembly code for these
architectures and can be effectively re-used when the new enhanced
versions of the processor architectures are adopted in complex SoC
designs. System development time can be saved with this re-use of
existing embedded code routines. Therefore, in this case we have a
dual case of re-use: hardware re-use (the processor core itself) and
software re-use (the embedded routines).
The majority of chip designers, and system architects has at least a
basic knowledge of the Instruction Set Architecture (ISA) of some

14

Chapter 2 - Design of Processor-Based SoC

c1assical processors and are therefore able to quickly program in their

assembly language, and understand their architecture (registers,
addressing modes, interrupt handling, etc). Even if a designer has
experience in assembly language programming of a previous member
of a processor's family, it is very easy for hirnlher to program in the
assembly language of a later member of the processors' family.
Limited man-power needs to be consumed for learning any new
instruction set architecture and assembly language.
Classical processors usually consist of a small number of gates and
memory elements, and equivalently occupy small silicon area,
compared with high-end RISC or CISC architectures with multiple
stages of pipelining and other performance-enhancing features.
Therefore, classical sm all processors provide an cost-effective
solution for embedding a processing element in a SoC with small
area, sufficient processing power and small electrical power
consumption. This solution may of course not be suitable in all
applications because of more demanding performance requirements.
But for a large number of applications the performance that a
c1assical cost-effective processor delivers is more than adequate.
Classical processors of small word lengths, such as 8 or 16 bits, have
a well-defined instruction set consisting of the most frequently used
instructions. Such an instruction set can lead to small, compact
programs with reduced memory requirements for code and data
storage. This is a very important point in SoC architectures where the
size of embedded memory components is a serious concern.

There are several examples of modem SoC architectures, used in

different application domains, which inc1ude one of more c1assical
embedded processors. This fact proves the wide adoption of classical
processors as embedded cores in complex designs, in cases where the
performance they offer to the system is sufficient. In all other cases,
optimized, high-performance modem embedded processors are utilized to
provide the system with the necessary performance for the application.
Table 2-2, Table 2-3 and Table 2-4 list a set of commercial embedded
processors that are commonly used in many applications. Both categories of
small, cost-effective, c1assical processors and also high-performance,
modem processors are inc1uded in these Tables. For each embedded
processor we give the company or companies that develop it and some
available processor characteristics, inc1uding core type (soft, hard), core size
and core performance.

Embedded Processor-Based Self-Test

Characteristics
8-bit RISC processor.
Synthesizable soft core.
http://www.arc.com
Less than 3500 gates for basic CPU.
40 MHz in FPGA implementation.
160 MHz in 0.18j..1m ASICprocess.
16-bit x86 compatible CISC processors.
V8086N186
ARC International
Synthesizable soft cores.
http://www.arc.com
15,000 gates (V8086) 22,000 gates (V186).
80 MHz in 0.25 j..Im ASIC process.
Turb086/Turbo186
16-bit x86 compatible CISC processors.
ARC International
Synthesizable soft cores.
http://www.arc.com
20,000 gates (Turb086) 30,000 gates
(Turbo186)
80+ MHz in 0.35 j..Im ASIC process.
C68000
16/32-bit microprocessor.
CAST and Xi/inx
Motorola MC68000 compatible.
http://www.cast-inc.com
Hard, firm or synthesizable soft core.
http://www.xilinx.com
2,200 to 3,000 logic slices in various Xilinx
FPGAs.
20 MHz to 32 MHz frequency.
VZ80
8-bit CISC processor.
ARC International
Z80 compatible.
http://www.arc.com
Synthesizable soft core.
Less than 8,000 gates.
V6502
8-bit CISC processor.
ARC International
6502 compatible.
http://www.arc.com
Synthesizable soft core.
Less than 4,000 gates.
V8-j..IRISC
8-bit RISC processor.
Synthesizable soft core.
ARC International
http://www.arc.com
3,000 gates.
100 MHz in 0.25 j..Im ASIC process.
Y170
8-bit processor.
Zilog Z80 compatible.
Systemyde
http://www.systemyde.com Synthesizable soft cores.
7,000 gates.
Y180
8-bit processor.
Systemyde
Zilog Z180 compatible.
http://www.systemyde.com Synthesizable soft core.
8,000 gates.
8-bit microcontroller.
DW8051
803x1805x compatible.
Synopsys
http://www.synopsys.com
Synopsys DesignWare core
Synthesizable soft core.
10,000 to 13,000 gates.
250 MHz frequency.
Table 2-2: Embedded processor cores (1 of3).
Embedded Processor
ARClite
ARC International

15

16

Embedded Processor
DW6811
Synopsys

Chapter 2 - Design of Processor-Based SoC

Characteristics
8-bit microcontroller.
6811 compatible.
http://www.synopsys.com
Synopsys DesignWare core
Synthesizable soft core.
15,000 to 30,000 gates.
200 MHz frequency in 0.13 IJm ASIC process.
SAM80
8-bit microprocessor.
Samsung Eleetronies
Zilog Z180 compatible.
http://www.samsung.com
Hard core.
0.6 IJm, 0.5 IJm ASIC processes.
SM8A02/SM8A03
8-bit microcontroller.
Samsung Eleetronies
80C51/80C52 subset compatible.
http://www.samsung.com
Hard core.
0.8 IJm ASIC processes.
eZ80
8-bit microprocessors family.
Zi/og
Enhanced superset of the Z80 family.
http://www.zilog.com
50 MHz frequency.
KL5C80A12
8-bit high speed microcontroller.
Kawasaki LSI
Z80 compatible.
http://www.klsi.com
10 MHz frequency.
R8051
8-bit RISC microcontroller.
Altera and CAST Ine.
Executes all ASM51 instructions.
http://www.altera.com
Instruction set of 80C31 embedded controller.
http://www.cast-inc.com
Synthesizable soft core.
2,000 to 2,500 Altera family FPGA logic cells.
30 to 60 MHz frequency.
8-bit microcontroller.
C8051
Evatronix
Executes all ASM51 instructions.
http://www.evatronix.pl
Instruction set of 80C31 embedded controller.
Synthesizable soft core.
Less than 1OK gates depending on technology.
80 MHz in 0.5 IJm ASIC process
160 MHz in 0.25 IJm ASIC process.
DF6811CPU
8-bit microcontroller CPU.
Altera and Digital Core
Compatible with 68HC11 microcontroller.
Design
Synthesizable soft core.
http://www.altera.com
2,000 to 2,300 Altera family FPGA logic cells.
http://www.dcd.com.pl
40 to 73 MHz frequency.
MIPS32 M4KTM
32-bit RISC microprocessor core of MIPS32'M
MIPS
architecture.
http://www.mips.com
Synthesizable soft core.
300 MHz typical frequency in 0.13 IJm process.
0.3 to 1.0 mm 2 core size.
Table 2-3: Embedded processor cores (2 of3).

Embedded Processor-Based Self-Test

Embedded Processor
FlexCore MIPS32 4Kec
LSI Logic

Xtensa
Tensilica

Characteristics
32-bit RISC CPU core of MIPS32 'M architecture.
Hard core.
167 MHz in 0.18 IJm ASIC process.
200 MHz in 0.11 IJm ASIC process.
32-bit RISC CPU core of ARM v4T architecture.
Hard core.
133 MHz frequency in 0.13 IJm ASIC process.
0.26 mm 2 core size.
32-bit RISC CPU core of ARM v4T architecture.
Synthesizable soft core.
100 to 133 MHz frequency in 0.13 ASIC
process.
0.32 mm 2 core size.
32-bit RISC CPU core of MIPS32 1IVI
architecture.
DesignWare Star IP Core.
Synthesizable soft core.
240-300 MHz frequency.
0.4 - 1.9 mm 2 core size.
32-bit unified microcontroller-DSP processor
core.
DesignWare Star IP Core.
Synthesizable soft core.
166 MHz frequency in 0.181Jm ASIC process.
200 MHz frequency in 0.13 IJm ASIC process.
32-bit superscalar RISC processor core.
Hard core.
550 MHz / 1000 MIPS in 0.15 IJm ASIC process.
4.0 mm 2 core size.
8-bit AVR-based RISC microcontroller.
Inciudes 2 Kbyte flash memory.
10 MHz freQuency.
8-bit AVR-based RISC microcontroller.
Inciudes 1 Kbyte flash memory.
12 MHz freQuency.
16/32-bit microprocessor.
Motorola MC68000 compatible.
Synthesizable soft core (VHDL and Verilog).
16-bit fixed point Digital Signal Processor.
TMS320C25 compatible.
Synthesizable soft core (VHDL and Verilog).
32-bit RISC configurable processor.
Up to 300 MHz on 0.13 IJm.

SHARC
Analog Devices

32-bit floating point DSP core.

300 Mhz/1800 MFLOPs.

http://www.lsilogic.com

ARM7TDMI
ARM

http://www.arm.com

ARM7TDMI-S
ARM

http://www.arm.com

MIPS4KE
Synopsys and MIPS

http://www.synopsys.com

TC1MP-S
Synopsys and Infineon

http://www.synopsys.com
http://www.infineon.com

PowerPC 440
IBM

http://www.ibm.com

AT90S2313
Atmel

http://www.atmel.com

AT90S1200
Atmel

http://www.atmel.com

C68000
Evatronix

http://www.evatronix.pl

C32025
Evatronix

http://www.evatronix.pl

http://www.tensilica.com

http://www.analog.com

Table 2-4: Embedded processor cores (3 of 3).

17

18

Chapter 2 - Design of Processor-Based SoC

The infonnation presented in Table 2-2, Table 2-3 and Table 2-4 has
been retrieved by the companies' public available documentation. It was our
effort to cover a wide range of representative types of embedded processors,
but of course not all available embedded processors today could be listed.
The intention of the above list of embedded processors is to demonstrate that
classic, cost effective processors and modem, high-perfonnance processors
are equally present today in the embedded processors market.
Apparently, when the perfonnance that a classical, sm all 8-bit or 16-bit
processor architecture gives to the system, is not able to satisfy the particular
perfonnance requirements of a specific application, other solutions are
always available, such as the high-end RISC embedded processors or DSPs
which can be incorporated in the design (several such processors are listed in
the Tables of the previous pages). The high perfonnance of modem
processor architectures, enriched with deep, multi-stage pipeline structures
and complex perfonnance-enhancing circuits, is able to meet any demanding
application needs (communication systems, industrial control systems,
medical applications, transportation and others).
As a joint result of the recent advances in very deep submicron
manufacturing technologies and design methodologies (EDA tools, HDLs
and SoC design methodology), today's complex processor-based SoC
devices offer complex functionality and high perfonnance that is able to
meet the needs ofthe demanding users ofmodem technology.
Unfortunately, the sophisticated functionality and the high perfonnance
of electronic systems are not offered at zero expense. Complex modem
systems based on embedded processors and realized as SoC architectures
have many challenges to face and major hurdles to overcome. Many ofthese
challenges are related to the design phases of the system and others are
related to the tasks of:

verifying the circuit's correct design;

testing the circuit's correct manufacturing; and
testing the circuit's correct operation in the field.

These tasks have always been difficult and time consuming, even when
electronic circuits size and complexity was much smaller than today. They
are getting much more difficult in today's multi-core SoC designs, but they
are also of increasing importance for the system's quality. An increasing
percentage ofthe total system development cost is dedicated to these tasks of
design verification and manufacturing testing. As a result, cost-reduction
techniques for circuit testing during manufacturing and in the field of
operation are gaining importance today and attract the attention of
researchers. As we see in this book, the existence of (one or more) embedded
processors in an SoC, although add to the chip's complexity, also provide a

Embedded Processor-Based Self-Test

19

powerful embedded mechanism to assist and effectively perform testing of

the chip at low cost.
In the next Chapter we discuss testing and testable design issues of
embedded processors and modem processor-based SoC architectures.

Chapter

3
Testing ofPro cess or-Bas ed SoC

3.1

Testing and Design for Testability

The problem of testing complex SoC architeetures has attracted

researchers' interest the recent years because it is a problem of increasing
difficulty and importance for the electronic circuits development community.
Testing, in the electronic devices world, is the systematic process used to
make sure that an IC has been correctly manufactured and is free of defects.
This correctness is verified with the application of appropriate inputs (calIed
test patterns or test vectors) and the observation of the circuit's response
which should be equal to the expected one, previously known from
simulation. This process is called manufacturing testing and is performed
before the IC is released for mounting to larger electronic systems.
Manufacturing testing is applied once to each manufactured IC and occupies
a significant part ofthe IC development cost.
The testing process may be also applied subsequently - after the IC has
been released for use - to make sure that the IC continues to operate
correctly when mounted in the final system. This is called periodic testing,
in-field testing or on-line testing and is a necessary process because a
correctly manufactured circuit that has been extensively tested during
manufacturing and found to be defect-free can later malfunction because of
several factors that appear in the field. Such factors are the aging of the
Embedded Processor-Based Self-Test
D.Gizopoulos, APaschalis, Y.Zorian
Kluwer Academic Publishers, 2004

22

Chapter 3 - Testing of Processor-Based SoC

device as weIl as extern al factors such as excessive temperatures, vibrations,

electromagnetic fields, induced particles, etc. Particles of relatively low size
and energy can still be harmful in today's very deep submicron technologies
even at the ground level because of the extremely sm all dimensions of the
ICs being designed and manufactured today. Therefore, testing for the
correct operation of a chip is not furthermore a one-time effort to be only
applied during manufacturing. Testing must be repeated in regular intervals
during the normal operation of the chip in its natural environment for the
detection of operational faults 2
The testing complexity of an electronic system (a chip, a board or a
system) is conceptually decomposed in two parts, which are closely related.
The first part is the complexity to generate sufficient tests, or test patterns,
for the system under test - the test generation complexity. The second part is
the complexity to actually apply tests - the test application complexity. Test
generation is a one-time effort and constant cost for all identical
manufactured devices while test application is a cost that must be accounted
for each tested device. We elaborate on how these two parts of the testing
complexity are related.
Test generation is usually performed as a combination of manual effort
and EDA tools assistance. Of course, the increasing size of electronic
systems has already led to increased needs for as much as possible
automation of the test generation process. Nevertheless, even today, expert
test engineers can be more effective than automatic tools in special
situations. Test generation can be a very time consuming process and it may
require a large number of repetitions and refinements, but in all cases, it is a
one-time effort for a particular design. When a sufficient fault coverage
leveP has been reached by a sequence of test patterns, all identical ICs that
will be manufactured subsequently will be tested using the same test
sequence and no more test generation process is required, unless the design
is changed. For complex designs, it may take a serious amount of person
power and computing time to develop a test sequence, but after this sequence
2

Operational faults in deep submicron technologies are classified into the following
categories. Permanent faults are infinitely aetive at the same loeation and refleet
irreversible physieal ehanges. Intermittent appear repeatedly at the same loeation and
eause errors in bursts only when they are aetive. These faults are indueed by unstable or
marginal hardware due to proeess variations and manufaeturing residuals and are aetivated
by environmental ehanges. In many eases, intermittent faults precede the oeeurrenee of
permanent faults. Transient faults appear irregularly at various loeations and last short
time. These faults are induced by neutron and alpha particles, power supply and
interconneet noise, electromagnetie interferenee and electrostatic discharge.
Fault eoverage obtained by a set of test patterns is the pereentage of the total faults of the
ehip that the test set ean detect. Faults belong to a fault model which is an abstraetion of
the physical defect mechanisms.

Embedded Processor-Based Self-Test

23

is developed and is proven to guarantee high fault coverage, test generation

process is considered successful.
The really hard part of the test generation process is not the actual time
necessary for the development of a test set (it may vary from a few hours or
days to several months). Rather, it is the ability ofthe test generation process
itself to obtain high fault coverage for the particular design even after a long
test generation time or with a large test set. In many situations, there are
complex ASIC or SoC designs that even the most sophisticated sequential
circuits test generation EDA tools can't handle. For such hard-to-test
designs, sufficient fault coverage can't be obtained unless serious Designfor-Testability changes are applied to the circuit. Design-for-Testability
(DfT) refers to design modifications that help test patterns to be easier
applied to the circuit's intern al nodes and node values to be easier observed
at the circuit outputs.
DfT modifications are not always easily adopted by circuit designers and
incorporated in the chip. The ultimate target of test generation is to obtain a
high fault coverage with an as small as possible test set (to reduce test
application time - discussed right after) with minimum DfT changes in the
circuit. DfT changes are usually avoided by circuit designers because they
degrade the performance and power behavior of the circuit during normal
operation and increase the size of the circuit. Circuit testability must be
considered as one of the most important design parameters and must be
taken into account as early as possible in its design phases. After a designer
has applied intelligent design techniques and has reached a circuit
performance that satisfies the product requirements, it is very difficult to
convince himlher that new design changes are necessary (with potential
impact on the circuit size, performance and power consumption) to improve
its testability (accessibility to internal nodes; easiness to apply test patterns
and observe test responses). It is much easier for circuit designers to account
for DfT structures early in the design process.
On the other hand, the second part of testing complexity, the test
application complexity is related to the time interval that is necessary to
apply a set of tests to the IC (by setting logic values to its primary inputs)
and to observe its response (by monitoring the logic values of its primary
outputs). The result of this process is the characterization of a chip as faulty
or fault-free and its final rejection or delivery for use, respectively. Test
application time for each individual chip depends on the number of test
patterns applied to it and also on the frequency at which they are applied
(how often a new test pattern is applied).
Test application time has a significant impact on the total chip
development cost. This means that a smaller test application time leads to
smaller device cost. On the other hand, a small test application time (testing

24

Chapter 3 - Testing of Processor-Based SoC

with a smaller test set) may lead to relatively poor fault coverage. If a test set
obtains small fault coverage, this means that only a small percentage of the
faults that may exist in the circuit will be detected by the test set. The
remaining faults of the fault model may exist in the circuit but they will not
be detected by the applied test set. Therefore, the insufficiently tested device
has a higher probability to malfunction when placed in the target system than
a device which has been tested to higher fault coverage levels.
A discussion on the details of test generation and test application is given
in the following paragraphs. Both tasks are becoming extremely difficult as
the complexity of ICs and in particular processor-based SoCs increases.
Test generation for complex ICs ean't be easily handled even by the most
advanced commercial combinational and sequential circuit Automatic Test
Pattern Generators (ATPG). ATPG tools can be used, ofcourse, only when a
gate level netlist of the circuit is available. The traditional fault models used
in ATPG flows are the single stuck-at fault model, the transition fault model
and the path delay fault model. The number of gates and memory elements
(flip-flops, latches) to be handled by the ATPG is getting extremely high and
in some cases relatively low fault coverage, of the selected fault model, can
only be obtained after many hours and many backtracks of the ATPG
algorithms. As circuit sizes increase this inefficiency of ATPG tools is
becoming worse.
The difficulties that an ATPG faces in test generation have their sources
in the reduced observability and controllability of the internal nodes in
complex architectures. In earlier years, when today's embedded IP cores
were used as packaged components in a System-on-Board design (packaged
chips mounted to Printed Circuit Board - PCB), the chips inputs and outputs
were easily accessible and testing was significantly simpler and easier
because ofthis high accessibility. The transition to the SoC design paradigm
and to miniaturized systems it develops, has given many advantages like
high performance, low power consumption, small size, small weight, ete, but
on the other side it imposed serious accessibility problems for the embedded
cores and, therefore, serious testability problems for the SoC. Deeply
embedded functional or storage cores in an SoC need special mechanisms
for the delivery of the test patterns from SoC inputs to the core inputs and
the propagation of their test responses from core outputs to the SoC
boundaries for externaiobservation and evaluation.
It is absolutely necessary that a complex SoC architecture includes
special DIT structures to improve the accessibility to its internal nodes and
thus improve its testability. The inclusion of DIT structures in a chip makes
the test generation process for it much more easy and effective, and the
required level of test quality can be obtained. We discuss alternative DIT

Embedded Processor-Based Self-Test

25

approaches and their advantages and disadvantages when applied to a

complex SoC.
Structured scan-based Off approaches are employed to help reducing the
complexity of test generation problem and accessibility of internal circuit
nodes. Scan-based Off links the memory elements (flip-flops, latches) of a
digital circuit in one or more chains. Each of the memory elements can be
given any logic value during the scan-in process (insertion of logic values in
the scan chain) and the content of each memory element can be observed
outside the chip during the scan-out process (extraction of the logic values
out ofthe scan chain). The scan-in and scan-out processes can be performed
in parallel: while a new test vector is scanned-in, the circuit response to the
previous test vector is scanned-out. Scan-based Off offers maximum
accessibility to the circuit internal nodes and is also easily automated
(mature commercial tools have been developed for scan-based Off and are
extensively used in industry). On the negative side, scan-based Off suffers
ftom the hardware overhead that it adds to the circuit and the excessive test
application time that is due to long scan-in and scan-out intervals,
particularly when the scan chains are very long. Scan-based Sff may have a
full-scan or partial-scan architecture (all memory elements or a subset of
them are connected in scan chains, respectively) [1], [23], [39], [170] and
helps reducing the test generation process by an ATPG tool, by giving the
ability to set values and observe internal circuit nodes. This way, the
problem of sequential circuits testing is significantly simplified or even
reduced to combinational circuit testing (when full scan is used).
Boundary scan architecture [124], a scan-like architecture at the chip
boundaries, has been successfully used for years for board testing, for testing
the chips boundaries and the interconnection between chips on a board.
Boundary scan testing is continuously applied to chips today while its
applications and usefulness are increasing.
At the SoC level, test communication between embedded cores, delivery
of test patterns ftom SoC inputs to core inputs and transfer of test responses
ftom core outputs to SoC outputs, are supported by the new IEEE 1500
Standard for Embedded Core Testing (SECT) which is being finalized [73].
IEEE 1500 SECT standardizes the test access mechanism for embedded
core-based SoC, defines the test interface language (Core Test Language CTL which is actually IEEE 1450.6, an extension to IEEE 1450 Standard
Test Interface Language - STIL), as well as a hardware architecture for a
core test wrapper to support the delivery of test patterns and the propagation
oftest responses of cores in a scan-based philosophy.
All the above scan-based test architectures (boundary scan at the chip
periphery, fulllpartial scan at the chip or core level, and IEEE 1500
compliant scan-based testing of cores at SoC level) as well as other

26

Chapter 3 - Testing of Processor-Based SoC

structured Off techniques are very successful inreducing the test generation
costs and efforts because of the existence of EOA tools for the automatic
insertion of scan structures. Manual effort is very limited and high fault
coverage is usua11y obtained. Furthermore, other structured Off techniques
like test point insertion (control and observation points) are widely used in
conjunction with scan design to further increase circuit nodes accessibility
and ease the test generation difficulties.
The major concems and limitations of a11 scan-based testing approaches
and in general of a11 structured or ad-hoc Off approaches that make them not
directly applicable to any design without serious considerations and planning
are the fo11owing. As we see subsequently, processors are types of circuits
where structured Off techniques can't be applied in a straightforward
manner.

Hardware overhead.
Off modifications in a circuit (multiplexers for test point
insertion, multiplexers for the modification of normal flip-flops
to scan flip-flops, additional primary inputs and/or outputs, etc)
always lead to substantial silicon area increase. In some cases
this overhead is not acceptable, for example when the larger
circuit size leads to a package change. Thus, Off modifications
may directly increase the chip development costs.
Performance degradation.
Scan based design and other Off techniques make changes in the
critical paths of a design. In a11 cases, at least some multiplexing
stages are inserted in the critical paths. These additional delays
may not be a problem in low-speed circuits when a moderate
increase in the delay of the critical paths which is
counterbalanced with better testability of the chip is not a serious
concem. But, in the case of high-speed processors or highperformance processor-based
SoCs
such performance
degradations, even at the level of 1% or 2% compared to the nonOff design, may not be acceptable. Processor designs, carefu11y
optimized to deliver high performance, of course, belong to this
class of circuits which are particularly sensitive to performance
impact due to Off modifications.
Power consumption increase.
The increase of silicon area wh ich is due to Off modifications is
also related to an increase in power consumption, a critical factor
in many low-cost, power-sensitive designs. Scan-based Off
techniques are characterized by large power consumption
because of the high circuit activity when test patterns are scanned

Embedded Processor-Based Seit-Test

27

in the chain and test responses are scanned out of it. Circuit
activity during the application of scan tests may be much more
than the circuit activity during normal operation and may lead to
peak power consumption not foreseen at the design stage. This
can happen because scan tests apply to the circuit not functional
input patterns that do not appear when the circuit operates in
normal mode. Therefore, excessive power consumption during
scan-based testing may seriously impact the manufacturing
testing of an IC, as its package limits may be reached because of
excessive heat dissipation.
Test data size (patterns and responses) and duration oftest.
Scan-based testing among the structured Dff techniques is
characterized by a large amount of test data: test patterns to be
inserted in the scan chain and applied to the circuit, and test
responses captured at the circuitlmodule outputs and then
exported and evaluated externally. The total test application time
(in clock cycles) in scan-based testing is many times larger
compared with the actual number of test patterns, because of the
large numbers of cycles required for the scan-in of a new test
pattern and the scan-out of a captured test response. Test
application time related to scan-in and scan-out phases is getting
larger when the scan chains get longer.

The outcome of the discussion so far is that modem ATPG tools,

supported by structured Dff strategies (scan chains, test points, etc) are
meant to produce, in "reasonable" test generation time, a test set for the
circuit under test that can be applied in "reasonable" test time (therefore, the
test set size should be "sufficiently" smalI) to obtain a sufficiently high fault
coverage of the targeted fault models. It is obvious that the meaning and
value of the words "reasonable" and "sufficient" depends on the type of
circuit under test and the restrictions of a particular application (total
development cost, requested quality ofthe system, etc).
The level of hardware overheads and performance penalties that are
allowed for Dff depends on the required test quality, and have a direct
impact on the test application time itself: more Dff modifications of a circuit
- thus more area overhead and more performance degradation - can lead to
smaller test application time, in other words, to a more easily testable circuit.
ATPG-based test generation time itself may not be a serious concern in
many situations. Even if ATPG-based test generation lasts very long time but eventually leads to a reasonable test set with acceptable fault coverage it is no more than a one-time cost and effort. All the identical manufactured
devices will be tested after manufacturing with the same set of test patterns.

28

Chapter 3 - Testing of Processor-Based SoC

As long as the A TPG-based test generation produces high quality tests with
acceptable hardware and performance overheads, the primary concern is the
test application time, i.e. the portion of the manufacturing phases that each
IC spends being tested. Therefore, it is important to obtain a small test set at
the expense of large test generation time, because a sm all test set will lead to
a smaller test application time for each device.
Test application time per designed chip is a critical factor which has a
certain impact on the production cycle duration, time-to-market and
therefore, to some extend, its market success. Scan-based tests and other DIT
supported test flows may lead to very large test sets that, although reach high
fault coverage and test quality levels, consist of enormous amounts of test
data for test pattern application and test response evaluation. This problem
has already become very severe in complex SoC architectures where scanbased testing consists of: (a) core level tests that are applied to the core itself
in a scan fashion (each core may include many different scan chains); and
(b) SoC level tests that are used to isolate a core (again by scan techniques)
and initiate the testing of the core or test the interconnection between the
cores. The sizes of scan-in and scan-out data for such complex architectures
have pushed the limits of modem Automatic Test Equipment (ATE, Figure
3-1) traditionally used for external manufacturing testing ofICs, because the
memory capacity of A TE is usually not enough to store such huge amounts
oftest data.
Automatie Test Equipment
(A TE, external tester)

Test
Patterns

Test
Responses

ATE memory
fATE

------

IC under
test

f lC = IC frequen cy

=ATE frequency

Figure 3-1: ATE-based testing.

A TE is the main mechanism with which high-volume electronic chips are

tested after manufacturing. "Tester" or "extern al tester" are terms also used
for ATE. An IC under test receives the test patterns previously stored in the
tester memory and operates under this test input. The IC response to each of
the test patterns is captured by the tester, stored back in the tester memory

Embedded Processor-Based Seit-Test

29

and finally compared with the known, correct response. Subsequently, the
next test pattern is applied to the IC, and the process is repeated until all
patterns of the test set stored in the tester memory have been applied to the
chip.
The idea of ATE-based chip testing is outlined in Figure 3-1. The tester
operates in a frequency denoted as f ATE and the chip has an operating
frequency (when used in the field mounted to the final system) denoted as
f re . This means that the tester is able to apply a new test pattern at a
maximum rate of f ATE while the IC is able to produce correct responses when
receiving new inputs at a maximum rate of f rc. The relation between these
two frequencies is a critical factor that determines both the quality of the
testing process with extern al testers and also the test application time and
subsequently the test cast. The relation between these two frequencies is
taken in serious consideration in all cases, independently of the quality and
cost of the utilized ATE (high-speed, high-cost ATE or low-speed, low-cost
ATE). We will study this relation further in this book when the use of lowcost ATE in the context of software-based self-testing will be analyzed.
The essence of the relation between the tester and the chip frequencies is
that if we want to execute a high quality testing to a chip and detect all (or
most) physical failure mechanisms of modem manufacturing technologies,
we must use a tester with a frequency f ATE which is elose or equal to the
actual chip frequency f rc. This means, in turn, that for a high frequency
chip, a very expensive, high-frequency tester must be used and this fact will
increase the overall test and development cost of the IC. A conflict between
test quality and test cost is apparent.
Another cost-related consideration for external testing is the size of the
tester's physical memory where test patterns and test responses are stored. If
the tester memory is not large enough to store all the patterns of the test set
and the corresponding test responses, it is necessary to perform multiple
loadings of the memory, so that the entire set of test vectors is eventually
applied to each manufactured chip. A larger test set requires a larger tester
memory (a more expensive tester) while multiple loadings of the tester
memory mean higher testing costs.
When a multi-million transistor chip is planned to be tested during
manufacturing with the use of an external tester, the size of the tester
memory should be large enough to avoid the application of many loadings of
new chunks of test patterns in the tester memory and the multiple unloading
of test responses from it. In most cases this is really infeasible: a modem
complex SoC architecture can be sufficiently tested (to obtain sufficient fault
coverage and fault quality) only after severalloadings of new test patterns to
tester memory. These multiple loadings lead to a significant amount of time
of the tester being devoted to each of the devices under test. Of course the

30

Chapter 3 - Testing of Processor-Based SoC

overall per-chip test application time can be significantly reduced if parallel

testing of many ICs is performed but this requires the availability of even
more expensive testers with this capability.
Even the use of sophisticated A TPG tools for the cores of a complex SoC
and the utilization of DfT strategies (even with high hardware overhead and
performance impact) are not able to significantly reduce the amount of test
patterns and test responses below some level still consisting of many loadapply-store sessions in the high-end tester. Therefore, the test application
time of complex SoC architectures tends to be excessive and this fact has a
direct impact on the test costs and total chip development costs.
The time that an IC spends during testing under the control of an external
tester adds to its total manufacturing time and final cost. Only high-end,
expensive ATE of our days consisting of a huge number of channels, and a
very high capacity memory for test patterns and test responses storage, and
operating in very high frequencies, are capable to face the testing
requirements of modem complex SoC architectures.
When a complex chip design has to be tested by an external ATE with
test patterns generated by ATPG tools and possibly applied in a scan-based
fashion, several factors must be taken under serious considerations. We
discuss these factors in the following paragraphs in a summary of the
analysis so far. The bottlenecks created by these considerations lead to new
testing strategies, such as self-testing, that will be discussed right after. An
updated discussion of all recent challenges for test technology can be found
in the International Technology Roadmap for Semiconductors [76].

Test cast: test data vo/urne.

Not all electronic applications and not all IC designs can afford
the high cost of a high-end IC tester with high memory capacity
and high frequency. Comprehensive testing of an IC requires the
application of a sufficiently large test set in such a tester. A wide
variety of IC design flows can only apply a moderate complexity
test strategy to obtain a "sufficient" level of test quality at an as
low as possible test development and test application cost. The
test strategy followed in each design and the overall test cost that
can be afforded (including the DfT cost, the test generation cost
and the test application cost: use of the tester) depends on the
production volume. If the volume is high enough then higher test
costs are reasonable, since they are shared among the large
number of chips produced and lead to small per-chip test cost; if,
on the other side, volume is low then a high-end ATE-based
solution is not a cost-effective solution. A tremendous number of

Embedded Processor-Based Self-Test

31

IC designs today have a relatively small production volume to

justify the use of an expensive tester.
Test quality and test effectiveness: at-speed testing.
Even in the case that the costs related to tester purchase,
maintenance, and use are not a concern in high cost, demanding
applications, the test quality level seeked in such applications
can't be easily reached with external testing. This is because high
fault coverage tests may not be easily developed for complex
designs but also because the continuously widening gap between
the tester operating frequency and the IC operating frequency
does not allow the detection of a large percentage of physical
failures in CMOS technology that manifest themselves as delay
faults (instead of logical faults). A very large set of physical
mechanisms that lead to circuits not operating in the target
speeds can only be detected if the chip is tested in the actual
frequency in which it is expected to operate in the target system.
This is called at-speed testing. Even for the best A TE available at
any point of time, there will always be a faster IC in which
performance-related
circuit
malfunctions
will
remain
undetectable 4
Therefore,
the
fundamental
target
of
manufacturing testing - detection of as many as possible physical
defects that may lead the IC to malfunction - can't be met under
these conditions.
Yield lass: tester inaccuracy.
Testers are external devices that perform measurements on
manufactured chips, and thus they suffer from severe
measurement inaccuracy problems wh ich for the high-speed
designs of our days lead to serious production yield loss. A
significant set of correctly operating ICs are characterized as
faulty and are rejected just due to ATE inaccuracies in the
performed measurements. This part of yield loss is added to the
already serious yield loss encountered in VDSM technologies
because of materials imperfections and equipment misses.
Yield lass: avertesting.
Another source of further yield loss is overtesting. Scan-based
testing, as weIl as other DtT techniques, put the circuit in a mode
of operation which is substantially different from its normal
mode of operation. In many cases, the circuit is tested for
potential faults that, even if they exist, they will never affect the

This is simply because, usually, the chips of a technology generation are used in the testers
of the same generation, but these testers are used to test the chips of the next generation.

32

Chapter 3 - Testing of Processor-Based SoC

normal circuit operation. The rejection of chips that have nonfunctional faults (like non-functionally sensitizable path delay
faults) leads to further yield loss in addition to yield los ses due to
tester inaccuracy.
External testing of chips relying on ATE technology is a traditional
approach followed by most high-volume chip manufacturers. Lower
production volumes do not justifY very high testing costs and the extra
problems analyzed above paved the way to the self-testing (or built-in selftesting - BIST) technology which is now well-respected and widely applied
in modem electronic devices as it overcomes several of the bottlenecks of
external, ATE-based testing. Development of effective self-testing
methodologies has always been achallenging task, but it is now much more
challenging than in the past because of the complexity of the electronic
designs to which it is expected to be applied successfully.
In the following two subsections we focus on self-testing in both its
flavors: classical hardware-based self-testing and emerging software-based
(processor-based) self-testing. Software-based self-testing which is the focus
ofthis book is analyzed in detail in subsequent Chapters.

3.2

Hardware-Based Self-Testing

Hardware-based self-testing or built-in self-testing (BIST) techniques

have been proposed since several decades aga [3], [4], [9], [46] to resolve
the bottlenecks that extern al A TE-based testing can not. Self-testing is an
advanced DIT technique which is based on the execution of the testing task
of a chip almost completely intemally, while other DIT techniques although
they modifY the chip's structure they perform the actual testing extemally.
Self-testing does not only provide the means to improve the accessibility of
internal chip nodes, like any other DIT technique, but also integrate the test
pattern generation and test response collection mechanisms inside the chip.
Therefore, the only necessary extern al action is the initiation of the selftesting execution.
The problems of extemal, ATE-based testing, discussed in the previous
subsection, have become much more difficult in the complex modem chip
architectures (ASIC or SoC) and therefore the necessity and usefulness of
self-testing methodologies is today much higher than in the past when
designs were much smaller and more easily testable. Several testability
problems of complex chips, that justifY the extensive use of self-testing, have
been identified in [133] (see also [23]). The factors that make testing (in
particular external testing) more difficult when circuits' complexity and size
increase are:

Embedded Processor-Based Self-Test

33

The increasingly high logic-to-pin (gate-to-pin or transistor-topin) ratio, that severely affects the ability to control and observe
the logic values of internal circuit nodes.
The operating frequencies of chips which are increasing very
quickly. The gigahertz (GHz) frequency domain has been
reached and devices like microprocessors with multi-GHz
operating frequencies are already a common practice.
The increasingly long test pattern generation and test application
times (due to increased difficulty for test generation and the
excessively large test sets).
The extremely large amount of test data to be stored in A TE
memory.
The difficulty to perform at-speed testing with external ATE. A
large population of physical defects that can only be detected at
the actual operating speed of the circuit escape detection.
The unavailability of gate-level netlists and the unfamiliarity of
designers with gate-level details, which both make testability
structures insertion difficult. Especially, in the SoC design era
with the extensive use ofhard, black-box cores, gate-level details
are not easy to obtain.
Lack of skilled test engineers that have a comprehensive
understanding oftesting requirements and testing techniques.

Seif-testing is defined as the ability of an electronic IC to test itself, i.e. to

excite potential fault sites and propagate their effect to observable locations
outside ofthe chip (Figure 3-2). The tasks oftest pattern application and test
response capturing/collection are both performed by internal circuit
resources and not by external equipment as in ATE-based testing.
Obviously, the resources used for test pattern application and test response
capturing/collection should also test themselves and faults inside them must
be detected. In other words, the extra circuit that is used for self-testing must
also be testable. This last requirement always adds extra difficulties to selftesting methodologies.

34

Chapter 3 - Testing of Processor-Based SoC

seit-test
pattern
generation
seit-test
response
evaluation

r+-

module
under
test

IC under test
f 1c = IC trequency

Figure 3-2: Self-testing of an le.

In a self-testing strategy, the test patterns (as weIl as the expected test
responses) are either stored in a special storage area on the chip (RAM,
ROM) and applied during the self-testing session (we call this approach
stored patterns self-testing), or, alternatively, they are generated by special
hardware that takes over this task (we call this approach on-chip generated
patterns self-testing). Furthermore, the actual test responses are either stored
in special storage area on the chip or compressedlcombined together to
reduce the memory requirements. The latter case uses special circuits for test
responses compaction and produces one or a few self-test signatures. In
either case (compacted or not compacted test responses) the analysis that
must eventually take place to decide ifthe chip is fault-free or faulty can be
done inside the chip or out of it. In the extreme case that comparison with
the expected correct response is done internaIly, a single-bit error signal
comes out of the chip to denote its correct or faulty operation. The opposite
extreme is the case where all test responses are extracted out of the chip for
external evaluation (no compaction). The middle case (most usual in
practice) is the one where a few self-test signatures (sets of compacted test
responses) are collected on-chip and at the end of self-test execution are
externally evaluated.
The advantages of self-testing strategies when compared to external,
ATE-based testing are summarized below.

The costs related to the purchase, use and maintenance of highend A TE are almost eliminated when self-testing is used. There
is no need to store test patterns and test responses in the tester
memory, but both tasks oftest application and response capturing
can be performed inside the chip by on-chip resources.
Self-testing mechanisms have a much better access to interna I
circuit nodes than external test mechanisms and can more likely
lead to test strategies with higher fault detection capabilities.
Self-testing usually obtains higher fault coverage than external

Embedded Processor-Based Self-Test

35

testing for a given fault model because a larger number of test

patterns can be usually applied.
Physical failure mechanisms that can only be detected when the
chip operates in its actual frequency can be indeed detected with
self-testing. This is the so-called at-speed testing requirement
and obviously leads to test strategies of higher quality and
effectiveness compared to others that do not detect performancerelated faults (transition faults, path delay faults, etc). The
superiority of self-testing at this point makes it in most cases the
only choice, because in today' s very deep submicron
technologies an increasingly large number of failure mechanisms
can only be detected by at-speed testing.
There is no yield loss due to external testers' measurement
inaccuracies simply because the chip tests itself and the test
responses are captured by parts of the same silicon piece. This is
a very serious concem today in the large sized chips being
manufactured, which already suffer from serious yield loss
problems because of their size and manufacturing process
imperfections. Tester inaccuracy problems that lead to further
yield loss are avoided by the utilization of appropriate selftesting techniques. The yield loss that is due to overtesting
(testing for faults that can't affect the normal circuit operation)
may still exist in self-testing, if the self-testing methodology tests
the chip in not normal operation modes.
The on-chip resources built for hardware-based self-testing can
be re-used in later stages of the chip's life cycle, while
manufacturing testing based on external equipment can not do
that. A manufacturing testing strategy based on external testing is
used once for each chip. On the other side, an existing selftesting mechanism that is built inside the device is an added
value for the product and can be employed later during the
circuit's normal operation in the field, to detect faults that can
appear because of aging of the device or several environmental
factors. This is known as periodic/on-line testing or in-field
testing.

The use of self-testing in today's complex Ies offers significant reduction

in test application time. If a large amount of test patterns is externally
applied from a tester they will be applied to the chip at the tester's lower
frequency and will also require multiple loadings ofthe tester memory which
will add further delays in test application for each chip. When, on the other
side, an equally large test set is applied to a chip using self-testing, not only

36

Chapter 3 - Testing of Processor-Based SoC

multiple loading sessions are avoided, but also all test patterns are applied at
the chip's actual frequency (usually higher than the tester frequency) and a
higher test quality is achieved. In external, A TE-based testing, performancerelated faults (delay/transition faults) may remain undetected because ofthe
frequency difference between the chip and the tester, and serious portion of
yield may be lost because of the tester's limited measurement accuracy. An
implicit assumption made for the validity of the above statement which
compares self-testing with external testing, is that the same test access
mechanisms (like scan chains, test points insertion and other Dff means) are
used in both cases. Only under this assumption, comparison between
external testing and self-testing is fair.
Self-testing methodologies are in some situations the only feasible testing
alternative strategy. These are cases where access to expensive ATE is not
possible at all or when test costs associated with the use of ATE are out of
question for the budget of a specific design. Dff modifications to enable
self-testing may be more reasonable for the circuit designers compared with
the excessive external testing costs. Many such cases exist today, for
example in low-cost embedded applications where a reasonably good test
methodology is needed to reach a relatively high level oftest quality, but on
the other side, no sophisticated solutions or expensive ATE can be used
because of budget limitations. Self-testing needs only an appropriate test
pattern generation flow and the design of on-chip test application and test
response collection infrastructure. Hardware and performance overheads due
to the employment of these mechanisms can be tailored to the specific cost
limitations of a design.
Hardware-based self-testing although a proven successful testing
technology for different types of small and medium sized digital circuits, is
not claimed to be a panacea, as a testing strategy, for all types of
architectures. A careful self-testing strategy should be planned and applied
with careful guidance by the performance, cost and quality requirements of
any given application. More specifically, the concerns that a test engineer
should keep in mind when hardware-based self-testing is the intended testing
methodology, are the following.

