Anum Imran

SOFTWARE IMPLEMENTATION AND PERFORMANCE OF UMTS

TURBO CODE
Masters of Science Thesis


















Examiners: Professor Jari Nurmi
Professor Mikko Valkama
Examiner and topic approved in the
Faculty of Computing and Electrical
Engineering Council meeting on 15
August, 2012.
II



ABSTRACT
TAMPERE UNIVERSITY OF TECHNOLOGY
Masters Degree Program in Information Technology
Faculty of Computing and Electrical Engineering
Imran, Anum: The UMTS turbo code implementation.
Master of Science Thesis: 74 pages, 22 appendix pages.
March 2013
Major: Communication Engineering
Examiners: Prof. Jari Nurmi and Prof. Mikko Valkama.
Keywords: Coding, turbo codes, UMTS, MATLAB simulation.

In the recent years, there has been a proliferation of wireless standards in television,
radio and mobile communications. As a result, compatibility issues have emerged in
wireless networks. The size, cost and competitiveness set limitations on implementing
systems compatible with multiple standards. This has motivated the concept of software
defined radio which can support different standards by reloading the software and
implementing computationally intensive parts on hardware, e.g., iterative codes. In a
typical communication system, all the processing is done in the digital domain. The
information is represented as a sequence of bits which is modulated on an analog
waveform and transmitted over the communication channel. Due to channel induced
impairments, the received signal may not be a true replica of the transmitted signal.
Thus, some error control is required which is achieved by the use of channel coding
schemes that protect the signal from the effects of channel and help to reduce the bit
error rate (BER) and improve reliability of information transmission.
Shannon gave the theoretical upper bound on the channel capacity for a given
bandwidth, data rate and signal-to-noise ratio in 1940s but practical codes were unable
to operate even close to the theoretical bound. Turbo codes were introduced in 1993
where a scheme was described that was able to operate very close to the Shannon limit.
Turbo codes are widely used in latest wireless standards e.g. UMTS and LTE. A basic
turbo encoder consists of two or more component encoders concatenated in parallel and
separated by an interleaver. The turbo decoder uses soft decision to decode the bits and
the decoding is done in an iterative fashion to increase reliability of the decision.
In this thesis, the turbo code for the UMTS standard is implemented in MATLAB. Four
versions of the Maximum Aposteriori Probability (MAP) algorithm are used in the
implementation. The simulation results show that the performance of the turbo code
improves by increasing the number of iterations. Also, better performance can be
achieved by increasing the frame size or the interleaver size and increasing the signal
power. Overall, the designing of turbo codes is a trade-off between energy efficiency,
bandwidth efficiency, complexity and error performance.
III



PREFACE
I would like to thank my supervisors, Professor Jari Nurmi and Professor Mikko
Valkama, for giving me the opportunity to research on this interesting topic for my
masters thesis. I wish to offer special thanks to Professor Jari Nurmi for his valuable
feedback, comments and encouragement during the course of the thesis.
I would also like to acknowledge the ideal and productive research environment
provided by Tampere University of Technology to work on my thesis.
In the end, I would like to thank my husband, my parents and especially my son for
being helpful, patient and supportive during the whole period of my studies.
Tampere, February 24, 2013.
Anum Imran
IV



LIST OF ABBREVIATIONS
A/D Analog-to-Digital
ADC Analog-to-Digital Converter
ACS Add Compare Select
APP A Posteriori Probability
ASIC Application Specific Integrated Circuit
ATSC Advanced Television Systems Committee
AWGN Additive White Gaussian Noise
BCH Bose Chaudhuri Hocquenghem
BER Bit Error Rate
BPSK Binary phase Shift keying
CML Coded Modulation Library
D/A Digital-to-Analog
DAC Digital-to-Analog Converter
DAB Digital Audio Broadcasting
DCS Digital Communication System
DSP Digital Signal Processor
DVB Digital Video Broadcasting
FPGA Field Programmable Gate Array
gcd Greatest Common Divisor
GSM Global System for Mobile Communication
IF Intermediate Frequency
LLR Log Likelihood Ratio
LTE Long Term Evolution
V



MAP Maximum A Posteriori
MLD Maximum Likelihood Decoder
OTA Over the Air
PCCC Parallel Concatenated Convolutional Code
PCM Pulse Code Modulation
pdf Probability density function
RF Radio Frequency
RS Code Reed Solomon Code
RSC Recursive systematic Convolutional
SDR Software Defined Radio
SISO Single Input Single Output
SoC System on Chip
SOVA Soft Output Viterbi Algorithm
UMTS Universal Mobile Telecommunication System
WLAN Wireless Local Area Network