Hardware overhead.
This is the amount or percentage of hardware overhead devoted
to hardware-based self-testing which is acceptable for the
particular design. Self-testing techniques based on scan-based
architectures have been widely applied to several designs. In
addition to the hardware overhead that the scan design (either full
or partial) adds to the circuit (regular storage elements modified
into scan storage elements), or the extra multiplexing for internal

Embedded Processor-Based Seit-Test

37

nodes access, hardware-based self-testing requires additional

circuits to be synthesized and integrated in the design for the test
pattern generation, the test pattern application and the test
response collection tasks.
A usual tradeoff that a test engineer faces when applying a
hardware-based self-testing technique is whether he/she will
have the test patterns stored in a memory unit on the chip or if
the test patterns will be generated by some on-chip machine
specially designed for this purpose. The final decision strongly
depends on the number of test patterns of the test set. If the test
set consists of just a few test patterns, they can be stored in a
small on-chip memory and thus the hardware overhead can be
small. Otherwise, if the number of test patterns is large and their
storage in an on-chip memory is not a cost-effective solution
(this is the most common situation), then it is necessary that a
small, "clever" on-chip sequential machine generates this large
number of test patterns (in many situations necessary, for
example in pseudorandom based self-testing) with a much
smaller hardware overhead compared to the memory storage
case. We discuss these two alternatives in the following.
Performance degradation.
This is the performance impact that the self-testing methodology
is allowed to have on the circuit. In many cases, carefully
designed and optimized high-performance circuits can't afford
any change at a11 in their critical paths. In this case, hardwarebased self-testing should be carefully applied to meet this strict
requirement and have minimal (or, ideally, zero) impact on
circuit performance during normal operation. The only way to do
self-testing in such cases without any impact on the circuit's
performance is to leave the critical paths of the design unaffected
by the DfT changes required to apply self-testing. This is areal
cha11enge since critical paths usually contain difficult to test parts
that need DfT modifications for testability improvement.
Power consumption.
This is the additional power that the circuit may consume during
hardware-based self-testing. It is a critical factor when the circuit
is self-tested in a not normal operation mode, Le. paths not
activated in normal circuit operation are activated during selftesting intervals. Increased power consumption during selftesting is a usual side-effect when scan-based testing is applied
and in cases that pseudorandom test patterns are used to excite
the faults in a circuit. Therefore, when power consumption really

38

Chapter 3 - Testing of Processor-Based SoC

matters, scan-based self-testing and pseudorandom patterns

based self-testing are not really good candidates. Power
consumption is a serious concern in power-sensitive, batteryoperated systems when chips are tested in the field using existing
self-testing mechanisms. In this case, excessive power
consumption for self-testing reduces the effective li fe cycle of
the system's battery.
As we have already mentioned in the first paragraphs of this subsection,
self-testing can be applied either using a dedicated memory that stores the
test patterns as weIl as the test responses collected from them, or can be
applied by a dedicated hardware test pattern generation machine and a
separate test response analysis machine. These two different self-testing
configurations are depicted in Figure 3-3 and Figure 3-4, respectively.
test
patters

under
test

test
responses

seil-test
memory

JC under test

Figure 3-3: Self-testing with a dedicated memory.

In hardware-based self-testing with a dedicated memory (Figure 3-3) the

actual application of the test patterns is performed by part of the circuit
which reads the test patterns from the memory, applies them to the module
under test and collects the test responses back to memory. The self-test
memory is a dedicated memory which is only used for self-testing purposes
and not the main chip's memory.
In hardware-based self-testing with dedicated hardware (Figure 3-4) the
test patterns do not pre-exist anywhere, but are rather generated on-the-fly
by a self-test pattern generation special circuit. Every test pattern generated
by this circuit is immediately applied to the module under test and the
module's response is collected and driven to another special circuit that
performs test response analysis. All test reponses of the module under test
are compacted by the response analyzer and a final self-test signature is
eventually sent out ofthe chip for external evaluation.

39

Embedded Processor-Based Seit-Test

on-chip
test
generator

I
gnder

module

on-chip
response
analyser

test

le under test
Figure 3-4: Self-testing with dedicated hardware.

A hardware-based self-testing strategy can be based on deterministic test

patterns or pseudorandom test patterns. Usually, in the deterministic case the
total number of test patterns is some orders of magnitude smaller than in the
pseudorandom case but the on-chip test generation of the pseudorandom
patterns is easier and performed by smaller circuits (calIed pseudorandom
patterns generators) compared with the deterministic case.
Deterministic test patterns can be either previously generated by an
ATPG tool (ifthe gate level information ofthe circuit under test is available)
or may be, in general, previously computed or known test patterns for a
module of the circuit under test. For example, several types of arithmetic
circuits can be comprehensively tested with known pre-computed small test
sets (even of constant size independently of the word length of the arithmetic
circuit5). In general, deterministic test patterns are relatively few and
uncorrelated/irregular. They can be stored in on-chip memory and applied
during self-testing sessions as shown in Figure 3-3. The expected circuit
responses are also stored in on-chip memory and the actual response of the
circuit under test can be compared inside the chip with the expected faultfree responses to obtain the final pass/fail result of self-testing. The size of
the dedicated on-chip memory for storing test patterns and test responses is a
critical factor for deterministic self-testing and for this reason, the total
number of test patterns must be rather small in order for the approach to be
applicable. Such cases are rather rare in complex circuits.
On the other side, pseudorandom self-testing uses on-chip pseudorandom
test patterns generators like Linear Feedback Shift Registers (LFSR) [1],
[23], or Cellular Automata (CA) [1], [23]. Arithmetic circuits such as adders,
subtracters and multipliers have also been shown to be successful candidates
for pseudorandom pattern generation during hardware-based self-testing

If a circuit that operates on n-bit operand(s) can be tested with a test set of constant size
(number oftest patterns) which is independent on n, we call the circuit C-testable and the
test set aC-test set.

40

Chapter 3 - Testing of Processor-Based SoC

[133] because they produce pseudorandom number sequences with very

good properties.
Test sets applied to circuits during pseudorandom self-testing consist of
large numbers of test patterns and it usually takes quite a long test
application time to reach a sufficient fault coverage (if it is at all feasible).
Pseudorandom self-testing is an aggressive self-testing approach that does
not rely on any algorithm for test generation but rather on the inherent ability
of some circuits to be easily tested with long, random test sequences.
Unfortunately, there are many circuits that are not random testable but on the
contrary are strongly random-pattern resistant, i.e. high fault coverage can't
be reached for them even with very long test sequences. Even in cases that
relatively high fault coverage can be obtained by a first set of random
patterns, the remaining faults of the circuit that should be tested to reach
acceptable fault coverage are very difficult to be detected by pseudorandom
test sequences.
The major advantage of pseudorandom-based self-testing is that on-chip
circuits that generate the test patterns are very small and thus hardware
overhead is very sm all too. For example, LFSRs, Cellular Automata, or
Arithmetic pseudorandom patterns generators, such as accumulators, consist
of simple modifications to existing circuit registers and thus have a
minimum impact on circuit area.
A well-known strategy for the improvement of the efficiency of
pseudorandom testing is the one where a few deterministic test patterns
capable to detect the random-pattern resistant faults of a circuit are carefully
embedded in long pseudorandom sequences. The resulting, "enhanced"
pseudorandom test sequence reaches high fault coverage in a relatively
smaller time than in pure pseudorandom testing. This improved capability is
counterbalanced by the extra hardware area required for the embedding of
the deterministic patterns in the pseudorandom sequence.
When the discussion comes to hardware-based self-testing of complex
ASICs or embedded core-based SoC architectures, an ideal self-testing
scenario is looked for. Such a scenario would be one that combines the
benefits of deterministic self-testing and pseudorandom self-testing. This
ideal solution would apply a relatively small set of test patterns capable to
quickly obtain acceptable high fault coverage. The test patterns of the set
would be also related to each other, so that they can be efficiently generated
on-chip by sm all hardware resources and need not be stored along with their
test responses in on-chip memory. In other words, the highest possible test
quality and effectiveness is seeked with the smallest possible test cost.
We will have the opportunity to elaborate more on the topic of test cost
through the pages of this book and show how software-based self-testing,
introduced in the next subsection (also called processor-based self-testing),

Embedded Processor-Based Self-Test

41

can be a low-cost but also high-quality self-testing methodology primarily

for embedded processors but also for processor-based SoC architectures. A
detailed discussion of software-based self-testing for processors is given in
Chapter 5 and further.

3.3

Software-Based Self-Testing

Classical hardware-based self-testing techniques have the limitations

described in the previous subsection. There are many situations where the
hardware or the performance overheads that a hardware-based self-testing
technique adds are not acceptable and go beyond the restrictions of the
design. There are also situations where self-testing applied with the use of
scan-based architectures leads to excessive power consumption because the
chip is tested in a special mode of operation which is different from the
normal operation for which it has been designed. It is likely that a circuit
designer is reluctant to adopt even the smallest design change for hardwarebased self-testing which, although it is beneficial for the circuit's testability,
it will also affect the original design's performance, size and power
consumption.
The existence of embedded processors in core-based SoC architectures
opens the way to an alternative self-testing technique that has great potential
to be very popular among circuit designers. This alternative is known as
software-based self-testing or processor-based self-testing and is the focus of
this book.
In software-based self-testing the test generation, test application and test
response capturing are all tasks performed by embedded software routines
that are executed by the embedded processor itself, instead of being assigned
to specially synthesized hardware modules as in hardware-based self-testing.
Processors can, therefore, be re-used as an existing testing infrastructure for
manufacturing testing and periodic/on-line testing in the field. Softwarebased self-testing is a "natural", non-intrusive self-testing solution where the
processor itself controls the flow of test data in its interior in such a way to
detect its faults and no additional hardware is necessary for self-testing.
The inherent processing power that embedded processors lend to SoC
designs, allows the application of any flavor and algorithm of self-testing
like deterministic self-testing, pseudorandom self-testing or a combination of
them. In software-based self-testing, the embedded processor executes a
dedicated software routine or collections of routines that generate a sequence
of test patterns according to a specific algorithm. Subsequently, the
processor applies each of the test patterns of the sequence to the component

42

Chapter 3 - Testing of Processor-Based SoC

under test6 , collects the component responses and finally stores them either
in an unrolled fashion (each response is stored in aseparate data memory
word) or in a compacted form (one or more test signatures). In a multiprocessor SoC design, each of the embedded processors can test itself by
software routines and then they can then apply software-based self-testing to
the remaining cores of the SoC.
The concept of software-based self-testing for the processor itself is
illustrated in Figure 3-5. Self-test routines are stored in instruction memory
and the data they need for execution are stored in data memory. Both
transfers (instructions and data) are performed using external test equipment
which can be as simple as a personal computer and as complex as a high-end
tester. Tests are applied to components of the processor core (CPU core)
during the execution of the self-test programs and test responses are stored
back in the data memory.

I data ~ self-test
code
Iresponse(s)
seit-test J
self-test

CPU core

1
CPU bus

Data Memory

Instruction Memory

External Test
Equipment
Figure 3-5: Software-based self-testing concept for processor testing.

It is implied by Figure 3-5 that the external test equipment has access to
the processor bus for the transfer of self-test programs and data in the
processor memory. This is not necessary in all cases. In general, there should
be a mechanism that is able to download self-test program and data in the
processor memory for the execution of the software-based self-testing to
detect the faults ofthe processor.
As a step further to software-based self-testing of a processor, the
concept of software-based self-testing for the entire SoC design is depicted
in Figure 3-6. The embedded processor core supported with appropriately

The component under test may be either an internal component ofthe processor, the entire
processor itself or a core of the SoC other than the processor.

Embedded Processor-Based Self-Test

43

developed software routines, is used to test other embedded cores of the

SoC. In a SoC, the embedded processor (or processors) have very good
access to all cares of the SoC and therefore can access, in one way or the
other, the inputs and outputs of every core. In Figure 3-6, it is again implied
that there is a mechanism for the transfer of self-test code and data in the
processor memory for further execution and detection of the core faults.
seit-test
patterns!
responses

CPU core

Data
Memory
seit-test
code

CPU subsystem

apply test !

Instruction
Memory

capture respons e

Core under test

Figure 3-6: Software-based self-testing concept for testing a SoC core.

Software-based self-testing is an alternative methodology to hardwarebased self-testing that has the characteristics discussed here. At this point,
we outline the overall idea of software-based self-testing that clearly reveals
its low-cost aspects. Appropriately scaled to different processor sizes and
architectures, software-based self-testing is a generic SoC self-testing
methodology.

Software-based self-testing is a non-intrusive test methodology

because it does not rely on DIT modifications of the circuit
structure (the processor or the SoC). It does not add hardware
and performance overheads, but on the contrary it tests the circuit
using software routines executed just like all normal programs of
the processor. Software-based self-testing does not affect at all
the original circuit structure and does not require any

44

Chapter 3 - Testing of Processor-Based SoC

modifications in the processor instruction set architecture (ISA)

or the carefully optimized processor design.
Software-based self-testing is a low cost test methodology
because it does not rely on the use of expensive external testers
and does not add any area, delay or power consumption
overheads during normal operation. Low-cost, low-speed, lowpin-count testers can be perfectly utilized by software-based selftesting during manufacturing testing of a processor or an SoC,
simply to download self-test routines to the processor's on-chip
memory (ifthese routines are not already permanently stored in a
flash or ROM memory), and to upload test responses or test
signatures for external evaluation (if this is necessary). If the
self-test program is sufficiently small then the overall test
application time is minimally affected by the downloading and
uploading times performed at the lower speed of the tester. The
actual test application phase - i.e. the self-test code execution will be performed at the normal operating speed of the processor
which is usually much higher than the tester's speed. The
usefulness of this self-testing scenario to low cost applications
and low volume production flows is more than apparent. The
chip manufacturing testing is not any more hooked to an
expensive tester but only to low-cost equipment that simply
transfers small amounts of embedded code, test patterns and test
responses to and from the processor memory.
Software-based self-testing performs at-speed testing of the
circuit, because all test patterns are applied at the actual
frequency that the circuit performs during its normal operation.
This is a general characteristic of all self-testing strategies
(hardware-based or software-based). Therefore, all physical
failures can be detected no matter how they alter the functionality
of the circuit (logic faults) or its timing behavior (delay or
performance faults). Therefore, the resulting test quality is very
high.
Software-based self-testing is a low power test methodology
because the circuit is never excited by tests that never appear
during normal operation. Software-based self-testing applies test
patterns that can be only delivered with normal processor
instructions and not during any special test mode (like scan).
Therefore, the consumed average electrical power during the
execution of the self-testing pro grams does not differ from the
average power consumption of the chip during normal operation.
If the total duration of the self-testing interval is short enough

Embedded Processor-Based Self-Test

45

then the total power consumed during self-testing will have a

minimal impact on the power behavior of the chip. This is a
particularly important aspect of software-based self-testing when
it is used during the entire life cycle of the circuit for on-line
periodic testing in the field. Excessive power consumption in
self-testing periods is a serious problem in battery-operated,
hand-held and portable systems, but not if software-based selftesting is used.
Software-based self-testing is a very flexible and programmable
test strategy for complex systems because simple software code
modifications or additions are sufficient to extent the testing
capability to new components added to the SoC, change the
processor's target fault model to another one that needs more test
patterns (seeking for even higher defect coverage), or even
change the purpose of execution of the routines from testing and
detection of physical faults, to diagnosis and localization in order
to identify the location of malfunctioning areas and cores of the
design.
The flexibility of software-based self-testing is not available in
hardware-based self-testing where fixed hardware structures are
only able to apply specific sets of patterns and no changes can be
done after the chip is manufactured 7 The software-based selftesting flexibility is a result of the programmability of the
embedded processors and the accessibility they have to all
internal components and cores of the SoC. The only factor that
may block the extension and augmentation of a self-testing
program is the size of the embedded memory in which it is
stored. If software-based self-testing is used in manufacturing
testing only, the memory size is not a problem because the
normal on-chip memory (cache memory or RAM) can be used
for the storage of the pro grams and the data. For other
applications such as peridodic/on-line testing in the field, a
dedicated memory (ROM, flash memory, etc) can be embedded
for use of the software-based self-testing process and may need
to be expanded or replaced if new or larger self-test routines are
added to the system.

As we see in the next Chapter a number ofhardware-based and softwarebased self-testing approaches have been proposed in the past along with
7

Minor flexibility may apply to hardware-based self-testing, such as the loading of different
seeds in pseudorandom patterns generators (LFSR).

46

Chapter 3 - Testing of Processor-Based SoC

several external testing approaches for proeessors arehiteetures. The targets

of eaeh approach are diseussed and the works are c1assified aeeording to
their eharaeteristies.
We eonc1ude the present Chapter in the following seetions by giving
detailed answers to three important questions that summarize the motivation
of research in the area of proeessor self-testing and in particular softwarebased self-testing.
The answers that are given to these questions justify the importanee of
software-based self-testing for today's proeessor and proeessor-based SoCs.
The questions are:

How ean software-based self-testing be an effeetive test resource

partitioning teehnique that reduees all eost faetors of
manufaeturing and field testing of a proeessor and an IC?
Why is embedded proeessor testing important for the overall
quality ofthe SoC that eontains it?
Why is embedded proeessor testing and proeessor-based testing
diffieult and ehallenging for a test engineer?

The three following subseetions elaborate on these three items.

3.4

Software-Based Self-Test and Test Resource

Partitioning

Test resource partitioning (TRP) is a term whieh was reeently introdueed

[72] to deseribe the effeetive partitioning of several different types of test
resourees for IC testing (see also [25]). Test resourees that must be
eonsidered for effeetive partitioning inc1ude:

hardware resourees (A TE hardware and built-in hardware

dedicated to testing);
time resourees (test generation time and test application time);
power resourees (power eonsumed during manufaeturing and online testing in the field);
pin resourees (number ofpins dedieated to testing), ete.

Software-based self-testing using embedded proeessors is an effeetive

TRP approach when several of these test resourees are to be optimized, and
this is one of its major advantages. This is the reason why we separately
diseuss it in this seetion and we elaborate on how software-based self-testing
optimizes several test resourees.
When hardware is the test resouree to be optimized, it refers both to
external equipment used far testing (ATE hardware and memory) and to
built-in hardware dedieated to testing. Software-based self-testing relaxes the

Embedded Processor-Based Seit-Test

47

close relation of testing with high-cost A TE and reduces the external test
equipment requirements to low-cost testers only for downloading of test
programs to on-chip memory and uploading of final test responses from onchip memory to the outside (the tester) for external evaluation.
In terms of special built-in hardware dedicated to test, software-based
self-testing is a non-intrusive test method that does not need extra circuits
just for testing purposes. On the contrary, it relies only on existing processor
resources (its instructions, addressing modes, functional and control units)
and re-uses them for testing the chip (either during manufacturing or during
on-line periodic testing in the field). Therefore, compared to hardware-based
self-testing that needs additional circuitry, software-based self-testing is
definitely a better TRP methodology.
When test application time is the test resource to be optimized, softwarebased self-testing has a significant contribution to this optimization too. Selftesting performed by embedded software routines is a fast test application
methodology for two simple reasons:
(a)

(b)

testing is executed at the actual operating speed of the chip and

not at the slower speed of the external tester (this does not only
decreases the test application duration but also improves the test
quality);
no scan-in and scan-out cycles are required as in other hardwarebased self-testing techniques that add significant delays in the
testing ofthe device.

Test time also depends on the nature and the total number of the applied
test patterns: in pseudorandom testing the number of test patterns is much
larger than in the deterministic one. Software-based self-testing that applies
deterministic test patterns executes in shorter time than in the case of
pseudorandom-based self-testing.
When power consumption is the test resource under optimization,
software-based self-testing never excites the processor or the SoC in other
mode than its normal operating mode for which it is designed and analyzed.
Other testing techniques like external testing using scan design or hardwarebased self-testing with scan design again or test points insertion, test the
circuit during a special mode of operation (test mode), which is completely
different from normal mode. During test mode, circuit activity is much
higher than in normal mode and therefore more power is consumed. It has
been observed that circuit activity (and thus power consumption) during
testing and self-testing can be up to three times higher than during normal
mode [118], [176]. Apart from this, excessive power consumption during
testing leads to energy problems in battery operated products and mayaiso

48

Chapter 3 - Testing of Processor-Based SoC

stress the limits of device packages when peak power exceeds specific
thresholds.
Finally, when pin count is the test resource considered for optimization,
software-based self-testing again provides an excellent approach. Embedded
software test routines and data that just need to be down loaded from a lowcost extemal tester, only require a very small number of test-related chip
pins. An extreme case is the use of existing JTAG boundary scan interface
for this purpose. This serial downloading of embedded self-test routines may
require more time than in the case of parallel download but if the routines
are sufficiently small, this is not a serious problem. Self-test routines size is a
metric that we will extensively consider when discussing software-based
self-testing in this book.

3.5

Why is Embedded Processor Testing Important?

Embedded processors have played for a long time a key role in the
development of digital circuits and are constantly the central elements in all
kinds of applications. Processors are today even more important because of
their increasing usage in all developed embedded systems. The answer to the
question of this subsection (processor testing importance) seems to be really
easy but all details must be pointed out.
The embedded processors in SoC architectures are the circuits that will
execute the critical algorithms of the system and will co-ordinate the
communication between all the other components/cores. The embedded
processors are also expected to execute the self-test routines during
manufacturing testing and in the field (on-line periodic testing), as weIl as,
other functions like system debug and diagnosis.
As a consequence, the criticality and importance of processor testing is
equivalent to the criticality and importance of its own existence in a system
or a SoC. When a fault appears in an embedded processor, for example in
one of its registers, then all programs that use this specific register (maybe
all pro grams to be executed) will malfunction and give incorrect results.
Although the fault exists only inside the processor (actually in a very small
part of it), the entire system is very likely to be completely useless in the
field because the system functionality expected to be executed in the
processor will give erroneous output.
Other system components or SoC cores are not as critical as the processor
in terms of correct functionality of the system. For example, if a memory
word contains a fault, only writes and reads to this specific location will be
erroneous, and this will lead to just a few programs (if any program uses this
memory location at a specific point of time) to malfunction. This may not be
of course a perfectly performing system but the implication of a fault in the

Embedded Processor-Based Self-Test

49

memory module does not have so catastrophic results in the system as a fault
in the embedded processor. The same reasoning is true for other cores like
peripheral device controllers. If a fault exists in some peripheral device
controller then the system may have trouble to access the specific device but
it will be otherwise fully usable.
The task of embedded processor testing is very important because if the
processor is not free of manufacturing faults it can 't be used as a vehicle for
the efficient software-based self-testing of the surrounding modules, and as a
result the entire process of software-based self-testing will not be applicable
to the particular system.
The importance of comprehensive and high quality processor testing is
not at all related either with the size of an embedded processor used in a SoC
architecture or its performance characteristics. A small 8-bit or 16-bit
microcontroller is equally important for a successful software-based selftesting strategy when compared with a high-end 32-bit RISC processor with
an advanced pipeline structure and other built-in performance-improving
mechanisms. Both types of processors must be comprehensively tested and
their correct operation must be guaranteed before they are used for selftesting the remaining parts of the system.
Apparently, the equal importance of all types and architectures of
embedded processors for the purposes of software-based self-testing does
not mean that all embedded processor architectures are tested with the same
difficulty.

3.6

Why is Embedded Processor Testing Challenging?

Testing or self-testing of a processor's architecture before it can be used

to perform self-testing of other components in the software-based self-testing
flow, is achallenging task. From many points of view, processor
architectures symbolize the universe of all design techniques [122]. All
general and special design techniques are used for the components of a
processor with the ultimate target to obtain the best possible performance
under additional design constraints like circuit size or power consumption.
For example, in many cases, the best performance is seeked under the
restriction that the processor circuit size does not exceed a specific limit
(probably imposed by chip packaging limitations and costs). In other cases,
the limitation comes from the maximum power that can be consumed in the
target applications and this is the factor that determines an upper bound to
the achievable performance of a processor design. Such upper limits are
usually set by the cost of the available cooling or heat rem oval mechanisms.
A processor is not a simple combinational unit, nor a simple finite state
machine that implements astate diagram. A processor uses the best design

50

Chapter 3 - Testing of Processor-Based SoC

techniques for each of its components (arithmetic units, storage elements,

interconnection modules) and the best available way to make them work
together under the supervision of the control unit to obtain the best
performance at the instruction execution level and not the function level.
This means that design optimization techniques in processors do not
primarily focus on the optimization of each function (although this is useful
too in most cases) but on the overall delivered performance ofthe processor.
Embedded processor testing and self-testing techniques are applied with
several difficulties and restrictions due to optimized architecture of
processors wh ich allow only marginal (if any) changes in circuit and
marginal impact on performance and power consumption. We discuss these
limitations in the following. The discussion actually puts in perspective the
problem of processor testing and self-testing by setting the
properties/requirements that a successful processor testing technique must
satisfy. The same properties/requirements are valid not only for self-testing
of the processor itself but also when the processor is used to ron self-testing
on the remaining components of a complex SoC. Note that, software-based
self-testing meets these requirements as we analyze in this book.

Embedded processors are well-optimized designs in terms of

performance and therefore the performance degradation that
structured Dff (like scan design) or ad-hoc Dff techniques
impose is not acceptable in most practical cases. Software-based
self-testing satisfies this requirement for zero performance
overhead, being a non-intrusive self-testing strategy that does not
require any circuit modifications.
Embedded processors are carefully optimized with respect to
their size and gates count so that they occupy as sm all as possible
silicon area when integrated in an SoC. Software-based selftesting also satisfies the requirement of virtually zero hardware
overhead, because self-test is performed by the execution of seIftest routines utilizing the processor internal resources for test
generation and response capturing.
Power consumption is a critical parameter in embedded
processors designed for low-power applications used in handheld, portable, battery operated devices. Software-based selftesting has the characteristic that it tests the processor and the
overall system during normal operation mode. No special test
mode is used and the average power consumption is the same as
in normal circuit operation. The additional requirement is that a
self-test pro gram must have as short as possible test execution
time so that the total consumed power does not have significant

Embedded Processor-Based Self-Test

51

affect on the battery charge available for the system. Of course

this last requirement applies only for in-field test applications
and on-line testing and not for manufacturing testing of the
processor and the SoC.
The self-test program used to test the embedded processor and
the other SoC cores must have as sm all as possible size (both
code and data) because it determines the amount of on-chip
memory that is necessary for the storage of self-test code and
data (test patterns and/or test responses). During manufacturing
testing the relation between the self-test pro gram size and the
available on-chip memory determines the downloading sessions
that are necessary to execute the self-testing using a low cost
tester. If the self-test program size is large and multiple loadings
of the memory are necessary then the test application time for
manufacturing testing will be much longer. On the other side,
during on-line testing the self-test pro gram size has a direct
impact on the system cost because a larger memory (ROM, flash,
etc) necessary for self-test program storage will increase the
system size and cost. The reduction of the self-test program size
is a primary goal in software-based self-testing.
Apart from determining the memory requirements of a softwarebased self-testing technique, the size of a self-test program
specifies a significant portion of the total test application time (in
manufacturing testing) because it is directly connected to the
downloading time from the low-cost, low-speed tester to the onchip memory. This time may be larger than the time for the
execution of the self-test program because downloading is done
in the low frequency of the tester while the program is executed
in the higher frequency of the chip. Of course, the actual
difference between the two frequencies give a beUer idea of
which time is most important for the total test application time of
the chip. As a result, the size of the self-test program is a crucial
factor both for the memory requirements of the design and for
the test application time and is one ofthe parameters that must be
carefully examined and evaluated in every software-based selftesting approach.
Another parameter related to the duration of the test application
per tested chip is the execution time of the self-test pro gram after
it has been downloadedin the on-chip memory. If the difference
between the tester frequency and the chip frequency is not so
large then the test execution time is very important for the overall
test application time. The big difference exists in cases that a

52

Chapter 3 - Testing of Processor-Based SoC

software-based self-testing approach applies a few deterministic

test patterns and another approach applies a large set of
pseudorandom ones. In the latter case the test execution time is
much larger than in the former while, usually, smaller fault
coverage is achieved.
Last but not least of course, is the requirement for high test
quality and high fault coverage. The target fault coverage for a
software-based self-testing strategy must be as high as possible
for the embedded processor itself first and then for the remaining
SoC cores. An embedded processor (small, medium or large)
should be tested up to a very high level of fault coverage to
increase the confidence that it will operate correctly when used.
This requirement is in some sense mandatory because, as we
have already mentioned above, a malfunctioning processor will
lead to almost all programs working incorrectly and the system
being useless. Therefore, a fault that escapes manufacturing
testing in a processor is much more important than a fault in a
memory array or other SoC component and the obtained fault
coverage for the processor itself must be larger than at the other
components.

We note that processor faults that are not functionally detectable i.e. they
can 't be detected when circuit operates in normal mode, will not be detected
by software-based self-testing. This is natural, since software-based selftesting only applies normal processor instructions for fault detection.
To give an idea of the criticality and difficulty of processor testing in
software-based self-testing it is useful to outline a small but informative
example. Consider a SoC consisting of an embedded processor that occupies
only 20% of the total silicon area, several memory modules occupying 70%
of the total area (usual situation to have large embedded memories) and
other components for the remaining 10% of the SoC area.
First, it is much more important to develop a self-testing strategy that
reaches a fault coverage of more than 90% or 95% for the processor than for
the embedded memories although the processor size is more than three times
smaller that the memories. As we explained earlier, faults in the processor
that escape detection will lead to the vast majority of the programs to
malfunction (independently of the exact location of the fault in the
processor: it may be a register fault, a functional unit fault or a fault in the
control unit). On the contrary, faults that escape detection in any of the
memory cores will only lead to a sm all number of pro grams not operating
correct1y. Moreover, one should not forget that the overall idea of processorbased self-testing or software-based self-testing for SoC architectures can be

Embedded Processor-Based Self-Test

53

useful and operational only when a comprehensively tested embedded

processor exists in the system and it is found to be fault-free.
Secondly, apartfrom its importance, processor testing in our sm all
example is much more difficult compared to memory self-testing. It is wellknown that large memory arrays can be tested for very comprehensive
memory fault models with small hardware machines (hardware-based selftesting, memory BIST) that produce the test patterns and compress the test
responses. Memory self-testing is a successful, well-proven technology for
many years now. Memory test algorithms can also be applied in the
software-based self-testing framework since the embedded processor has
very good access to all embedded memories and is capable to apply the
necessary test sequences to them. In the above sense, therefore, although the
memory arrays occupy more than three times larger area compared to the
processor, they are much more easily tested than the processor.
The combination of the importance and difficulty of testing an embedded
processor and subsequently use it for testing the rest of the system, reveals
the need for the development of effective software-based self-testing
techniques that fulfill to the maximum extent the requirements described
above.

Chapter

4
Processor Testing Techniques

Intensive research has been performed since the appearance of the first
microprocessor in the field of processor testing. A variety of generic
methodologies as weH as several ad hoc solutions has been presented in the
literature. In this Chapter we provide an overview of the open literature in
processor testing, putting emphasis on the different characteristics of the
approaches and the requirements that each of them tries to meet.
This Chapter consists of two parts. In the first part, we discuss the
characteristics of a set of different classes of processor testing techniques
(external testing vs. self-testing; functional testing vs. structural testing; etc)
along with the benefits and drawbacks of each one. In the second part, we
briefly discuss the most important works in the area in a chronological order
of publication. The Chapter concludes with the linking of the two parts,
where each of the works presented in the literature is associated with the one
or more classes that it belongs to aiming to provide a quick reference for
those that are interested to study the area of processor testing.

4.1

Processor Testing Techniques Objectives

Each processor testing technique has different objectives and restrictions

depending on the application in which it is used and the quality and budget
constraints that should be met.
Embedded Processor-Based Self-Test
D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic Publishers, 2004

56

Chapter 4 - Processor Testing Techniques

In the following sections, we elaborate on the different classes that a

processor testing methodology may belong to, describing their
characteristics and explaining the cases in which each of them is considered
to be an effective solution. As in any other digital system testing case, there
is no single solution that is applicable to all processor architectures. In many
cases, more than one testing strategies are combined to provide the most
efficient and suitable testing solution for a particular system's configuration.
The testing strategy that is eventually applied to a processor depends on the
specific processor's architecture and its instruction set, the characteristics of
the particular SoC where the processor is embedded in and the system design
and test cost constraints.
The classification of the variety of processor testing techniques in
different classes with different objectives provides a systematic process for
selecting the appropriate technique for a specific system configuration. This
can be done by matching the benefits of each class with the particular
requirements (test time, test cost, overheads, etc) of the particular
application. In this sense, the purpose of this Chapter is to be a
comprehensive survey on processor testing techniques and outline the
contribution of each approach.
4.1.1

External Testing versus Self-Testing

External testing of a processor (or any IC) means that test patterns are
applied to it by an external tester (ATE). The test patterns along with the
expected test responses have been previously stored in the ATE memory.
This is the classical manufacturing testing technique used in digital circuits.
Functional test patterns previously developed for functional verification can
be re-used in this scenario, potentially enhanced with ATPG patterns to
increase the fault coverage.
On the other side, self-testing of a processor means that test patterns are
applied to it and test responses are evaluated for correctness without the use
of external A TE, but rather using internal resources of the processor. Internal
resources may be either existing hardware and memory resources, or extra
hardware particularly synthesized for test-related purposes (on-chip test
generation and response capturing).
The benefits and drawbacks of external testing and self-testing of
processors are common to those of any other digital circuit and are
summarized in Table 4-1.

57

Embedded Processor-Based Self-Test

Benefits
Small on-chip
hardware overhead
Small chip
performance impact
Self Testing
At-speed testing
Low-cost ATE
Re-usable during
product's life cycle
Table 4-1: External testing vs. self-testing.
External Testing

4.1.2

Drawbacks
Not at-speed testing
High-cost ATE
Only for manufacturing
testing
Hardware overhead
Performance impact

DIT-based Testing versus Non-Intrusive Testing

Off techniques (ad-hoc, scan-based or other structured ones) can be used

in a processor to increase its testability. Such Off techniques can be
employed either in the case that external ATE-based testing is applied to the
processor or a self-testing strategy is used instead. As an example, Logic
BIST (LBIST) is a classical self-testing methodology based on
pseudorandom pattern generators and scan-based design. Scan based testing
as well as other structured Off techniques (like test points insertion) require
many design changes and in the case of processors, although applied in some
cases, it is not the selected testing style, simply because processor designers
are very reluctant to adopt any major test-related design changes.
On the other hand, non-intrusive testing techniques do not require any
Off design changes in the processor and are therefore more "friendly" to the
processor designers, since these techniques do not add any hardware,
performance or power consumption overheads. The question with nonintrusive testing techniques is whether they are able to reach sufficient fault
coverage levels since it is restricted to faults that can be detected during the
normal operation of the processor (sometimes called functionally detectable
faults).
The benefits and drawbacks of Dff-based testing and non-intrusive
testing are summarized in the following Table 4-2.
DfT-based
testing
Non-intrusive
testing

Benefits
High fault coverage
Extensive use of EDA
tools

No hardware,
performance or power
consumption overhead
Table 4-2: DfT-based vs. non-intrusive testing.

Drawbacks
Non-trivial hardware,
performance and
power consumption
overheads
Limited EDA use
Low fault coverage

58

4.1.3

Chapter 4 - Processor Testing Techniques

Functional Testing versus Structural Testing

Functional testing of microprocessors and processor cores has been

extensively studied the last decades. Functional testing does not try to obtain
high coverage of a partieular physical or structural fault model. It rather aims
to test for the correctness of all known functions performed by the digital
circuit. In the case of processors, functional testing aims to cover the
different functions implemented by the processor's instruction set, and for
this type of circuits it seems to be a very "natural" choiee. Therefore,
functional testing of processors needs only the instruction set architecture
(ISA) information for the processor to develop test pattern sets for the
processor testing and no other lower level (like gate level) model of the
processor. Functional test sets may be applied either externally or internally
in a self-test mode.
The drawback of functional testing is that it is not directly connected to
the actual structural testability of the processor, which is related to the
physieal defects. The structural testability that a functional testing achieves
strongly depends on the set of data (operands ) whieh are used to test the
functions of a processor. In most cases, pseudorandom operands are
employed in functional testing and they lead to test sets or test programs
with excessively large test application time not capable to reach high
structural fault coverage. In functional testing, previously developed test
sequences or test programs for design verification can be re-used for testing,
and therefore test development cost is very low.
Structural testing on the other side, targets a specific structural fault
model. EDA tools can be used for automatie generation of test sequences
(ATPG tools) with the possible support of structured DIT techniques like
scan chains or test points insertion. Structural test generation can be
performed only if a gate-level model of the processor is available. If such
information is available, high fault coverage can be possibly obtained for the
target structural fault model with a sm all test set or a sm all test program that
is executed in short time.
Table 4-3 summarizes the benefits and drawbacks of functional testing
and structural testing.

59

Embedded Processor-Based Self-Test

Benefits
No low-Ievel details
required
Functional verification
patterns can be reused
Small test
development cost
Structural
EDA tools can be
testing
used
High fault coverage
Small test sequences
Fast test programs
Table 4-3: Functional vs. structural testing.
Functional
testing

Drawbacks
No relation with structural
faults
Low defect coverage
Pseudorandom operands
Longtestsequences
Long test programs
Needs gate-level model of
processor
Higher test development
cost

Functional testing of processors has been extensively studied in the

literature as it is a straightforward approach which is built upon existing test
sequences of design verification and needs relatively small test development
cost. The major problem with the application of functional testing is that as
the complexities of processors increase, the distance between comprehensive
functional testing and actual structural testability is getting longer.
4.1.4

Combinational Faults versus Sequential Faults Testing

Processor testing may focus on the detection of the faults that belong
either to a combinational fault model 8 like the industry standard single stuckat fault model, or a sequential fault modeP, i.e. a fault model which faults
lead to a sequential behavior of the faulty circuit and require two-pattern
tests for fault detection. Delay fault models like the path delay fault model
are the most usual sequential fault models. Appropriate selection of the
targeted delay faults (such as the selection of the path delay faults) must be
done to reduce the ATPG and fault simulation time. A TPG and fault
simulation times for sequential fault models are much larger than for
combinational faults. In some cases, the number of path delay faults that are
to be simulated are so many that require an excessively long fault simulation
time.
EDA tools for combinational fault models (particularly for stuck-at fault
model) have been used since many years ago. Their sequential fault model

In general, combinational faults alter the behavior of a circuit in such a way that
combinational parts of it still behave as combinational ones but with a different - faulty fimction instead of the correct one.
In general, sequential faults change the behavior of combinational parts of a circuit into a
sequential one: outputs depend on the current inputs as weil as on previous inputs.

60

Chapter 4 - Processor Testing Techniques

counterparts are less mature but are continuously improving their

performance and efficiency.
Both combinational fault testing and sequential fault testing for
processors belong to the structural testing dass previously defined since they
both require a gate-level model of the circuit to be available for test
generation and fault coverage calculation. Obviously, testing a processor for
sequential (such as delay) faults using either external test patterns or built-in
hardware or software routines is a testing strategy that offers higher defect
coverage and higher testing quality. On the other side, combinational fault
testing like stuck-at fault testing requires much smaller test sets and test
pro grams with much shorter test execution time than testing for delay faults
or other sequential faults.
Combinational testing detects the faults that change the logic behavior of
the circuit and for this reason does not need to be applied at the actual
frequency of the chip. There fore , low-cost, low-speed testers can be used.
On the other side, since sequential testing is executed to detect timing
malfunctions, it must be executed at the actual speed ofthe chip.
Table 4-4 summarizes the benefits and drawbacks of combinational fault
testing and sequential fault testing.
Combinational
faults testing

Sequential
faults testing

Benefits
Small test sets or test
programs
Short test application
time
Short test generation
and fault simulation
time
EDA tools maturity
Higher test quality
Higher defect coverage

Drawbacks
Needs gate-level
model of the processor
Less defect coverage

Large test sets or test

programs
Long test application
time
Needs gate-level
model of the processor
Long test generation
and fault simulation
time
Less mature EDA tools

Table 4-4: Combinational vs. sequential testing.

4.1.5

Pseudorandom versus Deterministic Testing

The fault coverage that a processor testing technique can obtain depends
on the number, type and nature of the test patterns applied to the processor.
Pseudorandom testing for processors can be based on pseudorandom

Embedded Processor-Based Self-Test

61

instruction sequences, pseudorandom operands or a combination of these

two. Pseudorandom processor testing, like in every other circuit, has the
drawback that it may require excessively long test sequences to reach an
acceptable fault coverage level. This is particularly true for some processor
components that are random pattern resistant. Random instruction sequences
are usually very unlikely to reach high fault coverage.
Despite its difficulties, pseudorandom testing of processors has been
extensively studied and applied because it is a simple methodology that
needs minimum engineering effort: no special test generation algorithm or
tool is necessary for pseudorandom testing. Moreover, the development of
pseudorandom test sequences or pro grams does not require a gate level
model of the processor to be available. Of course, fault coverage ca1culations
can only be done if such a model exists. Pseudorandom pattern based fault
simulations are repetitively executed and may need serious amounts of time
to determine an efficient seecf value for the pseudorandom sequence and a
suitable polynomiafIl for the pseudorandom patterns generator. Appropriate
selection of a seed and polynomial pair may lead to significant reductions of
the test sequences that are necessary to reach the target fault coverage.
On the other hand, deterministic testing for a processor is based on either
ATPG-based test sequences or other previously ca1culated test sets for the
processor or its components. As an example, for most of the functional
modules of a processor (like the ALU, multipliers, dividers, shifters, etc)
there exist carefully pre-developed test sets that guarantee high fault
coverage. These test sets can be applied via processor instructions to the
components ofthe processor to obtain high fault coverage compared with the
pseudorandom case. Also, high fault coverage (either for combinational or
sequential fault models) can be reached with a good ATPG tool, if of course,
the gate-level model of the processor is available. Therefore, the benefit of
deterministic (pre-computed or ad-hoc ATPG-based) testing is the high fault
coverage reached with short test sequences. On the other hand, ATPG-based
testing can be done when such an EDA tool (the ATPG) is available and the
success of the approach depends on the quality of the adopted tool set. Of
course, when A TPG is used for combinational components of the processor
the attained fault coverage is higher and is more easily obtained compared
with the case of a sequential component and the use of a sequential ATPG.
Table 4-5 summarizes the benefits and drawbacks of pseudorandom
testing and deterministic testing.

10
11

Seed is the initial value of a pseudorandom sequence generator like an LFSR.

The characteristic polynomial of an LFSR determines the sequence of pseudorandom
patters that are generated and also describes the connections of its memory elements.

62

Chapter 4 - Processor Testing Techniques

Pseudorandom
testing

Determ in istic
testing

Benefits
Easy development
of test sequences
No gate level
details needed for
test development
High fault coverage
Short test
sequences

Drawbacks
Longtestsequences
Low fault coverage

Gate level details

necessary
Needs special software
(the ATPG) to reach a
sufficient test result

Table 4-5: Pseudorandom vs. deterministic testing.

Combination of pseudorandom and deterministic testing for processors is

not an unusual situation. Components of the processor that can be effectively
targeted by an ATPG tool are covered by deterministic test sets while for
other components a pseudorandom approach may be used. A common
practice that is applied to enhance pseudorandom testing is the embedding of
a few deterministic patterns in pseudorandom sequences. This method
reduces the length of pseudorandom sequences and improves their detection
capability.
A final remark on pseudorandom testing is that it can be re-usable and
programmable in the sense that an existing pseudorandom test generator may
be fed with different seeds or may be reconfigured to a different polynomial
and thus it can used for the testing of different parts of the processor.
4.1.6

Testing versus Diagnosis

Processor testing techniques may be dedicated either solely to the

detection of defects/faults or, additionally, to the diagnosis process and the
localization of the defects/faults. Fault diagnosis is strongly connected with
the manufacturing process improvement because information collected
during diagnostic testing may be effectively used for the fine tuning of the
manufacturing process to improve the yield of a production flow.
As in all cases, diagnostic test sets or test programs have larger
complexity that their counterparts which are developed only for the detection
of faults and pass/fail manufacturing testing. The actual complexity of
diagnostic testing depends on the required diagnosis resolution, i.e. the
cardinality of fault sets in which an individually located fault belongs to. The
higher the diagnosis resolution (the smaller the fault sets) the larger the test
set size that is necessary.
Table 4-6 summarizes the benefits and drawbacks of testing only
methods and diagnosis methods.

63

Embedded Processor-Based Self-Test

Testing
Diagnosis

Benefits
Small test
sequences
EDA tools support
High test quality
Supports yield
improvement

Drawbacks
Only pass/fail
indication

Large test
sequences
Less EDA tools
support

Table 4-6: Testing vs. diagnosis.

4.1.7

Manufacturing Testing versus On-linelField Testing

Classical manufacturing testing techniques focus only on the detection of

faults/defects that exist after chip manufacturing in the fab. A manufacturing
testing strategy (ATE-based on not) is used once before the chip is released
for use in the final system.
On the other side, processor testing techniques like any other Ie testing
technique can not only be used during manufacturing testing but also for
testing the chip in the field. The latter is called on-line testing and is done
while the chip is mounted in the final system and is periodically tested.
Self-testing techniques, unlike external testing ones, are excellent
candidates for re-use during on-line testing because the embedded testability
enhancing features and circuit structures can be re-used several times and
definitely not only during the manufacturing phase of the chip. On-line
periodic testing can detect faults that appear in the field because of external
environmental factors or chip aging.
Table 4-7 summarizes the benefits and drawbacks of manufacturing
testing and on-line testing.
Benefits
Either external
Of built-in
Re-usable
during product
life cycle
Table 4-7: Manufacturing vs. on-line/field testing.
Manufacturing
testing
On-line/field
testing

4.1.8

Drawbacks
May not be reusable
More expensive

Microprocessor versus DSP Testing

Embedded processing elements may appear in the form of either a

classical von Neumann architecture with a single memory for instructions
and data storage, or Harvard architecture with separate instruction and data
memories. Harvard architectures are commonly used in Digital Signal
Processors (DSPs) where data manipulation requirements are higher than in
general purpose processors and also data transfer bandwidth must be also

64

Chapter 4 - Processor Testing Techniques

higher. On the other side, general purpose processors usually employ

performance enhancing mechanisms like branch prediction that are not
commonly used in DSPs due to the probabilistic nature of these mechanisms.
Finally, both types of architectures may use relatively simple of more
complex pipeline structures for performance increase.
Microprocessors and embedded processors with general purpose
architecture usually have a complex control structure, compared with DSPs
which usually have simpler control units but more complex and large data
processing modules.