VI



LIST OF FIGURES

Figure 1.1 Block diagram of a digital radio system. ......................................................... 3
Figure 1.2 Block diagram of a digital communication system (adapted from [6]). .......... 4
Figure 2.1 Bit error rate for iterations (1,218) of the rate turbo code presented in
1993 [10] . ......................................................................................................................... 8
Figure 3.1 Fundamental turbo code encoder. .................................................................. 10
Figure 3.2 Conventional convolutional encoder with r = 1/2 and K = 3. ....................... 11
Figure 3.3 RSC encoder obtained from the conventional convolution encoder with ..... 12
Figure 3.4 State diagram of rate r = 1/2 constraint length K = 3 convolutional encoder 13
Figure 3.5 Trellis diagram for the encoder in Figure 3.2 [13]. ....................................... 13
Figure 3.6 Trellis termination for the encoder in Figure 3.2 [14]. .................................. 14
Figure 3.7 The trellis termination strategy for RSC encoder. ......................................... 14
Figure 3.8 Non-recursive r=1/2 and K=2 convolutional encoder with input and output
sequences......................................................................................................................... 15
Figure 3.9 Recursive r = 1/2 and K = 2 convolutional encoder of Figure 3.8 with input
and output sequences....................................................................................................... 15
Figure 3.10 The state diagram of recursive and non recursive encoders. ....................... 16
Figure 3.11 Serial concatenated code.............................................................................. 17
Figure 3.12 Parallel concatenated code. .......................................................................... 18
Figure 3.13 The interleaving using random interleaver. ................................................. 20
Figure 3.14 The interleaving using circular shifting interleaver. .................................... 20
Figure 3.15 The graphical representation for calculating the forward state metric. ....... 25
Figure 3.16 The graphical representation for calculating reverse state metric. .............. 26
Figure 4.1 The UMTS turbo encoder [26]. ..................................................................... 29
Figure 4.2 The UMTS turbo decoder architecture [26]. ................................................. 36
Figure 4.3 The correction function fc used by log-MAP, constant-log-MAP and linear-
log-MAP algorithm. ........................................................................................................ 40
Figure 4.4 A trellis section for the RSC code used by the UMTS turbo code [26]. ....... 44
Figure 5.1 Control flow chart of UMTS turbo code MATLAB implementation. .......... 49
Figure 5.2 BER for frame size K = 40 UMTS turbo code over Rayleigh channel. ........ 50
Figure 5.3 BER for frame size K = 40 UMTS turbo code over an AWGN channel. ..... 51
Figure 5.4 BER for frame size K = 100 UMTS turbo code after 6 iterations. ................ 52
Figure 5.5 Performance comparison of UMTS turbo code after 10 decoder iterations. . 53
Figure 5.6 BER for frame size K=40 UMTS turbo code after 10 decoder iterations over
an AWGN channel. ......................................................................................................... 54
Figure 5.7 BER for frame size K=40 UMTS turbo code over a Rayleigh fading channel
after 10 decoder iterations. .............................................................................................. 56
Figure 5.8 BER for frame size K=40 UMTS turbo code after 10 decoder iterations. .... 57
Figure 5.9 The evolution of BER as a function of number of iterations. ........................ 58
Figure 5.10 EXIT chart showing mean and standard deviation of turbo code. .............. 59
Figure 5.11 Extrinsic information expressed as a function of a priori information. ...... 60
Figure 5.12 The Extrinsic information plot as a function of SNR. ................................. 61
Figure 5.13 Decoder trajectory for frame size K=40 for different number of iterations. 62
Figure 5.14 Decoder trajectory for frame size K=40 turbo code after 10 iterations. ...... 62
Figure 5.15 Decoder trajectory for K=320 turbo code after 20 decoder iterations. ........ 63
Figure 5.16 BER curves for frame size K 40 using log MAP obtained from CML
simulation. ....................................................................................................................... 64
VII