4.2

Processor Testing Literature

A relatively large number of processor testing approaches have been

proposed during the last four decades. The objectives of each approach
strongly depend on the type of application that the processor is used in, as
well as, on the testing strategy constraints. As in any device testing problem,
not all constraints and requirements can be met simultaneously.
Research activities in a particular field are sometimes c1ustered in small
time periods where the importance of the field is high. When technological
and architectural advances changed the characteristics of processors,
different testing methodologies appeared. There may be also cases where
research performed several years ago was found useful as an answer to
modem problems.
Each of the processor testing works that we summarize below belongs to
one or more of the c1asses of processor testing presented in the previous
sections. We first list and briefly discuss each ofthem in chronological order
and then we map each work to the c1asses it belongs to.
4.2.1

Chronological List of Processor Testing Research

One of the first papers in the open literature that addressed the subject of
microprocessor testing was B.Wi111iams' paper in 1974 [169]. The internal
operation of microprocessors was reviewed to illustrate problems in their
testing. LSI test equipment, its requirements, its application were discussed
as well as issues on wafer testing.
In 1975, M.Bilbault described and compared five methods for
microprocessor testing [17]. The first method was called 'Autotest' and it
assembled the microprocessor in its natural environment. A test pro gram was
running that could generate 'good' or 'bad' responses. The second method
was similar to the first one, but it was based on comparison of the responses
using a reference microprocessor whose output was compared with the
microprocessor under test after each instruction cyc1e. The third method was

Embedded Processor-Based Self-Test

65

called 'real time algorithmic' and used a prepared program to send aseries of
instructions to the microprocessor and to compare its response with that
expected. The fourth method was called 'recorded pattern' and had two
phases. In the first one, the microprocessor was simulated and the responses
were recorded while in the second one the responses of the microprocessor
under test were compared with the recorded responses created during the
first phase. The fifth method from Fairchild was called 'LEAD' (learn,
execute and diagnose) where the test pro gram was transferred to the memory
of the tester, together with the response found from a reference
microprocessor, and the memory also contains aIl the details of the
microprocessor's environment. Advantages of speed and thoroughness were
claimed for 'LEAD'.
In 1975, R.Regalado introduced the concept of user testing of
microprocessors claiming that microprocessor testing in the user's
environment does not have to be difficult or laborious, even though it is
inherently complex [135]. The basic concepts of the 'people' (user) oriented
approach for microprocessors testing were: (1) "The test system's computer
performs every function that it is capable of performing" - this is the concept
of functional testing; and (2) "The communication link between the
computer and people (test engineers for example) is interactive".
In 1976, E.C.Lee proposed a simple microprocessor testing technique
based on microprocessor substitution [107]. Because of its low cost in
hardware and software development, this technique was suitable for user
testing in a simple user environment. A tester for the Intel 4004
microprocessor was described.
In 1976, D.H.Smith presented a critical study for four ofthe more widely
accepted, during that period, methods for microprocessor testing, with a
view to developing a general philosophy which could be implemented as a
test with minimum effort [147]. The considered microprocessor testing
methods were: actual use of microprocessor, test pattern generation based on
algorithms to avoid test pattern storage, stored-response testing, and
structural verification.
The pioneering work of S.Thatte and J.Abraham in functional
microprocessor testing was first presented in 1978 [156]. In this paper, the
task of faults detection in microprocessors, a very difficult problem because
of the processors' always increasing complexity, was addressed. A general
microprocessor model was presented in terms of a data processing section
(simple datapath) and a control processing section, as weIl as, a functional
fault model for microprocessors. Based on this functional fault model the
authors presented a set of test generation procedures capable of detecting aIl
considered functional faults.

66

Chapter 4 - Processor Testing Techniques

S.Thatte and J.Abraham presented, in 1979 [157] and 1980 [158],

respectively, test generation procedures based on a graph-theoretic model at
the register transfer level (RTL). The necessary information to produce the
graph model for any processor consists only of its instruction set architecture
and the functions it performs. The functional testing procedures do not
depend on the implementation details of the processor. The complexity of
the generated tests as a function of the number of instructions of the
processor is given and experimental results are reported on a HewlettPackard 8-bit microprocessor. A total of 8K instructions were used to obtain
a 96% coverage of single stuck-at faults, which was a complete coverage of
all single stuck-at faults that affect the normal operation of valid processor
instructions for this benchmark. The test sequences generated by the
approach presented in [158] are very long in the case of the instruction
sequencing logic because of the complexity of the proposed functional test
generation algorithm, while they are sufficiently short for the register
decoding logic, the data path and the Arithmetic Logic Unit (ALU) of the
processor.
In 1979, G.Crichton proposed a test strategy for functional testing of
microprocessors where the internallogic is separated in two types: data logic
and control logic in order to simplif)r the development of functional test
vectors [34]. A practical example was presented in the form of a test
pro gram for the SAB 8080A microprocessor. The worst case of the derived
functional test program lasts only 130 ms when executed at 2.5 MHz.
In 1979, C.Robach, C.Bellon, and G.Saucier proposed an application
oriented test method of a microprocessor system [137]. The goal was to test
the microprocessor system 'through' the application program. Thus, this
pro gram was partitioned into segments according to the hardware access
points of the system and a diagnostic algorithm was proposed. The
efficiency of this method with regard to the functional error hypothesis was
discussed.
In 1979, P.K.Lala proposed a test method for microprocessor testing
which was based on the partitioning of the microprocessor's instruction set
into several instruction sets [106]. Instructions affecting the same modules
inside the processor were grouped into the same instruction set. An
appropriate sub-set of each instruction set was determined to form a test
sequence for the microprocessor. Stuck-at faults at the address lines, data
lines and the output of some internal modules were also detected by this test
sequence. As an illustration, the test sequence for 8080 microprocessor was
derived.
In 1980, C.Robach, G.Saucier, and R.Velazco proposed a test method for
functional testing of microprocessors based upon a high level functional
description ofthe microprocessors [139].

Embedded Processor-Based Self-Test

67

In 1980, C.Robach and G.Saucier considered the problem of testing a

dedicated microprocessor system that performs a specific application [138].
They presented a diagnosis methodology based on a correlated analysis of
the application program (modeled by a control graph) and the hardware
system (modeled by data graphs).
In 1981, T.Sridhar and J.P.Hayes presented a functional testing approach
for bit-sliced microprocessors [149]. A functional fault model is used in this
work and complete tests are derived for bit-sliced microprocessors
resembling the structure of AMD 2901 processor slice. Bit-sliced
microprocessors are treated as iterative logic arraysl2 for which C-tests (tests
of constant size independent on the slices/cells ofthe array) are sufficient to
obtain complete fault coverage.
In 1981, P.Thevenod-Fosse and R.David considered the case of random
testing of the data processing section of a microprocessor based on the
principle that a sequence of random instructions with random data is applied
simultaneously to both a processor under test and a golden processor [159].
They proposed a methodology to calculate theoretically the number of
required random instructions for given instruction probabilities in user
programs, based on a functional fault model for registers and operators. For
the case of Motorola 6800 microprocessor they provided a program
consisting of about 6,300,000 random instructions.
In 1981, B. Courtois set the basis for on-line testing of microprocessors
[33]. He proposed a methodology without using massive redundancy for online testing that requires the periodic execution of a watch-dog and test
programs, and he quantified the fault detection time (also known as on-line
fault detection latency).
In 1982, M.Annaratone and M.G.Sami proposed a functional testing
methodology for microprocessors [6]. They adopted microprogramming to
create their functional microprocessor model starting from user available
information. Their aim was to generate test procedures that detect errors
associated with a functional model instead of faults, for which structural
information is necessary.
In 1982, J.Jishiura, T.Maruyama, H. Maruyama, S.Kamata presented the
problems raised in VLSI microprocessor testing and described a new test
vector generator and timing system for external testing designed to solve
these problems [82]. The test vector generator used a vertically integrated
vector generation architecture to handle long arrays of test vectors, and the
timing system had cross-cycle clocking, cross-cycle strobing, and multiple
clocking capabilities to achieve accurate test timing.
12

Iterative Logic Arrays (ILAs) consists of identical circuits (combinational or sequential)

which are regularly interconnected in one or more dimensions.

68

Chapter 4 - Processor Testing Techniques

In 1983, C.Timoc, F.Stoot, K.Wickman, L.Hess presented simulation

results on processor self-testing using weighted random test patterns [161].
Input weights were optimized to obtain a shorter test sequence for the testing
ofthe microprocessor.
In 1983, S.KJain and A.K.Susskind proposed that microprocessor
functional testing is divided into three distinct phases: verification of the
control functions and data transfer functions; verification of the datamanipulation functions; and verification of the input-output functions [77].
They considered in detail only the first of these phases. To verifY control
functions, they proposed a DIT technique where appropriate additional
signals inside the chip were used and made observable at the terminal pins of
the microprocessor chip. Complete instruction sequences executed in a test
mode used to verifY control functions. Also, test procedures for verifYing
control functions and data transfer functions were presented.
In 1983, P.Thevenod-Fosse and R.David complemented their previous
work of [159], by considering the case of random testing of the control
processing section of a microprocessor, as weil [160].
In 1984, X.Fedi and R.David continued the work of [159] and [160] by
presenting a random tester for microprocessors and comparisons between
theoretical results and experimental results for random testing of Motorola
6800 microprocessors [42].
In 1984, the functional testing work of [158] was complemented by the
work ofD.Brahme and J.Abraham [21] which reduces the complexity ofthe
generated tests for the processor's instruction sequencing and execution
logic. A functional model based on a reduced graph is used for the
microprocessor and a classification of all faults into three functional
categories is given. Tests are first developed for the registers read operations
and then for all remaining processor instructions. The developed tests are
proposed for execution in a self-test mode by the processor itself.
In 1984, J.F.Frenzel and P.N.Marinos presented a functional testing
approach for microprocessors based on functional fault models and user
available information for the processor with the aim to reduce the number of
tests. N/2-out-of-N codes are employed for this purpose 13 for the data words
of the processor [45]. The authors applied their approach to the same
hypothetical processor used in [158] and show that a smaller number of
instructions are required by their approach to test the microprocessor.
In 1984, M.G.Karpovsky and R.G. van Meter presented how functional
self-test techniques are used to detect single stuck-at faults in a
microprocessor [86]. The derived test programs based on these techniques
was of practical interest since they had a relatively small number of
13

In general, an m-out-of-n code consists of code words which have m I's and n-m O's.

Embedded Processor-Based Seit-Test

69

instructions with a small number of responses stored in memory. The

efficiency of these techniques was demonstrated by applying them to a 4-bit
microprocessor and performing single stuck-at fault simulation.
In 1984, G.Roberts and J.Masciola examined and compared the most
popular techniques used in testing microprocessor-based boards and systems
[140]. In particular, they concentrated on the test step that immediately
follows in-circuit testing. A technique, based on memory emulation, was
presented which effectively met the requirements of this test step.
Application of this technique to board test, quick verification, and system
test was examined, and fault detection with diagnostic capability was
discussed.
In 1985, RKoga, W.A.Kolasinski, M.T.Marra, and W.A.Hanna studied
several test methods to assess the vulnerability of microprocessors to singleevent upsets (SEUs) [92]. The advantages and disadvantages of each ofthese
test methods were discussed, and the question of how the microprocessor
test results can be used to estimate upset rate in space was addressed. As an
application of these methods, test results and predicted upset rates in
synchronous orbit were presented for a selected group of microprocessors.
In 1985, RVelazco, H.Ziade, and E.Kolokithas presented a
microprocessor test approach that allowed fault diagnosis/localization [168].
The approach was implemented on a behavioral test dedicated system: the
GAPT system and it was illustrated by results obtained during testing of
80C86 microprocessors.
In 1985, RFujii and J.Abraham presented a methodology for functional
self-test at the board level of microprocessors with integrated peripheral
control modules [47]. An enhanced instruction execution fault model and
new fault models for peripheral controllers were presented. Also, test
program development for data compression was presented. The application
ofthis methodology to Intel 80186 microprocessor was described.
In 1986, X.Fedi and RDavid [41] completed their experimental work
first presented in [42], describing a random input pattern tester for MC-6800
microprocessors. The tester was based on an efficient input pattern generator
which generates the inputs with the required probability distribution. The
authors illustrated the statistical properties of the test latency as a random
variable. Comparisons were given between deterministic and random
experiments, as weIl as, between experimental and theoretical results of
random testing.
In 1986, P.Seetharamaiah and V.RMurthy presented a test generation
global graph model at micro operation level that included architecture and
organization details as parameters to be used for flexible testing of
microprocessors [143]. Every microprocessor instruction was represented by
its abstract execution graph, which forms a sub graph of the global graph. A

70

Chapter 4 - Processor Testing Techniques

tabular method was developed for a systematic ordering of the subgraphs in

order of complexity, aimed at the fuH fault coverage of the entire processor.
Based on this approach a test procedure was developed.
In 1986, B.Henshaw presented a test pro gram for user testing of the
Motorola MC68020 32-bit microprocessor [64]. The MC68020 had some
features not found in earlier vers ions of the MC68000 family or other earlier
microprocessors that affected the effective development of this test program.
Test program development for user testing is based on a functional block
diagram because of the limited information available for the processor
architecture and structure. In this paper, the application of functional test
methods to the MC68020 was examined.
In 1987, K.K.Saluja, L.Shen, and S.Y.H.Su proposed a number of
algorithms for functional testing of the instruction decoding function of
microprocessors [142] (earlier version of this work was presented in 1983
[141]). The algorithms were based on the knowledge of timing and control
information available to users through microprocessor manuals and data
sheets. They also established the order of complexity of the algorithms
presented in this paper.
In 1987, P.G.Belomorski proposed a simple theoretical method for
random testing of microprocessors employing the ring-wise testing concept
[12]. On the basis of a functional model of a microprocessor and its faults,
he dealed with the problem of establishing the length of the random test
sequence needed to detect all faults with a given certainty. The method was
applied to determine the random test sequence necessary for testing the
MC6800 microprocessor and it was shown that the length ofthe random test
sequence, obtained theoreticaHy, actuaHy covered the worst case in a
pseudorandom test procedure.
In 1988, H.-P.Klug presented a microprocessor testing approach based on
pseudorandom instruction sequences generated by an LFSR [91].
Pseudorandom tests generated by the LFSR are transformed into valid
processor instructions. The approach is a functional testing one but with
emphasis on the reduction of the test sequences size and duration.
Experiments were performed on a smaH (5,500 transistors) execution unit of
a signal processor. First, the functional testing approach of[158] was applied
and it took three man months to obtain a 2,300 instructions sequence that
reached 94% coverage ofthe single stuck-open faults for the datapath part of
the unit and only 44% for the control part. On the contrary, the LFSR-based
pseudorandom instruction sequences generation method that the author
presented obtained only within a week, different test sequences of around
2,500 instructions that reached a 94% fault coverage for the datapath and
64% for the control part. Unfortunately, such pseudorandom-based test
generation approach for larger processors will require the application of

Embedded Processor-Based Seit-Test

71

excessively large test sequences without being able to reach high fault
coverage.
In 1988, L.Shen and S.Y.H.Su presented in [146] (and in a preliminary
version of the work in 1984 [145]) a functional testing approach for
microprocessors based on a Register Transfer Level control fault model,
which they also introduced. As a first step, the read and write operations to
the processor registers are tested (these operations are called the kernel) and
subsequently all the processor instructions are tested using the kernel
operations. The k-out-of-m codes are utilized to reduce the total number of
functional tests applied to the processor. Details ofthe functional fault model
and the procedure to derive the tests are provided in this work.
In 1989, E.-S.A.Talkhan, A.M.H.Ahmed, A.E.Salama focused on the
reduction of test sequences used for microprocessor functional testing, so
that as short as possible instruction sequences are developed to cover all the
processor instructions [152]. Application of the method to TI's TMS32010
Digital Signal Processor shows the obtained reductions in terms of number
of functional tests that must be executed.
In 1990, A.Noore and B.E.Weinrich presented a microprocessor
functional testing method in which three test generation approaches were
given: the modular block approach, the comprehensive instruction set
approach and the microinstruction set approach [119]. These approaches
were presented as viable alternatives to the exhaustive testing of all
instructions, all addressing modes and all data patterns, a strategy that is
getting more infeasible and impractical as processors sizes increase. This
type of papers, make obvious how difficult the problem of functional testing
of processors be comes as processor sizes increase. In this work, some
examples ofthe approach application to Intel's 8085 processor are given but
with no specifics on test pro gram size and execution time or fault coverage
obtained.
In 1992, A.J. van de Goor, and Th.J.W. Verhallen [58] presented a
functional testing approach for microprocessors which extends the functional
fault model introduced in [21] and [158] in the 80's, to cover functional units
not described by the earlier functional fault model. Memory testing
algorithms were integrated into the functional testing methodology to detect
more complex types of faults (like coupling faults and transition faults) in
the memory units of the microprocessor. The approach has been applied to
Intel's i860 processor, but no data is provided regarding the number of
tests applied, and the size and execution time ofthe testing program.
J.Lee and J.H.Patel in 1992 [108] and 1994 [109] treated the problem of
functional testing of microprocessors as a high level test generation problem
for hierarchical designs. The test generation procedure was split in two
phases, the path analysis phase and the value analysis phase. In path

72

Chapter 4 - Processor Testing Techniques

analysis, instruction sequences are developed using an instruction sequence

assembling algorithm to avoid global path contlicts. The algorithm uses
behavioral information of the microprocessor. In the value analysis phase, an
exact value solution is computed for the module under test level and the
assignment of values to all internal buses and signals for test application.
Behavioral information is used to represent the internal processor
architecture in a graph model. Experimental results are provided for six high
level benchmarks but not for processor in total. The reported fault coverage
for single stuck-at faults at the components of the high level benchmarks
ranges from 48.5% up to 100.0%, depending on the benchmark and the
module under test type.
In 1995, U.Bieker, P.Marwedel proposed an automatic generation
approach for self-test programs of processors [18]. The approach is
retargetable, i.e. it can be applied to different processor architectures since
the processor specifications and instruction set are used as the algorithm
input. Constraint Logic Programming (CLP) is used in this paper at the
register transfer level of the processor for the generation of efficient self-test
programs. The approach has been applied to four simple processor examples
and no detailed structural fault coverage results are provided.
In 1996, J.Sosnowski, A.Kusmierczyk presented an interesting
comparison between deterministic and pseudorandom testing of
microprocessors using the Intel 80x86 processors architecture as a
demonstration vehicle [148]. The drawbacks of each of the two approaches
were discussed and a combined approach was claimed to be better than any
of the two alone. As in other types of circuits testing, deterministic testing is
limited by the unavailability of structural information of the circuit, while
pseudorandom testing may require very large test sequences without
obtaining sufficient fault coverage and thus must be combined with
deterministic tests.
In 1996, S.M.I.Adham and S.Gupta proposed a BIST technique, termed
DP-BIST, suitable for high performance DSP datapaths [2]. The BIST
session for the DSP was controlled via hardware without the need for a
separate test pattern generation register or test pro gram storage. Furthermore,
the BIST scenario was appropriately set-up so as to also test the register file
as weIl as the shift and truncation logic in the datapath. The use of DP-BIST
enabled at-speed testing with no performance degradation and little area
overhead for the hardware test control. Besides, they showed how DB-BIST
can be used as a centralized test resource to test other macros on the chip, as
well as, the integration ofDP-BIST with internal scan and boundary scan.
In 1997, K.Radecka, J.Rajski and J.Tyszer introduced Arithmetic Built-In
Self-Test (ABIST) as an effective pseudorandom-based technique for testing
the datapath of DSP cores using the functionality of existing arithmetic

Embedded Processor-Based Self-Test

73

modules [131]. Arithmetic modules are used to generate pseudorandom tests

and to compact test responses while they test themselves, other components
of the DSP core and external circuits. Test application and response
collection can be performed using software routines ofthe core.
The work of 1997 [165] by R.S.Tupuri, I.A.Abraham is a first
presentation of [166] were the authors described the approach (constraint
extraction, constraint-based ATPG). This first approach of [165] was relying
on commercial ATPG tools that were inadequate to reach high fault
coverage levels.
In 1997, K.Hatayama, K.Hikone, T.Miyazaki, H.Yamada, proposed
instruction-based test generation for functional modules in processors as a
constrained test generation problem [61]. An instruction-based test
generation system ALPS (ALU-oriented test pattern generation system) is
outlined. The approach targets functional units (ALUs) of processors, by
translating module-level tests into instruction sequences taking into
consideration the instruction set architecture imposed constraints.
Experimental results are given in [61] for two functional units, a floating
point adder and a floating point multiplier of a RISC processor (without
further details on the identity of the processor and its architecture
characteristics). The constrained test generation process for the floating point
adder reaches 89.10% fault coverage, while for the floating point multiplier
it reaches 89.15% fault coverage.
In 1998, J.Shen and I.Abraham [144] presented a functional test
generation approach for processors and applied it to manufacturing testing as
well as to design validation. No apriori structural fault model is considered
in the approach and the information that is necessary for the test generation
method (and prototype tool) to operate is the processor's instruction set and
the operations that the processor performs as a response to each instruction.
For the functional testing of each instruction, the approach generates a
sequence of instructions that enumerates all the combinations of the
instruction operation and systematically selected operands. Also, random
instruction sequences are generated for groups of instructions, to exercise the
instructions and propagate their effects. Experiments have been performed
on two processor benchmarks: the Viper 32-bit processor [36] and GL85
which is a model of Intel's 8085 processor. For Viper, the proposed test
generation method obtains 94.04% single stuck-at fault coverage. For GL85,
a fault coverage of 90.20% for single stuck-at faults is obtained again as a
combination of the two facets of the methodology: exhaustive testing of each
instruction with all combinations of operations and operands, and random
generation of instruction sequences for groups of instructions. This
approach, being a functional one that relies on exhaustive and/or
pseudorandom testing, applies very long instruction sequences. For example,

74

Chapter 4 - Processor Testing Techniques

It IS reported in [144] that a total of 360,000 instruction cycles were

necessary for the GL85 processor to execute a self-test program that reached
a 86.7% fault coverage for single stuck-at faults.
In 1998, DSP testing using self-test programs was considered by W.Zhao
and C.Papachristou [175]. This approach is based on the application of
random patterns to the components of the DSP using self-test programs
developed based on behavioral and structural level testability analysis. Such
analysis is not performed in functional self-testing where neither information
on the RTL description of the processor is required nor a specific structural
fault model is considered. Experimental results are provided in [175],
showing that a 94.15% single stuck-at fault coverage is obtained by this
pseudorandom approach in a simple DSP designed by the authors for the
experiments. No details were given regarding the size of the self-test
program, the number of test patterns applied to the DSP datapath modules or
the total number of clock cycles for the execution ofthe self-test program.
In 1999, R.S.Tupuri, A.Krishnamachary, and J.A.Abraham proposed an
automatie functional constraint extraction algorithm that transforms modules
of a processor by attaching virtual logic which represents the functional
constraints [166]. Then, automatic test generation can be executed on the
transformed components to derive tests that can be realized with processor
instructions. The implemented algorithm has been applied to different
modules of three processor models (Viper, DLX and ARM) and fault
coverage for single stuck-at faults between 81.14% and 97.17% was
reported. These fault coverages are much higher when compared with the
case that ATPG is done at the processor level for the same processor
components without the use of the constraint extraction of the proposed
algorithm.
In 1999, K.Batcher and C.Papachristou proposed Instruction
Randomization Self Test (IRST) for processor cores, a pseudorandom selftesting technique [10]. Self-test is performed with processor instructions that
are randomized by a special circuitry designed out of the processor core for
this purpose. IRST does not add any performance overhead to the processor
and the extra hardware is relatively small compared to the processor size
(3.1% hardware overhead is reported for a DLX-like RISC processor core).
The obtained fault coverage for the processor core after the execution of a
random instructions sequence running for 50,000 instruction cycles is
92.5%, and after the execution of 220,000 instruction cycles is 94.8%
(processor size is 27,860 and it contains 43,927 single stuck-at faults).
In 2000 W.-C.Lai, A.Krstic and K.-T.Cheng proposed a software-based
self-testing technique with respect to path delay faults [105]. The proposed
approach is built upon constraint extraction, classification of the paths of the
processor, constraint structural ATPG (thus, deterministic path delay test

Embedded Processor-Based Seit-Test

75

patterns are used) and automatic test program synthesis. The target is the
detection of delay faults in functionaIly testable paths of the processor. The
entire flow requires knowledge of the processor's instruction set, its microarchitecture, RTL netlist as weIl as the gate-level netlist for the identification
of the functionaIly testable paths. Experiments have been performed for the
Parwan educational processor [116] as weIl as for the DLX RlSC processor
[59]. In Parwan, as self-test program of 5,861 instructions (bytes) obtained a
99.8% coverage of aIl the functionaIly testable path delay faults, while in
DLX, a self-test program of 34,694 instructions (32-bits words) obtained a
96.3% coverage of aIl functionaIly testable path delay faults.
The contribution of the work presented by L.Chen and S.Dey in 2001
[28] (preliminary version was presented in [27]) is twofold. First, it
demonstrates the difficulties and inefficiencies of Logic BIST (LBIST)
application to embedded processors. This is shown by applying Logic BIST
to a very simple 8-bit accumulator-based educational processor (Parwan
[116]) and a stack-based 32-bit soft processor core that implements the Java
Virtual Machine (picojava [127]). In both cases, Logic BIST adds more
hardware overhead compared to fuIl scan, but is not able to obtain
satisfactory structural fault coverage even when a very high number of test
patterns are applied. Secondly, a structural software-based self-testing
approach is proposed in [28] based on the use of self-test signatures. Self-test
signatures provide a compacted way to download previously prepared test
patterns for the processor components into on-chip memory. The self-test
signatures are expanded by embedded software routines into test patterns
which are in turn applied to the processor components and test responses
(either individuaIly or in a compacted signature) are coIlected for external
evaluation. The component test sets are either previously generated by an
ATPG and then embedded in pseudorandom sequences or are generated by
software implemented pseudorandom generators (LFSRs). Experimental
results on the Parwan educational processor show that 91.42% fault coverage
is obtained for single stuck-at faults with a test pro gram consisting of 1,129
bytes, running for a total of 137,649 clock cycles.
In 2001, the problem of processor testing and processor-based SoC
testing was addressed by W.-C.Lai and K.-T.Cheng [104], where instructionlevel DfT modifications to the embedded processor were introduced. Special
instructions are added to the processor instruction set with the aim to reduce
the length of a self-test pro gram for the processor itself or for other cores in
the SoC, and to increase the obtained fault coverage. Experimental results on
two simple processor models, the Parwan [116] and the DLX [59], show that
complete coverage of aIl functionaIly testable path delay faults can be
obtained with smaIl area DfT overheads that also reduce the overall self-test
program length and its total execution time. In Parwan, test program is

76

Chapter 4 - Processor Testing Techniques

reduced by 34% and its execution time reduced by 39% with an area
overhead of 4.7% compared to the case when no instruction-level OfT is
applied to the same processor [105]. Moreover, complete 100% fault
coverage is obtained while in [l05] fault coverage was a bit lower 99.8%. In
the OLX case, the self-test program is 15% smaller and its execution time is
reduced by 21 % with an area overhead due to OfT of only 1.6%. Fault
coverage is complete (100%) for the OfT case while it was 96.3% in the
design without the OfT modifications [105]. All fault coverage numbers
refer to the set of functionally testable path delay faults.
In 2001, F.Como, M.Sonza Reorda, G.Squillero and M.Violante [32]
presented a functional testing approach for microprocessors. The approach
consists of two steps. In the first step, the instruction set of the processor is
used for the construction of a set of macros for each instruction. Macros are
responsible for the correct application of an instruction and the observation
of its results. In the second step, a search algorithm is used to select a
suitable set of the previously developed macros to achieve acceptable fault
coverage with the use of ATPG generated test patterns. A genetic algorithm
is employed in the macros selection process to define the values for each of
the parameters of the macros. Experimental results are reported in [32] on
the 8051 microcontroller. The synthesized circuits consist of about 6,000
gates and 85.19% single stuck-at fault coverage is obtained, compared to
80.19% of a pure random-based application of the approach. The number of
macros that actually contributed to the above low fault coverage, lead to a
test program consisting of 624 processor instructions.
In 2002, L.Chen and S.Oey exploited the fault diagnosis capability of
software-based self-testing [29]. A large number of appropriately developed
test programs are applied to the processor core in order to partition the fault
universe in sm all er partitions with unique pass/fail pattern. A sufficient
diagnostic resolution and quality was obtained when the approach is applied
to the simple educational processor Parwan [116].
In 2002, N.Kranitis, O.Gizopoulos, A.Paschalis and Y.Zorian [94], [95],
[96] introduced an instruction-based self-testing approach for embedded
processors. The self-test programs are based on small deterministic sets of
test patterns. First experimental results for the methodology were presented
in these papers. The approach was applied to Parwan, the same small
accumulator-based processor used in [28] and a 91.34% single stuck-at faults
coverage was obtained with a self-test pro gram consisting of around 900
bytes and executing for about 16,000 clock cycles (these numbers gave
about 20% reduction of test pro gram size and about 90% reduction in test
execution time compared to [28]).
In 2002, P.Parvathala, K.Maneparambil and W.Lindsay presented an
approach called Functional Random Instruction Testing at Speed (FRITS)

Embedded Processor-Based Seit-Test

77

which applies randomized instruction sequences and tries to reduce the cost
of functional testing of microprocessors [125]. To this aim Dff
modifications are proposed that enable the application of the functional selftesting methodology using low-cost, low-pin count testers. Moreover,
automation of the self-testing programs generation is considered. The basic
feature of FRITS which is also its main difference when compared with
classical functional processor self-testing is that a set of basic FRITS
routines (caIled kerneis) are loaded to the cache memory of the processor
and are responsible for the generation of several programs consisting of
random instruction sequences and are used to test parts of the processor.
External memory cycles are avoided by appropriate exception handling that
eliminates the possibility of cache misses that initiate main memory
accesses. The FRITS methodology is reported to be applied in the Intel
Pentium 4 processor resulting in around 70% of single stuck-at fault
coverage. Also, application of the approach to the Intel Itanium processor
integer and floating point units led to 85% single stuck-at fault coverage.
The primary limitation of this technique, like any other random-based,
functional self-testing technique, is that an acceptable level of fault coverage
can only be reached if very long instruction sequences are applied - this is
particularly true for complex processor cores with many components.
In 2003, L.Chen, S.Ravi, A.Raghunathan and S.Dey focused on the
scalability and automation of software-based self-testing [31]. The approach
employs RTL simulation-based techniques for appropriate ranking and
selection of self-test pro gram templates (instruction sequences for test
delivery to each of the processor components) as weIl as techniques from the
theory of statistical regression for the extraction of the constraints imposed
to the application of component-Ievel tests by the instruction set of the
processor. The constraints are modeled as virtual constraint circuits (VCC)
and automated self-test pro gram generation at the component level is
performed. The approach has been applied to a relatively large
combinational sub-circuit of the commercial configurable and extensible
RISC processor Xtensa from Tensilica [174]. A self-test program of
20,373 bytes, running for a total of 27,248 clock cycles obtained 95.2% fault
coverage of the functionaIly testable faults of the component. A total of 288
test patterns generated by an ATPG were applied to the component.
In 2003, N.Kranitis, G.Xenoulis, D.Gizopoulos, A.Paschalis and
Y.Zorian [99], [100], showed the applicability of software-based self-testing
to larger RISC processor models while a classification scheme for the
different processor components was introduced. The processor self-testing
problem is addressed by a solution focusing on the test costs reduction in
terms of engineering effort, self-test code size and self-test execution time,
aIl together leading to significant reductions of the total device test time and

78

Chapter 4 - Processor Testing Techniques

costs. In particular, the work of [100] gives an extensive application of the

low-cost, deterministic self-testing methodology to several different
implementations of the same RISC processor architecture. The processor
model consists of 26,000 to 30,000 logic gates depending on the
implementation library and synthesis parameters. A self-test program of 850
instructions (words) executing in 5,800 clock cycles obtained more than 95%
single stuck-at fault coverage for all different processor implementations.
In 2003, G.Xenoulis, D.Gizopoulos, N.Kranitis and A.Paschalis
discussed the use of software-based self-testing as a low-cost methodology
for on-line periodic testing of processors [173]. The basic requirements of
such an approach were discussed and the co ding styles for software-based
self-testing that are able to satisfy the requirements were presented.
4.2.2

Industrial Microprocessors Testing

Apart from the research described in the previous sections which focused
on more or less generic methodologies for processor testing and self-testing,
a large set of industrial case studies papers were presented at the IEEE
International Test Conference (ITC) over the last years. In this type of
papers, authors of major microprocessor design and manufacturing
companies summarized the manufacturing testing methodologies applied to
several generations of successful, high-end microprocessors of our days.
Several exciting ideas were presented in these papers although, for obvious
reasons, only a small set of the details were revealed to the readers.
A list of papers of this category is included in the bibliography of this
book including microprocessor testing papers from Intel, IBM, Motorola,
Sun, HP, AMD, DEC, ARM, TI. The list includes industrial papers
presented at ITC between years 1990 and 2003.

4.3

Classification of the Processor Testing

Methodologies

Each of the techniques briefly discussed in the previous sections belongs

to one or more ofthe categories discussed in the beginning ofthe Chapter. In
this section we make the association of each methodology with the processor
testing categories it belongs to. This association is sometimes tricky because
a methodology originally developed for one objective mayaiso be applicable
with other objectives in mind and in a different application field. Therefore,
the main intention of the classification of this section is to give a rough idea
of the work performed over the years with emphasis on several aspects of
processor testing.

79

Embedded Processor-Based Self-Test

Table 4-8 presents the classification of the works analyzed above into the
categories described earlier. Some explanations are necessary for the better
understanding ofTable 4-8.
A paper appearing in the self-testing category means that it specifically
focused on self-testing but in many cases the approach can be applied to
external testing as well.
A paper appearing in the OfT-based testing category means that some
design changes are required to the processor structure while all other works
do not change the processor design.
A paper appearing in the sequential faults testing category means that the
approach was developed andJor applied targeting sequential faults such as
de1ay faults. All other works were either applied for combinational faults
testing or did not use any structural fault model at all (functional testing).
Category

Works

Seit Testing
(others are on extern al testing only)

[17] , [161] , [21] , [86] , [47] ,

[18] , [2 ], [131] , [175] , [10] ,
[28] , [27] , [104] , [94] , [96] ,
[125] , [31] , [97] , [100] , [99]
[77] , [2] , [104]

DfT-based Testing
(others are non-intrusive)
Functional Testing

Structural Testing
(including register transfer level)
Sequential Faults Testing
(others are on combinational faults)
Pseudorandom Testing
Deterministic Testing
(including ATPG-based)

[91] ,
[105] ,
[95] ,

[135] , [156] , [158] , [157] , [34] ,

[137], [139] , [149] , [ 6], [77] , [21] ,
[45] , [86] , [47] , [64] , [141] , [142] ,
[12] , [91] , [146] , [145] , [152] ,
[119] , [58] , [109] , [108] , [165] ,
[166] , [144] , [32] , [125]
[149] , [143] , [146] , [145] , [2 ],
[131], [61] , [175] , [105] , [28] , [27] ,
[104] , [29] , [94] , [95] , [96] , [31 ],
[99] , [97] , [100]
[105] , [104]
[158] , [157], [159] , [161] , [160] ,
[41] , [12] , [91] , [148] , [131] , [144] ,
[175], [10] , [32] , [125]
[148] , [28] , [27] , [94 ], [96] , [95] ,
[97] , [100] , [99] , [166] , [165] ,
[105] , [28] , [27] , [32] , [ 31]
[137], [138] , [140] , [168] , [29]

Diagnosis
(others are on testing only)
[33] , [92] , [64] , [107] ,
Field Testing - On-Une Testing
(others are on manufacturing
testing)
[152] , [2 ], [131] , [175]
DSP Testing
(others are on microprocessor
testing)
Table 4-8: Processor testmg methodologles classlflcatlOn.

[135] ,

[173]

Chapter

5
Software-Based Processor Self- Testing

In this Chapter we discuss a processor self-testing approach which has

recently attracted the interest of several test technologists. The approach is
based on the execution of embedded self-test programs and is known as
software-based selftesting. We present software-based self-testing as a lowcost or cost-effective self-testing technique that aims to high structural fault
coverage of the processor at a minimum test cost.
The principles of software-based self-testing and the methodology outline
are presented in this Chapter in a generic way that does not depend on any
particular processor's architecture. Different alternatives in the application of
software-based self-testing are also discussed.
Through a detailed discussion of the idea, its requirements and objectives
and its actual implementation in several practical examples of embedded
processors that follows in the subsequent Chapter, it is shown that softwarebased self-testing, if appropriately realized and carefully tailored for the
needs of a particular processor's design, can be an effective and efficient
strategy for low-cost self-testing of processor cores.
Low-cost processor testing has several different facets: it may refer to the
test generation time and man effort dedicated to the development of a selftesting strategy for a processor core; it may also refer to the total test
application time for the processor chip, while it may also refer to the extra
hardware dedicated to self-testing, and the performance overhead or the
power consumption overhead of a particular self-testing flow. We elaborate
Embedded Processor-Based Self-Test
D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic Publishers, 2004

82

Chapter 5 - Software-Based Processor Self-Testing

on these aspects of test cost in the context of software-based self-testing for

embedded processors.
In this Chapter, we present the steps of the software-based self-testing
philosophy, each one of them with different alternatives, depending on the
processor architecture and the application requirements. The presentation
style is generic, and different approaches found in the literature (already
analyzed in Chapter 4) can be selectively applied as parts of the
methodology.
The analysis of software-based self-testing given in this chapter is
followed, in the next Chapter, by a comprehensive set of experimental
results on several different publicly available processor architectures.
Different Instruction Set Architectures (ISA) are studied and the
effectiveness of software-based self-testing on each of them is discussed.

5.1

Software-based self-testing concept and flow

The concept of software-based self-testing has been introduced in the

previous Chapters, but for completeness of this Chapter, we present again
the basic idea and we subsequently proceed to the details and the steps of its
application. Figure 5-1 depicts the basic concept of processor testing by
executing embedded self-test software routines.

CPU bus
External Test
Equipment

Figure 5-1: Software-based self-testing for a processor (manufacturing).

Application of software-based self-testing to a processor core's

manufacturing testing consists ofthe following steps:

The self-test code is downloaded to the embedded instruction

memory of the processor via external test equipment which has

Embedded Processor-Based Seit-Test

83

access to the internal bUS I4 The embedded code will perform the
self-testing ofthe processor. Alternatively, the self-test code may be
"built-in" in the sense that it is permanently stored in the chip in a
ROM or flash memory (this scenario is shown in Figure 5-2). In this
case, there is no need for a downloading process and the self-test
code can be used many times for periodic/on-line testing of the
processor in the field.
The self-test data is downloaded to the embedded data memory of
the processor via the same extern al equipment. Self-test data may
consist, among others, of: (i) parameters, variable and constants of
the embedded code, (ii) test patterns that will be explicitly applied to
internal processor modules for their testing, (iii) the expected faultfree test responses to be compared with actual test responses.
Downloading of self-test data does not exist if on-line testing is
applied and the self-test program is permanently stored in the chip.
Control is transferred to self-test pro gram which starts execution.
Test patterns are applied to internal processor components via
processor instructions to detect their faults. Components' responses
are collected in registers and/or data memory locations. Responses
may be collected in an unrolled manner in the data memory or may
be compacted using any known test response compaction algorithm.
In the former case, more data memory is required and test
application time may be longer, but, on the other side, aliasing
problems are avoided. In the latter case, data memory requirements
are smaller because only one, or just a few, self-test signatures are
collected, but aliasing problems may appear due to compaction.
After self-test code completes execution, the test responses
previously collected in data memory, either as individual responses
for each test pattern or as compacted signatures are transferred to the
external test equipment far evaluation.

In the discussion of software-based self-testing of this Chapter, we

mainly focus on manufacturing testing and in several locations we point out
the differences when software-based self-testing is used for periodic, on-line
testing in the field. In the case ofperiodic, on-line testing there is no need to
transfer self-test code, data and responses to and from external test
equipment (see Figure 5-2). Self-test code, data and expected response(s)
(compacted or not) are stored in the chip as part ofthe design (in a ROM for
example). Execution of self-test programs leads to a pass/fail indication
14

In general, it is assumed that a mechanism exists for the transfer of self-test code and data
to the embedded instruction and data memory. This mechanism can be for example a
simple serial interface or a fast Direct Memory Access (DMA) protocol.

84

Chapter 5 - Software-Based Processor Selt-Testing

which can be used subsequently for further actions on the system (repair, reconfiguration, re-computation, etc).
seit-test
code
seit-test
response(s)
CPU core
Data Memory

seit-test
data
Seit-test Memory
(ROM . flash)

CPU bus

Figure 5-2: Software-based self-testing for a processor (periodic).

The steps of software-based self-testing described in the list above can be

applied in a different way depending on the processor-based system
architecture and configuration; we elaborate on this in the rest of the
Chapter. Two factors that have significant impact on the actual
implementation details of software-based self-testing are the test quality that
is looked for and the test cost that can be afforded in a particular system.
These factors determine the permissible limits for the self-test code size and
execution time as well as other details of software-based self-testing.
Let us study in more detail, at the processor level, how test patterns are
applied to internal processor components by the execution of processor
instructions, how faults in these components are excited and how their
effects are propagated outside the processor core for observation and
eventually for fault detection. We present this process in a generic way that
does not depend on any particular processor's instruction set.
Application of test patterns to processor component via processor
instructions consists ofthe following three steps:

15

Test preparation: test patterns are placed in locations (usually

registers but also in memory locations) from which they can be
easily applied to a processor component (the component under
test). This step may require the execution of more than one
processor instructions 15.
Test application and response collection: test patterns are applied
to the processor's component under test and the component's
response(s) are collected in locations (usually registers but may

For example, a test pattern for a two-operand operation consists oftwo parts, the first and
the second operand value. Application of such a test pattern requires two register writes.

Embedded Processor-Based Seit-Test

85

also be memory locations). This step usually takes one

instruction.
Response extraction: responses collected internally are exported
towards data memory (if not already in memory by the test
application instruction). This step mayaiso require the execution
ofmore than one instruction l6

We note that the above steps can be partially merged depending on the
particular coding style used for a processor and also the details of its
instruction set. For example, if the processor instruction set contains
instructions that directly export the results of an operation (ALU operation,
etc) to a memory location then the steps of response collection and response
extraction are actually merged. Figure 5-3 shows in a graphical way the three
steps of software-based self-testing application to an internal component of a
processor core. All three steps are executed by processor instructions; no
additional hardware has to be synthesized for self-testing since the
application of self-testing is performed by processor instructions that use
existing processor resources. The processor is not placed in a special mode
of operation but continues its normal operation executing machine code
instructions.

16

This is the case when the test application instruction gives a multi-word result (two-word
results are given by muItiplication and division), or the extraction of a data word and a
status word (a status register) is necessary.

86

Chapter 5 - Software-Based Processor Self-Testing

from memory

fault

CPU core

fault
effect

Test Preparation

CPU core
Test Applicationi Response Collection

faull
effecl

faull

CPU_
core
L-_
_ _ _ _--'

10 :;7

Response Extraction

Figure 5-3: Application of software-based self-testing: the three steps.

The proeess of applying a test pattern to a proeessor eomponent using its

instruetion set ean be further clarified if we have a look at a very simple
example using assembly language pseudoeode whieh is verymueh alike the
assembly language of most modem RISC proeessors (a classie RISC loadstore arehiteeture). The following assembly language pseudoeode portion
shows the steps of Figure 5-3 in the ease that the eomponent under test is an
arithmetie logic unit (ALU) and the instruetion applied to it is an addition
instruetion. The two operands of the operation eome from two proeessor
registers while the result of the addition is stored to a third proeessor
register.
load rl, X
load r2, Y
add r3, rl, r2
store r3, Z

test preparation step 1

test preparation step 2
test application/response collection
response extraction

The registers used in the example are named rl, r2, r3, while x, Y are
memory loeations eaeh of whieh eontain half of the test patterns to be
applied to the ALU and Z is the memory loeation where the test response of

Embedded Processor-Based Seit-Test

87

the ALU is eventually stored 17 The test preparation step consists of loading
the two registers rI, r2 with part of the test pattern to be applied to the
ALU. Memory locations X and Y contain these two parts of the ALU test
pattern. Registers r1 and r2 (like all general purpose registers of the
processor) have access to the ALU component inputs. The test application
and response collection step consists of the execution of the addition
instruction where the test pattern now stored in rl and r2 is applied to the
ALU and its response (result of the addition) is collected into register r3 (in
a single instruction add). Finally, the store instruction is the response
extraction step where the ALU response which is stored in register r3 is
transferred from inside of the processor to the outside external memory
location z. As we have already mentioned, test responses may either be
collected in memory locations individually, or compacted into a single or a
few signatures to reduce the memory requirements.
Of course, not every module of the processor can be tested with such
simple code portions as the one above, but this assembly language code
shows the basic simple idea of software-based self-testing. The effectiveness
of software-based self-testing depends on the way that test generation (i.e.
generation of the self-test programs) is performed to meet the specific
requirements of an application (test code size, test code execution time, etc).
Moreover, it strongly depends on the Instruction Set Architecture of the
processor under test. The requirements for an effective application of
software-based self-testing are discussed in the following subsection.

5.2

Software-based self-testing requirements

Software-based self-testing is by its definition an effective processor selftesting methodology. Its basic objective is to be an alternative self-testing
technique that eliminates or reduces several bottlenecks of other techniques.
Specifically, software-based self-testing is effective because:

17

it does not add hardware overhead to the processor but rather it

uses existing processor resources (functional units, registers,
control units, etc) to test the processor during its normal mode of
operation executing embedded code routines;
it does not add performance overhead and does not degrade the
circuit normal operation because nothing is changed in its welloptimized structure and critical paths remain unaffected;

In this example, the test patterns are stored in locations ofthe data memory (words X and
Y). As we will see, test patterns can be also stored in the instruction memory, as parts of
instructions when the immediate addressing mode is used in self-test programs.

88

Chapter 5 - Software-Based Processor Self-Testing

it does not add extra power dissipation during the processor's

nonnal operation;
it does not add extra power dissipation when the self-test
programs are executed, simply because self-test programs are
just like nonnal programs executed when the processor perfonns
its nonnal operation l8 ;
it does not rely on expensive extemal testers, but rather on lowcost, low pin-count, testers which are only necessary for the
transfer of the self-test programs (code and data) to the
processor's memory and for the transfer ofthe self-test responses
out ofthe processor's memory for extern al evaluation.

Software-based self-testing applies tests that can appear during the

processor's nonnal operation. Therefore, the faults that it can detect are the
functionally testable faults of the processor (either the logic faults or the
timing faults). This property of software-based self-testing has a positive and
a negative impact. On the positive side, no overtesting is applied, i.e. only
faults that could affect the nonnal operation of the processor are targeted. On
the negative side, there is no mechanism (no Dff circuitry) that increases the
testability ofthe processor in its hard-to-test areas.
The actual efficiency of software-based self-testing on specific processor
architectures depends on system parameters such as the total system cost
(and thus its test cost) and the required manufacturing testing quality. A
number of questions must be answered before a test engineer applies
software-based self-testing to a processor. The answers to these questions
construct a framework for the development of self-test programs. In the next
sections, we elaborate on these subjects as weil as on their impact on the
success of software-based self-testing.
5.2.1

Fault coverage and test quality

The primary and most important decision that has to be taken in any test
development process is the level of fault coverage and test quality that is
required. All aspects of the test generation and test application processes are
affected by the intended fault coverage level: is 90%, 95% or 99% fault
coverage for single stuck-at faults sufficient?
Another issue related to test quality is the fault model that is used for test
development. Comprehensive sequential fault models, such as the delay fault
18

Power consumption concems during manufacturing testing are basically related with the
stressing of the chip's package and thermal issues. Power consumption concems during
on-line periodic testing are related with the duration ofthe system's battery which may be
exhausted faster if long self-test pro grams are executed.

Embedded Processor-Based Seit-Test

89

model, are known to lead to higher test quality and higher defect coverage
than traditional, simpler combinational fault models such as the stuck-at fault
model. On the other hand, sequential fault models consume significantly
more test development time and also lead to much larger test application
time because of the larger size of the test sets they apply to the circuit.
In summary, in software-based self-testing a hunt for higher fault
coverage and/or a more comprehensive fault model under consideration
means:

more test engineering effort and more CPU time for manual and
automatic (ATPG-based) self-test program generation,
respectively;
larger self-test programs and thus longer downloading time from
external test equipment to the processor's memory;
longer test application time, i.e. longer self-test program
execution intervals, thus more time that each chip spends during
testing.

When a self-test program guarantees very high fault coverage for a

comprehensive fault model, this denotes a high quality test strategy.
In software-based self-testing, there is also a fundamental question that
must be answered by any potential methodology: what extent of fault
coverage and test quality is at all feasible?
Software-based self-testing being a non-intrusive approach may not be
able to achieve fault coverage levels that test approaches based on structured
DtT techniques can obtain. Software-based self-testing is capable to detect
faults that can possibly appear in normal operation of the circuit and
therefore performs only the absolutely necessary testing to the chip. This
way the problem of overtesting the processor is avoided. Overtesting can
happen when a chip is rejected as faulty as a consequence of the detection of
faults that can never happen while the chip operates in its normal mode. In
structured DtT based testing techniques like scan-based testing, chips are
tested under non-functional mode of operation and therefore overtesting can
happen with severe impact on production yield. On the other side, the
existence of DtT infrastructure in a chip makes its testing easier.
A basic issue throughout this Chapter is the total test cost of softwarebased self-testing. Therefore, in subsequent sections, we will try to quantify
the test cost terms and thus explain the reasons why software-based selftesting is a low-cost test methodology.

90

5.2.2

Chapter 5 - Software-Based Processor Self-Testing

Test engineering effort for self-test generation

A major concern in test generation for electronic systems is the manpower and related costs that are spent for test development. This cost factor
is an increasingly important one as the complexity of modem chips
increases. In processors and processor-based designs, software-based selftesting provides a low-cost test solution in terms oftest development cost.
Any software-based self-testing technique can theoretieally reach the
maximum possible test quality in terms of fault coverage (i.e. detection of all
faults of the target fault model that can appear during system's normal
operation) if an unlimited time can be spent for test development. Of course,
unlimited test development time is never allowed! A limited test engineering
time and corresponding labor costs can be dedicated during chip
development for the generation of self-test programs.
The ultimate target of a testing methodology is summarized in the
following sentence: obtain the maximum possible fault coverage/test quality
under specijic test development and test application cost constraints.
Therefore, if a methodology is capable to reach a high fault coverage level
with small test engineering effort in short time, it is indeed a cost effective
test solution and particularly useful for low-cost applications. This aspect of
software-based self-testing is the one that is analyzed more in this Chapter.
Of course, if a particular application's cost analysis allows an unlimited
(or very large) effort or time to be spent far test development, then higher
fault coverage levels can be always obtained.
When discussing software-based self-test generation in subsequent
sections we focus on the importance of different parts of the processor for
self-test programs generation so that high fault coverage is obtained by
software self-test routines as quickly as possible. By applying this "greedy",
in some sense, approach, limited test engineering effort is devoted to test
generation to get the maximum test quality under this restrietion.
Figure 5-4 presents in agraphie way the main objective of softwarebased self-testing as a low-cost testing methodology. Between Approach A
and Approach B both of whieh may eventually be able to reach a high fault
coverage of 99%, the best one is Approach A since it is able to obtain a
95%+ fault coverage quicker than Approach B. "Quicker" in this context
means with less engineering effort (man power) and/or with less test
development costs. This is basic objective of software-based self-testing as
we describe it in this Chapter.

Embedded Processor-Based Seit-Test

.
!

100% - ---- --- --- ----------- -1- --- - -- -- -

91

Approach A
t-------------------------

95%

Approach B

CD

Cl

CD

>
o
U

""
~

t\I

U.

Effort or Cost

Figure 5-4: Engineering effort (or cost) versus fault coverage.

When devising a low-cost test strategy, the most important first step is to
identify the cost factor or factors, if more than one, that are the most
important for the specific test development flow. Low test engineering effort
may be one of the important cost factors because it is calculated in personnel
cost for test development (either manual or EDA assisted). But, on the other
side, test development is always a one-time effort that results to a self-test
program which is eventually applied to all the same processors. Therefore,
the test development cost and corresponding test engineering efforts are
divided by the number of devices that are fmally produced. In high volume
production lines test development costs have marginal contribution to the
overall system costs. If production volume is not very high then test
development costs should be more carefully considered in the global picture.
This is the case of low-cost applications.

Test application time

5.2.3

Total test application time for each device of a production line has a
direct relation with the total chip manufacturing cost. In particular, test
application time in software-based self-testing consists ofthree parts:

self-test pro gram download time from extemal test equipment

into embedded memory;
self-test pro gram execution and responses collection time;
self-test responses upload time from embedded memory to
extemal test equipment for evaluation

92

Chapter 5 - Software-Based Processor Self-Testing

The relation between the frequency ofthe external test equipment and the
frequency of the processor under test determines the percentage of the total
test application time that belongs to the downloading/uploading phases or
the self-test execution phase. The basic objective of software-based selftesting is to be used as a low-cost test approach that utilizes low-cost, lowmemory and low pin-count extern al test equipment. To this aim the most
important factors that must be optimized in test application is the time of the
downloading of the test program and the uploading of the test responses, i.e.
the first and third parts of software-based self-testing as given above. A
simple analysis of the total test application time of software-based selftesting is useful to reveal the impact of all three parts oftest application time.
Let us consider that the processor under test has a maximum operating
frequency of f up while the external tester l9 has an operating frequency of
ftester- This means that a self-test program (code and data) can be
down loaded at maximum rate of ftester and it can be executed by the
processor at a maximum rate of f uP'
There are two extremes for the downloading of the self-test pro gram into
the embedded memory of the processor. One extreme is the parallel
downloading of one self-test program word (instruction on data) per cyc1e.
Fast downloading of self-test programs can be also assisted if the chip
contains a fast Direct Memory Access (DMA) mechanism which is able to
transfer large amounts of memory contents without the participation of the
processor in this transfer. The other extreme is the serial downloading of a
self-test program bit per cycle. This is applicable is case that aserial
interface is only available in the chip for self-test program downloading but
this is a rare situation. If the total size of self-test pro gram (code and data) is
sm all enough, then even the simple serial transfer interface will not severely
impact the total self-test application time.
Let us also consider that a self-test program consisting of C instruction
words and D data words must be downloaded to the processor memory
(instruction and data memory) and that a final number of R responses must
be eventually uploaded from processor data memory to the tester memory
for external evaluation. Finally, we assume that the execution of the entire
self-test pro gram and the collection of all responses take in total K c10ck
cyc1es of the processor. For simplicity, we assume that the K c10ck cyc1es
inc1ude any stall cycles that may happen during self-test pro gram execution
either for memory accesses (memory stall cyc1es) or for processor internal
reasons (pipeline stall cyc1es).

19

We use the term "tester" denoting any extemal test-assisting equipment. It can be an
expensive high-frequency tester, a less expensive tester or even a simple personal
computer, depending on the specific application and design.

Embedded Processor-Based Self-Test

93

The total test application time for software-based self-testing is roughly

given by the following formula:
T

(C+D) 1ftester + K/f up + R/ftesten

(C+D+R) 1ftester + K/fup,

= W/ftester +

K/fup,

or

or

where W

C+D+R

It is exact1y this last formula that guides the process of test generation for
software-based self-testing and the embedded software co ding styles, as we
will see in subsequent sections. The relation between the ftester and fup
frequencies reveals the upper limits for the self-test program size
(instructions - C, and data - D) and responses size (R) as weIl as for the seIftest program execution time (number of cycles K).
Figure 5-5 and Figure 5-6 use the above formula to show in a graphical
way, how the total test application time (T) of software-based self-testing is
affected by the relation between the frequency of the chip (fup) and
frequency of the tester (ftester) as weIl as by the relation between the
number of clock cycles ofthe pro gram (K) and its total size (w).
Figure 5-5 presents the application time of software-based self-testing as
a function ofthe ratio K/w, for three different values ofthe ratio fup/ftester.
A large value of the K/w ratio means that the self-test program
(instructions+data+responses) is small and compact (smaller value for w)
andJor it consists of loops of instructions executed many times each (small
code but large execution time). Obviously as we see in Figure 5-5, in all
three cases of fupl ftester ratios (2, 4 and 8), when the K/w ratio increases
the total test application time increases. The most important observation in
the plot is that when the fupl ftester ratio is smaller (2 for example), this test
application time increase is much faster compared to larger fupl ftester ratios
(4 or 8). In other words, when the chip is much faster than the tester then an
increase in the number of clock cycles for self-test execution (K) has a small
effect on the chip's total test application time. On the other side, ifthe chip is
not very fast compared to the tester, then an increase in the number of clock
cycles leads to a significant increase in test application time.

94

Chapter 5 - Software-Based Processor Self-Testing

Q)

c:
0

.~

fu p l f tester

Q.

~
(;j
Q)

upl

tester

16

KJW ratio
Figure 5-5: Test application time as a function ofthe K/W ratio.

From the inverse point of view, Figure 5-6 shows the application time of
software-based self-testing as a function of the ratio fup/ ftesten for three
different values ofthe ratio K/W. An increasing value ofthe fup/ftester
ratio denotes that the processor chip is much faster than the tester. This
means that for the same ratio K/W the test application time will be smaller
since the pro gram will be executed faster. This reduction becomes sharper
and more significant when the K/W ratio has smaller values (2 for example),
i.e. when the self-test pro gram is not so compact in nature and the number of
elock cyeles is elose to the size ofthe program.

95

Embedded Processor-Based Seit-Test

K/W
Q)

E
i=
c:
0

Q.

~
Ci)

K/W

16

f~ftester ratio
Figure 5-6: Test application time as a function of the fup/ftester ratio.

In the above discussion, W is the sum of the self-test code, self-test data
and self-test responses, C, D and R, respectively. When test responses of the
processor components are compacted by the self-test program (a special
compaction routine is part ofthe self-test pro gram) then R is a sm all number,
possibly 1, when a single signature is eventually calculated. This leads to a
reduction of time fot uploading the test responses (signatures) but on the
other side, the self-test pro gram execution time (number of clock cycles K) is
increased because of the execution of the compaction routines.
We have to point out that the above calculations and values ofFigure 5-5
and Figure 5-6 are simple and rough approximations of the test application
time of software-based self-testing based only on the time required to
downloadJupload the self-test program (code, data and responses) and the
time required to execute the self-test pro gram and collect the test responses
from the processor. A more detailed analysis is necessary for each particular
application, but the overall conclusions drawn above will still be valid for
the impact of the fup/ ftester and K/W ratios on the test application time of
software-based self-testing.
In the case of on-line periodic testing of the processor, the
downloadJupload phases either do not exist at all or are executed only at

96

Chapter 5 - Software-Based Processor Self-Testing

system start-up20. Therefore, the self-test pro gram size does not have an
impact on the test application time in on-line testing. On the other side, the
size of the self-test pro gram is important in on-line periodic testing because
it is related with the size of the memory unit where it will be stored for
periodic execution.
5.2.4

A new self-testing efficiency measure

In external testing or hardware-based self-testing, the number of applied

test patterns that are necessary to attain certain fault coverage gives a
measure of the test efficiency. The number of test patterns multiplied by the
test application frequency (one test pattern per clock cycle or one test pattern
per scan) determines the overall test application time.
In software-based self-testing the number of applied test patterns do not
directly correspond to the overall test application time because this depends
on the number of clock cycles than a self-test pro gram needs before each test
pattern is applied. Additionally, the test application time for (manufacturing)
software-based self-testing has another significant portion: the self-test
program downloading time.
Based on the brief analysis for the test application time of software-based
self-testing given in the previous section 5.2.3, we define a new measure for
the efficiency of software-based self-testing that includes all parameters that
determine test application time.
The new measure is called sbst-duration (SD) and is defined as:

so

= W+ K

where, W=C+ 0+ Rand Kare as defined in section 5.2.3 (number of words

for code, data and responses, respectively) and Q=ftesterl fup (the ration
between the tester and the processor frequencies).
The sbst-duration measure gives a simple, frequency-independent idea of
the duration of software-based self-testing, just like the number of test
patterns do for external testing or hardware-based self-testing. The actual test
application time can be obtained by multiplying sbst-duration with the
external tester's frequency.
Using the sbst-duration measure, different software-based self-testing
methodologies can be compared independently of the actual frequencies of
the processor or the extern al test equipment. The same fault coverage for the
processor can be possibly by self-test programs with different sbst-duration

20

Ifthe self-test code and data are stored in a memory unit such as a ROM, then there is no
download/upload phase. If they are transferred trom a disk device into the processor
RAM, this will only happen once at system start-up.

Embedded Processor-Based Self-Test

97

measures. The one with the smallest measure can be easily considered more
efficient at a given ration between the frequencies of the external tester and
the processor.

5.2.5

Embedded memory size for self-test execution

Embedded software routines that are executed in a processor are fetched

from embedded memory that stores the instruction sequences as weIl as their
data. In software-based self-testing, the maximum size of a self-test program
(code and data) may be restricted by the size ofavailable on-chip memory.
In the case of manufacturing testing of processors using software-based
self-testing, the memory used for self-test pro gram execution is the on-chip
cache memory. Self-test programs in this case must be designed so that
external memory cycles are eliminated (caused by cache misses). In a SoC
application where limited cache memory is available on-chip, software selftesting routines should not exceed in size the memory size. If software selftesting routines can't be accommodated in the cache memory entirely, then
multiple loadings ofthe memory with portions ofthe self-test code and data
are necessary with a corresponding impact on the total test application time
of the chip. If self-test code and data are stored in main memory instead of
cache memory, then the available memory size is larger but the self-test
execution time will be longer.
On-chip memory limitations are more critical when software-based selftesting is used for on-line periodic testing while the system operates in its
normal environment. In this case, software self-testing programs are regular
"competitors" of system resources with normal programs executed by the
user and thus available instruction and data memory is usually much less
than during manufacturing testing where the entire memory is available can
be used for software-based self-test execution.
In on-line periodic testing, self-test code and data may be permanently
stored in the system, so that it can be executed periodically (during idle
intervals or in intervals where the system normal execution is paused).
Therefore, the size of the self-test program and data is a serious cost factor
for the system in terms of extra silicon area occupied by the dedicated
memory (a ROM or flash memory are both appropriate candidates for such a
configuration). If a permanently stored self-test pro gram leads to serious cost
increase, then an alternative way to apply on-line, software-based self-testing
is to load the self-test pro gram to the chip's memory units at system's startup. This happens once and self-test program is available for periodic
execution but it permanently occupies part ofthe system's memory.

98

Chapter 5 - Software-Based Processor Self-Testing

5.2.6

Knowledge of processor architecture

The effectiveness and efficiency of any test development strategy is

related to the level of information that is necessary to be available for the
circuit under test. In the case of processor cores, if only one piece of
information is available this must be the processar's Instruction Set
Architecture, which includes:

the processor's instruction set: instruction types, assembly

language syntax and operation ofthe machine language
instructions;
theprocessor's registers which are visible to the assembly
language programm er;
the addressing modes for operands reference used by the
processor's instructions;

The Instruction Set Architecture information is always available for a

processor in the programmer's manual and can be considered a minimum
basis for software-based self-test generation. As we mentioned in Chapter 4,
functional testing methodologies for processors can be based only on this
information.
At the other end of the spectrum, structural testing techniques (including
structural software-based self-testing techniques) for processors, require
availability of the low-Ievel details of the processor. This means that gate
level implementation details of the processor components are necessary for
test generation21 If such low-Ievel information for the processor under test
actually exists and is available to the test engineer and EDA tools for
(possibly constraints-based) test generation, then this flow can lead to very
high fault coverage with limited engineering effort. In many cases, gate level
information of an embedded processor is not available far automatic test
generation at the structural level, but even if it was, automatie test pattern
generation for complex sequential machines like processor can 't be handled
even by the most sophisticated EDA tools 22
The objective of software-based self-testing, as a low-cost self-testing
methodology, is to be able to reach relatively high fault coverage and test
quality levels at the expense of small test engineering effort and cost. In
order to preserve the generality of the approach and make it attractive to

21

22

Either a gate-level netlist must be available or a synthesizable model ofthe processor from
which a gate-level netlist can be generated after synthesis.
Even non-pipelined processors are very deep sequential circuits, and sequential ATPG
tools usually fail to generate sufficiently small test sets that reach high fault coverage for a
processor.

Embedded Processor-Based Seit-Test

99

most applications, as few as possible information ab out the processor's

architecture and implementation should be required.
Software-based self-testing is generic processor testing methodology
which can be used independently of the level of information know for the
processor core (soft core, firm core or hard core). Its efficiency may vary
from case to case depending on the available information. The more
structural information is available the higher fault coverage can be obtained.

5.2.7

Component based self-test code development

Like every divide-and-conquer approach, if component-based self-test

routines development is used in software-based self-testing it is able to
manage the complexity of test generation for embedded processors.
Processors are sophisticated sequential machines where all known design
methods are usually applied to get the best performance of the circuit. Even
the most advanced ATPG tools are not able to develop sufficient sets of test
patterns for such complex architectures. Dividing the problem of processor
test generation and in particular self-test routines development into smaller
problems for each of the components, makes the solution of the problem
feasible.
Component-based test development, allows a selective approach where
the most important components are targeted for test generation, so that the
highest possible structural fault coverage is obtained with the smallest test
development effort and cost. As we will see in subsequent sections of this
Chapter, the most important components of the processor in terms of
contribution to the total fault coverage are considered first.
If the test generation problem is considered at the top-level of the
processor hierarchy (the entire processor) without using ATPG tools but
following a pseudorandom philosophy (use of pseudorandom instruction
sequences, pseudorandom operations and pseudorandom operands ), then
serious drawbacks, such as low fault coverage and long test application time,
have been extensively reported in the literature. These drawbacks are due to
the fact that processors are inherently pseudorandom resistant circuits.
A last but not least indication that component-based self-test routines
development is indeed an effective approach to software-based self-testing,
is that the most successful works in the literature (see Chapter 4) that report
successful application and relatively high structural fault coverage results on
software-based self-testing for processors are those that are componentbased in nature.

100

Chapter 5 - Software-Based Processor Self-Testing

5.3

Software-based self-test methodology overview

In this section we give an overview of software-based self-testing and its

breakdown into different phases. Subsequently, we analyze each of the
phases along with simple informative examples.
As we have already stated, the intention of this book is not to describe a
single approach for software-based self-testing of processor cores, but on the
contrary to discuss the overview and specific details of software-based selftesting with the final aim to make clear that it is a low-cost self-test solution
that targets high structural fault coverage for processors.
Although several functional self-testing techniques have been applied in
the past to processors (see the bibliography review given in Chapter 4), they
are not always suitable for low-cost self-testing because of two main
reasons:

they don't target high fault coverage for a particular structural

fault model but rather they focus on coverage of a functional
fault model or a function-related metric, therefore, they don't
focus on structural testing and fail to obtain high structural fault
coverage;
they are mostly pseudorandom-based and rely on the application
of pseudorandom instruction sequences, pseudorandom
operations, pseudorandom operands or a combination of these
three; due to this pseudorandom nature functional testing
approaches for processors require very long instruction
sequences andJor very large pro gram sizes for self-testing, while
at the same time, they are unable to reach the high levels of
structural fault coverage because of the random pattern resistance
of several processor components.

We present our view of software-based self-testing as a high test quality

methodology for low-cost self-testing that achieves its objectives being:

oriented to structural testing adopting weil known structural fault

models; software-based self-testing has been and can be used
with respect to combinational (stuck-at) or sequential (delay)
fault models;
oriented to component based test generation, i.e. separate selftest routines are developed for selected processor components,
actually for the most important components of the processor first;
a significant part of the methodology is spent on prioritization of
the processor components according to their importance for
testing;

Embedded Processor-Based Self-Test

101

focused on the low-cost characteristics of the self-test routines

that are developed, i.e. small size of the self-test pro grams (code
and data), sm all execution time, small power consumption, all of
them under the guidance of the primary goal which is always the
highest possible structural fault coverage.

As we will see, component level test generation can be of different

flavors depending on the information that is available for the processor core,
the individual characteristics of the processor's Instruction Set Architecture,
as well as the constraints of a particular application. Therefore, component
level test generation can be:

based on a combinational or sequential A TPG tool which can be

guided by a constraints extraction phase executed before ATPG;
the extracted constraints describe the effect that the processor's
Instruction Set Architecture has on the possible values that can
be assigned to processor's components inputs 23 ;
based on pseudorandom test patterns sequences which are
generated by software-emulated pseudorandom patterns
generators 24 ;
based on known, pre-computed deterministic test sets for the
components of the processor; such pre-computed test sets are
available for a set of functional processor components such as
arithmetic and storage components.

Moreover, these different approaches for component level self-test

generation can be combined together for a specific processor. A subset ofthe
processor's components may be targeted one way and others with an other.
We outline software-based self-test development as a process consisting
of four consecutive phases: Phases A, B, C and D, which are analyzed in the
following paragraphs, along with visual representation of each ofthem.
The starting point of software-based self-testing as a low-cost testing
methodology is the availability of two pieces of information as a basis for
the methodology:

the Register transfer Level (RTL) description ofthe processor;

the Instruction Set Architecture (ISA) ofthe processor.

ISA is in all cases available in the programmer's manual ofthe processor

and describes the instruction set of the processor and assembly language, the
23

24

ISA-imposed constraints extraction is not always possible. Obviously, ATPG-based test

generation for processor components can be used when a gate-level model ofthe processor
is available.
ISA-imposed constraints extraction may be necessary in this case too.

Chapter 5 - Software-Based Processor Self-Testing

102

visible registers it includes as weIl as the addressing modes used for

operands access. Detailed knowledge and understanding of the instruction
set and assembly language of the processor, consists a key in the successful
application of software-based self-testing to it. Particularly, in terms of lowcost testing, "clever" assembly language coding is very important in both
cases when self-test code is manually or automatically generated. Moreover,
the particular details and restrictions of an instruction set architecture playa
significant role on the applicability of software-based self-testing. For
example, the same component-Ievel test set can be more efficiently
transformed to a self-test routine in the assembly language of one processor
than in the assembly language of another. Different instruction sets may lead
to different self-test program sizes and execution times.
RTL description of the processor can be either available in a very high
level only indicating the most important processor components and their
interconnections, or it may be available in more detailed and accurate form if
an RTL model of the processor is provided in some hardware description
language like VHDL or Verilog. Such a description is sometimes available
either in a synthesizable form when the processor is purchased by the
designer as a soft core, or at least in a simulatable form for high level
simulation of the processor model when integrated in the SoC architecture.
Any R TL description of the processor is useful since it allows an easy way
to identifY the processor's building components, the existence of which
would otherwise be only speculated.
A low-cost software-based self-testing methodology can consist of the
following four phases A to D as summarized in Figure 5-7 and further
detailed subsequently.
PhaseA

Information Extraction

!
Phase B

Components
Classitication/Prioritization

1
Phase C

Component-Ievel Seit-test
Routines Development

!
Phase D

Processor-Ievel Seit-test
Program Optimization

Figure 5-7: Software-based self-testing: overview ofthe four phases.

Embedded Processor-Based Seit-Test

103

In the following paragraphs, each of the four phases A through D

presented in a visual way, to enhance readability.

IS

Phase A (Figure 5-8).

During this phase, the R TL information of the processor as weH as its
Instruction Set Architecture information is used to identifY information that
will subsequently be used for the actual deve10pment of the self-test
routines. In particular, during Phase A, the methodology:

identifies all the components that the processor consists of; this is
an essential part of the component-based software-based selftesting because test development is performed at the component
level;
identifies the operations that each of the components performs;
and
identifies instruction sequences consisting of one or more
instructions for controlling the component operations, for
applying data operands to the operations and for observing the
results of operations to processor outputs.
Phase A
Identity Processor
Components

--

'RT-Levellnfo

"-

--

ISA Info

11--......

Identify Component
Operations

Identity Instruction
Sequencesto
Control/Apply/Observe
each Operation

Figure 5-8: Phase A of software-based self-testing.

The finaloutcomes ofPhase Aare the following:

a set ofprocessor components C;

a set of component operations Oe for each ofthe components;
a set of instruction sequences I e, 0 for the controlling, the
application and the observation of operation 0 to a component C.

104

Chapter 5 - Software-Based Processor Self-Testing

These outcomes of Phase Aare subsequently used for Phases B, C, and D

for the generation of efficient self-test routines for the processor's
components.

Phase B (Figure 5-9).

During this second phase of software-based self-testing, processor
components (identified in Phase A) are classified in different categories
depending on their role in the processor's operation. Each component
class/category has different characteristics and different importance in the
processor's testability. After classification, the processor components are
prioritized to reflect this importance in the overall processor's testability.
This prioritization stage is very important in the context of low-cost selftesting because it helps the self-test pro gram development process to quickly
reach sufficient fault coverage levels with as sm all as possible test
development effort and with an as small and fast as possible self-test
program.
Phase B

--

-...,

'-

---

Classify Processor
Components Types

.-"

Info from
Phase A

Prioritize Processor
Components for Test
Development

Figure 5-9: Phase B of software-based self-testing.

Phase C (Figure 5-10).

This third phase of software-based self-testing is the actual component
level self-test routine development phase. It uses as input the following
pieces of information:

the information extracted during Phase A: set of processor's

components, set of operations for each component and
instruction sequences for the controlling, application and
observation of each component operation;
the information extracted during Phase B: components
classification and components prioritization;
a components test library that contains sufficient sets of test
patterns and test generation algorithms in a unified assembly
pseudocode for processor components, previously derived using
any available method (combinational or sequential A TPG-

105

Embedded Processor-Based Self-Test

generated,
pseudorandom-based,
deterministic tests, etc)

known

pre-computed

Phase C
I'--

./

Info from
Phases A, B

"r

r.......

Develop Self-Test
Routines for HighPriority Components

.....

--"

Component
Test Library

Stop when Sufficient

Fault Coverage is
Obtained

......

Figure 5-10: Phase C of software-based self-testing.

The objective of Phase C is the development of self-test routines for

individual processor's components based on the test sets provided in the
component test library. The coding style of the self-test routines may differ
from one component to another and also depends on the type of test set or
test algorithm available for the component in the component test library
(deterministic or pseudorandom).
Component self-test routines development starts from the "most
important" components to be tested, i.e. the components that have been
previously (during Phase B) identified to have higher priority for self-test
routines development. The criteria that determine which components are of
higher priority are analyzed in the corresponding subsection below, but
basically are the size (gate count) of each ofthe components and how easily
the component is accessible (easy controlling of its inputs and easy
observation of its outputs).
Higher priority components are first targeted, because the primary
objective of software-based self-testing is to reach high structural fault
coverage as soon as possible, without even targeting each and every of the
processor components.
When test development for one targeted component is completed, the
overall processor structural fault coverage is the criterion that determines if
software self-test routines development has to stop or to continue. As we will
see in the next Chapter where extensive experimental results are given, very
high fault coverage can be obtained with just a few important processor
components going through self-test routines development. The remaining
components are either sufficiently tested as a side-effect or their overall
contribution to the missing total fault coverage is very small.

Chapter 5 - Software-Based Processor Self-Testing

106

The final result of Phase C of software-based self-testing is a set of

component self-test routines developed in the order of their priority and
assembled together in a self-test program for the processor. The following
code shows just the outline of the overall self-test program for the processor
consisting of component's self-test routines.
self-test program for processor:
seIf-test routine for component 1
seIf-test routine for component 2
seIf-test routine for component k

After the execution of each self-test routine for a component

(where k:O:n the number of all processor components) the
test responses of all the components have been sent out of the processor in
either an unrolled fashion (in several memory words, each one containing a
single response to a component test pattern) or in a compacted fashion (one
or a few signatures combining all test responses). Test response compaction
routines may be either integrated to each of the components' self-test
routines, or aglobaI compaction routine may be appended to the processor's
self-test program with the task of reading the unrolled test responses from
the memory and compacting them in one signature. Details are deiscussed in
subsequent section.
Ci, i=l, 2, ... , k,

Phase D (Figure 5-11).

This last phase of software-based self-testing is an optional one and can
be utilized if a particular application has very strict requirements either for
sm aller size of the self-test pro gram or for smaller self-test execution time. It
is the self-test program optimization phase where the self-test routines
developed for each of the targeted components in Phase C are optimized
according to a set of criteria. The most usual way of self-test routines
optimization is the merging of self-test routines developed for different
components. Moreover, coding style changes can alter the format of a seIftest routine, optimizing it for one criterion.

Embedded Processor-Based Self-Test

107

Phase 0

-.....

I'--Routines from
Phase C

""'"
r

---.....11

-......,

Optimize Self-Test
Routines

I'--Optimization
Criteria

""'"

---.....11

Figure 5-11: Phase D of software-based self-testing.

The following subsections elaborate on the above four different phases of

software-based self-testing and in particular on the most critical parts of each
of them. The emphasis of the analysis is to justity several choices made in
low-cost software-based self-testing. Whenever suitable, examples are given
in terms of components types and classes, as well as, in terms of assembly
language routines.
The examples throughout this Chapter are simple and informative and are
derived using a popular Instruction Set Architecture of a well-known RISC
processor, the MIPS RISC processor architecture [22], [85]. Two of the
publicly available processor benchmarks that we study in the next Chapter
implement the MIPS architecture.
This classical load-store architecture is used to demonstrate several
aspects of software-based self-testing, while other different processor
architectures are also studied and software-based self-testing is applied to
them. Detailed experimental results are presented in the next Chapter.

5.4

Processor components classification

The components that an embedded processor consists of can be classified

in different categories according to their use and contribution to the
processor operation and instructions execution. In the following subsections,
we define generic classes of processor components and elaborate on their
characteristics. Figure 5-12 shows the classification of processor components
discussed in the following.

108

Chapter 5 - Software-Based Processor Self-Testing

Figure 5-12: Classes of processor components.

5.4.1

Functional components

The functional components are the components of a processor that are

directly and explicitly related to the execution of processor instructions. The
functional components are, in some sense, "visible" to the assembly
language programmer, or in other words their existence is easily implied by
the instruction set architecture of the processor. Functional components may
belong to one ofthe following sub-classes:

Computational fonctional components which perform specific

arithmetic/logic operations on data as instructed by the
functionality of processor instructions. Such components are:
Arithmetic Logic Units (ALUs), adders, subtracters,
comparators, incrementers, shifters, barrel shifters, multipliers,
dividers, or compound components consisting of some of the
previous. The computational functional components that realize
arithmetic operations may either deal with integer or floating
point arithmetic.
Storage functional components which serve as storage elements
for data and control information. Storage components contain
data that is fed to the inputs of computational functional
components or is captured at their output after their computation
is completed. This sub-class of storage functional components
includes all
assembly programmer visible registers,
accumulators, the register file (s), several pointers and special
processor registers that store control and status information
visible to the assembly programmer.

Embedded Processor-Based Seit-Test

109

Interconnect functional components which are components that

implement the interconnection between other types of processor
functional components and control the flow of data in the
processor's datapath, mainly between the previous two subclasses of computational and storage functional components.
Interconnect components include multiplexers controlled by
appropriate control signals generated after instruction decoding,
or bus control elements like tri-state buffers realizing the same
task.

The information on the number and types of the processor's functional

components can be directly derived from the RTL description of the
processor or, in the worse case, can be implied by the programmer's manual
and instruction set. Simply stated, the existence of the functional
components analyzed above is easily implied by a careful study of the
Instruction Set Architecture of the embedded processor. We outline some
simple examples to clarifY this statement.
If the instruction set of the processor includes an integer multiplication
instruction, this means that the processor contains a hardware multiplier of
either serial, multi-cycle architecture or of a parallel, single-cycle
architecture. The existence of the multiplication instruction itself delivers
this information, even if an RTL description ofthe processor is not available.
The integer multiplier, in this case, is a computational functional component.
For example, in the MIPS instruction set the integer multiply instruction25 :
mult Rs, Rt

implies that an integer multiplier component exists in the processor to realize

the multiplication of the contents of the 32-bit general purpose registers Rs
and Rt (both of the integer register file of the processor). According to the
MIPS instruction set architeeture description, the product of the
multiplication is stored in two special 32-bit integer registers named Hi and
Lo, not explicitly mentioned in the syntax ofthe multiplication instruction. If
the instruction takes 32 cycles to be completed - information available in the
programmer's manual - this means that the multiplier component
implements serial multiplication, while if the instruction takes 1 cycle to be
completed, the multiplier is a parallel one 26

25

26

In the examples of this Chapter, we use the assembly language of the MIPS processors.
We denote the registers as Rs, Rt, Rd or as RI, R2, etc, for simplicity reasons, although
traditionally, these registers are denoted as $sO, $sl, etc in the MIPS assembly language.
Multiplication using a parallel multiplier may take more than 1 clock cycle, if for
performance reasons of the other instructions, the multiplication is broken in more than
one phases and takes more (usually two) c10ck cyc1es.

110

Chapter 5 - Software-Based Processor Self-Testing

As a second example we consider the case in which the instruction set

contains an arithmetic or logic operation where the second, for example,
operand of the operation (in the assembly language description) can be a
general purpose register (from the register file), a memory location or an
immediate operand (coming directly from the instruction itself). In this case,
a multi-input (three-input in this case) multiplexer should exist at the second
operand input of the component that performs the operation. This
multiplexer is an interconnect functional component which existence is
easily extracted from the instruction set architecture even if an R TL
description is not available. For example, in the MIPS instruction set
architecture, the following two instructions perform the bitwise-OR
operation between two general purpose registers (first instruction) or a
register and an immediate operand (second instruction). In both cases, the
result is stored in a general purpose register Rd. The second operand of the
operation is in the first case another register R t and in the second case it is
an immediate operand directly coming from the instruction itself (the
immediate operand Irnrn in the MIPS machine language format consists of 16
bits ).
er Rd, Rs, Rt
eri Rd, Rs, Imm

A two-input multiplexer that feeds the second input of the ALU which
implements the bitwise-OR operation is in this case an identified
interconnect functional component.
Storage functional components are the easiest case of the three subclasses of functional components since in most cases their existence is
directly included in the assembly language instructions or explicitly stated in
the programmer's manual (like in the case of the Hi and La registers of
MIPS which although are not referred in assembly language are explicitly
mentioned in the programmer's manual of the processor). Usually, an
accumulator's or general purpose register's name is part of assembly
language format, while also a status register that saves the status information
after instructions are executed, is also directly implied by the format and
meaning of special processor instructions used for its manipulation. Getting,
again, an example from the MIPS instruction set architecture, the following
assembly language instruction:
and R4, R2, R3

identifies three general purpose registers from the integer register file: R2,
and R4. Moreover, an additional piece of information that can be
extracted from the existence of this instruction is that the processor contains
a register file with at least 2 read ports and at least 1 write port.

R3

Embedded Processor-Based Self-Test

111

In data-intensive processor architectures where data processing

parallelism is critical, like in the case of Digital Signal Processors (DSPs),
the class of functional processor components dominates the processor circuit
size more than in any other case of general purpose processor or controller.
This is true for example, because more than one computational functional
components of the same type may co-exist to increase parallelism, or
because more general purpose registers than in typical processors exist to
increase the available storage elements in the DSP architecture.
Functional components and in particular the computational sub-class of
functional components are usually large components in size and thus in
number of faults. In modern bus widths of 32 and 64 bits, and in DSP
applications with increased precision requirements and larger internal
busses, the functional units that perform ca1culations like addition,
multiplication, division or combined ca1culations like multiply-accumulate
in DSPs, are very large in size and consist of some tens of thousands of logic
gates. Such components have high importance in software-based self-test
routines development as we will see in a while after completing the
description of the different classes of processor components.
5.4.2

Control components

The control components are those that control either the flow of
instructions and data inside the processor or the flow of data from and to the
external environment (memory subsystem, peripherals, etc).
A classical control component is the component that implements the
instruction decoding and produces the control signals for the functional
components of the processor: the processor control unit. The processor
control unit may be implemented in different ways: as a Finite State
Machine (FSM), or a microprogrammed control unit. If the processor is
pipelined a more complex pipeline control unit is implemented.
Other typical control components are the instruction and data memory
controllers that are related to the task of instruction fetching from the
instruction memory and are also related to the different addressing modes of
the instruction set.
The common characteristic of control components is that they are not
directly related to specific functions of the processor or directly implied by
the processor's instruction set. Control components existence is not evident
in the instruction format and micro-operations of the processor and the actual
implementation of the control units of a processor may significantly differ
from one implementation to another. On the contrary, the functional
components of different processors are implemented more or less with the
same architecture.

112

Chapter 5 - Software-Based Processor Self-Testing

The control components of a processor are usually much sm aller in size

(gate count) compared to the functional components, but their testing is also
important because if they malfunction it is very unlikely that any instruction
of the processor can be correctly executed. Control components are
moreover, more difficult to be tested compared with functional components
because of their reduced accessibility and variety of internal
implementations.
With the scaling of processor word lengths towards 32 and 64 bits
functional components have grown in size while control components have a
smaller increase.

5.4.3

Hidden components

The hidden components of the processor are those that are included in a
processor's architecture usually to increase its performance and instruction
throughput.
The hidden components are not visible to the assembly language
programmer and user programs should functionally operate the same way
under the existence or absence of the hidden components. The only
difference in program execution should be its performance since in the case
of existence of hidden components performance must be higher than without
them.
A classical group of hidden components is the one consisting of the
components that implement pipelining, the most valuable performanceenhancing technique devised the last decades to increase processor's
performance and instruction execution throughput. The hidden components
related to pipelining include the pipeline registers between the different
pipeline stages 27 , the pipeline multiplexers and the control logic that
determines the operation of the processor's pipeline. These include
components involved with pipeline hazards detection, pipeline interlock and
forwardinglbypassing logic.
The hidden components of a processor may be either of storage,
interconnect or control nature. Storage hidden components have a similar
operation with the sub-class of storage functional components. Pipeline
registers belong to this type of storage hidden components. The pipeline
control logic has control characteristics similar to the control components
class. Finally, there are hidden components that implement pipelining which
have an interconnect nature. These are the multiplexers of the pipeline
structure of a processor which realize the forwarding of data in the pipeline
'27

Pipeline registers do not belong to the class of storage functional components because they

are not visible to the assembly language programmer.

Embedded Processor-Based Self-Test

113

when pipeline hazards are detected. The logic which detects the existence of
pipeline hazards consists of another type of hidden components, the pipeline
comparators which can be considered part ofthe pipeline controllogic.
Other cases of hidden components are those related to other performance
increasing mechanisms like Instruction Level Parallelism (ILP) and
speculative mechanisms to improve processor performance such as branch
prediction schemes. Such prediction mechanisms can be added to a
processor to improve programs execution speeds but their malfunctioning
will only lead to reduced performance and not to functional errors in
pro grams execution.
It is obvious from the description above that self-test routines
development for hidden processor' s components (or any other testing means)
may be the most difficult among the different components classes. The
situation is simplified when the processor under test does not contain such
sophisticated mechanisms (processor without pipelining). This is true in
many cases today where previous generations microprocessors are being
implemented as embedded processors. In a large set of SoC designs, the
performance delivered by such non-pipelined embedded processors is
sufficient, and therefore software-based self-testing for them has to deal only
with functional and control components.
In cases that the performance of classical embedded processors is not
enough for an application, modem embedded processors are employed. The
majority of modem embedded processors include performance-enhancing
mechanisms like a k-stage pipeline structure (k may range from 2 to 6 or
even more in embedded processors).
Although direct self-test routines development for hidden components is
not easy, the actual testability of some of them (like the pipeline-related
components) when self-test routines are developed for the functional
components of the processor can be inherendy very high. Intuitively, this is
true due to the fact that pipeline structure is a "transparent" mechanism that
is not used to block the execution of instructions but rather to accelerate it.
The important factor that will determine how much of test development
effort and cost will be spent on pipelining, is their relative size and
contribution to overall processor's testing.