Figure 5.17 BER curves for frame size K 40 using log MAP obtained from simulated
Turbo code. ..................................................................................................................... 65
Figure 5.18 BER for frame size K = 40 using max-log MAP obtained from CML
simulation. ....................................................................................................................... 66
Figure 5.19 BER for frame size K = 40 using max-log MAP obtained from simulated
Turbo code. ..................................................................................................................... 66
Figure 5.20 BER for frame size K = 40 over Rayleigh channel obtained from CML
simulation. ....................................................................................................................... 67
Figure 5.21 BER for frame size K = 40 over Rayleigh channel obtained from simulated
turbo code. ....................................................................................................................... 67
Figure 5.22 BER for frame size K = 100 using log MAP from CML simulation. ......... 68
Figure 5.23 BER for frame size K = 100 using log MAP from simulated turbo code. .. 68
Figure 5.24 BER for frame size K = 320 over AWGN channel from CML simulation. 69
Figure 5.25 BER for frame size K = 320 using simulated Turbo code. .......................... 69


VIII



LIST OF TABLES

Table 3.1 Input and output sequences for convolutional encoders of Figure 3.8 and
Figure 3.9. ....................................................................................................................... 16
Table 3.2 Input and Output Sequences for Encoder in Figure 3.9. ................................. 18
Table 4.1 The list of prime number p and associated primitive root v. .......................... 32
Table 5.1 BER for K = 40 UMTS turbo code over Rayleigh channel. ........................... 51
Table 5.2 BER for frame size K = 100 UMTS turbo code. ............................................ 52
Table 5.3 BER for frame size K = 40, 100 and 320. ....................................................... 53
Table 5.4 BER for frame size K = 40 after 10 decoder iterations. .................................. 54
Table 5.5 BER for K = 40 over a Rayleigh fading channel after 10 decoder iterations. 55



IX



Table of Contents
1. INTRODUCTION .................................................................................................... 1
1.1 Software defined radio ....................................................................................... 1
1.2 Architecture of SDR ........................................................................................... 2
1.2.1 The RF section ............................................................................................ 3
1.2.2 The IF section .............................................................................................. 3
1.2.3 The baseband section .................................................................................. 3
1.3 Digital communication system ........................................................................... 4
1.4 Thesis overview .................................................................................................. 5
2. CHANNEL CODING ............................................................................................... 6
2.1 Introduction ........................................................................................................ 6
2.2 Types of channel codes ...................................................................................... 6
2.3 Need for better codes .......................................................................................... 7
3. TURBO CODES ....................................................................................................... 9
3.1 Turbo code encoder ............................................................................................ 9
3.1.1 Recursive Systematic Convolutional (RSC) encoder ............................... 10
3.1.2 Representation of turbo codes ................................................................... 12
3.1.3 Trellis termination ..................................................................................... 14
3.1.4 Recursive convolutional encoders vs. Non-recursive encoder ................. 15
3.1.5 Concatenation of codes ............................................................................. 17
3.1.6 Interleaver design ...................................................................................... 18
3.2 Turbo code decoder .......................................................................................... 21
3.2.1 Map decoder .............................................................................................. 21
4. THE UMTS TURBO CODE .................................................................................. 29
4.1 UMTS turbo encoder ........................................................................................ 29
4.1.1 Interleaver ................................................................................................. 31
4.2 Channel ............................................................................................................. 34
4.3 Decoder architecture ......................................................................................... 36
4.4 RSC decoder ..................................................................................................... 38
4.4.1 Max* operator ........................................................................................... 39
4.4.2 RSC decoder operation ............................................................................. 43
4.5 Stopping criteria for UMTS turbo decoders ..................................................... 46
5. MATLAB IMPLEMENTATION AND ANALYSIS OF RESULTS .................... 48
5.1 MATLAB implementation ............................................................................... 48
5.2 Turbo code error performance analysis ............................................................ 50
5.2.1 BER performance for frame size K=40 .................................................... 50
X



5.2.2 BER performance for frame size K=100 .................................................. 52
5.2.3 BER as a function of frame size K (Latency) ........................................... 53
5.2.4 BER performance over AWGN channel for max* algorithm ................... 54
5.2.5 BER performance over Rayleigh channel for max* algorithm ................. 55
5.2.6 Performance comparison over Rayleigh and AWGN channel ................. 56
5.2.7 BER as a function of number of iterations ................................................ 57
5.3 EXIT chart analysis .......................................................................................... 58
5.3.1 Extrinsic information as a function of a priori information ..................... 60
5.3.2 Decoder trajectory ..................................................................................... 61
5.4 Comparison with simulation results from MATLAB coded modulation library
(CML) ......................................................................................................................... 64
5.4.1 Turbo code performance comparison for frame size K = 40 using log
MAP algorithm ....................................................................................................... 64
5.4.2 Turbo code performance comparison for frame size K = 40 using max-log
MAP algorithm ....................................................................................................... 65
5.4.3 Turbo code performance comparison for frame size K = 40 over Rayleigh
channel 67
5.4.4 Turbo code performance comparison for frame size K = 100 .................. 68
5.4.5 Turbo code performance comparison for frame size K = 320 .................. 69
6. SUMMARY AND CONCLUSIONS ..................................................................... 70
6.1 Summary .......................................................................................................... 70
6.2 Conclusions ...................................................................................................... 71
REFERENCES ................................................................................................................ 72
Appendix: MATLAB code ............................................................................................ 75