5.5

Processor components test prioritization

It has been pointed out so far, that software-based self-testing is

considered as a low-cost test methodology for embedded processors and
corresponding SoC designs that contain them. Therefore, the primary goal of
reaching high structural fault coverage must be achieved at the expense of an
as low as possible engineering effort and cost for test generation and as sm all

114

Chapter 5 - Software-Based Processor Self-Testing

as possible test applicatin time. In software-based self-testing, the low-cost

test target can be achieved when small and fast test programs are developed
in the smallest possible test development cost. Towards this aim, the
processor components previously classified in the three classes of functional,
control and hidden must be prioritized in terms of their importance for test
generation. Self-test program development for each component will start
from the most important components and will then continue to other
components with a descending order of importance. Component level selftest routines development continues until sufficient fault coverage is
reached. The following flow diagram of Figure 5-13 depicts this iterative
process of software-based self-testing.
Prioritized list 1-_-1 Component self-test
routine development
of Component

No
Yes

Figure 5-13: Prioritized component-level self-test pro gram generation.

The prioritization criteria for a processor's components are analyzed in

this subsection. After prioritization is finished, component level self-test
routines development is performed. The components with higher priority are
first targeted and self-test routines are developed for them. Fault simulation
evaluates the overall obtained fault coverage and determines if further test
development is necessary to reach the required fault coverage for the
processor. A gate level netlist of the processor is required fr gate level fault
simulation. We must note that a gate level netlist at this point is necessary
only to decide if test development must continue. The cmponent level test
development prcess on the other side may or may not need such a piece of
information to be available as we see later.

Embedded Processor-Based Seit-Test

115

The criteria that are used for component prioritization for low-cost
software-based self-testing are discussed and analyzed in the subsequent
paragraphs. The criteria are in summary the following:

Criterion 1 - component size and percentage of total processor

faults set that belongs to the component.
Criterion 2 - component accessibility and ease of testing using
processor instructions.
Criterion 3 - correlation of the component's testability with the
testability of other components.

We elaborate on the importance of the three criteria in the following

subsections.

5.5.1

Component size and contribution to fault coverage

This criterion, simply stated, gives the following advice: component level
self-test routines development should give higher priority to large
components that contain large number of faults.
When a processor component is identified from the instruction set
architecture and the RTL model of the processor, its relative size and its
number of faults as a percentage of the total number of processor faults is a
valuable piece of information that will guide the development of self-test
routines. Large processor components containing large numbers of faults
must be assigned higher priority compared to sm aller components because
high coverage of the faults of this component will have a significant
contribution to the overall processor's fault coverage. For example,
developing a self-test routine that obtains a 90% fault coverage for a
processor component that occupies the 30% ofthe processor gate count (and
faults number), contributes a 30% x 90% = 27% to the total processor fault
coverage, while a 99% fault coverage on a sm aller component that occupies
only 10% ofthe processor gate count (and faults as well) will only contribute
by 10% x 99% = 9.9% to the overall processor fault coverage. Needless to
note, that reaching 90% fault coverage in a component is always much easier
than to reach 99% in another.
This criterion, although simple and intuitive, is the first one to be
considered if low-cost testing is the primary objective. Large components
must definitely be given higher priority than smaller ones since their
effective testing will sooner lead to large total fault coverage.
The actual application of this criterion for the prioritization of the
processor's components requires that the information for the gate count of
the processor components is known. Unfortunately, this information is not

Chapter 5 - Software-Based Processor Self-Testing

116

always available to the test engineer28 In this case, software-based self-test

development can be only based on speculations for the relative sizes of the
processor components.
Two speculations are in almost all cases true and can be easily followed
as a guideline for test devclopment:

functional components of all sub-classes (computational,

interconnect and storage) are larger in size and thus is faults
count than control and hidden components; therefore, they
should be assigned high er priority for self-test routines
development;
among the three types of functional components those with the
largest size are the computational ones that perform arithmetic
operations like adders, ALUs, multipliers, dividers and shifters
and also the storage components, the most important of which
are the register files.

A safe way to verify the correctness of the above statements is by

providing data from representative processors and their components gate
counts. The experimental results presented in Chapter 6 are towards this aim.
As a first indication we mention that, depending on the synthesis library and
the internal architecture of the components, the computational functional
components of the PlasmaIMIPS processor model [128] occupy from
24.06% up to 48.98% of the total gate count of the processor. The
computational functional components in this processor model include an
Arithmetic Logic Unit (ALU), a multiplier (that can be either aserial or a
parallel one), a divider (in serial implementation) and a shifter component.
Moreover, the register file ofthe embedded processor occupies from 37.98%
up to 56.74% of its total gate count. Combined, the computational functional
components and the register file (a storage functional component) of the
processor occupy at least 80.89% and as much as 87.72% of the total
processor area. Again, the different percentages depend on the synthesis
library, the synthesis options and also the different internal architectures of
the processor components (e.g. parallel vs. serial multiplier). The total
number of gates in the several different implementations of Plasma range
between 17,500 and 31,000 gates.
The above gate counts and corresponding percentages give a clear
indication ofthe importance ofthe computational functional components and

28

Gate counts are available either when a gate-level netlist of the component is given or
when a synthesizable model of the processor (and thus the components) is available for
synthesis.

Embedded Processor-Based Self-Test

117

the storage functional components and at least the first reason for which they
should be first addressed by software-based self-test routines development.

5.5.2

Component accessibility and ease of test

The second, but equally important with the first one, criterion for
prioritization of processor components for low-cost self-test program
generation, is the component's accessibility from outside the processor using
processor instructions. The higher this accessibility is, the easier the testing
of the component iso
The development of self-test routines is much easier and requires less
engineering effort and cost when the component under test is easily
accessible by programmer-visible, general purpose registers ofthe processor.
This means that the component inputs are connected to such registers and
component outputs drive such registers as well.
In this case, the application of a single test pattern to the component
under test by a self-test program simply consists of the following three steps:

execute instruction(s) to load input register(s) with test pattern

from outside (data memory or instruction memory29);
execute an instruction to apply the test pattern to the component;
and
execute instruction(s) to store the result from register(s) to
outside the processor (memory)

These three steps correspond to the software-based self-testing steps we

first mentioned in Figure 5-3. A simple example below shows such an easy
application of a test pattern to a shifter component of a MIPS-like embedded
processor.
lw R2, offsetl(R4)
lw R3, offset2(R4)
srlv Rl, R2, R3
sw Rl, offset3(R4)

The component under test is a shifter and the operation tested by this
portion of assembly code is the right logical shift of the shifter. Register R2
contains the original value to be right-shifted (first part of the test pattern to
be applied to the shifter) and register R3 contains the number ofpositions for
right-shifting (second part ofthe test pattern for the shifter). In our example,
both these values are derived from memory (base address contained in
29 When the immediate addressing mode is used, test patterns are actually stored in the
instruction memory, i.e. as partofthe instructions.

118

Chapter 5 - Software-Based Processor Self-Testing

register R4 incremented by the amount offsetl or offset2, respectively).

The loading of these two values is done by the first two lw (load word)
instructions. The third instruction of the code, applies the test pattern to the
shifter (s r 1 v = shift right logical by variable; the shift amount is stored in a
variable/register) and the shifted result (the output of the shifter) is stored
into register Rl. Finally, the last instruction (sw - store word), stores the
content of register Rl (the component's test response) to a memory location
outside the processor. In the sw instruction the memory location where the
test response will be stored is aga in identified by a base register content (R4)
and an offset to be added to it (offset3).
Such assembly language code portions for the application of component
level test patterns can be constructed only for components which have direct
access from/to general purpose registers. In most cases such components are
only the computational functional components of the processor such as the
shifter component discussed in the example. The inputs of the computational
functional components are directly accessible by programmer visible
register(s) which in turn can be easily loaded with the desired values (test
patterns) and also the components outputs (result of the calculation - test
response) are driven to other programmer visible register(s) which values in
turn can be easily transferred outside of the processor and stored to memory
locations for further evaluation.
Equally weIl accessible and easily testable are most of the storage
functional components of the processor because of the availability of
processor instructions that directly write values to such components (like
general purpose registers, accumulators, etc) and also instructions that
directly read their values and transfer them out of the processor. For
example, a test pattern can be applied to the general purpose register R6 of a
MIPS-like processor with the following lines of assembly language code30
li R6, test-pattern
sw R6, offset(Rl)

30

We remind that Ri is not the usual notation of registers in MIPS assembly language, but
rather, it is $50, $51, $tO, $t1, etc. We use the Ri notation for simplicity.

119

Embedded Processor-Based Seit-Test

The first instruction is the load immediate (li) instruction31 which loads
the register R6 with the test pattern value to be applied to the component
under test: register R6. The li instruction does not apply a test pattern which
is stored in data memory (as an 1 w instruction does) but a test pattern which
is stored in instruction memory (the instruction itself). The second
instruction (sw) stores the content of the register R6 (which is now the
component's test response) to a memory location addressed by the base
address in register Rl increased by the affset.
The same simplicity in self-test program development and application
does not apply to other components than computational and storage
functional components because (a) they are either not connected to
programmer-visible registers but are rather connected to special registers not
visible to the programm er; and/or (b) they can not be directly accessible by
processor instructions for test pattern application.
Therefore, computational and storage functional components must be
given priority for self-test routine development since they can quickly
contribute to a very high fault coverage for the entire processor with simple
self-test routines. If we combine this second criterion of component
accessibility and ease of testing with the previous criterion of relative
component size, we can see that functional components are very important
for self-test pro gram development in a low-cost software-based self-testing
approach. The third criterion discussed in the following subsection supports
further this importance.
5.5.3

Components' testability correlation

The third criterion is related to the ability to test some components as a

side effect of testing other ones. This means that when a set of test patterns is
applied to a processor component to obtain sufficient fault coverage, another
processor component is tested to some degree as weIl. Let us examine
situations where this applies.

31

When a functional component (a computational or a storage one)

is being tested by an instruction sequence specially developed for

Actually, the li instruction is not areal machine instruction ofthe MIPS architecture but
rather an assembler pseudo-instruction (also called macro) that the aseembler decomposes
to two instructions lui (load upper immediate) and ori (or immediate). Load upper
immediate loads a 16-bit quantity to the high order (most significant) half of the register
and ori loads the low order 16 bits of it. Therefore, the instruction li R6, testpattern is translated to:
lui R6, high-half
ori R6,

where

R6,

low-half

test-pattern~high-half & low-half (&

denotes concatenation).

120

Chapter 5 - Software-Based Processor Self-Testing

this purpose, then part ofthe controllogic ofthe processor is also

tested in parallel. Such part is for example the instruction decode
part of the control logic. In particular, when the instruction
sequence that tests the functional component contains a sufficient
variety of processor instructions then it is likely that significant
fault coverage for the instruction decode part is obtained as weIl.
When a functional component is being tested, then part of
interconnect functional components is also tested in parallel.
Multiplexers at the inputs or the outputs of the functional
component under test will be partially tested.
When a functional component is being tested, then part of the
pipeline logic of the processor is also tested in parallel. At least
parts of the pipeline registers and parts of the multiplexers of
forwarding paths will be tested to some extent. Therefore, if the
instruction sequence that tests the functional component contains
instructions that pass through different paths of the pipeline
forwarding logic, then the pipeline logic is also sufficiently
tested.

We remark at this point that the criteria described in this and the previous
subsections are only meant to be used to prioritize the importance of
processor components for test development and are not in any sense
statements that are absolutely true in any processor architecture and
implementation. These criteria are good indications that some components
must be given priority over others for a low-cost development of self-test
routines.
Concluding with this third criterion, we mention that when other
components than functional components are being tested, for example
control components, it is very unlikely that other components are being
sufficiently tested as weIl. For example, when executing a self-test program
for the processor instruction decode component, what is necessary is to pass
through all (or most) of the different instructions of the processor. When
such a self-test pro gram is executed only a few of its instructions detect also
some faults in a computational functional unit, like an adder, which requires
a variety of data values to be applied to it (the adder). Therefore, sufficient
testing of the decoder component does not give, as a side effect, sufficient
fault coverage for the adder component or other functional components. On
the contrary, when sufficient fault coverage is obtained by a self-test routine
for the adder component then, in parallel, the decode unit is also sufficiently
tested at its part that is dedicated for the decoding of the addition-related
instructions.

Embedded Processor-Based Self-Test

121

In the global view, when separate self-test routines have been developed
for all the processors functional units (or at least for the computational and
the storage components) then the other components (control components like
the instruction decode and the instruction sequencing components and
hidden components like the pipeline registers, multiplexers and control
logic) are also very likely to be sufficiently tested. The opposite situation is
not true: when self-test routines have been developed targeting the control
components or the hidden components then the functional components are
not sufficiently tested as well, simply because the variety of data required to
test them is not included in the self-test routines far the control and hidden
components.
After having described the criteria for the prioritization of the processor
components far self-test routines development, we elaborate in the next two
subsections on the identification and selection of component operations to be
tested, as well as, on the selection of appropriate operands to test the selected
operations with processor instructions.

5.6

Component operations identification and selection

Phase A of software-based self-test development identifies a set of

instruction sequences 1e,o which consists of processor instructions I that,
during execution cause component C to perform operation o. The
instructions that belong to the same set 1e,o have different
controllability/observability properties since, when operation 0 is performed,
the inputs of component c are driven by internal processor registers with
different controllability characteristics while the outputs of component C are
forwarded to internal processor registers with different observability
characteristics.
Different controllability and observability for processar registers refers to
the ease of writing values to a register and transferring its contents to the
outside of the processor. Higher controllability means that a smaller number
of processor instructions are required to assign a value to a register; while
higher observability means that a smaller number of instructions are required
to transfer the contents ofthe register out ofthe processor.
Therefore, for every operation Oe of a component c, derived in phase A,
an appropriate instruction sequence I must be selected from the set of
instruction sequences 1e,o that apply the same operation to the same
component. The objective is to end up with the shortest instruction sequence
required to apply any particular operand to the component inputs and the
shortest instruction sequence required to propagate the component outputs to
the processor primary outputs.

122

Chapter 5 - Software-Based Processor Self-Testing

Let us see an example to demonstrate the instruction selection for the

computational functional component Arithmetic Logic Unit (ALU) of the
MIPS-like architecture. Such an ALU has two 32-bit inputs, one 32-bit
output and a control input wh ich specifies the operations that the component
performs under the control of the processor instructions. Figure 5-14 shows
the ALU component and its inputs and outputs. The operands are 32-bits
wide, the result is also 32-bits wide and the operation (control) input consists
of 3 bits (there are eight different operations that the component performs).
operand 1

operand 2

ALU
operation

result
Figure 5-14: ALU component ofthe MIPS-like processor.

The set of operations

following:
OALU

={

OALU

that the ALU component performs

IS

the

add,
subtract,
and,
or,
nor,
xor,
set_on_less_than_unsigned,
set_on_less_than_signed

For every operation 0 belonging to the OALU set, we identify the

corresponding set of processor instructions IALU,o that during execution
cause the ALU component to perform operation O. As the ALU can perform
8 operations, the sets of instructions are the following:
IALU,ADD
I ALu , SUBTRACT
IALU,AND
IALu,oR
I ALu , NOR
IALu,xoR

123

Embedded Processor-Based Self-Test

IALu,sET ON LESS_THAN_UNSIGNED
IALU,SET_ON_LESS_THAN_SIGNED

For the development of self-test routine for the ALU, one instruction I
from each set IALU,o (eight different sets for the eight different operations of
the ALU) is needed to apply a data operand for each operation o.
For example, the set IALU,NoR has only one processor instruction and thus
the selection is straightforward. The following instruction can only be used
to test for the NOR operation ofthe ALU:
nor Rd, Rt, Rs

The

sets

of instructions

IALu,oR,

I ALU , XOR,

I ALu , AND,

I ALu , SUBTRACT,

I ALu , SET_ON_LESS_THAN_UNSIGNED'
I ALU , SET_ON_LESS_THAN_SIGNED, all consist of two
instructions, one in the R-type (register) format ofMIPS where the operation
is applied between two registers of the processor and the other in the I-type
(immediate) format of MIPS where the operation is applied between a
register and an immediate operand.
The instructions in the I-type format have less controllability than the
instructions in the R-type format. Thus, the instructions in the R-type format
must be selected because they provide the best controllability and
observability characteristics due to the use of the fully controllable and
observable general purpose registers of the register file. Therefore, from
these sets, the following instructions will be selected to test the
corresponding operation ofthe ALU.

or Rd, Rt, Rs
xor Rd, Rt, Rs
and Rd, Rt, Rs
sub Rd, Rt, Rs
subu Rd, Rt, Rs
slt Rd, Rt, Rs
sltu Rd, Rt, Rs

Finally, the IALU,ADD set of instructions consists of a large number of

instructions since the ALU is also used in memory reference instructions that
calculate sums of a base register content and an offset. Instructions included
in this set are the following among others.
add Rd, Rs, Rt
addu Rd, Rs, Rt
addi Rt, Rs, Imm
addiu Rt, Rs, Imm
lw Rt, offset(Rs)
sw Rt, offset(Rs)

Chapter 5 - Software-Based Processor Self-Testing

124

In this case, the selected instruction would be any of the first two of the
listed instructions above that belongs to the R-type format because it
possesses the best controllability and observability since it uses the general
purpose registers of the register file ofthe processor.

5.7

Operand selection

As we saw in the previous subsection, each of the processor components

performs a set of different operations under the control of the processor
instructions. Of course, arithmetic and logic computational functional
components have the largest variety of operations, like in the case of a multifunctional ALU. The implementation ofthe different operations can be done
internally in many different ways, such as separate implementation of the
arithmetic and logic operations and their combination with a set of
multiplexers. At the instruction set level, and under the assumption that the
only available structural information of the processor may be the RT level
description of it, software-based self-testing concentrates on the application
of a set of test patterns at the component operand inputs so that any
particular operation is sufficiently tested.
To outline an example, we can again consider the case ofthe MIPS ALU
with the operations described in the previous subsection. Each of these
operations excites a different part ofthe processor ALU as Table 5-1 shows.
Operation

add
subtract
and
or
nor
xor
set on less than unsigned
set on less that signed

ALU part used

Arithmetic part (adder)
Arithmetic part (subtracter)
Logic part (and)
Logic part (or)
Logic part (nor)
Logic part (xor)
Arithmetic part (subtracter)
Arithmetic part (subtracter)

Table 5-1: Operations ofthe MIPS ALU.

We see in Table 5-1 that each operation excites a different part of the
ALU component (in this case that the component is an ALU, some
operations excite the arithmetic and some others excite the logic part of it).
Different sets of test patterns are required to excite and detect the faults in
these different parts of the component.
Appropriate selection of component-Ievel test patterns is an essential
factor for the successful development of self-test routines for components. In
this subsection the focus is on this aspect of software-based self-testing:
operand selection.

Embedded Processor-Based Self-Test

125

Component operand selection corresponds to component-level test

pattern generation and therefore the different styles oftest pattern generation
will be detailed. These different styles are discussed in the subsequent
paragraphs and will be related right after with the corresponding coding
styles for software-based self-testing.
5.7.1

Self-test routine development: ATPG

According to this approach, test generation for the processor components

is based on the use of a combinational or sequential Automatic Test Pattern
Generator (ATPG), depending on the type of component considered.
Application of ATPG-based self-test routines development requires the
availability of a gate level model of the processor (or at least for the
component under test) on which the ATPG will operate. Alternatively, a
synthesizable model ofthe processor may be used for gate-level ATPG after
synthesis has been performed on it. In the case, that neither a gate-level
netlist nor a synthesizable model of the processor is available (the processor
has been delivered as a hard core) ATPG-based self-test routine
development can't be supported.
A TPG-based test development for processor components may be or may
not be successful, depending on the complexity and size of the component
under test. In the case of simple combinational components, combinational
ATPG is usually successfully applied. On the contrary, in components with a
considerable sequential depth and with a large number of storage elements,
sequential A TPG algorithms and tools may be unable to reach sufficient
fault coverage for the component.
Moreover, as studied in many approaches presented in the literature,
component-level ATPG may require constraint extraction so that the derived
component test patterns can be applied to them using processor instructions.
Constraints that must be extracted are spatial constraints andlor temporal
constraints.
Constraint extraction is a promising direction related to processor testing
and in general sequential A TPG for large hierarchical designs, like
processors, but it is still under development and it is very likely that it will
lead to significant results and tools the next few years. Conceptually,
constraints extraction is a bottom-up approach, since low level constraints
are extracted for processor components and then transferred upwards to high
level processor instructions (called also realizable tests). This bottom-up
approach, although theoretically correct, does not always obtain sufficient
results. Constraints are either very difficult to be extracted, but even if this is
the case, mapping of the constraints to processor instructions is not a
straightforward task.

Chapter 5 - Software-Based Processor Self-Testing

126

In the case that a gate level model ofthe component is available (or it can
be obtained from synthesis) and combinational or sequential ATPG succeeds
to produce a test set of sufficient fault coverage (with or without the
assistance of constraints extraction), the result is a set of k component test
patterns:
atpg-test-pattern-l
atpg-test-pattern-2
atpg-test-pattern-k

The two important properties of the set of k test patterns for self-test
development are:

the cardinality ofthe test set (number k), and

their format and correlation.

A self-test routine that tests a processor component C for one of its

operations 0, using an instruction I selected among the instructions of the
set Ie,o and applying a test set of k test patterns is outlined below. We
assume, as in all places in this Chapter, a classical RISC load-store
architecture where arithmetic and logic operations are applied only to
general purpose registers.
atpg-tests-loop:
load register(s) with pattern(s) from memory
apply instruction I
store result(s) to memory
repeat atpg-tests-loop

The software loop in the above pseudocode will be repeated k times

where k is the total number of test patterns generated by the ATPG. In this
self-test routine style, the test patterns of the component have been stored in
data memory (as variables in the assembly language source code) and a loop
applies them consecutively to the component under test.
Alternatively, the component test patterns can be applied to it with
another style of self-test code. In this second style, the test patterns are stored
in the instruction memory of the processor (the instructions themselves) and
are applied using the immediate instruction format of the processor (also
called immediate addressing mode). The following pseudocode outlines this
self-test style for the application of ATPG-based test patterns.

Embedded Processor-Based Seit-Test

127

atpg-tests-no-loop:
load register(s) with ~ediate pattern 1
apply instruction I
store result(s) to memory
load register(s) with ~ediate pattern 2
apply instruction I
store result(s) to memory
load register(s) with ~ediate pattern k
apply instruction I
store result(s) to memory

In the loop-based application oftest patterns, they occupy part ofthe data
segment (thus data memory) of the self-test program as variables. In the
second case, the test patterns are not stored in variables (data memory) but
they rather occupy part ofthe code segment (instruction memory) ofthe seIftest pro gram.
As an example, consider that the functional component under test is a
binary subtracter, a gate-level netlist of the component is available and also
that the ATPG has generated a test set consisting of k test patterns. When the
ATPG-based test patterns are applied in a loop from data memory then the
following code of MIPS assembly language gives an idea of how the k test
patterns can be applied to the subtracte~2.
test-subtracter: andi R4, R4, 0
next-test:
lw R2, xstart(R4)
lw R3, ystart(R4)
sub Rl, R2, R3
sw Rl, rstart(R4)
addi R4, R4, 4
slti R5, R4, 4*k
bne R5, RO, next-test

Since the subtracter is a two-operand computational functional

component, the k test patterns that the ATPG generates have an X operand
part and a Y operand part. We assume that the k patterns are stored in data
memory in consecutive memory locations starting at xstart and ystart
addresses for the X operand and the Y operand, respectively. Register R4
and si ti are the immediate versions of the logic and, addition and set-onless-than instructions. Set-on-Iess-than sets the first register to 1 if the second operand is
less than the third. Branq if not equal (bne) instruction takes the branch if the compared
registers are not equal. Finally register RO (denoted $zero in MIPS assembly) has always
an all-zero content.

32 andi, addi

128

Chapter 5 - Software-Based Processor Self-Testing

counts the number of test patterns times fOU~3 to show the correct addresses
of data memory for loading the test patterns applied to the subtracter and
storing the test response. Register R4 also controls the number of repetitions
of the loop (k repetitions in total). Registers R2 and R3 are loaded with the
next pattern in each repetition and the result of the subtraction is put by the
subtracter in the Rl register. The Rl register that contains the test response
of the subtraction is finally stored to an array of consecutive memory
locations starting at address rstart as shown in the code. At the end of
each loop iteration, the counter R4 is incremented by 1 and acheck is
performed to see if the k test patterns have been exhausted. If this is not the
case, the loop is repeated.
The self-test code above consists of eight instructions (words) and 2k
words storing the k two-word test patterns (one word for operand X and one
word for operand Y). The execution time of this routine depends on the
number k of test patterns to be applied to the subtracter. The exact execution
time and number of clock cycles depends on whether the processor
implementation is pipelined or not and also on the latency of memory read
and write cycles.
For simplicity, le us consider a non-pipelined processor. If we assume
that each instruction executes in one clock cycle apart from memory reads
and writes which take 2 clock cycles, then a rough estimation of the number
of clock cycles required for the completion of above self-test routine is lOk
clock cycles 34
Figure 5-15 presents in a visual way the application of A TPG-generated
test patterns from a data memory array using load word instructions. Two
instructions are required to load the registers with the test patterns. The test
vectors are then applied by the subtract instruction to the subtracter
component as the code above shows.

33 In the case of MIPS, we consider words of 4 bytes and hence addresses of sequential
34

words differ by 4.
In a pipelined implementation, one instruction will be completed in each cycIe but with a
smaller period. Pipeline and memory stalls will increase the execution time of each loop.

Embedded Processor-Based Self-Test

129

test patterns
in data memory
words

data
memory
Figure 5-15: ATPG test patterns application from memory.

The alternative way to implement the self-test code for the application of
the k test patterns to the subtracter is to use the immediate operand
addressing mode that all (or most) processors' instruction set architectures
include. According to the immediate addressing mode, an operand is part of
the instruction. The MIPS architecture consists of 32-bit instructions where
the immediate operand can occupy a 16-bit part of the instruction. For
example, in the immediate addressing mode instruction:
andi Rt, Rs, Imm

which implements the logical AND operation between register Rs and

immediate operand Imrn and stores the result in register Rt, the immediate
operand Imrn is a 16-bits long.
In order to store a 32-bit value to a register using 16-bit immediate
operands an additional instruction is used in MIPS which loads the upper
half 16-bit part of the register. This instruction is called load upper
immediate (lui) and loads the 16-bits immediate value to the upper 16 bits
ofthe register Rt (bits 31 downto 16):
lui Rt, Imm-upper

In order to load the lower half part of the register, an or immediate (0 r i)

instruction can be used that loads the lower half (16-bit) part of the register
leaving the upper half part unchanged.

130

Chapter 5 - Software-Based Processor Self-Testing

ori Rt, Rt, Imm-lower 35

Therefore, for the application of the k test patterns to the subtracter the
following self-test routine can be applied.
test-subtracter: andi R4, R4, 0
lui R2, xtest-l-upper
ori R2, R2, xtest-l-lower
lui R3, ytest-l-upper
ori R3, R3, ytest-l-lower
sub Rl, R2, R3
sw Rl, rstart(R4)
addi R4, R4, 4
lui R2, xtest-2-upper
ori R2, R2, xtest-2-1ower
lui R3, ytest-2-upper
ori R3, R3, ytest-2-1ower
sub Rl, R2, R3
sw Rl, rstart(R4)
addi R4, R4, 4
lui R2, xtest-k-upper
ori R2, R2, xtest-k-lower
lui R3, ytest-k-upper
ori R3, R3, ytest-k-lower
sub Rl, R2, R3
sw Rl, rstart(R4)
addi R4, R4, 4

This self-test code consists of unrolled segments of code, each of which

applies one test pattern to the subtracter component. Registers R2 and R3 are
loaded with the 32-bit values in two instructions, the 1 ui instruction that
loads the upper half of the register and the 0 r i (or immediate) wh ich loads
the lower half of the register while leaving the upper half unchanged. The
test pattern is applied by the sub instruction while the test responses of the
subtracter are collected in an array of memory words starting at address

35

The sequence of 1 u i and 0 r i instructions is combined as an assembler macro (or

pseudoinstruction) called l i (load immediate) which loads a register with a full 32-bit
immediate value using 1 u i and 0 r i. We have decided not to incIude macros (or pseudoinstructions) in our code examples, so that the reader can quickly caIculate the total
number of words of the example, assuming one word per instruction. Macros, usually are
equivalent to more than one word each.

131

Embedded Processor-Based Seit-Test

rstart as in the previous case ofloop-based application and register R4 is

used as an index of the test responses array.
Figure 5-16 presents in a visual way the application of ATPG-generated
test patterns using immediate addressing mode instructions. Four instructions
are required to load the two input registers with the test patterns not using
any part of the data memory but instead of the instructions themselves (two
instructions for each register - the 1 u i and 0 r i). The test patterns are then
applied by the subtract instruction to the subtracter component as the code
above shows.

test patterns
in instruction
memory words
(part of the instructions)

instruction
memor

core

data
memory
Figure 5-16: ATPG test patterns application with immediate instructions.

The code using immediate addressing mode consists of 7k words (seven

words for each of the k test patterns). Assuming again a one cycle execution
for each instruction apart ftom memory reads and writes which take 2
cycles; the total execution time of the code is equal to about 8k clock cycles.
The size of the last self-test routine for immediate addressing application
ofthe ATPG-based test patterns is due to the fact that the uncorrelated k test
patterns for the subtracter can't be applied in any loop-based mann er (such
as in the case of application ftom the memory shown before). On the other
side, this unrolled coding style leads to a smaller routine execution time: 8k
compared with the lOk time of the loop-based routine we saw before.
Table 5-2 summarizes the characteristics of code size and execution time
for the two different cases of application of A TPG-based test patterns we
just studied. Moreover, the table shows the total test application time when
the two routines are applied using a 50 MHz tester and the operation

132

Chapter 5 - Software-Based Processor Self-Testing

frequency ofthe processor is 200 MHz (see the calculations given in Section
5.2.3). The number oftest responses is k.

Coding style

Processor frequency - 200 MHz

Tester frequency - 50 MHz
Words
Responses Cycles
Test Application
Time

ATPG-based
2k+B
o . 11 k+ 0 . 16
k
lOk
loop from
memory
ATPG-based
0.20 k )lsec
7k
k
Bk
with immediate
Tabte 5-2: ATPG-based self-test routines test application times (case I).

)lsec

We can see in Table 5-2 that in this case where the processor is faster
from the tester to a reasonable level (processor is 4 times faster than the
tester), the loop-based application of ATPG-based patterns from data
memory is about two times faster in test application time compared to the
immediate addressing mode application ofthe same k test patterns.
If, on the other side, the tester is ten times slower than the processor chip
(consider for example a tester of 100 MHz frequency used for a 1 GHz
processor) then the difference between the two approaches becomes more
significant, as Table 5-3 shows.

Coding style

Processor frequency - 1 GHz

Tester frequency -100 MHz
Words
Responses Cycles
Test Application
Time

ATPG-based
2k+B
o . 0 4 k + O. 0 B
k
lOk
loop from
memory
ATPG-based
7k
k
0.088 k )lsec
Bk
with immediate
Tabte 5-3: ATPG-based self-test routines test application times (case 2).

)lsec

It is also useful to make another remark regarding the ATPG-based

approach for component self-test routines development. The remark is that
when the value of k is small (this is the case for some relatively simple
processor components) this approach (either in the loop format from
memory or in the immediate format) gives a very good solution to softwarebased self-testing because the test routine size and its execution time will
both be very sm all and the obtained fault coverage will be very high.
Unfortunately, for the most important components of a processor it is not
always possible to derive a small test set using an ATPG. Therefore, when
the number k is large, either the size of the self-test routine is excessively
large (this is the case when the immediate addressing mode is used) or the

Embedded Processor-Based Self-Test

133

execution time of the pro gram of excessively high (this is the case in both
cases of loop-based approach and the immediate addressing mode approach).
Wehave to note at this point that the dock cydes that the load word and
store word instructions need (the CPI - docks per instruction) has an impact
on the exact values of the previous Tables which are only indicative. Also,
the self-test routines execution time is affected by the existence or not of a
pipeline structure.

5.7.2

Self-test routine development: pseudorandom

Pseudorandom testing is another popular approach in digital systems

testing, mainly because of its simplicity in test generation (actually, there is
no test generation phase! in the sense that ATPG defines it).
When pseudorandom testing is applied to a digital circuit, a long
sequence oftest patterns is usually applied to the circuit's primary inputs. In
hardware-based self-testing the pseudorandom test patterns are generated by
hardware state machines specially synthesized for this purpose36 The most
common hardware pseudorandom pattern generators are the Linear
Feedback Shift Registers (LFSRs), the Cellular Automata (CA) and the
pseudorandom pattern generators based on arithmetic circuits like adders,
subtracters, multipliers. All these three types of circuits along with several
modified versions presented in the literature have been proved to have
excellent properties in generating sequences of random-like patterns.
LFSR-based pseudorandom testing is the most popular pseudorandom
technique. The quality of the pseudorandom sequences depends on the initial
value loaded into the register (called seed) and the polynomial that the
internal connections ofthe LFSR realize.
LFSR-based pseudorandom self-testing can be also implemented using
embedded self-test routines. Self-test routines actually emulate the
generation of pseudorandom pattern sequences by realizing the polynomial
evaluation in software instead ofhardware. There are two different ways that
an LFSR-emulating pseudorandom self-testing approach can be applied to a
processor core to test its modules:

36

development of aseparate pseudorandom self-test routine for

each component that applies the generated test patterns directly
to the component;
development of a main, programmable LFSR-emulation self-test
routine that will generate pseudorandom sequences after being

The hardware-based self-testing scheme where test patterns are applied from the memory
does not apply to pseudorandom testing because ofthe large number oftest patterns.

134

Chapter 5 - Software-Based Processor Self-Testing

called by component self-test routines and apply them to

components
The advantage of the first approach of separate routines for
pseudorandom patterns generation is that the interaction with the memory is
very limited and patterns can be dircctly applied to the component under test
without having been put in the memory before. The drawback of this
approach is that the self-test code size is larger because a separate routine (or
even a set of routines) is developed for each component around the basic
processor instructions that apply the tests to the component operations.
An example of the code in this first approach of pseudorandom based
self-testing is given below again for testing the subtracter component of the
processor. We assume that an LFSR is emulated by processor instructions.
Other pseudorandom patterns generators schemes can be employed as weil
(like cellular automata or arithmetic circuits) but we show the LFSR-based
scheme in our example code since this is the most classical scheme in
hardware-based pseudorandom self-testing and has also been applied to
software-based pseudorandom self-testing for processor cores.

Embedded Processor-Based Self-Test

135

test-subtracter: sub RII, RII, RII

ori R8, RO, max-patterns
lw R9, polynomial-mask-x
lw RIO, polynomial-mask-y
lw R2, seed-x
lw R3, seed-y
sub RI, R2, R3
next-sub:
sw RI, rstart(RII)
addi RII, RII, 4
subi R8, R8, 1
beq R8, RO, exit
ori R4, R2, 0
andi R4, R4, 1
srl R5, R2, 1
beq R4, RO, complete-x
xor R5, R5, R9
complete-x:
andi R2, R5, FFFF
ori R4, R3, 0
andi R4, R4, 1
srl R5, R3, 1
beq R4, RO, complete-y
xor R5, R5, RIO
complete-y:
andi R3, R5, FFFF
j

next-sub

exit:

In the above self-test code, registers R2 and R3 are the pseudorandom

operands that will be applied to the subtracter at each iteration of the loop.
The result is collected in register Rl and stored to the responses array after
subtraction has been performed. Register RB counts the number of
pseudorandom patterns applied and initially contains the maximum number
(number of iterations of the main loop). Registers R9, RIO contain the masks
that implement the characteristic polynomial of the software-emulated
LFSRs. Registers R2 and R3 are initially loaded with the seeds. We assume
in our example that a different LFSR is implemented for each ofthe X and Y
operands (R2 and R3 registers). Registers R4 and RS are used to implement
the algorithm of the LFSR next value calculation (extract the rightmost bit;
look if it is a 1; right shift the previous value; xor with the mask if rightmost
bit is 1). Register Rll is the index to the test responses array.
We note that the routine above applies pseudorandom test patterns to a
subtracter using its basic instruction (sub). In order to test each component

136

Chapter 5 - Software-Based Processor Self-Testing

and each operation of every component, a separate routine must be designed

and applied to the component. The routine above consists of 24 words.
Alternatively, more than one instructions can be applied to the processor for
each new pseudorandom pattern (or pair of X, Y patterns) generated. This
means that after the sub instruction, another one will be inserted and applied
to the same (or other) component and response will be transferred out ofthe
processor. Each iteration of the basic loop above that generates and applies
an new pattern takes 16, 17 or 18 cycles (because of the taken or untaken
branches of the above code depending on whether the rightmost bit of the
current LFSRs value is 0 or 1).
The advantage of the second approach of a main, programmable routine
for pseudorandom patterns generator is that it avoids many copies of
instructions needed for each of the separate routines for each component.
The drawback of this approach is that it requires several calls to other
routines in the program, which will increase the overall duration of the seIftest program execution. Therefore, it is again adecision regarding the
tradeoffbetween longer test application time and larger self-test routine.
The outline of a self-test routine according to the second approach is the
following:
test-component:

ori R8, R8, max-patterns

# initialize LFSR routine (seed, mask)

next-pattern:

# call LFSR routine for first operand

# call LFSR routine for second operand
# apply the target instruction
sw Rl, rstart(Rll)
addi Rll, Rll, 4
subi R8, R8, 1
beq R8, RO, exit
j next-pattern

exit:

In the code above, we assurne that the test response is collected in Rl and
then stored to the responses array. The LFSR routine can be called as many
times as necessary for the specific operation being tested (usually two times
for two-operand operations, so that in the first time the one operand's next
random value is generated and in the second time the other operand's
random value is generated).
We mention at this point that all code examples given in this Chapter
using the MIPS instruction set may be differently applied to other processor

Embedded Processor-Based Self-Test

137

architectures and of course mayaiso be differently applied in MIPS-like

processors depending on the co ding style that an assembly programmer
prefers and the specific requirements of the application. Several changes can
be done to the code examples given in this book. The examples we present
are only indicative, and their purpose is to give an idea for how the related
approaches may be applied, using as a demonstration vehicle a popular
instruction set architecture of a successful processor.

5.7.3

Self-test routine development: pre-computed tests

A third component test development approach is based on the use of

regular sets of known, pre-computed test patterns for its key components.
Such pre-computer test patterns are very useful and effective when the netlist
ofthe component (or the entire processor) is not available for ATPG. Sets of
known, pre-computed test patterns can be stored in a components test
patterns library and accordingly applied to processor's components.
In the literature, there has been proposed several such test sets which are
either characterized by a small number of test patterns or by a regularity and
correlation between the test patterns they consist of. In the former case, such
pre-computed test sets were basically developed for external testing and this
is why the test set size is small. In the latter case, these test sets were
developed in such a way so that efficient hardware built-in self-test
generators can be designed for their on-chip production and delivery to the
components under test. Such test patterns are deterministic and not
pseudorandom and have been specially designed to obtain high fault
coverage for common architectures of several components. Some of these
test patterns sets have been shown to obtain very high structural fault
coverage for any word length of the components and any of its different
internal architectures 37 In that sense, the regular, deterministic test sets are
generic and when applied to a processor component, no fault simulation is
necessary to prove their effectiveness. Of course, the exact fault coverage
they obtain on a particular circuit can only be evaluated if a gate-level netlist
is available, but even with no fault simulation these test sets are known to
provide very high structural fault coverage.
Processor components which can be effectively tested with a set of
deterministic, regular test patterns are:

37

the arithmetic and logic components: adders, subtracters, ALUs,

multipliers, dividers, shifters, barrel shifters, comparators,
incrementers, etc; (computational fimctional components) [20],
[53], [54], [55], [115], [162];

See, for example, the works on multipliers testing [53], [54], [95], detailed in a few pages.

138

Chapter 5 - Software-Based Processor Self-Testing

the registers and register files (storage functional components)

[98];
the multiplexers of the data busses (interconnect fonctional
components).

The main advantages of the use of regular, pre-computed, deterrninistic

test patterns are the following:

component architecture independence: this applies when the

same test set is apriori known to be effective (at least to some
sufficient level) for different architectures of the component;
sm all test set size: regular deterrninistic test sets consist of a
sm all number oftest patterns;
ease of on-chip generation: this is true not only for hardware
based self-testing but also for software-based self-testing as we
see in the following paragraphs

Regular, deterrninistic test sets consist of test patterns that have a relation
between them, which makes them easily on-chip generated (either by
hardware or by software). Each test pattern of the test set can be derived
from the previous one by a simple operation such as:

an arithmetic operation (addition, subtraction, multiplication);

a shift operation
a logic operation
combination of the three above

One may think that pseudorandom tests are generated in a similar way
(LFSR emulation consists of shifts and exc1usive-or operations). The
difference between the two approaches is that pseudorandom testing requires
the generation of a large number oftest patterns, while regular, deterrninistic
relies on a sm all test set.
Regular, pre-computed, deterrninistic test sets combine the positive
properties of ATPG-based and pseudorandom based test development for
processor components. Therefore, in cases when ATPG-based test
development is not able to obtain high fault coverage with very few test
vectors, and also in cases when pseudorandom test development is not able
to obtain a high fault coverage within a reasonable amount of time (c1ock
cyc1es), regular pre-computed sets of test patterns is the solution. They reach
high fault coverage levels with a number of test patterns (and thus c10ck
cyc1es) much less than pseudorandom testing and a bit more that ATPGbased test development. On-chip generation of regular, deterrninistic test sets
is as easy as pseudorandom test development while the self-test program size
is reasonably small.

Embedded Processor-Based Seit-Test

139

As a key example, we mention the case of multipliers testing [53], [54]

where it was proven that a regular test set of 256 test patterns can obtain a
fault coverage of more than 99% for different types of array multipliers
(array multipliers, tree multipliers, Booth encoded or not) and for any word
length of the multiplier. This test set of 256 test patterns (which was later
shown to be equally sufficient when test patterns were reduced to 225 and
100 respectively [52], [93], is a little larger than the test set that an ATPG
can produce for an array or tree multiplier and much smaller that the test set
that is necessary to obtain such a high fault coverage with pseudorandom
testing (around 200 test patterns can reach a 95% fault coverage, while a
99% fault coverage needs more that 1000 or even 2000 test patterns,
depending on the multiplier structure and gate-level details. Moreover, the
regular test set can be very easily generated on-chip by a dedicated, counterbased circuit (equally easy as an LFSR-based hardware self-testing
approach).
In the case of software-based self-testing, such regular, deterministic test
sets can be also easily generated by compact self-test routines.
Table 5-4 summarizes the characteristics ofthe four different componentlevel test development techniques, where self-testing is performed using
embedded software routines.
Approach

Fault
coverage
High

Seit-test
code size
Small

Test application
time
Short

ATPG loop trom

memory
ATPG with
High
Large
immediate
Pseudorandom
Medium
Small
Regular
High
Small
Deterministic
Table 5-4: Characteristics of component self-test routines development.

5.7.4

Short
Long
Short

Self-test routine development: style selection

Now, that we have analyzed and discussed the different test development
approaches for theprocessor components, the question to be answered is:
which one of the approaches is selected for a particular component of a
processor? The answer may seem straightforward after the long analysis, but
it is always useful to summarize.

140

Chapter 5 - Software-Based Processor Self-Testing

We select the ATPG-based, self-test routine development approach jor

processor components when:

the component is a smalI, combinational or sequential component

and an ATPG can generate a sm all number (some tens) of test
patterns;
a gate-level netlist of the component is available or can be
obtained;
constraints extraction can be performed for the component, so
that ATPG is guided by the constraints and is able to generate
test patterns that can be applied only with processor instructions;
there are no pre-computed test sets known for this type of
processor component;

In A TPG-based test development, high fault coverage is guaranteed

(provided that the ATPG tool succeeds to obtain high fault coverage).
If the question is which of the two ATPG-based code styles to select
(fetching test patterns from memory or using the immediate addressing
format), this decision depends on the actual number of test patters. If they
are very few then immediate addressing mode is simpler to be applied
(simple writing ofthe assembly code).
We select the pseudorandom-based, self-test routine development
approachjor processor components, when:

the component is expected or known to be pseudorandom pattern

testable and not resistant (even if the gate level netlist of it is not
available for fault simulation);
a gate-level netlist of the component is not available or can't be
obtained;
there are no pre-computed test sets known for this type of
processor component;

The actual fault coverage obtained in the pseudorandom case, can only be
calculated if a gate-level netlist is available, but the test development phase
(development of the self-test routines) can be done without any gate-level
information available.
We select the regular, pre-computed pattern based, self-test routine
development approach jor processor components, when:

there are known, effective, pre-computed test sets for this type of
component (even if the gate level netlist of it is not available);
a gate-level netlist of the component is not available or can 't be
obtained by synthesis;

141

Embedded Processor-Based Seit-Test

5.8

Test development for processor components

After having finished the discussion on component-Ievel test

development for processor components and the comparison between them,
we elaborate further on the specifics of the different classes of processor
components (functional, control and hidden).

5.8.1

Test development for functional components

The analysis of the different component-Ievel self-test routine

development approaches, given in the previous sections, mainly focused on
the functional components of the processor (classified in the computational,
storage and interconnect sub-classes). For the functional components, all
three self-test routine development styles (ATPG-based, pseudorandom, precomputed/deterministic) can be applied provided that their requirements for
test development are met.
The most important components of a processor belong to the functional
category and, therefore, self-test routines development for them can
completely be done using one of the self-test routine development
approaches presented so far.
The experiment results presented in the next Chapter support our
argument that the most important components ofthe processor (Le. the larger
and more easily testable ones) are actually the functional components.
Software-based self-test has been applied to several publicly available
processor models of different architectures, assembly languages and word
lengths. The experimental results prove our claim that software-based selftesting can be an excellent, low-cost self-testing methodology which obtains
very high fault coverage levels with sm all self-test programs.

5.8.2

Test development for control components

The second category of processor components, the control components,

has also a central role in the correct execution of the user programs that run
on the processor and thus must be considered for software-based self-test
generation, right after the most important components: functional.
Control components operation is related to:

instruction fetching and decoding;

addressing modes implementation and
(instruction and data memory);
internal and external busses interfacing, etc;

memory

access

142

Chapter 5 - Software-Based Processor Self-Testing

In most cases, control components of the processor are relatively small in

size and therefore their contribution to the overall processor fault coverage is
relatively low.
Let us consider the case of the control unit of a processor, that produces the
control signals for the several computational, storage and interconnect
functional components. Such control units are in many cases implemented as
finite state machines, and can be realized in many different ways (microprogrammed units, hardwired units, etc).
Testing of control units can be based on:

functional software-based self-test: the control unit (state

machine) can be tested by ensuring it passes through all of its
states (if all of them are known) and it passes through all normal
transitions between states (this is an exhaustive testing approach,
which is infeasible for large FSM designs);
scan-based,hardware-based self-test: if structured DIT
modifications are allowed, this is an efficient approach in many
cases, because DIT changes in the control unit most likely do not
have an impact on the critical path of the processor and thus can
be easily adopted by the chip designers;

Moreover, it has been observed (also noted in the experimental results of

the next Chapter) that, while self-test routines are being developed for the
functional components of the processor, they, in parallel, reach sufficient
fault coverage levels for the control units as well. This is particularly true, if
a relatively rich variety of processor instructions are used in the self-test
routines for the functional components. Therefore, processor control
components may be tested to some extend, as a side effect of the testing of
the processor's functional components.
If a control unit component must be specifically targeted by a self-test
routine, then this routine should include an exhaustive excitation of the
component's operations. For example, all the different instructions and all
different addressing modes of the processor must be applied if we want to
efficiently test an instruction decoding component. Such self-test routines
can be based on already existing routines previously developed for design
verification of the processor by designers. In design verification of the
processor, it is very common that a self-test program executes all different
processor instructions to prove that the processor works correctly in all these
cases. Such a routine (or set ofroutines) can be re-used to effectively test the
processor's control unit.
The experimental results of the next Chapter prove both the arguments
that control components are small in size and that they can be tested to some

Embedded Processor-Based Self-Test

143

sufficient level while self-test routines dedicated to functional components

are applied to the processor.

5.8.3

Test development for hidden components

The third category of processor components, according to our

classification scheme, consists of the hidden components of the processor.
The role of the hidden components is to improve the processor's
performance, although the assembly language programmer is not aware of
their existence and their actual structure is not visible.
The most important types of hidden components related to performance
improvement are the components that implement pipelining. Pipelining has
been the most important and successful performance improving technique
for processors during the last decades. A pipeline mechanism implemented
in a processor consists usually of:

a set of large multi-bit pipeline registers for the transfer of

instructions and control signals down the pipeline stages; this
way each instruction carries with it all necessary information to
continue execution although new instructions have already been
issued in previous pipeline stages;
a set of large multiplexers for the selection of different sources of
data when forwarding is implemented; this way pipeline stalls
are avoided due to data dependencies between instructions;
a control logic at each pipeline stage, which is usually very
small; for example pipeline logic detects the existence of data
hazards and activates forwarding or stalls, etc;

In the context of low-cost, software-based self-testing for processors, it is

not completely suitable for self-test program generation to target hidden
components such as the components of the pipeline structure of a processor.
Aiming to detect, with small and fast self-test pro grams, faults in
components that are not visible and their existence can't be implied from the
known information, may either be infeasible or very costly. Therefore, seIftest routines development for such components is not of high priority for
software-based self-testing. All that can be done is discussed in the
following paragraphs.
Since pipeline existence (or at least its exact implementation and details)
is not known to the assembly language programm er, the components related
to it can't be targeted directly by software-based self-test routines. On the
contrary, they can be indirectly tested when other components of the
processor are being tested. In many cases, this indirect testing can be quite
effective.

Chapter 5 - Software-Based Processor Self-Testing

144

Let us elaborate further on each ofthe three pipeline components above:

Pipeline registers: these registers are tested while the other

components of the processor are being tested; this is because the
instructions of the self-test routines pass through all the stages of
the pipeline applying a variety of test patterns to them. Faults in
the registers are excited and also easily propagated to the primary
outputs of the processor since every instruction carries all
necessary bits all the way down the pipeline final stage. In this
sense, the pipeline registers can be considered easy-to-test
components. Our experimental results show that this statement is
valid at least for our benchmark processors.
Pipeline multiplexers: these multiplexers can be easily tested as
weil, if the self-test routines for other components of the
processor make sure that the different forwarding paths of the
pipeline are excited; this can only be done if RTL information for
the actual implementation of the pipeline structure is available
for the processor.
Pipeline control: it consists of a relatively sm all number of gates
for each pipeline stage (mainly a set of small comparators that
compare parts of the instructions in consecutive pipeline stages
to identity necessary forwarding and potential pipeline stalls),
and thus does not require special test development and is
expected to be tested to some extend when the pipeline registers
and pipeline multiplexers are tested.

Pipeline multiplexers are large multi-bit multiplexers that implement the

forwarding operation in pipelines, work together with the relatively sm all
comparators of the forwarding unit. In case that the self-test pro grams
developed previously for the functional components of the processor do not
contain appropriate instruction sequences that excite all paths through the
multiplexers and comparators of the pipeline login, special routines can be
developed. The development of such routines can only be done if some
information for the actual implementation of the pipeline structure of the
processor is available, so that all different forwarding paths are used and
therefore all multiplexers and comparators are sufficiently tested.

145

Embedded Processor-Based Seit-Test

A1 A2 A3

ALU

Figure 5-17: Forwarding logic multiplexers testing.

For example, Figure 5-17 shows part of the pipeline logic where a 3input multiplexer is used at the A input of the ALU to select from three
different sources. Signals Al, A2, A3 are connected to the ALU input A
depending on the forwarding path activated in the processor's pipeline (if
such a path is activated at this moment). A successful self-test program
generation process must guarantee (if RTL or structural information of the
processor is available) that aB paths via the multiplexer (Al to A, A2 to A,
A3 to A) are activated by corresponding instructions sequences so that the
multiplexer component is sufficiently tested.
As we see in the experimental results of the next Chapter, pipeline
structures of the processors are relatively easily tested while self-test
routines are being applied to the other processor components. Of course, we
point out that the publicly available benchmark processors we use may have
a relatively simple pipeline structure compared with high-end modem
microprocessors or embedded processor. Unfortunately, the unavailability of
such complex commercial processor models makes the evaluation of
software-based self-testing on them difficult. Fortunately, software-based
self-testing is attracting the interest of major processor manufacturers and it
is likely that results will become available in the near future. This way, the
potentiallimitations (if any) of software-based self-testing may be revealed.
Moreover, software-based self-testing must be evaluated for other
performance-improving mechanisms such as the branch prediction units of
processors. In such cases, due to the inherent self-correcting nature of these
mechanisms, faults inside them are not observable at aB to processor outputs,
but only performance degradation is expected. Performance measuring
mechanisms can be possibly used for the detection of faults in such units and
research in this topic is expected to ga in importance in the near future.

146

Chapter 5 - Software-Based Processor Self-Testing

5.9

Test responses compaction in software-based selftesting

So far, we concentrated on the test pattern generation and delivery part of

software-based self-testing, while we assumed that all test responses of the
components under test were collected in a responses array in the processor's
data memory. Under this assumption, each test response can be separately
compared with the expected fault-free response, either by the processor itself
or by external equipment.
If a compacted self-test signature must be collected for all test patterns
applied to the processor, or a signature for each component under test, then a
special response compaction routine must be developed either for each of the
components (if aper component signature is needed) or for the processor in
total.
Compaction of self-test responses can be performed in two different
ways:

two-step compaction: component test responses are stored in a

data memory array and are then separately processed by a
response compaction routine that produces a single signature for
the entire processor or a single signature for each of the targeted
components;
one-step compaction: component test responses are compacted
"on-the-fly" while they are generated; part of the components
self-test routines is specially designed to compact the new test
response with the previous value ofthe self-test signature;

In two-step compaction, the component self-test routines may be smaller

since they do not include code for response compaction and simply store a
response to data memory. As we know, smaller self-test routines require
smaller download time from the external low-cost testers. The disadvantage
of two-step compaction scheme is that is has a very intensive communication
with the data memory (one write for each new component test response), and
this may impact the overall test application time (routine execution) due to
the long response time ofmemory. Moreover, two-step response compaction
occupies a significant amount of memory where the test responses must be
collected.
On the other side, one-step compaction of test responses leads to larger
self-test routines for the processor component and thus larger download
time, but they have a reduced interaction (if any) with data memory and are
therefore, most probably, faster routines than those with two-step
compaction.

147

Embedded Processor-Based Self-Test

Figure 5-18 depicts two-step response compaction while Figure 5-19

depicts one-step response compaction.

component test routines

"""'- - '

global
compaction
routine

component
test responses

/L--.:r--.......-..!final

memory

producethese

"====. IC under test

data

signature(s)

memory

Figure 5-18: Two-step response compaction.

component test routines

1TI"- - '

final
signature(s)

under
test

instruction

memory

data

memory

"====~ IC under test

Figure 5-19: One-step response compaction.

The compaction schemes that can be used in software-based self-testing

are the same as in hardware-based self-testing since the same properties are
seeked. The primary concern in test responses compaction is always the fault
coverage loss due to fault masking and aliasing. A smaller number of
collected signatures leads to higher probability of aliasing, while more
signatures reduce this probability and related fault coverage loss.
In pseudorandom software-based self-testing, aliasing probability is
theoretically smaller than in A TPG-based or regular patterns based, simply
because the number of test patterns applied is much larger in pseudorandom
testing. The low aliasing properties of compaction schemes based either on
Multiple Input Signature Registers (MISRs) or in Accumulator-Based
Compaction (ABC) scheme, are valid when the number of test patterns are
very large. On the contrary, when a few deterministic (ATPG or pre-

148

Chapter 5 - Software-Based Processor Self-Testing

computed) test patterns are applied, aliasing due to a compaction scheme

strongly depends on the exact patterns that are applied and therefore no
general argument can be valid. Experimental results are required to
determine if a compaction scheme is effective or not for a particular
processor component.

5.10

Optimization of self-test routines

Software-based self-testing is usually applied at the component level, i.e.

separate self-test routines are developed and subsequently applied to each
one of the targeted processor components. Not all the components are
targeted, since a sufficient fault coverage may be reached when self-test
routines are developed only for a subset of them (usually all the functional
components and some of the control components). Under this model, if
components CI, C2, C3 and C4 have been targeted and self-test routines Pj,
P2, P3 and P4 have been developed for them with sizes (in bytes or words) Sj,
S2, S3 and S4 then the total self-test program for the processor has a total size
equal to Sj+S2+S3+S4. Ifthe execution time of each ofthe four routines is Tl,
T 2 , T 3 and T 4, respectively, then the total execution time for the program will
be approximately equal to T j+T2+T3+T4 (some extra time will be necessary
for invoking the four routines from a main routine).
In many cases, optimization of a self-test pro gram for a processor is
necessary to make it more efficient in terms of size (smaller program) or
execution time (faster program). This optimization is necessary in order to:

reduce the self-test pro gram download time from the low-cost
external tester;
reduce the self-test program execution time;

As a consequence, the total test application time (download + execution)

will be reduced for each processor being tested. Of course, this last phase, of
self-test code optimization is optional and may be skipped when the initial
self-test program size and execution time are sufficiently small for the
particular processor and application, and no extra engineering effort is
necessary (or can be afforded) to optimize it further. Self-test pro gram
optimization is the fourth step of software-based self-testing (Phase D of
Figure 5-7.
In the two next sections we discuss two different self-test code
optimization approaches that may be applied separately or together. Other
techniques may be applied also.

Embedded Processor-Based Self-Test

149

"Chained" component testing

5.10.1

This self-test code optimization technique is based on the successive

testing of processor components where the test response of one of them is
used as a test pattern for the next component. Therefore, components are put
in a virtual "chain" and are tested one after the other. If, for example, three
processor components Cl, C2 and C3 are "chained" together in this order,
then an optimized self-test program repeats several cycles where a test
pattern is applied to Cl, its test response (stored in a register) is applied as a
test pattern for C2, its test response (stored again in a register) is applied as a
test pattern to C3 and the test response of C3 is transferred out of the
processor as the test response of the "chained" testing of the three
components.
This type of self-test code optimization has the potential to develop very
compact self-test routines that test all grouped components to a high level of
fault coverage. Its successful application depends on the following factors.

The function that a component performs must be able to give a

sufficient variety of responses that can be used as test patterns for
the subsequent components in the chain.
Errors at the outputs of the first components of a chain (due to
faults inside them) must be propagated to processor primary
outputs after they pass through subsequent components of the
chain. This must be true throughout the entire chain of
components.

Self-test code size and execution time reduction is almost guaranteed by

this technique although in some case there may be some compromise in the
fault coverage level because of observability problems, due to the second
factor given above.
The following simple example gives an idea of this optimization
technique considering that two components: an adder and a shifter are tested
together in a chained fashion. Figure 5-20 depicts this idea. Test instruction
I applies a test pattern to the adder. The adder test response is captured in a
register. The content of this register is used by test instruction 2 as a test
pattern for the shifter component. The shifter output is the combined test
response of the two components.

150

Chapter 5 - Software-Based Processor Self-Testing

Test Instruction 1

Test Instruction 2

adder test pattern

adder

shifter

shifter test response

Figure 5-20: "Chained" testing of processor components.

In this example, successful application of the chained testing of the two

components requires the following:

A sufficient test set for the shifter can be generated at the outputs
of the adder. In other words, appropriate inputs must be supplied
to the adder so that they test the adder itself and also provide a
sufficient test set for the shifter.

The shifter does not mask the propagation of error at the adder
outputs (because of faults in the adder) towards primary outputs
ofthe processor.

Let us consider that, if separate self-test routines are developed for the
two components, adder and shifter, each of them consists of a basic loop that
applies a set of test patterns to each component. Let us assume also, that the
test set for the adder applies 70 test patterns to it (this can be the case for a
carry lookahead adder) and that the test set for the shifter applies 50 test
patterns to it. Also, let us consider that the basic loop of the self-test routine
for the adder executes each iteration in 30 dock cydes, and that the basic
loop of the self-test routine for the shifter also executes each iteration in 30
dock cydes. Therefore, a self-test program that uses these two routines one
after
the
other
will
execute
in
approximately
70 x 30 + 50 x 30 = 3,600 dock cycles. Applying the "chained" testing
optimization technique will most probably lead to a combined loop that
applies to the adder and shifter a total of 80 test patterns and executes each
iteration in 31 dock cydes (one more instruction is added to the loop). The
total number of test patterns is larger since it may be necessary to "expand"
the set of adder tests so that their responses produce a sufficient test set for
the shifter which is tested subsequently. Moreover the larger loop execution

Embedded Processor-Based Seit-Test

151

time for each execution (31 instead of 30) is due to the fact that an extra
instruction is added to it for the application of the test pattern to the shifter.
In the combined, "chained" routine the total execution time of the combined
loop will be 80 x 31 = 2,480 clock cycles. The numbers used in this small
analysis are picked to show the situation in which the optimization technique
can lead to a more efficient self-test code. There may be, of course,
situations where the optimized code is not better than the original. For
example, this may be the case when the total iterations of the combined loop
are too many. This may be caused by the inability to reasonably "expand"
the adder test set so that a sufficient test set for the shifter is produced at the
adder outputs. If, in our simple adderlshifter example the total number of
combined loop iterations is 120 instead of 80 then the number of clock
cycles for the new loop will be 120 x 31 = 3,720 which is larger from the
original back-to-back execution ofthe two component routines.
The following pseudocodes show how "chained" component's testing
can be applied. First, two separate self-test routines for component's Cl and
C2 are given. Then, the combined routine for the "chained" testing of the
components is shown. In this example we assume that the components are
both originally tested with ATPG-based test patterns applied in a loop from
data memory.
atpg-loop-Cl:
load register(s) with next test pattern for Cl
apply instruction lCl
store result(s) to memory
if applied-patterns < Kl
repeat atpg-loop-Cl
atpg-loop-C2:
load register(s) with next test pattern for C2
apply instruction lC2
store result(s) to memory
if applied-patterns < K2
repeat atpg-loop-C2
atpg-loop-Cl-C2:
load register(s) with next test pattern for Cl
apply instruction lCl
apply instruction lC2
store result(s) to memory
if applied-patterns < max(Kl, K2)+m
repeat atpg-loop-Cl-C2

152

Chapter 5 - Software-Based Processor Self-Testing

Although we use an ATPG-based self-test routine style in the pseudocode

where patterns come from data memory, any of the other self-test routine
styles can be applied in "chained" components testing.
Chained testing of processor components can lead to self-test code size
and execution time reduction because:

Only one self-test routine in necessary for all components in the

chain. The routine may be sm aller than the sum of all the
separate routines.
The single self-test routine will execute faster than the
combination of the separate routines because it will contain only
one slightly slower loop instead of several ones, and also it will
have a much smaller number of interactions with the memory
system (to store test responses). Only the combined test response
of the final component in the chain is stored to memory.
The combined self-test routine will produce a much sm aller
number of self-test responses into processor memory and
therefore the uploading of these responses towards low-cost
tester memory will be shorter and total test application time will
be reduced.

In summary, "chained" testing of processor components has several

advantages over individual component-Ievel self-testing. It requires some
more sophisticated analysis to be done (on the feasibility of "chained"
testing for a particular processor) but it leads to significantly smaller self-test
programs with less loops, less memory interaction and less test responses.
Therefore, a sm aller test application time ofthe device is gained.
5.10.2

"Parallel" component testing

According to this second alternative for the optimization of the self-test

routines, the same set of test patterns is applied to more than one component
and therefore, the application of separate self-test routines to each of them is
avoided. We call this technique "parallel" testing of the processor
components, to denote that the components are tested with the same set of
patterns. Figure 5-21 shows "parallel" testing of processor components for a
pair of an adder and a subtracter component. The same test pattern is applied
to both of them with separate instructions (addition instruction for the adder
and subtraction for the subtracter).
This second optimization technique does not have the drawback of fault
masking or fault blocking that the "chained" testing technique has. In
"parallel" testing of processor components, each test response from the

153

Embedded Processor-Based Seit-Test

components is stored to the memory and not used as a test pattern for
another component.
Test Instruction 1

Test Instruction 2
same
test pattern

adder

to memory

subtracter

to memory

Figure 5-21: "Parallel" testing of processor components.

The benefit of the "parallel" testing of processor components is that the

loops that generate the component tests are combined together to form a
new, globalloop for the components that are now tested in "parallel". There
is no need for a separate generation of the next test pattern for each of the
components. Therefore, the self-test code size is significantly reduced and
also the total execution time for the new loop is much smaller than the
combined execution time ofthe separate loops. The number oftest responses
is not reduced as in the case of "chained" testing of components but may
rather be increased. This can happen when a smaller test set for a component
is expanded to a larger one that contains the small one (or obtains fault
coverage similar or same to the small one).

5.11

Software-based self-testing automation

Every testing and self-testing methodology requires a set of supporting

algorithms for its application and also requires a successful implementation
of EDA tools that realize these algorithms efficiently. Manual test
development, although useful and very efficient in sm all circuit sizes, can 't
be easily applied to large circuits.
Software-based self-testing of embedded processor cores requires
corresponding automatie tools for the support of the generation of selftesting routines to test the processor components separately and the
processor in total. Figure 5-22 shows a graph with parts of the softwarebased self-testing flow for a processor that can be automated.

154

Chapter 5 - Software-Based Proeessor Self-Testing

automatie generation
of eomponent
self-test routines

automatie generation
and optimization
of proeessor
self-test programs
Figure 5-22: Software-based self-testing automation.

A first part that can be automated is the component test development

part, in which a set of test patterns that can be applied by processor
instructions is derived for each of the components that the processor consists
of. In some cases, this part has been implemented using constraints
extraction for the components and feeding the extracted constraints to an
ATPG. If successful, this process leads to a set of ATPG-generated test
patterns for the target processor component which can be applied to it using
processor instructions. Several attempts to this direction in the open
literature show the importance ofthis path (for arecent, excellent analysis of
this problem, the reader may refer to [31]). Automatie (as opposed to
manual) extraction of instruction set imposed constraints for components
testing will lead to significant reduction of the test generation (and self-test
routines generation) time. The advantage of this approach is that, if
successful, can be universally applied to different processors and different
instruction set architectures.
A second part of software-based self-testing that can be automated is the
part of extraction of the information required to develop the self-test
routines. This information has been discussed earlier in this Chapter and
consists of:

the set ofprocessor components from its RTL description;

the set of operations that each component of the processor
executes (as instructed by processor instructions);
the set of different instructions of the processor that excite a
particular operation of a component;

Embedded Processor-Based Self-Test

155

the controllability and observability of processor registers and

processor components.

The third part of software-based self-testing than can be automated is the

development ofthe component (and processor) self-test routines themselves.
In particular, a tool for the automatie generation of self-test routines for
processor testing should be supplied with:

the instruction set architecture (ISA) information of the

processor;
the register transfer level (RTL) information ofthe processor;
the set oftest patterns that must be applied to each ofthe targeted
processor components;
the self-test coding style that must be applied to each component.

The expected output of such an automatie tool is a set of per-component

self-test routines, or a combined, optimized self-test routine for the entire
processor which:

apply the given test patterns to the processor components using

instructions;
guarantee propagation of fault effects to the processor's primary
outputs.

F or smaIl processor models manual extraction of constraints or manual

extraction of processor components information (dass, operations, etc) as
weIl as self-test routines development can be quite efficient. It requires the
availability of an expert assembly language programmer supplied with the
information ab out the processor instruction set and RT level architecture.
This can 't be efficiently performed for larger processor models and therefore
automation of the software-based self-testing process is necessary for this
technology to penetrate the complex SoC market.
We believe that automation of software-based self-testing for processor
and SoC architectures will be a research direction with increasing activity
within the next few years.

Chapter

6
Case Studies - Experimental Results

In this Chapter we discuss the application of software-based self-testing

to several embedded processor designs that are publicly available. As in
other cases of engineering research, one of the most difficult problems in
software-based self-testing is the availability of reasonably complex and
representative benchmark cases to demonstrate the practical value and
applicability ofthe methodology.
Today, the development of embedded processor designs is a profitable
business for several fabless companies (see Table 2-2, Table 2-3 and Table
2-4). Therefore, it is very difficult to obtain, for research purposes, a modem
processor model with complete functionality. On the other side, it is not
necessary, for the demonstration of a methodology, to have an exactly
equivalent model of a commercial processor core. It is sufficient to work on
a fully functional processor model that realizes the instruction set
architecture of a known processor.
The selected processor models described in this Chapter have a wide
range of complexities and different instruction set architectures and can be
used as benchmarks to demonstrate in a reasonable and persuasive manner
the benefits of the software-based self-testing methodology. Their common
characteristic is that they are available in a synthesizable, soft core model
(VHDL or Verilog). A vailability of a synthesizable processor model gives
the ability to apply software-based self-testing in different synthesized
versions of the processor (optimized for area or delay) and also gives the
Embedded Processor-Based Self-Test
D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic Publishers, 2004

158

Chapter 6 - Case Studies - Experimental Results

ability to calculate the relative sizes of the processor's components. Of

course, availability of a synthesizable processor model gives also the ability
to perform fault simulation on the synthesized design and obtain fault
coverage results that demonstrate the effectiveness ofthe methodology.
The set of benchmark processor models used in this Chapter can be
valuable as a common basis for comparisons in future research in the area of
software-based self-testing for processors and processor-based SoC
architectures. In any case, the more complex processor models are available,
the better for the demonstration of any processor self-testing methodology.
For each of the benchmark processors, after a first brief description, we
discuss its implementationlsynthesis results providing statistics of the
processor's component sizes. Then, we present, for each processor, the fault
simulation results from the application of low-cost, software-based selftesting to it. These experimental results include embedded self-test program
sizes, self-test execution time, as weIl as, fault coverage with respect to
single stuck-at faults.
We have implemented aIl benchmark processors using a classic flow of
synthesis and simulation from their original VHDL or Verilog source code.
A 0.35 um ASIC library38 has been used for synthesis and gate count is
presented in gate equivalents where one gate equivalent is the two-input
NAND gate.
We have applied software-based self-testing to aIl the selected
benchmark processors. For each one of the processors, we discuss self-test
routines development and we give the self-test pro gram size in words, the
response data size in words and execution time in CPD cycles, as weIl as, the
fault coverage obtained for the entire processor and its individual
components (those that have been targeted for test development or not).

6.1

Parwan processor core

Parwan [116] is a very simple 8-bit accumulator-based processor that has

been developed for educational purposes and is briefly mentioned in this
section only because it has been used as a demonstration vehicle in the
software-based self-testing methodologies presented in [27], [28], [29], [94],
[95], [96], [104], [105]. Parwan is an 8-bit CPD with a 12-bit address bus
able to access a 4K memory. The Parwan instruction set includes common
instructions like load from and store to memory, arithmetic and logical
operations, jump and branch instructions and is therefore able to implement
several real algorithms. Parwan also supports direct and indirect addressing
38 A second 0.50 um ASIe library has been also used for comparisons between two different
libraries. This comparison has been performed for the PlasmaIMIPS processor benchmark.

Embedded Pracessor-Based Self-Test

159

modes for memory operands. Considering that each addressing mode leads
to different instructions, Parwan's instruction set consists of 24 different
instructions.
The Parwan processor model is available in VHDL synthesizable format
and its architecture includes the components shown in Table 6-1. The
classification of each component to the classes described in the previous
Chapter is also shown in Table 6-1.
Component Name

Component Class

Arithmetic Logic Unit (ALU)

Shifter Unit (SHU)
Accumulator (ACC)
Pragram Counter (PC)
Status Register (SR)
Memory Address Register (MAR)
Instruction Register (IR)
Control Unit (CRTL)

Functional computational
Functional computational
Functional storage
Contra I
Contral
Control
Contra I
Contra I

Table 6-1: Parwan processor components.

Out of the eight processor components only the Arithmetic Logic Unit
and the Shifter are combinational circuits and also the only functional
computational components. It should also be noted that the only processor
data register is the accumulator. This single data processor register is fully
accessible in terms of controllability and observability by processor
instructions.
We have synthesized Parwan from its VHDL source description and the
resulting circuit consists of 1,300 gates including 53 flip-flops.

6.1.1

Software-based self-testing of Parwan

The Parwan processor components that have been targeted for self-test
program development are the ALU, the Shifter and the Status Register. SeIftest routines development has been done in three phases: Phase A for the
ALU, Phase B for the Shifter and Phase C for the Status Register. We have
selected this sequence because the third functional unit of the processor, the
Accumulator, is already sufficiently tested after Phase A.
Table 6-2 shows the statistics of the self-test code for the three
consecutive phases A, Band C.

160

Chapter 6 - Case Studies - Experimental Results

Phase A
target
ALU
311
631
72
9,154

Number of Instructions
Self-Test Program Size (bytes)
Response Data Size (bytes)
Execution Time (cycles)

Phase B
target
Shifter
440
881
122
16,545

Phase C
target
Status
Register
463
9;>3
124
16,667

Table 6-2: Self-test pro gram statistics for Parwan.

Table 6-3 shows the fault coverage for single stuck-at faults for each of
the Parwan processor components after each ofthe three phases.
Component Name

Arithmetic Logic Unit {ALU}

Shifter Unit {SHU}
Accumulator {ACe}
Program Counter (PC}
Status Register {SR)
Memor}:: Address Register (MAR)
Instruction Register (IR)
Control Unit {CRTL}
Total CPU

Fault
Coverage
after
Phase A
98.31%
75.56%
98.67%
87.05%
92.13%
97.22%
98.26%
82.93%

Fault
Coverage
after
Phase B
98.48%
93.82%
98.67%
88.10%
92.13%
97.22%
98.26%
85.52%

Fault
Coverage
after
Phase C
98.48%
93.82%
98.67%
88.10%
92.13%
97.22%
98.26%
87.68%

88.70%

91.10%

91.34%

Table 6-3: Fault simulation results for Parwan processor.

Experimental results on Parwan for single stuck-at faults were given in

[28]. Single stuck-at fault coverage of 91.42% is reported in [28] which was
obtained by a self-test program of 1,129 bytes that needs 137,649 clock
cycles for execution (compare with the 923 bytes and the only 16,667 clock
cycles reported in Table 6-3).

6.2

Plasma/MIPS processor core

Plasma [128] is a publicly available RISC processor model that

implements the MIPS I instruction set architecture. Plasma supports
interrupts and a11 MIPS I user mode instructions except unaligned load
and store operations (which are patented) and exceptions.
The synthesizable CPU core is implemented in VHDL with a 3-stage
pipeline structure. The Plasma processor architecture consists of the
components shown in Table 6-4. We can also see in the list below the
characterization of each of the components into the classes described in the
previous Chapter.

161

Embedded Processor-Based Self-Test

Component Name

Component Class

Register File
Multiplier
Divider
Arithmetic-Logic Unit
Shifter
Memory Control
Program Counter Logic
Control Logic
Bus Multiplexer
Pipeline

Functional
Functional
Functional
Functional
Functional
Control
Control
Control
Functional
Hidden

storage
computational
computational
computational
computational

interconnect

Table 6-4: Plasma processor components.

The internal architecture of each of the processor components can be of

many different types. In particular, the components that perform arithmetic
operations such as the ALUs, adders, subtracters, multipliers, dividers,
incrementers, etc can be designed in many different ways. For the case ofthe
PlasmalMIPS processor benchmark, we have implemented two different
architectures of the multiplier component, aserial one and a parallel one,
each of which leads to different gate count for the processor. As we see in
the experimental results, the low-cost, software-based self-testing
methodology is able to obtain very high fault coverage in both cases.
For comparison purposes, we have synthesized the PlasmaiMIPS
processor core into three different designs which we call Design I, 11 and III,
respectively. Design I contains aserial multiplier, Design 11 contains a
parallel multiplier and synthesis has been performed for area optimization,
while Design III also contains a parallel multiplier but it is a delay optimized
design. Table 6-5, Table 6-6 and Table 6-7 show the synthesis results for
these three implementations of Plasma. Design I operates in a 66.0 MHz
frequency, Design 11 in a 57.0 MHz frequency and Design III in a 74.0 MHz
frequency.
The total gate count of the synthesized processor is slightly larger than
the sum of gate counts of the individual components due to the existence of
glue logic among the processor components, at the top level of hierarchy,
which can't be identified as aseparate processor component.

162

Chapter 6 - Case Studies - Experimental Results

Component Name
Register File
Multiplier/Divider 39
Arithmetic-Logic Unit
Shifter
Memory Control
Pragram Counter Logic
Control Logic
Bus Multiplexer
Pipeline

Total CPU

Gate Count
9,906
3,044
491
682
1,112
444
223
453
885
17,459

Table 6-5: Plasma processor synthesis for Design I.

Component Name

Gate Count

Register File
Multiplier/Divider 40
Arithmetic-Logic Unit
Shifter
Memory Contral
Program Counter Logic
Control Logic
Bus Multiplexer
Pipeline

9,905
11,601
491
682
1,119
444
230
453
885

Total CPU

26,080

Table 6-6: Plasma processor synthesis for Design 11.

39

40

In Design T, multiplier/divider are together implemented in a single component and in

serial architecture.
In Designs IT and III, multiplier/divider are again a single component hut the multiplier has
a parallel implementation while the divider keeps its serial structure. This is a very natural
choice hecause the division operation is a very rare one compared to multiplication, and
therefore only the multiplication operation deserves a parallel, more efficient
implementation.

163

Embedded Processor-Based Seit-Test

Component Name

Gate Count

Register File
Multiplier/Divider
Arithmetic-Logic Unit
Shifter
Memory Contral
Program Counter Logic
Control Logic
Bus Multiplexer
Pipeline

11,905

Total CPU

30,896

13,358
900

834
1,163
493
361
623
961

Table 6-7: Plasma processor synthesis for Design IIl.

We note that in the three implementations of the Plasma processor the

total gate counts of its functional components represent an 83.48% (Design
I), 88.69% (Design 11) and 89.39% (Design IIl) of the entire processor's
area, respectively. Therefore, in this processor benchmark, our claim that the
functional components of the processor are the largest and therefore those
that must be initially targeted by low-cost, software-based self-testing, is
easily justified. We will see that in the other processor benchmarks the same
argument is still valid to a smaller or larger extend.

6.2.1

Software-based self-testing of PlasmalMIPS

We have applied low-cost, software-based self-testing to PlasmaIMIPS

processor benchmark. For Design I we have developed a self test program
targeting the Register File, the Arithmetic Logic Unit, the
MultiplierlDivider, the Shifter and the Memory Control components. The
self-test pro gram consists of965 words (32-bits) and is executed in a total of
3,552 clock cycles. The fault coverage obtained for each of the processor
components is shown in Table 6-8.

164

Chapter 6 - Case Studies - Experimental Results

Component Name

Fault Coverage

tor Design I

Register File
Multiplier/Divider
Arithmetic-Logic Unit
Shifter
Memory Contral
Pragram Counter Logic
Control Logic
Bus Multiplexer
Pipeline

97.7%
87.5%
96.6%
98.4%
88.3%
53.1%
78.9%
65.7%
91.9%

Total CPU

92.2%

Table 6-8: Fault simulation results for the Plasma processor Design 1.

The next two implementations of the Plasma processor (Design 11 and

Design III) contain a parallel multiplier instead of aserial one. The inclusion
of a parallel multiplier in the processor increases its size (see Table 6-6 and
Table 6-7) by more than 8,000 gates (32 bit array multiplier) but it speeds up
the multiplication operation which is now executed in a single cycle instead
of 32 of the serial implementation. Design 11 is optimized for area and
Design III is optimized for speed, so the area of Design III is larger by about
4,000 gates for the entire processor.
For these two implementations of Plasma, we have developed the same
self-test routines for the following components: Register File, Parallel
Multiplier, Serial Divider, Arithmetic-Logic Unit, Shifter, Memory Control,
and Control Logic. The self-test routine sizes and execution times for each
component and the overall processor are shown in Table 6-9.
Targeted Component

Register File
Parallel Multiplier
Se rial Divider
Arithmetic-Logic Unit
Shifter
Memory Control
Contral Logic
Total CPU

Seit-Test
Routine Size
(words)
319
28
41
79
210
76
100

Execution
Time
(cycles)
582
3,122
1,154
275
340
160
164

853 41

5,797

Table 6-9: Self-test routine statistics for Designs II and III of Plasma.

~I In Design I with the serial multiplier the total pro gram was 965 words and total cycles

were 3,552 because the serial multiplier needs a larger test program that is executed for
more clock cycles than the one for the parallel multiplier.

Embedded Processor-Based Self-Test

165

We notice in the self-test routine statistics of Table 6-9 that there exist
components with very small self-test routines, such as the multiplier and the
divider, which routines take a very large percentage of the overall test
execution time. This is because these routines consist of small, compact
loops that are executed for a large number oftimes and apply a large number
of test patterns to the component they test. On the other side, there. are
components like the Shifter with a relatively large self-test code which is
executed very fast because it is a code that does not contain loops but rather
applies a sm all set of test patterns using immediate addressing mode
instructions for all shifter operations. The self-test routine for the ArithmeticLogic Unit consists of segments for every ALU operation that combine small
compact loops and immediate addressing mode instructions. The self-test
routine of the Memory Control consists of load instructions with specific
data previously stored in data memory, and store instructions that generates
the final responses in data memory as weIl. Finally, the self-test routine of
the Control Logic component is based on an exhaustive application of all the
processor's instruction opcodes not already applied in the routines of the
previous components. This functional testing approach at the end of the
methodology is very natural to be applied to the control unit of the
processor.
At this point we remark that a similar self-test routine development
strategy has been adopted for the remaining benchmark processors, for their
components that are similar with the components ofPlasmaIMIPS.
The fault simulation results for the two designs of PlasmaIMIPS, Design
11 and III are shown in Table 6-10.
Component Name

Register File
Multi~lier/Divider

Arithmetic-Logic Unit
Shifter
Memo~ Control
Program Counter Logic
Control Logic
Bus Multiplexer
Pipeline
Total CPU

Fault
Coverage for
Desian 11
97.8%
96.3%
96.8%
98.4%
87.9%
54.9%
89.3%
71.8%
98.4%

Fault
Coverage for
Desian 111
97.8%
95.2%
95.8%
95.3%
90.3%
55.9%
85.3%
71.3%
96.0%

95.3%

94.5%

Table 6-10: Fault simulation results for Designs II and III ofPlasma.

We see that in the two implementations with the parallel multiplier the
overall fault coverage is higher than in the case of Design I which contains a
serial implementation of the multiplier. This fault coverage increase is

166

Chapter 6 - Case Studies - Experimental Results

simply due to the fact that a large component like the parallel multiplier with
very high testability has been inserted in the processor. The same design
change could be done for the division operation and another large
component, the parallel divider, could be added. This would lead to further
increase of the processor's fault coverage. We didn't implement this design
change because the division operation is not as common as the multiplication
and therefore the cost of adding a large parallel component that is
infrequently used is not justified in low-cost applications. Of course in a
special implementation of the processor in an application with many
divisions, the inc1usion of a parallel divider will lead to an increased
performance of the processor, as well as, an increased fault coverage
obtained by software-based self-testing. In such a case, where a parallel
divider is also implemented, fault coverage can be as high as 98% for single
stuck-at faults.
The overall processor fault coverage is, in both designs, very high (more
than 94.5%) while the particular fault coverage levels of each component
may slightly differ because of the different synthesis optimization
parameters that lead to a different gate-level implementation of the
component.
We also note that the pipeline logic is tested as a side-effect oftesting the
remaining processor components achieving very high fault coverage. This
fact is due to the simple pipeline structure that Plasma realizes.

Application to different synthesis libraries

Wehave performed another experiment with the Plasma processor
benchmark in order to evaluate the effect of synthesizing the processor with
a different implementation library. A 0.50 um library has been used and
synthesis has been done with area optimization. We call this synthesized
design as Design IV. Design IV contains a parallel multiplier, has a
frequency of 42.0 MHz and its synthesis results in gate counts are given in
Table 6-11.
6.2.1.1

Embedded Pracessor-Based Self-Test

167

Component Name

Gate Count

Register File
Multiplier/Divider
Arithmetic-Logic Unit
Shifter
Memory Contral
Program Counter Logic
Contral Logic
Bus Multiplexer
Pipeline

11,558
11,654
558
636
1,120
449
244
431
876

Total CPU

27,824

Table 6-11: Plasma processor synthesis for Design IV.

As Design IV is an area optimized one, it is directly comparable with

Design 11. We applied to Design IV exactly the same self-test routines (Table
6-9) that were applied to Design 11 (and also to Design III) and the
comparison of fault coverages for all components in Designs 11 and IV are
summarized in Table 6-12.
Component Name

" Register File

Multiplier/Divider
Arithmetic-Logic Unit
Shifter
Memory Contra I
Program Counter Logic
Control Logic
Bus Multiplexer
Pipeline

Total CPU

Fault
Coverage tor
Desisn 11
97.8%
96.3%
96.8%
98.4%
87.9%
54.9%
89.3%
71.8%
98.4%

Fault
Coverage tor
Desisn IV
97.8%
96.1%
97.5%
99.9%
88.5%
54.9%
88.3%
72.0%
96.9%

95.3%

95.3%

Table 6-12: Comparisons between Designs II and IV ofPlasma.

We note that the total fault coverage for the processor is exactly the same
95.3% of all single stuck-at faults. For each of the components of the
processor, fault coverage results may slightly differ to a maximum of 1.5%
ofthe component's faults (shifter, pipeline). These small differences are due
to the different cells that each implementation library contains. The same
self-test program reaches slightly different structural fault coverage in each
ofthe processor components.
Some useful conclusions can be drawn from this first application of
software-based self-testing to the reasonable size processor model:

Self-test code size is very small, less than 1,000 words.

168

Chapter 6 - Case Studies - Experimental Results

6.3

Self-test routines execution time is also very small, less than

6,000 clock cycles.
The fault coverage for the largest functional storage component,
the Register File, is always very high.
The fault coverage for all thc functional computational
components that have been targeted is very high.
The pipeline logic is sufficiently tested although it has not been .
targeted for self-test code development. This fact is due to the
simple pipeline structure ofPlasma.

Meister/MIPS reconfigurable processor core

Our next selected CPU benchmark that belongs to the class of

reconfigurable CPUs which have gained importance today because they
offer the ability to tailor an instruction set and corresponding functional units
of a processor for a particular application. Based on the Meister Application
Specific Instruction set Processor (ASIP) environment [114], we have
designed a MIPS R3000 (MIPS I ISA) compatible processor with a 5stage pipeline structure. A subset of 52 instructions of the MIPS R3000
instruction set [85] was implemented while co-processor and interrupt
instructions were not implemented in this experiment. It must be mentioned
that in the current educational release version of the ASIP design
environment of [114] data hazard detection and forwarding are not
completely implemented. Therefore, although the pipeline structure exists
(pipeline registers, multiplexers, etc) its complete functionality is not used.
An RTL VHDL model was generated by the tool for this MIPS compatible
processor.
The MeisterlMIPS processor consists of the identifiable components
shown in Table 6-13. We can also see in the list below the characterization
of each ofthe components into the classes described in the previous Chapter.
Component Name
Register File
Parallel MultiplierlSerial Divider

ALU

Shifter
Hi-Lo Registers
Controller
Data Memory Controller
Program Counter Logic
Instruction register
Pipeline registers
Table 6-13: Meister/MIPS processor components.