1



1. INTRODUCTION
In the recent years, there has been a proliferation of wireless standards in television,
radio and mobile communications. As a result, compatibility issues have emerged in
wireless networks. Some of the most popular standards include wireless local area
network (WLAN), i.e., IEEE 802.11; 2.5G/3G mobile communications, i.e., Global
System for Mobile Communication (GSM), Universal Mobile Telecommunication
System (UMTS), Long Term Evolution (LTE); digital television standards, i.e., Digital
Video Broadcast (DVB), Advanced Television Systems Committee (ATSC) standards;
digital radio standards, i.e., Digital Audio Broadcasting (DAB). The inconsistency
between the wireless standards is causing a lot of problems to equipment vendors,
network operators and subscribers. Equipment vendors face difficulties in airing new
technologies because of short time-to-market requirements. The subscribers are forced
to change their handsets to upgrade themselves to the new standards whereas the
network operators face the dilemma during upgrade of network from one generation to
another due to large number of subscribers with handsets incompatible with the new
generation of standards [1].
The design of highly flexible digital communication has become an area of great
interest in the recent years as the inconsistencies between wireless standards is
inhibiting deployment of global roaming facilities and causing problems in introducing
new features and services. Also, there are demands of improved services and cheaper
rates from customers. The rapid evolution of communication technology is pushing the
service providers to keep pace with latest technology trends to survive in the market.
Traditional wireless systems with hard-coded capabilities are no longer able to keep step
with the rapid growth rate of communication technologies. Such devices, e.g. cell
phones and two way radios, etc, with fixed embedded software have significant
limitations and they become obsolete with the introduction of new services and
technologies in market, thus requiring expensive upgrades of total replacement.
Software defined radio (SDR) is one way to address these issues.
1.1 Software defined radio
In mobile wireless transmissions, size, cost and competitiveness set limitations on
implementing systems compatible with multiple standards. This has motivated the
concept of software defined radio which can support different standards by reloading
the software. SDR term refers to reconfigurability and adaptability of radio modules
using software. It is basically a collection of hardware and software technologies where
2



the radio functionalities are implemented as software modules running on generic
hardware platforms such as field programmable gate arrays (FPGA), Digital Signal
Processors (DSP) and Programmable System on Chip (SoC). Multiple software modules
implementing different standards can be present in the radio system. The radio system
can take up different personalities depending on the software module being used [2].
The SDR technology facilitates implementation of future-compliant systems. All the
concerned parties, i.e. subscriber, network operator and handset manufacturer, benefit
from this scheme. Both network infrastructure and subscriber handsets can be
implemented based on SDR concept. It makes the system very flexible and allows easy
migration of networks from one generation to another. In the same way subscriber
handset can adapt to various network protocols by choosing the suitable software
module. New software modules can be transferred to the subscriber handset using over
the air (OTA) upload [3]. Network operator can roll out new services at a much faster
pace by mass customization of subscriber handsets. Manufacturers can improve the
quality of their products and remove bugs by software upgrades.
SDR is thus a technology for building systems which support multiple services, multiple
standards, multiple bands, multiple modes and offer diverse services to its user.
Functional modules in a radio system such as modulator/demodulator, signal generator,
channel coding, multiplexing and link-layer protocols can be implemented based on
SDR concept. This helps in building reconfigurable software radio systems where
dynamic selection of parameters for each of the above-mentioned functional modules is
possible [4] [5].
1.2 Architecture of SDR
The digital radio system is composed of three main functional blocks: Radio Frequency
(RF) section, Intermediate Frequency (IF) section and baseband section [1]. To achieve
the flexibility and adaptability required by SDR, the boundary of digital processing
should be moved as close as possible to the antenna. Thus for an SDR system, the
analog-to-digital (A/C) and digital-to-analog (D/A) wideband conversion is done in the
IF section instead of baseband section as done for conventional radio. Figure 1.1 shows
the block diagram of a digital radio system [1].
3