Component Class
Functional storage
Functional computational
Functional computational
Functional computational
Functional storage
Control
Control
Control
Hidden
Hidden

169

Embedded Processor-Based Seit-Test

The usefulness of experimenting with MeisterlMIPS is that we can

compare the results of software-based self-testing in this architecture which
is another implementation of about the same instruction set of the previous
model, PlasmaIMIPS. MeisterlMIPS configurable processor is another
implementation of the classical, popular instruction set architecture of MIPS
that the PlasmaIMIPS benchmark also implements. We performed our
experiments with the MeisterlMIPS too, to verifY the effectiveness of the
approach to another implementation of the same instruction set architecture
where the components of the processor are intemally implemented in a
different way than PlasmaIMIPS. Table 6-14 shows the synthesis results of
the MeisterlMIPS processor. The clock frequency of the MeisterlMIPS
implementation is 44.0 MHz.
Component Name
Register File
Parallel Multiplier/Serial Divider
ALU
Shifter
Hi-Lo Registers
Controller
Data Memory Controller
Program Counter Logic
Instruction register
Pipeline registers
Total CPU

Gates Count
11,414
12,564
658
633
536
2,352
1,086
644
275
5,693
37,402

Table 6-14: Meister/MIPS processor synthesis.

As in the case of Designs II and III of Plasma, the component for

multiplication and division is a single component that contains a parallel
multiplier and aserial divider. A slight difference between PlasmaIMIPS and
MeisterlMIPS is that the Hi and Lo registers that hold the results of
multiplication and division in the MIPS architecture are implemented as a
separate component in MeisterIMIPS, while in Plasma/MIPS they are part of
the multiplier/divider component.
We can see that in Meister the functional components of all types occupy
a total of 68.99% of the total processor's area. This is still a very large
percentage which is smaller than in the case of PlasmalMIPS because
MeisterlMIPS implements a more complex pipeline structure where pipeline
registers and other logic occupy a significantly larger area of the processor
than in the case ofPlasmalMIPS.

170

6.3.1

Chapter 6 - Case Studies - Experimental Results

Software-based self-testing of MeisterlMIPS

We have developed self-test routines for the list of the components

shown in Table 6-15 which together compose a self-test program of 1,728
words executed in 10,061 clock cycles for the MeisterlMIPS processor. We
can also see the self-test routine size and execution time for each component.
Targeted Component

Register File
Parallel Multiplier
Serial Divider
Arithmetic-Logic Unit
Shifter
Hi-Lo Registers
Control Logic
Total CPU

Self-Test
Routine Size
(words)
720
68
65
192
378
30
275

Execution
Time
(cycles)
859
5,855
1,396
1,188
437
35
291

1,728

10,061

Table 6-15: Self-test routines statistics for Meister/MIPS processor.

The fault coverage obtained by the above self-test routines for

MeisterlMIPS after evaluation of 600 responses in data memory is given in
Table 6-16.
Component Name
Register File
Parallel Multiplier/Serial Divider
ALU
Shifter
Hi-Lo Registers
Controller
Data Memory Controller
Program Counter Logic
Instruction register
Pipeline registers

Fault Coverage
99.8%
95.2%
98.4%
99.8%
100.0%
79.2%
58.2%
58.5%
97.4%
91.0%

Total CPU

92.6%

Table 6-16: Fault simulation results for Meister/MIPS processor.

We note that the MeisterlMIPS processor benchmark implements almost

the same MIPS I instruction set architecture as the Plasma model, but with
a more complex control and pipeline. This is due to the fact that the VHDL
RTL design generated by the ASIP design environment does not support
data hazard detection and forwarding. In the MeisterlMIPS processor,
careful assembly instruction scheduling is required along with insertion of
nop (no operation) instructions wherever it is necessary. The result is an

Embedded Processor-Based Self-Test

171

increased test program size (almost double) and test execution time (almost
double) and sm aller fault coverage (92.6% compared to more than 94.5% of
Plasma). Otherwise (if the pipeline logic was completely implemented) test
program statistics would have been very similar to the case of PlasmaJMIPS
and fault coverage would have been much higher.

6.4

Jam processor core

Jam CPU [78] has been implemented in Chalmers University of

Technology. It follows a 32-bit RISC CPU architecture called Concert'02.
Jam CPU has a five-stage pipeline structure which is very similar to the
pipeline structure ofthe DLX architecture [63]. Jam CPU is implemented in
VHDL synthesizable format and its five-stage pipeline structure includes the
following stages: Instruction Fetch, Instruction Decode, Execute, Memory
and Write Back. Jam has multi-cycle operations, pipeline hazard checking
and pipeline forwarding.
Jam processor consists of the identifiable components shown in Table
6-17. We can also see in the list below the characterization of each of the
components into the classes described in the previous Chapter.
Component Name
Register File (REGS)
Integer Unit (lU) - includes the ALU
Immediate Extender (IMM EXT)
Memory Access Unit (MAU 1) for Instruction Memory
Memory Access Unit (MAU 2) tor Data Memory
Control Logic (CONTROL)
Pipeline Registers
Table 6-17: Jam processor components.

Component Class
Functional storage
Functional computational
Functional computational
Control
Control
Control
Hidden

The usefulness of experimenting with Jam CPU is that is a step forward

to the study of software-based self-testing in RISC architectures. It is a
different architecture than the one implemented by PlasmaJMIPS and
Meister/MIPS and it contains a fully implemented, five-stage pipeline
structure with hazard detection and forwarding.
The pipeline structure of the Jam processor core is implemented at the
top level of the VHDL design. It is therefore difficult to identify components
that implement the pipeline mechanism other than the pipeline registers. The
rest of the pipeline logic (control logic for hazard detection and multiplexers
for forwarding) is not accounted as separate components. The synthesis
results ofthe Jam processor from its original VHDL source code are given in
Table 6-18. The dock frequency of the Jam processor implementation is
41.8 MHz.

Chapter 6 - Case Studies - Experimental Results

172

Component Name
Register File (REGS)
Integer Unit (lU)
Immediate Extender (IMM EXT)
Memory Access Unit (MAU 1) tor Instruction Memory
Memory Access Unit (MAU 2) tor Data Memory
Control Logic (CONTROL)
Pipeline Registers
Total CPU

Gate Count
22,917
5,698
269
576
576
388
3,771
43,208

Table 6-18: Jam processor synthesis.

The register file consists of 32 registers of 32 bits and is the largest

component of the processor that dominates the processor area (more than
50%). The Integer Unit in the Jam processor core implements the following
integer operations: multiplication, addition, subtraction, bitwise OR, bitwise
AND, bitwise XOR, shift. All operations apart from multiplication take one
cycle, while multiplication takes 33 cycles (multiplication in the Jam
processor integer unit is implemented in aserial fashion, like in the case of
the Plasma Design I). The Integer Unit contains control logic, a shift register
used for multiplication and an ALU that handles addition, subtraction and
the logical operations. The Immediate Extender participates in the immediate
addressing mode instructions. The Memory Access Units 1 and 2 are used
for the communication of the processor with the instruction memory and
data memory, respectively.
The remaining gates in the Jam processor not given in Table 6-18 are part
of the top-level entity and include not only glue logic, but pipeline
multiplexers and control used for hazard detection and forwarding. They
represent a total of about 9,000 gates.
We can see that the functional components in the case of the Jam
processor occupy a total of66.85% ofthe processor's area. This fact is again
(as in MeisterlMIPS processor) a result of the complex pipeline structure of
the processor and all the logic gates needed to implemented it (only the
pipeline registers, which are the identifiable part of the processor's pipeline
architecture occupy more than 8% ofthe processor).
6.4.1

Software-based self-testing of Jam

The Jam processor benchmark gives us the ability to evaluate softwarebased self-testing in a more complex RISC processor model with a fully
implemented pipeline architecture which realizes hazard detection and
forwarding.
We have developed self-test routines for the list of components shown in
Table 6-19. In this Table we can also see the self-test routine size and

173

Embedded Pracessor-Based Self-Test

execution time for each component. These routines together compose a seIftest program of 897 words executed in 4,787 clock cycles for the Jam
processor.
Component Name

Register File (REGS)

Integer Unit (lU)
Immediate Extender (IMM EXT)
Memory Access Unit (MAU 2)
Control Logic (CONTROL)

Self-Test
Routine
Size
(words)
478
147
32
120
120

Total CPU

897

Execution
Time
(cycles)
550
3,920
38
135
144
4,787

Table 6-19: Self-test routine statistics for Jam processor.

We have not developed any special self-test routines for the pipeline
logic since pipeline forwarding is already activated many times during the
execution of the se1f-test pro gram and the pipeline logic, multiplexers and
registers are sufficiently tested as a side-effect of testing the remaining
components. The fault simulation results for the Jam processor after
evaluation of 454 responses in data memory are shown in Table 6-20. The
achieved overall processor fault coverage is very high, 94% with respect to
single stuck-at faults.
Component Name
Register File (REGS)
Integer Unit (lU)
Immediate Extender (IMM EXT)
Memory Access Unit (MAU 1) for Instruction Memory
Memory Access Unit (MAU 2) for Data Memory
Contra I Logic (CONTROL)
Pipeline Registers

Total CPU
Table 6-20: Fault simulation results for Jam processor.

6.5

Fault
Coverage
98.1%
98.9%
98.5%
69.4%
81.7%
81. 2%
89.7%
94.0%

oc8051 microcontroller core

The 8051 microcontroller is a member of MCS-51 family, originally

designed in the 1980's by Intel. The 8051 has gained great popularity since
its introduction and, it is estimated that it is used in a large percentage of all
embedded system products today.

174

Chapter 6 - Case Studies - Experimental Results

We have selected as a benchmark the oc8051 model of the 8-bit 8051

microcontroller [121]. This implementation of 8051 has a two stage pipeline
structure (major difference with other 8051 implementations). At the first
pipeline stage instruction fetching and decoding takes place while during
second pipeline stage the instructions are executed and the results are stored
in memory. The oc8051 model is implemented in Verilog HDL.
The oc8051 processor consists of the identifiable components shown in
Table 6-21. We can also see in the list below the characterization of each of
the components into the classes described in the previous Chapter.
Component Name
Arithmetic Logic Unit (ALU)
ALU Source Select (ASS)
Decoder (DEC)
Special Function Registers (SFR)
Indirect Address (INDI ADDR)
Memory Interface (MEM)
Pipeline Registers

Component Class
Functional computational
Functional interconnect
Control
Functional storage
Functional storage
Control
Hidden

Table 6-21: oc8051 processor components.

The usefulness of experimenting with the oc8051 model is that it is a

very popular architecture used extensively as an embedded processor core in
SoCs. It follows a classical Intel architecture and is a substantially different
design from the RISC processors described in the previous sections. The
synthesis results of oc8051 from its original Verilog source code are
presented in following Table 6-22. The clock frequency of the oc8051
processor implementation is 83.1 MHz.
Component Name
Arithmetic Logic Unit (ALU)
ALU Source Select (ASS)
Decoder (DEC)
Special Function Registers (SFR)
Indirect Address (INDI ADDR)
Memory Interface (MEM)
Total CPU

Gate Count
1,147
269
970
4,507
635
2,703
10,305

Table 6-22: oc8051 processor synthesis.

The Arithmetic Logic Unit in the oc8051 processor core implements the
following integer operations: addition, subtraction, bitwise OR, bitwise
AND, bitwise XOR, shift and multiplication.
The ALU Source Select component selects the ALU input sources. The
Special Function Registers component contains 18 registers of one or two
bytes each (accumulator, B register, Pro gram Status Word, Stack Pointer,
etc). The Indirect Address component implements the indirect addressing

175

Embedded Processor-Based Seit-Test

mode instructions. The Memory Interface component is used for the

communication of the processor with the memory, while the Decoder
component is the controllogic ofthe processor. The remaining logic gates of
oc8051 not included in Table 6-22 are again part of the top level of the
hierarchy, constitute the logic for the implementation ofthe 2-stage pipeline
structure of oc8051 and other surrounding logic to the above components.
We see that in the case ofthe oc8051 processor a total of63.64% ofits
area is occupied by the functional components. Targeting these components
for self-test routines development and also considering the simple pipeline
structure of the processor, we have good chances to obtain high fault
coverage with small self-test code.
6.5.1

Software-based self-testing of oc8051

In the case of the oc8051 microcontroller benchmark we have developed

self-test routines for the list of components shown in Table 6-23. In Table
6-23 we can also see the self-test routine size and execution time for each
component and the overall processor. These routines together compose a
self-test program of 3,760 bytes executed in 5,411 clock cycles for the
oc8051 microcontroller.
Component Name

Arithmetic Logic Unit {ALU}

ALU Source Select (ASS}
Special Function Registers (SFR)
Indirect Address (INDI ADDR)
Decoder (DEC}
Memor~ Interface {MEM}
Total CPU

Seit-Test
Code Size
~bxtes!
1,452
512
548
560
360
328

~cxcles!
2,964
541
598
614
324
370

3,760

5,411

Seit-Test
Code Time

Table 6-23: Self-test routine statistics for oc8051 processor.

The fault simulation results after evaluation of 416 test responses m

memory are shown in Table 6-24.

Chapter 6 - Case Studies - Experimental Results

176

Component Name
Arithmetic Logic Unit (ALU)
ALU Source Select (ASS)
Decoder (DEC)
Special Function Registers (SFR)
Indirect Address (INDI ADDR)
Memory Interface (MEM)

Fault Coverage
98.4%
97.1%
81.5%
96.2%
90.9%
89.9%

Total CPU

93.1%

Table 6-24: Fault simulation results for oc805l processor.

We see that for oc8051 as weIl high fault coverage of 93.1 % with respect
to single stuck-at faults is obtained with a relatively small and fast pro gram.

6.6

RISC-MCU microcontroller core

RISC-MCU [136] is a RISC microcontrol unit designed after the AVR

8-bit RISC microcontroller from Atmel. It is a synthesizable processor core
model implemented in VHDL. RISC-MCU has the same instruction set with
the Atmel A T90S 1200 processor (89 instructions).
RISC-MCU microcontroller consists of the identifiable components
shown in Table 6-25. We can also see in the list below the characterization
of each ofthe components into the dasses described in the previous Chapter.
Component Name
Shift Register (SR)
General Purpose Registers (GPR)
Arithmetic Logic Unit (ALU)
Control Unit (CTRL)
Program Counter (PC)
10 Port (IOP)
Timer
Pipeline

Component Class
Functional computational
Functional storage
Functional computational
Control
Control
Control
Control
Hidden

Table 6-25: RISC-MCU processor components.

The usefulness of experimenting with the A VR model provided by

RISC-MCU is to compare it with the previous RISC processor models
described in the previous sections. RISC-MCU is a smaller word (8-bit)
RISC architecture as opposed to the other larger word (32-bit) architectures
that we studied so far. The results obtained from its synthesis from the
original VHDL source files are given in the following Table 6-26. The dock
frequency ofthe RISC-MCU processor implementation is 249.9 MHz (much
faster than the 32-bit RISC processors).

Embedded Processor-Based Self-Test

Component Name
Shift Register (SR)
General Purpose Registers (GPR)
Arithmetic Logic Unit (ALU)
Control Unit (CTRL)
Program Counter (PC)
10 Port (IOP)
Timer
Pipeline

177
Gate Count

TotalCPU

127
1, 513

406
777
645
185

294
650
4,693

Table 6-26: RISC-MCU processor synthesis.

This smaller RISC processor benchmark has almost the same functional
components of larger 32-bit RISC processors but for a smaller word length.
We can see that the functional components of all types in RISC-MCU
occupy a total of 43.60% of the entire processor area. This percentage is
much smaller than in the case of larger 32-bit RISC processors described in
the previous sections. This is due to the fact that although the control logic
components remain of the same order of magnitude for a small word length
(8 bits) or a large word length (32 bits), the datapath components that
perform the operations (computational functional components) and store the
data (storage functional components) are significantly larger in the 32-bits
word length. For example, the General Purpose Registers (GPR) component
of RISC-MCU occupies only l,513 gates, while the corresponding register
files of Plasma, Meister and Jam occupy more than 9,000, 11,000 or even
22,000 gates (Jam).

6.6.1

Software-based self-testing ofRISC-MCU

For the RISC-MCU benchmark we have developed self-test routines for

the list of components shown in Table 6-27. In Table 6-27 we can also see
the self-test routine size and execution time for each component and the
overall processor. These routines together compose a self-test program of
1,258 words executed in 2,446 clock cycles for the RISC-MCU
microcontroller.

Chapter 6 - Case Studies - Experimental Results

178
Component Name

Shift Register (SR)

General Purpose Registers (GPR)
Arithmetic Logic Unit (ALU)
Control Unit (CTRL)
10 Port (IOP)
Total CPU

Seit-test
Code Size
(words)
360
510
128
200
60
1,258

Seit-test
Code Time
(cycles)
410
674
1,032
240
90
2,446

Table 6-27: Self-test routine statistics for RISC-MCU processor.

The fault simulation results for the RISC-MCU microcontroller are

shown in Table 6-28.
Component Name
Shift Register (SR)
General Purpose Registers (GPR)
Arithmetic Logic Unit (ALU)
Control Unit (CTRL)
Program Counter (PC)
10 Port (IOP)
Timer
Pipeline

Fault Coverage
99.2%
99.0%
98.4%
91.1%
72.5%
91.4%
74.0%
92.4%

Total CPU

91.2%

Table 6-28: Fault simulation results for RlSC-MCU processor.

We see that high fault coverage of 91.2% with respect to single stuck-at
faults is obtained with a relatively sm all and fast self-test pro gram. The fault
coverage is less than in the case of larger 32-bit RISC processors due to the
smaller scale ofthe functional components.

6.7

oc54x DSP Core

The oc54x DSP core is a synthesizable 16/32, dual-16 bit DSP core
which is available in synthesizable Verilog format [120]. Oc54x is an
implementation of a popular family of DSPs designed by Texas Instruments
and is software compliant with the original TI C54x DSP.
The oc54x DSP processor consists of the identifiable components shown
in Table 6-29. We can also see in the list below the characterization of each
of the components into the classes described in the previous Chapter.

179

Embedded Processor-Based Seit-Test

Component Name
Accumulator (ACC)
Arithmetic Logic Unit (ALU)
Barrel Shifter (BSFT)
Compare/Select and Store Unit (CSSU)
Exponent Decoder (EXP)
Multiply/Accumulate Unit (MAC)
Temporary Register (TREG)
Control

Component Class
Functional computational
Functional computational
Functional computational
Functional interconnect
Functional computational
Functional computational and storage
Functional storage
Control

Table 6-29: oc54x processor components.

The usefulness of experimenting with the oc54x DSP model is that it is

an implementation of a popular DSP architecture. DSPs are very widely used
in SoC architectures either alone or in conjunction with other general
purpose processor cores. The oc54x DSP core, like all DSPs are excellent
candidates for the application of low-cost, software-based self-testing,
because they contain large word length functional components (mostly
computational and storage ones). The synthesis results for oc54x from its
original Verilog source code are given in Table 6-30. The clock frequency of
the oc54x DSP implementation is 48.8 MHz.
Component Name
Accumulator (ACC)
Arithmetic Logic Unit (ALU)
Barrel Shifter (BSFT)
Compare/Select and Store Unit (CSSU)
Exponent Decoder (EXP)
Multiply/Accumulate Unit (MAC)
Temporary Register (TREG)
Control
Total CPU

Gates Count
687
3,145
1,682
444
799
3,819
130
905
11,611

Table 6-30: oc54x DSP synthesis.

We can see that the synthesis statistics ofthe oc54x DSP make it a very
suitable architecture for the applicatin of software-based self-testing
because of the existence of many and large functional components. A total of
92.21 % of the entire DSP area is occupied by the functional units of all
subtypes. All of them are well accessible units that can be easily targeted
with software-based self-testing routines.

6.7.1

Software-based self-testing of oc54x

Our DSP benchmark oc54x has very high fault coverage because it
consists of many and large functional components and a small control logic.
We have developed self-test routines for the following functional

Chapter 6 - Case Studies - Experimental Results

180

components of oc54x as shown in Table 6-31. In this list we can also see the
self-test routine size and execution time for each component and the overall
DSP. These routines together compose a self-test program of 1,558 words
executed in 7,742 clock cycles for the oc54x DSP benchmark.
Component Name

Seit-test
Routine Size

Execution
Time (cycles)

lwords~

Accumulator (ACC)
Arithmetic Logic Unit (ALU}
Barrel Shifter (BSFT)
Compare/Select and Store Unit (CSSU)
Exponent Decoder (EXP}
Multipl:i/Accumulate Unit (MAC)
Temporary Register (TREG)
Control
Total CPU
Table 6-31: Self-test routines statistics for oc54x DSP.

64
102
700
180
256
24
112
120

78
956
778
264
340
5,020
152
154

1,558

7,742

The targeted functional components of the oc54x DSP are very classical
components in a DSP datapath and the generation of self-test routines for
them is similar to other components like these for the previous processors'
benchmarks. The fault simulation results after evaluation of 572 test
responses in data memory are shown in Table 6-32.
We see that very high fault coverage of 96.0% with respect to single
stuck-at faults is obtained with a relatively sm all and fast self-test program.
The very high fault coverage is achieved because the vast majority of
processor area is occupied by its functional components.
Component Name
Accumulator (ACe}
Arithmetic Logic Unit (ALU)
Barrel Shifter (BSFT)
Compare/Select and Store Unit (CSSU)
Exponent Decoder (EXP)
Multiply/Accumulate Unit (MAC)
Temporary Register (TREG)
Contral
Total CPU
Table 6-32: Fault simulation results for oc54x DSP.

Fault Coverage
99.2%
98.0%
99.4%
89.0%
91.1%
98.9%
98.0%
84.0%
96.0%

Embedded Processor-Based Seit-Test

6.8

181

Compaction of test responses

For each of the processor benchmarks we have also performed

compaction of the test responses which are stored in data memory, so that a
single final signature is collected. For this reason, an additional compaction
routine of few instructions has been developed (the size of the routine is
from 10 to 20 words depending on the benchmark). When compaction is
applied, the self-test pro gram is slightly modified by changing the store
instructions that store the components' test responses in the data memory
with a call instruction that transfers the control to the compaction routine.
The application of the compaction scheme increases significantly the test
execution time ofthe modified test program. Table 6-33 shows the execution
times for the benchmark processors with and without compaction routines
executed. The times for the execution without compaction are collected from
the Tables of the previous sections.
Benchmark
Processor
Plasma
(Designs 11, 111, IV)
Meister
Jam
oc8051
RISC-MCU
oc54x

Execution time
without compaction
(clock cycles)
5,797
10,061
4,787
5,411
2,446
7,742

Execution time
with compaction
(clock cycles)
9,874
23,865
8,876
9,513
5,122
12,893

Table 6-33: Execution times of self-test routines.

The application of compaction slightly affects the size of the self-test

pro gram (by the 10 to 20 instructions we mentioned), but it seriously
increases the test application time. The application of compaction during
manufacturing testing depends on the frequency of the external tester. If the
time for uploading the responses is less than the time required for
compaction then compaction is not performed. During on-line testing, where
no external tester exists in field, the application of compaction is necessary.

6.9

Summary of Benchmarks

The following Table 6-34 summarizes the processor cores discussed in

this Chapter. The last column of Table 6-34 gives a statement for the
usefulness of experimenting with each ofthese processor benchmarks.

Chapter 6 - Case Studies - Experimental Results

182
Oescription

HOL

Usefulness

8-bit, accumulator-based
processor
32-bit RISC processor

VHDL
VHDL

Meisterl
MIPS

32-bit RISC processor

VHDL

Jam

32-bit RISC processor

VHDL

oc8051

8-bit microcontroller

Verilog

RISC-MCU

8-bit RISC microcontroller

VHDL

oc54x

16/32, dual 16-bit DSP

Verilog

Proof of concept. Has been

used in related literature.
C/assic 32-bit RISC
architecture. Used for
different synthesis
objectives.
Same c/assic 32-bit RISC
architecture implemented by
an ASIP environment.
Another 32-bit RISC
architecture. Larger design
because of more complex
pipeline structure.
C/assic, 8-bit accumulatorbased architecture. Very
popular embedded core.
A modern, sm all and fast
8-bit RISC architecture for
comparison with 32-bit RISC
ones.
A DSP architecture from a
successful model.

Benchmark
Processor
Parwan
Plasmal
MIPS

Table 6-34. Summary ofbenchmark processor cores.

We re mark that all the benchmarks we used in this Chapter are publicly
available and represent some very good efforts to implement some common,
classic and popular instruction set architectures. The success of the
application of software-based self-testing in these benchmarks gives a strong
evidence for the practicality and usefulness of the methodology, but this
success does not mean that the approach can be applied straightforwardly to
any commercial embedded processor, microprocessor or DSP, neither that
the same self-test programs will obtain the same fault coverage in the real,
commercial implementations ofthe same instructions set architectures.
Table 6-35 demonstrates the effectiveness of software-based self-testing
on each of the selected benchmark processors that can be used in practical
low-cost embedded systems. Table 6-35 summarizes for each benchmark:
the processor size in gate equivalents, the functional components percentage
with respect to the total processor area, the size of the self-test program, the
execution time of the self-test pro gram (without responses compaction) and
the total single stuck-at fault coverage for the entire processor.

Embedded Processor-Based Seit-Test

Benchmark
Processor

Gate
Count

Plasma I
Plasma 11
Plasma 111
Plasma IV
Meister
Jam
oc8051
RISC-MCU
oc54x

17,458
26,080
30,896
27,824
37,402
43,208
10,305
4,693
11,611

Functional
Components
Percentalile
83.48%
88.69%
89.39%
89.26%
68.99%
66.85%
63.64%
43.60%
92.21%

183
Self-test
program
size
965 w
853 w
853 w
853 w
1,728 w
897 w
3,760 b
1,258 w
1,558 w

Execution
Time

Fault
coverage

lc~clel
3,552
5,797
5,797
5,797
10,061
4,787
5,411
2,446
7,742

92.2%
95.3%
94.5%
95.3%
92.6%
94.0%
93.1%
91. 2%
96.0%

Table 6-35: Summary of application of software-based self-testing.

From the contents of Table 6-35 we can draw the following conclusions
that outline the effectiveness of software-based self-testing on the selected
processor benchmarks.

The self-test programs sizes are very small at the range of just a
few kilobytes.
The execution time of the self-test programs is, in all cases, less
than 10,000 clock cycles.
High fault coverage is obtained in all cases and the highest
figures are obtained far the benchmarks that have a larger
percentage of functional components.

Chapter

7
Processor-Based Testing ofSoC

In this Chapter, we discuss the concept of software-based self-testing of

SoC designs as an extension of software-based self-testing of embedded
processors.
The basic idea of software-based self-testing of SoC is that an embedded
processor that has already been tested by the execution of embedded
software routines is subsequently used to perform testing of the remaining
embedded cores of the SoC. Because of the central role of the embedded
processor in SoC self-testing, the technique is also called processor-based
self-testing.
In this Chapter first, we present the concept, some details and the
advantages of software-based self-testing of SoC, and then we describe
recent software-based self-testing approaches from the literature.

7.1

The concept

Software-based self-testing for SoC is shown in Figure 7-1 (this is an

augmented version of Figure 2-1 where we show the basic mechanism of
software-based self-testing of SoC). Software-based self-testing of SoC uses
an embedded processor core supplied with self-test routines which have been
downloaded from an external tester. The actual execution of testing is not
hooked to this extern al tester but it is performed by the processor itself at the
actual speed of operation of the Soc. The embedded processor core applies
Embedded Processor-Based Self-Test
D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic PUblishers, 2004

186

Chapter 7 - Processor-Based Testing of SoC

tests to other cores of the SoC to which it has access and captures the core
responses, playing the role of an "internai" tester.
External Test
Equipment

Memory

Figure 7-1: Software-based self-testing for SoC.

Self-testing a SoC using an embedded processor is not a straightforward

task. Several aspects of this problem must be taken into consideration while
different SoC architectures require different application of this generic
strategy.
An embedded core in a SoC may be delivered from its IP provider
supported with testing infrastructure (scan chains, OfT structures, P1500
wrapper) as well as a set of test patterns to be applied to it (when the core
comes as a hard core). In this case, the embedded processor core must be
supplied by the external test equipment with all necessary information (test
patterns, protocol for their application) so that it can effectively apply the
core test patterns and capture its test responses.
There mayaiso be cases where an embedded core in a SoC is not
supported by an existing test strategy. If the core is delivered in a gate level
netlist form or a synthesizable HOL form (firm and soft core, respectively),
then the core can be targeted for test generation either by a commercial
A TPG tool (combinational or sequential) or by a pseudorandom test strategy.
It mayaiso be modified to inc1ude OfT structures (scan, test points, etc). The
final testing approach of the core can be assigned for application to the

Embedded Processor-Based Self-Test

187

processor core which will be supplied by the core test patterns and will
effectively apply them to the core at the actual operating speed of the SoC
(at-speed testing).
In other cases, an existing core may come with an autonomous testing
mechanism, such as hardware-based self-testing. This is a very common
situation in embedded memory cores which usually contain a memory BIST
mechanism that applies a comprehensive memory test set to the core.
Memory BIST mechanisms do not occupy excessive silicon area and most
memory core providers deliver their cores with such embedded BIST
facilities. In this case, the only role that the embedded processor is assigned
in software-based self-testing of SoC, is to start and stop (if necessary) the
memory BIST execution at the embedded memory core and capture the final
signature of BIST for further evaluation. Availability of an autonomous
mechanism in an embedded processor is not, of course, restricted only in the
case of memories.
A serious factor that determines the efficiency of software-based selftesting in SoC designs is the mechanism that is used for the communication
between the embedded processor and the external test equipment. Since this
communication is restricted to the frequency of the external test equipment,
it has an impact to the overall SoC test application time. If the self-test codes
and self-test data/patternslresponses that must be transferred to the processor
memory are downloaded with a low-speed serial interface, the total
download time will me long and will seriously affect the overall test
application time. lmprovements can be seen if a parallel downloading
mechanism is used, or even better if the external test equipment can
communicate with the processor's memory using a Oirect Memory Access
(OMA) mechanism which does not interfere with the processor.
Apart from performing self-testing on the embedded cores of the SoC,
testing can be performed by software-based self-testing to the SoC buses and
interconnects. The overall performance of the SoC strongly depends on the
performance of the interconnections and hence detection of defects/faults in
these due to crosstalk must be achieved during manufacturing testing. The
use of software-based self-testing for crosstalk testing has been studied in the
literature as we will see in the next section.
Apart from the "as-is" use of an embedded processor for the self-testing
of the cores and interconnects of a SoC, two other approaches may be used
(have also been studied in the literature): (a) an existing embedded processor
may be modified to include additional instructions (instruction-level Off) to
assist the testing task of the processor itself and the SoC overall; (b) a new
processor, dedicated to SoC testing may be synthesized and used in the SoC
just for the application of the SoC test strategy. In both scenaria (a) and (b)
the objective is the reduction of the overall test application time of the SoC.

188

Chapter 7 - Processor-Based Testing of SoC

80th solutions may be very efficient if they do not add excessive area and
performance overheads in the SoC architecture.

7.1.1

Methodology advantages and objectives

Software-based self-testing for SoC is an extension of software-based

self-testing for a processor alone. In some sense, self-testing the SoC can be
considered an "easier" problem than self-testing the processor itself because:
(a) a processor is the most difficult design/core that is embedded in an SoC;
(b) using a processor to test itself has the fundamental difficulty that the test
pattern generator, the circuit under test and the test responses analyzer are all
the same entity: the processor; in testing other SoC cores using the
processor, there is a clear separation between the core that is being tested
and the core that applied the test patterns and captures the test responses.
As the two problems (software-based self-testing for processor and for
SoC) are closely related, they share the same list of advantage and have a
common set of objectives. We briefly discuss about both in the following
paragraphs.
Software-based self-testing for SoC is a low-cost test strategy which, on
the other side, delivers high test quality. The advantages of software-based
self-testing for SoC architectures are the following:
At-speed testing. The SoC is tested at the actual frequency of
operation and not in the - possibly lower - frequency of an
external
tester.
Therefore,
performance-related failure
mechanisms (such as delay faults) can be also detected. These
failure mechanisms escape detection when the chip is tested at
lower frequencies.
Small hardware and performance overheads. No additional
circuit modifications are necessary other than those required by
the particular testing strategy of each core. Even if instructionlevel DIT is applied, the overall area and performance overhead
to the entire SoC is very smalI, if it exists at all.
Yield improvement. As already analyzed in the software-based
self-testing analysis for processors, this methodology eliminates
the yield loss which is due to external tester's inaccuracies.
Being a test methodology that is applied by the chip itself, it
reduces such inaccuracies and chip rejections related to them.
Flexibility, re-usability. Software-based self-testing of SoC is a
flexible self-testing methodology which gets the advantages of
the flexibility of its main performer: the processor. For a SoC
architecture, the software-based self-testing strategy can be
revised at any stage to include new test sets for cores (for other

Embedded Processor-Based Self-Test

189

fault models 42 for example) without the need to change anything

in the SoC design or test infrastructure. Moreover, an embedded
processor can be re-used during the SoC life cycle to perform
on-line/periodic testing and detect faults that appear after some
usage ofthe system.
Low-cost testing. Most of all, software-based self-testing for SoC
is a low-cost test approach that is loosely hooked with expensive
external test equipment. Extensive analysis of this aspect of the
approach was given in previous Chapters.

SoC testing based on self-test routines execution on an embedded

processor inherently possesses the previous advantages. In order to obtain a
maximum of efficiency from the methodology, the following objectives
must be satisfied by a particular software-based self-testing methodology for
SoC.
Minimum interaction with extern al testers. This objective, if met,
leads to significant reduction ofthe SoC application time because
the exchange of data between the SoC and tester is done at the
frequency of the tester which is usually sm aller than the SoC
operating frequency.
Reduced self-test execution time. This is a second objective that
is related to SoC test application time. If a self-test routine for a
core of the SoC is executed in less time, then the overall test
application time for the core (and thus the SoC) is also reduced.
Small engineering effort. This objective (obviously the ultimate
objective of every testing approach) can be satisfied if softwarebased self-testing proceeds in two ways: (a) automation of seIftest routines development; (b) targeting the processor cores that
are most "important" for the total fault coverage of the SoC. The
"importance" of a core in a SoC can be determined by its size
compared with the size of other cores and other factors as weil.
Software-based seJf-testing of processor-based SoC architectures is a
recent direction in electronic testing which will ga in importance as time
passes. This is a prediction which is based on the increasing difficulty that
SoC test engineers face today in multi-core, complex SoC architectures. The
amount of test data that must be transferred from a tester to the SoC and vice
versa is getting enormous and the direct impact that the use of expensive,
high-memory, high-performance, high pin-count, testers has on the SoC
42

The fault models used for the different cores of the SoC may significantly defer. While
digital core testing is based on the single stuck-at fault model or a more comprehensive
structural fault model such as the path delay fault model, memory cores may be tested for
a variery of memory-related fault models.

Chapter 7 - Processor-Based Testing of SoC

190

manufacturing cost is recognized unanimously [76]. Solutions that will

reduce the test costs associated with modem SoC designs and will also have
a positive impact on yield improvement are requested. Software-based or
processor-based self-testing is such a solution.
We present in the next section abrief discussion on recent works
presented in the literature on SoC testing based on an existing embedded
processor. Several different aspects of the problem have been tackled in
these papers. Several others remain to be addressed by the methodologies of
the future years.

7.2

Literature review

We don't use any particular order of the papers mentioned in this

section, other than their chronological order, as we have already done in the
processor testing literature review Chapter 4. The literature review of this
section gives an idea ofthe current state ofresearch in this field.
S.Hellebrand, H.-J.Wunderlich and A.Hertwig in 1996 [62] discussed a
mixed-mode (combined pseudorandom and deterministic) self-testing
scheme using embedded processors and software routines. A scan-based
self-test scheme is the basis of this work. The approach tries to reduce the
length of pseudorandom sequences combining them with appropriate
deterministic test patterns in such a way that the memory requirements ofthe
deterministic part are not excessive. Experimental results on several ISCAS
circuits with software routines using Intel's 80960CA assembly language are
reported.
J.Dreibelbis, J.Barth, H.Kalter and R.Kho proposed in 1998 [38] a builtin self-test architecture for embedded DRAMs based on a custom
"processor-like" test engine. The test processor of this approach has two
separate instruction memories and is designed so that the DRAM test
dedicated pins are as few as possible. A reasonable area overhead for the test
processor is reported.
R.Rajsuman in 1999 [134] presented an approach for SoC testing using
an embedded microprocessor. At the first phase of the approach, the
microprocessor is tested for the correct operation of all its instructions in a
pseudorandom, functional BIST manner. Pseudorandom operands are
generated by an LFSR and responses are compacted by aMISR. At the
second phase, embedded memory cores are tested by software application of
classic memory testing algorithms. At the third phase of the approach, other
types of SoC cores are targeted. An example is given for the testing of an
embedded Digital to Analog Converter (DAC). The approach is finally
combined with an Iddq testing mechanism which is not related with the
embedded microprocessor.

Embedded Processor-Based Seit-Test

191

C.A.Papachristou, F.Martin and M.Nourani presented in 1999 [123] their

approach for microprocessor-based testing of core-based SoC designs. The
flexibility of processor-based SoC testing is outlined in this paper. A DMA
mechanism is assumed for the transfer of test programs to the processor
memory for subsequent execution by the processor. The access mechanism
used to apply core tests is described. The approach has a pseudorandom
philosophy with test patterns generated by LFSR and responses compacted
in MISR. A detailed analysis of the entire SoC process (downloading, core
access, test application, etc) is given and combined with a set of
experimental results on a simple test SoC. A 96.75% fault coverage is
reported for the simple SoC (5,792 faults in total).
A.Jas and N.A.Touba in 1999 [80] presented the use of an embedded
processor for efficient compression/de-compression of test data in a SoC
design. The SoC architecture is scan-based and the objective is the reduction
of test data stored in tester memory as well as the over all test application
time. Experimental results on ISCAS circuits demonstrate the improvements
obtained from this approach.
X.Bai, S.Dey and J.Rajski proposed in 2000 [8] a self-test methodology
to enable on-chip at-speed testing of crosstalk defects in SoC interconnects.
The self-test methodology was based on the maximal aggressor fault model,
that enables testing of the interconnect with a linear number of test patterns.
Efficient embedded test generators that generate test vectors for crosstalk
faults and embedded error detectors that analyze the transmission of the test
sequences received from the interconnects and detect any transmission
errors, were presented. Also test controllers to initiate and manage test
transactions by activating appropriate test generators and error detectors, and
having error diagnosis capability, were presented. The proposed self-test
methodology was applied to test for crosstalk defects the buses of a DSP
chip.
M.H.Tehranipour, Z.Navabi, and S.M.Fakhraie proposed in 2001 [153] a
software-based self-test methodology that utilizes conventional
microprocessors to test their on-chip SRAM. A mixture of existing memory
testing techniques was adopted that covers all important memory faults. The
derived test program implemented the "length 9N" memory test algorithm.
This method can be implemented for embedded memory testing of all
microprocessors, microcontrollers and DSPs. The proposed methodology
was applied to the case of the 32K SRAM of the Texas Instruments
TMS320C548 DSP.
J.-R.Huang, M.K.Iyer and K.-T.Cheng applied in 2001 [68] the concept
of software-based self-testing on a bus-based SoC design. Cores are supplied
with configurable test wrappers and self-test routines are executed by an

192

Chapter 7 - Processor-Based Testing of SoC

embedded processor to test the cores. Fault coverage results are provided in
the paper for the testing ofISCAS circuits used as example cores ofthe SoC.
C.-H.Tsai and C.-W.Wu proposed in 2001 [164] a processorprogrammable built-in self-test scheme suitable for embedded memory
testing in a Soc. The proposed self-test circuit can be programmed via an on
chip microprocessor. Upon recelvmg the commands from the
microprocessor, the test circuit generates pre-defined test patterns and
compares the memory outputs with the expected outputs. Most popular
memory test algorithms can be realized by properly programming the seIftest circuit using the processor instructions. Compared with processor-based
memory self-testing schemes that use a test pro gram to generate test patterns
and compare the memory outputs, the test time of the proposed memory
BIST scheme was greatly reduced.
C.Galke, M.Pflanz and H.T.Vierhaus proposed in 2002 [49] the concept
of designing a dedicated processor for the self-test of SoCs based on
embedded processors. A minimum sized test processor was designed in
order to perform on-chip test functions. Its architecture contained special
adopted registers to realize LFSR or MISR functions for test pattern decompaction and pattern filtering. The proposed test processor architecture is
scalable and based on a standard RISC architecture in order to facilitate the
use of standard compilers on it.
A.Krstic, W.-C.Lai, L.Chen, K.-T.Cheng and S.Dey presented in 2002
[101], [102] a review ofthe group's work in the area ofsoftware-based selftesting for processors and SoC designs. The software-based self-testing
concept is presented in detail in these works and its advantages are clearly
outlined. Discussion covers self-testing of the embedded processor (for
stuck-at and delay faults), self-diagnosis of the embedded processor, selftesting of buses and interconnects, self-testing of other SoC cores using the
processor and also instruction-Ievel DIT. The authors have worked in these
subtopics of software-based self-testing and a reference in more detail has
been given in Chapter 4 for those that are related to processor self-testing
only.
S.Hwang and J.A.Abraham discussed in 2002 [71] an optimal BIST
technique for SoC using an embedded microprocessor. The approach aims to
the reduction of memory requirements and test application time in scanbased core testing using pseudorandom and deterministic test patterns. This
is achieve using a new test data compression technique using the embedded
processor of the SoC. Experimental results are given using Intel's x86
instruction set (pro grams are deve10ped in C and a compiler is used for their
translation in assembly/machine language) on several ISCAS circuits.
Comparisons are also reported with the previous works of [62], [80], and the

Embedded Processor-Based Seit-Test

193

superiority of the approach of [71] is shown in terms of the total test data
size ofthe three approaches on the ISCAS circuits studied.
L.Chen, X.Bai, S.Dey proposed in 2002 [26] a new software-based seIftest methodology for SoC based on embedded processors that utilizes an onchip embedded processor to test at-speed the system-level interconnects for
crosstalk. Testing of long interconnects for crosstalk in SoCs is important
because crosstalk effects degrade the integrity of signals traveling on long
interconnects. They demonstrated the feasibility of the proposed method by
applying it to test the interconnects of a processor-memory system. The
defect coverage was evaluated using a system-level crosstalk defect
simulation environment.
M.H.Tehranipour, M.Nourani, S.M.Fakhraie and A.Afzali-Kusha in 2003
[154] outlined the use of embedded processor as a test controller in a SoC to
test itself and the other cores. It is assumed that a DMA mechanism is
available for efficient downloading of test programs in the processor
memory. The flexibility of the approach of processor-based SoC testing is
discussed. The approach is based on appropriate use of processor
instructions that have access to the SoC cores. The approach is evaluated on
a SoC design based on a DSP core called UTS-DSP which is compatible
with TI's TMS320C54x DSP family. The SoC also contains an SRAM, a
ROM, aSerial Port Interface and a Host Port Interface. A test program has
been developed for the entire SoC consisting of a total of 689 bytes and a
test execution time of 84.25 msec. The test pro gram reached a 95.6% fault
coverage for the DSP core, 100% for the two memories and 86.1 % and
81.3% for the two interface cores, respectively. Although, in the work of
[154], no sufficient details are given for the approach, the interest of this
paper is that it gives an idea of the overall SoC testability that can be
obtained by a very small embedded software program.

7.3

Research focus in processor-based SoC testing

Research in the area of processor-based or software-based self-testing of

SoC architectures is expected to attract the interest of test researchers and
engineers in the near future. Aspects of the approach that need special
consideration are:

Self-test optimization including: test application time reduction,

test data volume reduction, compression/de-compression
techniques of test data and parallelization of cores testing.
Diagnostic capabilities of software-based self-testing for SoC.
Performing diagnosis and identifying the malfunctioning parts of
the SoC (cores or interconnects) will provide valuable

194

Chapter 7 - Processor-Based Testing of SoC

infonnation for SoC manufacturing processes that will lead to

yield improvements.
Automation of self-test programs generation for different types
of embedded SoC cores and different instruction sets of the
embedded processors.
On-fine testing for SoC using an embedded processor as well as
fault tolerance techniques utilizing embedded processors and
software routines.

The advances on these topics will eventually prove the effectiveness of

software-based self-testing as a generic, low-cost, high-test quality selftesting methodology for complex SoC architectures built around embedded
processors.

Chapter

8
Conclusions

The emerging approach of software-based self-testing of embedded

processors as weIl as processor-based SoC architectures has been discussed
in this book. Software-based self-testing performs self-testing of processors
and SoC based on the execution of embedded software routines, instead of
assigning this task to dedicated hardware resources.
The definition of software-based self-testing sets as its essential objective
the reduction of test costs for the processor and the SoC. Therefore,
software-based self-testing is a low-cost test methodology. It does not add
area or performance overheads to the design and it reduces the interaction
with extemal test equipment which increases the test costs. Loose relation
with extemal testers has another cost-related positive impact: yield loss due
to the inherent tester's measurement inaccuracies is eliminated since testing
is executed by the chip itself.
Moreover, software-based self-testing is a high test quality methodology
because it allows the detection of failure mechanisms that can only be
detected if the chip is tested at the actual operating frequency (at-speed
testing).
Software-based self-testing is a flexible and re-usable self-testing
technique that can be tuned to the specifics of a design, can be augmented
for more comprehensive fault models and can also be re-used during the
system's life cyde for on-line/periodic testing.

Embedded Processor-Based Self-Test

D.Gizopoulos, A.Paschalis, Y.Zorian
Kluwer Academic Publishers, 2004

Chapter 8 - Conclusions

196

We have presented these aspects ofthe methodology in this book and we

have described the requirements that must be satisfied as weIl as details on
how the methodology objectives can be met.
By setting the requirements for efficient software-based self-testing, we
want to set a common framework for comparisons in the topic. We discussed
a flow for low-cost, software-based self-testing of embedded processors by
tackling the most important components of the processor and by developing
efficient self-test routines for each of them. Alternative styles in self-test
routines development were discussed along with potential optimization
techniques.
The software-based self-testing framework is also supported in a more
quantitative manner, with a set of experimental results on publicly available
embedded processor models of different complexities and architectures.
The future aims of research in the field of software-based self-testing for
processors and SoC that will prove the long term usefulness of the approach
and its applicability to a wide range of processor architectures are the
following:

Scalability: applicability of the approach to large, industrial

embedded processors including several performance enhancing
techniques.
Automation: development of self-test routines for the
components, self-test programs for the processor and the SoC in
an automated flow that reduces test engineering effort and cost.

FinaIly, the application of software-based self-testing for on-line/periodic

testing of processors and SoC, as weIl as for low-cost fault tolerance using
embedded software routines will attract the interest oftechnologists.

References

[1]

[2]

[3]

[4]

[5]

[6]

[7]

M.Abramovici, M.Breuer, A.D.Friedman, Digital Systems Testing

and Testable Design, Piscataway, New Jersey: IEEE Press, 1994.
Revised Printing.
S.M.I.Adham, S.Gupta, "DP-BIST: a built-in self-test for DSP
data paths-a low overhead and high fault coverage technique",
Proceedings of the Fifth Asian Test Symposium (ATS) 1996,
pp.205-512.
V.D.Agrawal, C.R.Kime, K.K.Saluja, "A Tutorial on Built-In SelfTest, Part 1: Principles", IEEE Design & Test of Computers, vol.
10, no. 1, pp. 73-82, March 1993.
V.D.Agrawal, C.R.Kime, K.K.Saluja, "A Tutorial on Built-In SeIfTest, Part 2: Applications", IEEE Design & Test of Computers,
vol. 10, no. 2, pp. 69-77, June 1993.
D.Amason, A.L.Crouch, R.Eisele, G.Giles, M.Mateja, "A case
study of the test development for the 2nd generation ColdFire R
microprocessors", Proceedings of the IEEE International Test
Conference (ITC) 1997, pp. 424-432.
M.Annaratone, M.G.Sami, "An Approach to Functional Testing of
Microprocessors", Proceedings of the Fault Tolerant Computing
Symposium (FTCS) 1982, pp. 158-164.
P.J.Ashenden,
J.P.Mermet,
R.Seepold,
System-on-Chip
Methodologies & Design Languages, June 2001, Kluwer
Academic Publishers.

198

References

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]
[18]

[19]

X.Bai, S.Dey, J.Rajski, "Self-Test Methodology for At-Speed Test

of Crosstalk in Chip Interconnects", Proceedings of the
ACM/IEEE Design Automation Conference (DAC) 2000, pp. 619624.
P.H.Bardell, W.H.McAnney, J.Savir, Built-In Test for VLSl:
Pseudorandom Techniques, John Wiley and Sons, New York,
1987.
K.Batcher, C.Papachristou, "Instruction randomization seIftest for
processor cores", Proceedings of the IEEE VLSI Test Symposium
(VTS) 1999, pp. 34-40.
P.H.Bardell, M.J.Lapointe, "Production experience with Built-In
SeIf-Test in the IBM ES/9000 System", Proceedings of the IEEE
International Test Conference (lTC) 1991, pp. 28-37.
P.G.BeIomorski, "Pseudorandom self-testing of mieroprocessors",
Microprocessing & Microprogramming, vo1.19, no.l, pp. 37-47,
January 1987.
D.K.Bhavsar,
D.R.Akeson,
M.K.Gowan,
D.B.Jackson,
"Testability access of the high speed test features in the Alpha
21264 mieroprocessor", Proceedings of the IEEE International
Test Conference (ITC) 1998, pp. 487-495.
D.K.Bhavsar, J.H.Edmondson, "Testability strategy of the Alpha
AXP 21164 mieroprocessor", Proceedings of the IEEE
International Test Conference (ITC) 1994, pp. 50-59.
D.K.Bhavsar, J.H.Edmondson, "Alpha 21164 testability strategy",
IEEE Design & Test of Computers, vo1.14, no.l, pp. 25-33,
January-March 1997.
D.K.Bhavsav, et al, "Testability Access of the High Speed Test
Features in the Alpha 21264 Mieroprocessor", Proceedings of the
IEEE International Test Conference (ITC) 1998, pp. 487-495.
M.Bilbault, "Automatie testing of mieroprocessors", Electronique
& Microelectronique Industrielles, no. 208,1975, pp.31-33.
U.Bieker, P.Marwedel, "Retargetable Self-Test Program
Generation Using Constraint Logic Programming", Proceedings of
the ACM/IEEE Design Automation Conference (DAC) 1995, pp.
605-611.
P.E.Bishop, G.L.Giles, S.N.lyengar, C.T.Glover, W.-O.Law,
"Testability considerations in the design of the MC68340
Integrated Processor Unit", Proceedings of the IEEE International
Test Conference (lTC) 1990, pp. 337-346.

References

[20]

[21]

[22]
[23]

[24]

[25]
[26]

[27]

[28]

[29]

[30]

[31]

199

R.D.Blanton, J.P.Hayes, "Design of a Fast, Easily Testable ALU",

Proceedings ofthe IEEE VLSI Test Symposium (VTS) 1996, pp.
9-16.
D.Brahme,
J.A.Abraham,
"Functional
Testing
of
Microprocessors", IEEE Transactions on Computers, vol. C-33,
pp. 475-485, June 1984.
R.L.Britton, MIPS Assembly Language Programming, Pearson
Prentice Hall, Upper Saddle River, NJ, 2004.
M.L.Bushnell, V.D.Agrawal, Essentials of Electronic Testing for
Digital, Memory and Mixed-Signal VLSI Circuits, Kluwer
Academic Publishers, 2000.
A.Carbine, D.Feltham, "Pentium-Pro Microprocessor Design for
Test and Debug", Proceedings of the IEEE International Test
Conference (ITC) 1997.
K.Chakrabarty, V.lyengar, A.Chandra, Test Resource Partitioning
for System-on-a-Chip, Kluwer Academic Publishers, 2002.
L.Chen, X.Bai, S.Dey, "Testing for interconnect crosstalk defects
using on-chip embedded processor cores", Journal of Electronic
Testing: Theory and Applications, vo1.18, no.4-5, pp. 529-538,
August-October 2002.
L.Chen, S.Dey, "DEFUSE: a Deterministic Functional Self-Test
Methodology for Processors", Proceedings ofthe IEEE VLSI Test
Symposium (VTS) 2000, pp. 255-262.
L.Chen, S.Dey, "Software-Based Self-Testing Methodology for
Processor Cores", IEEE Transactions on Computer Aided Design
of Integrated Circuits and Systems, vol. 20, no. 3, pp. 369-380,
March 2001.
L.Chen, S.Dey, "Software-Based Diagnosis for Processors",
Proceedings of the ACMlIEEE Design Automation Conference
(DAC) 2002, pp. 259-262, June 2002.
T.Cheng, E.Hoang, D.Rivera, A.Haedge, J.Fontenot, A.G.Carson,
"Test Grading the 68332", Proceedings of the IEEE International
Test Conference (!TC) 1991, pp. 150-159.
L.Chen, S.Ravi, A.Raghunathan, S.Dey, "A Scalable SoftwareBased Self-Test Methodology for Programmable Processors",
Proceedings of the ACMIIEEE Design Automation Conference
(DAC) 2003, pp. 548-553, June 2003.

200

References

[32]

F.Corno, M.Sonza Reorda, G.Squillero, M.Violante, "On the Test

of Microprocessor IP Cores", Proceedings of the IEEE Design
Automation & Test in Europe Conference (DATE) 2001, March
2001, pp.209-213.
[33] B.Courtois, "A Methodology for On-Line Testing of
Microprocessors", Proceedings of the Fault-Tolerant Computing
Symposium (FTCS) 1981, pp. 272-274.
[34] G.Crichton, "Testing microprocessors", IEEE Journal of SolidState Circuits, June 1979, pp. 609-613.
[35] A.L.Crouch, M.Pressly, J.Circello, "Testability features of the
MC68060 microprocessor", Proceedings ofthe IEEE International
Test Conference (ITC) 1994, pp. 60-69.
[36] W.J.Culler, Implementing Safety Critical Systems: The VIPER
microprocessor, Kluwer Academic Publishers, 1987.
[37] L.L.Day,
P.A.Ganfield,
D.M.Rickert, F.J.Ziegler,
"Test
Methodology for a Microprocessor with Partial Scan",
Proceedings of the IEEE International Test Conference (lTC)
1998, pp. 708-716.
[38] J.Dreibelbis, J.Barth, H.Kalter, R.Kho, "Processor-Based Built-In
Self-Test for Embedded DRAM", IEEE Journal of Solid-State
Circuits, vol. 33, no. 11, pp. 1731-1740, November 1998.
[39] E.B.Eichelberger, E.Lindbloom, J.A.Waicukauski, T.W.Williams,
Structured Logic Testing, Englewood Cliffs, New Jersey: PrenticeHall,1991.
[40] S.Erlanger, D.K.Bhavsar, R.Davies, "Testability Features of the
Alpha 21364 Microprocessor", Proceedings of the IEEE
International Test Conference (ITC) 2003, pp. 764-772.
[41] X.Fedi, R.David "Some experimental results from random testing
of microprocessors", IEEE Transactions on Instrumentation &
Measurement, vol. IM-35, no.l, pp. 78-86, March 1986.
[42] X.Fedi, R.David "Experimental results from Random testing of
Microprocessors", Proceedings of the Fault-Tolerant Computing
Symposium (FTCS) 1984, pp. 225-230.
[43] R.S.Fetherston, I.P.Shaik, S.C.Ma, "Testability features of AMDK6 microprocessor", Proceedings ofthe IEEE International Test
Conference (ITC) 1997, pp. 406-413.

References

[44]

[45]

[46]

[47]

[48]
[49]

[50]

[51]
[52]

[53]

[54]

[55]

201

T.G.Foote,
D.E.Hoffman,
W.V.Huott,
T.J.Koprowski,
B.J.Robbins, M.P.Kusko, "Testing the 400 MHz IBM generation-4
CMOS chip", Proceedings of the IEEE International Test
Conference (lTC) 1997, pp. 106-114.
J.F.Frenzel, P.N.Marinos, "Functional Testing of Microprocessors
in a User Environment", Proceedings of the Fault Tolerant
Computing Symposium (FTCS) 1984, pp. 219-224.
R.A.Frohwerk, "Signature Analysis: A New Digital Field Service
Method", Hewlett-Packard Journal, vol. 28, no. 9, pp. 2-8, May
1977.
R.Fujii,
J.A.Abraham,
"Self-test
for
microprocessors",
Proceedings of the International Test Conference (ITC) 1985, pp.
356-361.
S.B.Furber, ARM System-on-Chip Architecture (2nd Edition),
Addison-Wesley, August, 2000.
C.Galke, M.Pflanz, H.T.Vierhaus, "A test processor concept for
systems-on-a-chip", Proceedings of the IEEE International
Conference on Computer Design (lCCD) 2002, pp.210-212.
M.G.Gallup,
W.Ledbetter,
R.McGarity,
S.McMahan,
K.C.Scheuer, C.G.Shepard, L.Sood, "Testability features of the
68040", Proceedings of the IEEE International Test Conference
(ITC) 1990, pp. 749-757.
S.Ghosh, Hardware Description Languages: Concepts and
Principles, New York: IEEE Press, 2000.
D.Gizopoulos, A.Paschalis and Y.Zorian, "An Effective BIST
Scheme for Datapaths", Proceedings of the IEEE International
Test Conference, (ITC) 1996, pp. 76-85.
D.Gizopoulos, A.Paschalis and Y.Zorian, "An Effective Built-In
Self-Test Scheme for Booth Multipliers", IEEE Design & Test of
Computers, vol. 15, no. 3, pp. 105-111, July-September 1998.
D.Gizopoulos, A.Paschalis and Y.Zorian, "An Effective Built-In
Self-Test Scheme for Array Multipliers", IEEE Transactions on
Computers, vol. 48, no. 9, pp. 936-950, September 1999.
D.Gizopoulos, A.Paschalis, Y.Zorian and M.Psarakis, "An
Effective BIST Scheme for Arithmetic Logic Units", Proceedings
ofthe IEEE International Test Conference, 1997, pp. 868-877.

202

References

[56]

[57]

[58]

[59]

[60]

[61]

[62]

[63]

[64]
[65]

[66]

[67]

N.Gollakota, A.Zaidi, "Fault grading the Intel 80486",

Proceedings of the IEEE International Test Conference (lTC)
1990, pp. 758-761.
F.Golshan, "Test and On-line Debug Capabilities of IEEE Std
1149.1 in UltraSPARCTM-III Microprocessor", Proceedings ofthe
IEEE International Test Conference (lTC) 2000, pp. 141-150.
A.J. van de Goor, Th.J.W. Verhallen, "Functional Testing of
Current Microprocessors", Proceedings of the IEEE International
Test Conference (ITC) 1992, pp. 684-695.
M.Gumm, VLSI Design Course: VHDL-Modelling and Synthesis
0/ the DLXS RISC Processor, University of Stuttgart, Germany,
December 1995.
R.Gupta, Y.Zorian, "Introducing Core-Based System Design",
IEEE Design & Test of Computers, October-December 1997, pp.
15-25.
K.Hatayama, K.Hikone, T.Miyazaki, H.Yamada, "A Practical
Approach to Instruction-Based Test Generation for Functional
Modules ofVLSI Processors", Proceedings ofthe IEEE VLSI Test
Symposium (VTS) 1997, pp. 17-22.
S.Hellebrand, H.-J.Wunderlich, A.Hertwig, "Mixed-Mode BIST
Using Embedded Processors", Proceedings of the IEEE
International Test Conference (ITC) 1996, pp. 195-204.
J.L.Hennessy, D.Patterson, Computer Architecture
A
Quantitative Approach, Second Edition, San Francisco, CA:
Morgan Kaufmann, 1996.
B.Henshaw, "An MC68020 users test program", Proceedings of
the International Test Conference (ITC) 1986, pp. 386-393.
G.Hetherington, G.Sutton, K.M.Butler, T.J.Powell, "Test
generation and design for test for a large multiprocessing DSP",
Proceedings of the IEEE International Test Conference (ITC)
1995, pp. 149-156.
K.Holdbrook, S.Joshi, S.Mitra, J.Petolino, R.Raman, M.Wong,
"MicroSPARC: a case-study of scan based debug", Proceedings of
the IEEE International Test Conference (ITC) 1994, pp. 70-75.
H.Hong, R.Avra, "Structured design-for-debug-the SuperSPARC
II methodology and implementation", Proceedings of the IEEE
International Test Conference (ITC) 1995, pp. 175-183.

References

[68]

[69]

[70]

[71]

[72]
[73]
[74]

[75]

[76]
[77]

[78]
[79]

[80]

203

J.-R. Huang, M. K Iyer, K-T. Cheng, "A Self-Test Methodology

for IP Cores in Bus-Based Programmable SoCs", Proeeedings of
the IEEE VLSI Test Symposium (VTS) 2001, pp. 198-203.
C.Hunter, E.KVida-Torku, J.LeBlane, "Balaneing struetured and
ad-hoe design for test: testing of the PowerPC 603
mieroproeessor", Proeeedings of the IEEE International Test
Conferenee (ITC) 1994, pp. 76-83.
C.Hunter, J.Gaither, "Design and implementation of the "G2"
PowerPC 603e-embedded mieroproeessor eore", Proeeedings of
the IEEE International Test Conferenee (ITC) 1998, pp. 473-479.
S.Hwang, J.AAbraham, "Optimal BIST Using an Embedded
Microproeessor", Proeeedings of the IEEE International Test
Conferenee (ITC) 2002, pp. 736-745.
IEEE International Workshop on Test Resouree Partitioning (TRP)
2000,2001,2002,2003.
IEEE P1500 SECT web site. http://grouper.ieee.org/groups/1500
IEEE, IEEE Standard Verilog Language Reference Manual, Std
1364-1995, New York: IEEE, 1995 (also available in
http://standards.ieee.org)
IEEE, IEEE Standard VHDL Language Reference Manual, Std
1076-1993, New York: IEEE, 1993 (also available in
http://standards.ieee.org)
International Teehnology Roadmap for Semieonduetors (ITRS),
2003 Edition, http://public.itrs.net.
S.KJain, AKSusskind, "Test strategy for mieroproeessors",
Proeeedings of the ACMIIEEE 20th Design Automation
Conferenee (DAC) 1983, pp.703-708.
Jam CPU model, http://www.etek.ehalmers.se/~e8mn/web/jam
A.Jantseh, Modeling Embedded Systems and SoC's: Concurrency
and Time in Models of Computation, Morgan Kaufmann, June
2003.
A.Jas, N.ATouba, "Using an embedded proeessor for effieient
deterministie testing of systems-on-a-ehip", Proeeedings of the
IEEE International Conferenee on Computer Design (ICCD) 1999,
pp. 418-423.

204

References

[81]

Y.Jen-Tien, M.Sullivan, C.Montemayor, P.Wilson, R.Evers,

"Overview of PowerPC 620 multiprocessor verification strategy",
Proceedings of the IEEE International Test Conference (ITC)
1995, pp. 167-174.

J.Jishiura, T.Maruyama, H. Maruyama, S.Kamata, "Testing VLSI

microprocessor with new functional capability", Proceedings of
the IEEE International Test Conference (ITC) 1982, pp. 628-633.
[83] D.D.Josephson, D.J.Dixon, BJ.Arnold, "Test Features of the HP
PA7100LC Processor", Proceedings ofthe IEEE International Test
Conference (ITC) 1993, pp. 764-772.
[84] D.Josephson,
S.Poehlman, V.Govan, C.Mumford, "Test
Methodology for the McKinley Processor", Proceedings of the
IEEE International Test Conference (ITC) 2001, pp. 578-585.
[85] G.Kane, J.Heinrich, MIPS RlSC Architecture, Prentice Hall, 1992.
[86] M.G.Karpovsky, R.G. van Meter, "An approach to the testing of
microprocessors", Proceedings of the ACM/IEEE 21st Design
Automation Conference (DAC) 1984, pp.196-202.
[87] S.Karthik,
M.Aitken,
L.Martin,
S.Pappula,
B.Stettler,
P.Vishakantaiah, M.d'Abreu, J.A.Abraham, "Distributed mixed
level logic and fault simulation on the PentiumR Pro
microprocessor", Proceedings of the IEEE International Test
Conference (ITC) 1996, pp. 160-166.
[88] M.Keating, P.Bricaud, Reuse Methodology Manual/or System-ona-Chip Designs, Third Edition, June 2002, Kluwer Academic
Publishers.
[89] A.Kinra, A.Mehta, N.Smith, J.Mitchell, F.Valente, "Diagnostic
techniques for the UltraSPARC microprocessors", Proceedings
of the IEEE International Test Conference (ITC) 1998, pp. 480486.
[90] A.Kinra, "Towards Reducing 'Functional-Only' Fails for the
UltraSP ARC Microprocessors", Proceedings of the IEEE
International Test Conference (ITC) 1999, pp. 147-154.
[91] H.-P.Klug, "Microprocessor Testing by Instruction Sequences
Derived from Random Patterns", Proceedings of the IEEE
International Test Conference (ITC) 1988, pp. 73-80.
[92] R.Koga, W.A.Kolasinski, M.T.Marra, W.A.Hanna, "Techniques of
microprocessor testing and SEU-rate prediction", IEEE
[82]

References

[93]

[94]

[95]

[96]

[97]

[98]

[99]

[100]

[101]

[102]

205

Transactions on Nuclear Science, vol. NS-32, no.6, pp.4219-24,

December 1985.
N.Kranitis, D.Gizopoulos, A.Paschalis, M.Psarakis and Y.Zorian,
"PowerlEnergy Efficient Built-In Self-Test Schemes for Processor
Datapaths", IEEE Design & Test of Computers, vol. 17, no. 4, pp.
15-28, October-December 2000. Special Issue on Microprocessor
Test and Verification.
N.Kranitis, D.Gizopoulos, A.Paschalis, Y.Zorian, "InstructionBased Self-Testing of Processor Cores", Proceedings of the IEEE
VLSI Test Symposium (VTS) 2002, pp. 223-228.
N.Kranitis, A.Paschalis, D.Gizopoulos, Y.Zorian, "Effective
Software Self-Test Methodology for Processor Cores",
Proceedings of the IEEE Design Automation and Test in Europe
Conference (DATE) 2002, pp. 592-597.
N.Kranitis, A.Paschalis, D.Gizopoulos, Y.Zorian, "InstructionBased Self-Testing of Processor Cores", Journal of Electronic
Testing: Theory and Applications, Kluwer Academic Publishers,
Special Issue on VTS 2003, vol. 19, no 2, pp 103-112, April 2003.
N.Kranitis, G.Xenoulis, D.Gizopoulos, A.Paschalis, Y.Zorian,
"Low-Cost Software-Based Self-Testing of RISC Processor
Cores", IEE Computers and Digital Techniques, Special Issue on
DATE 2003 Conference, September 2003.
N.Kranitis, G.Xenoulis, D.Gizopoulos, A.Paschalis, Y.Zorian,
"Software-Based Self-Testing of Large Register Banks in RISC
Processor Cores", Proceedings of the 4th IEEE Latin American
Test Workshop 2003.
N.Kranitis, y.xenoulis, D.Gizopoulos, A.Paschalis, Y.Zorian,
"Low-Cost Software-Based Self-Testing of RISC Processor
Cores", Proceedings of the IEEE Design Automation and Test in
Europe Conference (DATE) 2003.
N.Kranitis, G.Xenoulis, A.Paschalis, D.Gizopoulos, Y. Zorian,
"Application and Analysis of RT-Level Software-Based SelfTesting for Embedded Processor Cores", Proceedings of the IEEE
International Test Conference (nC) 2003, pp. 431-440.
A.Krstic, L.Chen, W.C.Lai, K.T.Cheng, S.Dey, "Embedded
Software-Based Self-Test for Programmable Core-Based
Designs", IEEE Design & Test of Computers, vol. 19, no. 4, pp.
18-26, July-August 2002.
A.Krstic, W.C.Lai, L.Chen, K.T.Cheng, S.Dey, "Embedded
Software-Based Self-Testing for SoC Design", Proceedings of the

206

References

[103J

[104J

[105J

[106J

[107J

[108J

[109J

[110J

[IIIJ

[112J

ACM/IEEE Design Automation Conference (DAC) 2002, pp. 355360.

M.P.Kusko, B.J.Robbins, T.J.Snethen, P.Song, T.G.Foote,
W.V.Huott, "Microprocessor test and test tool methodology for the
500 MHz IBM S/390 G5 chip", Proceedings of the IEEE
International Test Conference (ITC) 1998, pp. 717-726.
W.-C.Lai, K.-T.Cheng, "Instruction-Level DFT for Testing
Processor and IP Cores in System-on-a-Chip", Proceedings of the
ACM/IEEE Design Automation Conference (DAC) 2001, pp. 5964, June 2001.
W.-C.Lai, A.Krstic, K.-T.Cheng, "Test Pro gram Synthesis for Path
Delay Faults in Microprocessor Cores", Proceedings of the IEEE
International Test Conference (ITC) 2000, pp. 1080-1089.
P.K.Lala, "Microprocessor chip testing-a new method",
Proceedings of the Microtest. Soc. Electronic & Radio
Technicians, 1979, pp. 152-162.
E.C.Lee, "A simple concept in microprocessor testing", Digest of
Papers ofthe IEEE Semiconductor Test Symposium, 1976, pp. 1315.
J.Lee, J.H.Patel, "An Instruction Sequence Assembling
Methodology for Testing Microprocessors", Proceedings of the
IEEE International Test Conference (ITC) 1992, pp. 49-58.
J.Lee, J.H.Patel, "Architectural Level Test Generation for
Microprocessors", IEEE Transactions on Computer-Aided Design
ofIntegrated Circuits and Systems, vol. 31, no. 10, pp. 1288-1300,
October 1994.
M.E.Levitt, S.Nori, S.Narayanan, G.P.Grewal, L.Youngs, A.Jones,
G.Billus, S.Paramanandam, "Testability, debuggability, and
manufacturability features of the UltraSPARC-I microprocessor",
Proceedings of the IEEE International Test Conference (ITC)
1995, pp. 157-166.
J.A.Lyon, M.Gladden, E.Hartung, E.Hoang, K.Raghunathan,
"Testability Features ofthe 68HCI6Z1", Proceedings ofthe IEEE
International Test Conference (ITC) 1991, pp. 122-131.
T.L.McLaurin, F.Frederick, "The Testability Features of the
MCF5407 Containing the 4th Generation Coldfire Microprocessor
Core", Proceedings of the IEEE International Test Conference
(ITC) 2000, pp. 151-159.

References

207

[113] T.L.McLaurin, F.Frederick, R.Slobodnik, "The Testability

Features of the ARM1026EJ Microprocessor Core", Proceedings
of the IEEE International Test Conference (ITC) 2003, pp. 773782.
Processor,
[114] Meister
Application
Specific
Instruction
http://www .eda-meister.org
[115] B.T.Murray, J.P.Hayes, "Hierarchical Test Generation using
Precomputed Tests for Modules", IEEE Transactions on Computer
Aided Design of Integrated Circuits and Systems, vol. 19, no. 6,
pp. 594-603, June 1990.
[116] Z.Navabi, VHDL: Analysis and Modeling of Digital Systems, New
York: McGraw-Hill, 1993.
[117] W.Needham,
N.Gollakota,
"DFT
strategy
for
Intel
microprocessors", Proceedings of the IEEE International Test
Conference (ITC) 1996, pp. 396-399.
[118] N.Nicolici, B.M.AI-Hashimi, Power-Constrained Testing of VLSI
Circuits, Kluwer Academic Publishers, 2003.
[119] A.Noore, B.E.Weinrich, "Strategies for functional testing of
microprocessors", Proceedings of the 22 nd IEEE Southeastern
Symposium on System Theory, 1990, pp.431-435.
[120] oc54x DSP model, http://www.opencores.org/projects/oc54x.
[121] oc8051 CPU model, http://www .opencores.org/projects/oc8051.
[122] Panel Session, "Microprocessor Testing: Which Technique is
Best?", Proceedings of the ACMIIEEE Design Automation
Conference (DAC) 1994, p. 294, June 1994.
[123] C.A.Papachristou, F.Martin, M.Nourani, "Microprocessor Based
Testing for Core-Based System on Chip", Proceedings of the
ACMIIEEE Design Automation Conference (DAC) 1999, pp. 586591, June 1999.
[124] K.P.Parker, The Boundary-Scan Handbook, Third Edition, Kluwer
Academic Publishers, 2003.
[125] P.Parvathala, K.Maneparambil, W.Lindsay, "FRITS
A
Microprocessor Functional BIST Method", Proceedings of the
IEEE International Test Conference (ITC) 2002, pp. 590-598.
[126] R.Patel, K.Yarlagadda, "Testability features of the SuperSPARC
microprocessor", Proceedings of the IEEE International Test
Conference (ITC) 1993, pp. 773-781.
[127] Picojava Microprocessor Cores, Sun Microsystems [Online).
Available: http://www.sun.com/microelectronics/picoJava

208

References

[128] PlasmalMIPS
CPU
Model,
http://www.opencores.org/projects/mips
[129] c.Pyron, J.Prado, J.Golab, "Next generation PowerPC
microprocessor test strategy improvements", Proceedings of the
IEEE International Test Conference (ITC) 1997, pp. 414-423.
[130] C.Pyron, M.Alexander, J.Golab, G.Joos, B.Long, R.Molyneaux,
R.Raina, N.Tendolkar, "DFT Advances in the Motorola's
MPC7400, a PowerPC G4 Microprocessor", Proceedings of the
IEEE International Test Conference (ITC) 1999, pp. 137-146.
[131] K.Radecka, J.Rajski, J.Tyszer, "Arithmetic Built-In Self-Test for
DSP Cores", IEEE Transactions on Computer Aided Design of
Integrated Circuits and Systems, vol.
16, no.
11,
pp. 1358-1369, November 1997.
[132] R.Raina, R.Bailey, D.Belete, V.Khosa, R.Molyneaux, J.Prado,
A.Razdan, "DFT Advances in Motorola's Next-Generation 74xx
PowerPCTM Microprocessor", Proceedings of the IEEE
International Test Conference (ITC) 2000, pp. 132-140
[133] J.Rajski, J.Tyszer, Arithmetic Built-In Self-Test for Embedded
Systems, Prentice-Hall, Upper Saddle River, New Jersey, 1998.
[134] R. Rajsuman, "Testing A System-on-Chip with Embedded
Microprocessors", Proceedings of the IEEE International Test
Conference (ITC) 1999, pp. 499-508.
[135] R.Regalado, "A 'people oriented' approach to microprocessor
testing", Proceedings of the IEEE International Symposium on
Circuits and Systems (ISCAS) 1975, pp. 366-368.
[136] RISC-MCU
CPU
model,
http://www .opencores.org/proj ects/riscmcu/
[137] C.Robach, C.Bellon, G.Saucier, "Application oriented test for
dedicated microprocessor systems", Microprocessors and their
Applications, 1979, pp. 275-283.
[138] C.Robach, G.Saucier, "Application Oriented Microprocessor Test
Method", Proceedings of the Fault Tolerant Computing
Symposium (FTCS) 1980, pp. 121-125.
[139] C.Robach, G.Saucier, R.Velazco, "Flexible test method for
microprocessors", Proceedings of the 6th EUROMICRO
Symposium on Microprocessing and Microprogramming, 1980,
pp.329-339.

References

209

[140] G.Roberts, J. Masciola, "Microprocessor boards and systems. A

new approach to post in-circuit testing", Test, vol.6, no.2, pp. 3234, March 1984.
[141] KKSaluja, L.Shen, S.Y.H.Su, "A Simplified Algorithm for
Testing Microprocessors", Proceedings of the IEEE International
Test Conference (ITC) 1983, pp. 668-675.
[142] KKSaluja, L.Shen, S.Y.H.Su, "A simplified algorithm for testing
microprocessors", Computers & Mathematics with Applications,
vo1.13, no.5-6, 1987, pp.431-441.
[143] P.Seetharamaiah, V.R.Murthy, "Tabular mechanisation for flexible
testing of microprocessors", Proceedings of the International Test
Conference (ITC) 1986, pp. 394-407.
[144] J.Shen, J.Abraham, "Native Mode Functional Test Generation for
Microprocessors with Applications to Self-Test and Design
Validation", Proceedings of the International Test Conference
(ITC) 1998, pp. 990-999.
[145] L.Shen, S.Y.H.Su, "A Functional Testing Method for
Microprocessors", Proceedings of the Fault-Tolerant Computing
Symposium (FTCS) 1984, pp. 212-218.
[146] L.Shen, S.Y.H.Su, "A Functional Testing Method for
Microprocessors", IEEE Transactions on Computers, vol. 37, no.
10, pp. 1288-1293, October 1988.
[147] D.H.Smith, "Microprocessor testing - method or madness", Digest
of Papers of the IEEE Semiconductor Test Symposium, 1976, pp.
27-29.
[148] J.Sosnowski,
A.Kusmierczyk,
"Pseudorandom
versus
Deterministic Testing of Intle 80x86 Processors", Proceedings of
the IEEE Euromicro-22 Conference 1996, pp. 329-336.
[149] T.Sridhar, J.P.Hayes, "A Functional Approach to Testing BitSliced Microprocessors", IEEE Transactions on Computers, vol.
C-30, no. 8, pp. 563-571, August 1981.
[150] C.Stolicny, R.Davies, P.McKernan, T.Truong, "Manufacturing
pattern development for the Alpha 21164 microprocessor",
Proceedings of the IEEE International Test Conference (ITC)
1997, pp. 278-285.
[151] J.Sweeney, "Testability implemented in the VAX 6000 model
400", Proceedings of the IEEE International Test Conference
(ITC) 1990, pp. 109-114.

210

References

[152] E.-S.A.Talkhan, A.M.H.Ahmed, A.E.Salama, "Microprocessors

Functional Testing Techniques", IEEE Transactions on ComputerAided Design, vol. 8, no. 3, pp. 316-318, March 1989.
[153] M.H.Tehranipour, Z.Navabi, S.M.Fakhraie, "An efficient BIST
method for testing of embedded SRAMs", Proceedings of the
IEEE International Symposium on Circuits and Systems (lSCAS)
2001, vol. 5, pp.73-76.
[154] M.H.Tehranipour, M.Nourani, S.M.Fakhraie, A.Afzali-Kusha,
"Systematic Test Pro gram Generation for SoC Testing Using
Embedded Processor", Proceedings of the International
Symposium on Circuits and Systems (lSCAS) 2003, pp. V541V544.
[155] N.Tendolkar, B.Bailey, A.Metayer, B.Svrcek, E.Wolf, E.Fiene,
M.Alexander, R.Woltenberg, R.Raina, "Test Methodology for
Motorola's High-Performance e500 Core Based on PowerPC
Instruction Set Architecture", Proceedings of the IEEE
International Test Conference (ITC) 2002, pp. 574-583.
[156] S.M.Thatte, J.A.Abraham, "A Methodology for Functional Level
testing of Microprocessors", Proceedings of the Fault-Tolerant
Computing Symposium (FTCS) 1978, pp. 90-95.
[157] S.M.Thatte, J.A.Abraham, "Test generation for general
microprocessors architectures", Proceedings of the Fault-Tolerant
Computing Symposium (FTCS) 1979, pp. 203-210.
[158] S.M.Thatte, J.A.Abraham, "Test Generation of Microprocessors",
IEEE Transactions on Computers, vol. C-29, pp. 429-441, June
1980.
[159] P.Thevenod-Fosse, R.David "Random testing of the Data
Processing Section of a Microprocessor", Proceedings ofthe FaultTolerant Computing Symposium (FTCS) 1981, pp. 275-280.
[160] P.Thevenod-fosse, R.David, "Random Testing of Control Section
of a Microprocessor", Proceedings of the Fault Tolerant
Computing Symposium (FTCS) 1983, pp. 366-373.
[161] C.Timoc, F.Stoot, K.Wickman, L.Hess, "Adaptative self-test for a
microprocessor", Proceedings of the IEEE International Test
Conference (lTC) 1983, pp. 701-703.
[162] Q.Tong, N.K.Jha, "Design of C-testable DCVS binary array
dividers", IEEE Journal of Solid-State Circuits, vol. 26, no. 2, pp.
134-141, February 1991.

References

211

[163] O.A.Torreiter, V.Baur, G.Goecke, K.Melocco, "Testing the

enterprise IBM System/390 multi processor", Proceedings ofthe
IEEE International Test Conference (ITC) 1997, pp. 115-123.
[164] C-H.Tsai, C-W.Wu. "Processor-programmable memory BIST for
bus-connected embedded memories", Proceedings ofthe Asia and
South Pacific Design Automation Conference (ASP-DAC) 2001,
pp. 325-330.
[165] R.S.Tupuri, J.A.Abraham, "A Novel Functional Test Generation
Method for Processors using Commercial ATPG", Proceedings of
the IEEE International Test Conference (ITC) 1997, pp. 743-752.
[166] R. S. Tupuri, A.Krishnamachary, J .A.Abraham, "Test Generation
for Gigahertz Processor V sing an Automatie Functional Constraint
Extractor", Proceedings of the ACM/IEEE Design Automation
Conference (DAC) 1999, pp. 647-652, June 1999.
[167] G.Vandling, "Modeling and Testing the Gekko Microprocessor,
An IBM PowerPC Derivative for Nintendo", Proceedings of the
IEEE International Test Conference (ITC) 2001, pp. 593-599.
[168] R. Velazco, H.Ziade, E.Kolokithas, "A mieroprocessor test
approach allowing fault localisation", Proceedings of the
International Test Conference (ITC) 1985, pp. 73 7-743.
[169] B.Williams, "LSI automatie test equipment applied to dynamic
mieroprocessor testing", New Electronies, vol. 7, no. 21, 1974, pp.
50-52.
[170] M.J.Y.Williams, J.B.Angell, "Enhancing Testability of LargeScale Integrated Circuits via Test Points and Additional Logic",
IEEE Transactions on Computers, vol. C-22, no. 1, pp. 46-60,
January 1973.
[171] W.Wolf, Madern VLSI Design: System-an-Chip Design, 3rd
Edition, Prentice Hall, January 2002.
[172] T.Wood, "The Test and Debug Features of the AMD-K7
Microprocessor", Proceedings of the IEEE International Test
Conference (ITC) 1999, pp. 130-136.
[173] G.Xenoulis, D.Gizopoulos, N.Kranitis, A.Paschalis, "Low-Cost
On-Line Software-Based Self-Testing of Embedded Processor
Cores", Proceedings ofthe 9th IEEE International On-Line Testing
Symposium (IOLTS) 2003, pp. 149-154.
[174] Xtensa Microprocessor Overview Handbook, Tensilica Inc.,
http://www.tensilica.com/xtensa_ overview_handbook.pdf, August
2001

212

References

[175] W.Zhao, C.Papachristou, "Testing DSP Cores Based on Self-Test

Programs", Proceedings ofthe IEEE Design Automation & Test in
Europe Conference (DATE) 1998, pp. 166-172.
[176] Y.Zorian, "A distributed BIST control scheme for complex VLSI
devices", Proceedings of the IEEE VLSI Test Symposium (VTS)
1993, pp. 4-9.

Index

ATE
At-speed testing
Automatie test equipment

See Automatie test equipment

31,57
2,28

Boundary sean
Built-in self-test

25,48
32,37,137, 190

C
CAD
Computer aided design
Core type
firm
hard
soft

See Computer aided design

1,8
10,99,186
10, 99, 125, 186
10,99, 157

Design-for-Testability
Deterministie testing
Diagnosis
Digital signal proeessor
Direet Memory Aeeess

23,43,57,88,142,186
39,60, 72, 101
45,48,62,191
9,12,18,64,72,178
92, 187

Index

214

Embedded processor
benchmark
Engineering effort

1,9, 15,41,64, 75, 82, 98, 113, 145, 185

18, 110, 145
61,77,90,98,113,148,189,196

Fault coverage
Functional testing

22
55,58,98, 165

Hardware description languages

Hardware-based self-testing

8, 10
32,41,53, 133, 147, 187

Instruction set architecture

3, 13,44,58, 73, 82, 98, 115, 137
International Technology Roadmap for Semiconductors
7,30
ISA
See Instruction set architecture
ITRS
See International Technology Roadmap for Semiconductors
L

LFSR
Linear feedback shift register
Low-cost testing

See Linear feedback shift register

39, 133
3, 19,81,90, 101, 115, 189

o
On-line testing
Overhead
hardware
performance

3,41,51,63,83, 189
26, 72
36,50,81, 188

p
Power consumption
Pre-computed test
Processor component
"chained" testing
"parallel" testing
computational
control
functional
hidden
interconnect
operation

1, 14,23,37,44,57,81, 101
39,61, 101, 137
149
152
108, 116, 127, 137, 159
111,120,141,165
108, 124, 137, 141, 163
112,121, 143
109,112, 138, 161
103, 121

Index

prioritization
size
storage
Processor model
Jam
Meister
oc54x
oc8051
Parwan
Plasma
RISC-MCU
Pseudorandom testing

215

100, 104, 113

115, 158
108, 116, 138, 159
171
168
178
173
158
160
176
40,47,60, 133

R
Register file
Register transfer level
RTL

109, 116, 138, 161

101, 155
See Register transfer level

S
SBST
Scan design
Self-test execution time
Self-test routine
optimization
size
style
Sequential fault
Software-based self-testing
embedded memory
phases
requirements
sbst-duration
SoC
test application time
Standard
Core test language (CTL)
IEEE 1500
System-on-Chip

See Software-based self-testing

25,37,89,142,186
77,97,158,189
42,90,158,189,196
148
48,128,131,164
106, 125
59,79,88
3, 18,41, 74, 81, 113, 153
97
102
87
96
185
93
25
25
1, 7, 11,21,30, 155, 185

Test application
Test cost

2,22, 72, 81,91

2,29,40,56,81,84,89

Index

216

Test data volume

Test generation
Test resuree partitioning

30, 193
22,30,50,73,93
46

V
VDSM
Verilog
Very deep sub-mieron
VHDL

See Very deep sub-micron

8, 13, 102, 158

1, 7, 31
8, 102, 158

y
Yield
inaeeuraey
overtesting

35,62,89,188
31
31,89

About the Authors

Dimitris Gizopoulos is an Assistant Professor at the Department of Infonnatics,
University of Piraeus, Greece. His research interests include processor testing,
design-for-testability, self-testing, on-line testing and fault tolerance of digital
circuits. Gizopoulos received the Computer Engineering degree from the University
of Patras, Greece and a PhD from the University of Athens, Greece. He is author of
more than sixty technical papers in transactions, journals, books and conferences
and co-inventor of a US patent. He is member of the editorial board ofIEEE Design
& Test of Computers Magazine, and guest editor of special issues in IEEE
publications. He is a member of the Steering, Organizing and Program Committees
of several test technology technical events, member of the Executive Committee of
the IEEE Computer Society Test Technology Technical Council (TTTC), a Senior
Member ofthe IEEE and a Golden Core Member ofthe IEEE Computer Society.
Antonis Paschalis is an Associate Professor at the Department of Infonnatics and
Telecommunications, University of Athens, Greece. Previously, he was Senior
Researcher at the Institute of Informatics and Telecommunications of the National
Research Centre "Demokritos" in Athens. He holds a B.Sc. degree in Physics, a
M.Sc. degree in Electronics and Computers, and a Ph.D. degree in Computers, all
from University of Athens. His current research interests are logic design and
architecture, VLSI testing, processor testing and hardware fault-tolerance. He has
published over 100 papers and holds a US patent. He is a member of the editorial
board of JETTA and has served the test community as vice chair of the
Communications Group of the IEEE Computer Society TTTC and participating in
several organizing and program committees of international events in the area of
design and test.
Yervant Zorian is the Vice President and Chief Scientist of Virage Logic Corp.
Previously he was the Chief Technology Advisor of LogicVision Inc. and a
Distinguished Member of Technical Staff at AT&T Bell Laboratories. Zorian
received the MSc degree in Computer Engineering from the University of Southern
California and a PhD in electrical engineering from McGill University. He also
holds an executive MBA from Warthon School of Business, University of
Pennsilvenia. He is the author of over 200 technical papers and three books, has
received several best paper awards and holds twelve U.S. patents. Zorian serves as
the IEEE Computer Society's Vice President for Technical Activities and the Editorin-Chief Emeritus of the IEEE Design & Test of Computers. He participates in
editorial advisory boards of IEEE Spectrum, and JETTA. He chaired the Test
Technology Technical Council of IEEE Computer Society, and founded IEEE
P1500 Standard Working Group. He is a Golden Core Member of IEEE Computer
Society, Honorary Doctor of National Academy of Sciences of Annenia, and a
Fellow ofthe IEEE.