1.2.1 The RF section
The RF section is responsible for transmitting and receiving the RF signal and
converting the RF signal into an IF signal. The RF section consists of antennas and
analog hardware modules. The RF front-end on the receive side performs RF
amplification and analog-down-conversion (from RF to IF). On the transmit side, the
RF section performs analog-up-conversion (from IF to RF) and RF power amplification.
The RF front-end is designed in such a way to reduce interference, multipath and noise.
1.2.2 The IF section
The IF section performs two tasks in either direction. On the receiver path, the analog-
to-digital converter (ADC) functional block performs A/D conversion followed by
Digital-down-conversion (Demodulation) by the DDC functional block. On the transmit
path, the D/A conversion is done by digital to analog converter (DAC) functional block
and Digital-up-conversion (modulation) by DUC functional block. Digital filtering and
sample rate conversion are often needed to interface the output of the ADC and DAC to
the processing hardware at the receiver and transmitter respectively.
1.2.3 The baseband section
In the baseband section baseband operations like channel coding, source coding,
equalization, encryption, decryption, modulation, demodulation, frequency hopping and
timing recovery are carried out. To allow reconfigurability and flexibility, application-
specific integrated circuits (ASICs) are replaced with reprogrammable or reconfigurable
modules in software defined radio [3]. In this thesis, the main area of focus is the
baseband section in which all the major tasks of a basic digital communication system
(DCS) are performed.
Rx


Tx

Source coding
Encryption
Decryption
Channel coding
Time Recovery

ADC DDC
DAC DUC
RF Section IF Section Baseband Section
Figure 1.1 Block diagram of a digital radio system.
4



1.3 Digital communication system
Recent progress in DCS design rests mainly upon software algorithms instead of
dedicated hardware. In a typical communication system, all the processing is done in the
digital domain. Digital transmission offers data processing options and flexibilities not
available with analog transmission. The principle feature of a DCS is that during a finite
interval of time, it sends a waveform from a finite set of possible waveforms whereas
for an analog communication system, the waveform can take any shape. The objective
of the receiver is to determine which waveform from a finite set of waveforms was
transmitted. A typical DCS illustrating signal flow and the signal processing steps is
shown in Figure 1.2 [6].
The upper blocks denote signal transformations at the transmitter end whereas the lower
blocks indicate the receiver end. The information source inputs analog information into
the system which is converted into bits, thus assuring compatibility between the
information source and signal processing within the DCS. Source coding is then done to
remove redundant information by quantization and compression. Encryption is done to
prevent unauthorized users from understanding messages and injecting false messages
and thus ensure data privacy and integrity. Next channel coding is performed by adding
redundant bits to the data for error detection and correction. It increases the reliability of
data at the expense of transmission bandwidth or decoder complexity. Multiplexing and
multiple access procedure combine signals that might have different characteristics of
different sources so that they can share the same communication source.

Figure 1.2 Block diagram of a digital communication system (adapted from [6]).

C
h
a
n
n
e
l
Format
Source
Encode
Encrypt
Channel
Encode
Multiplex Modulate
Format
Source
Decode
Decrypt
Channel
Decode
Demulti-
plex
Demodu-
late
5



The pulse modulate block usually includes filtering to minimize the transmission
bandwidth. It converts the binary symbols to a pulse-code-modulation (PCM)
waveform. For applications involving RF transmission, the next important step is band
pass modulation which translates the baseband waveform to a much higher frequency
using carrier wave. Frequency spreading produces a signal that is invulnerable to
interference and noise thus enhancing privacy. The signal processing steps that take
place in the transmitter are reversed in the receiver.
This thesis focuses on channel coding for SDR. The thesis gives an overview of
different type of channel codes and then narrows the research work to Turbo Codes.
Turbo Codes are widely used in modern communication systems. In this thesis, Turbo
Codes are implemented for UMTS system and their performance is analysed under
various channel conditions, frame sizes, signal power and iterations.
1.4 Thesis overview
The first part of the thesis deals with the study of basic turbo codes. A detailed study of
the encoder, the encoding algorithm, interleaver design and MAP decoding algorithm
has been done. Then the UMTS turbo code implementation has been discussed.
The thesis is organized into six chapters. Chapter 1 discusses the essence of Software
Defined Radio with emphasis on baseband module. The elements of the basic digital
communication are discussed. Chapter 2 discusses the basic channel coding schemes
and their importance. The discussion includes an overview of the performance of
various channel codes and the need of better channel coding schemes. Turbo codes are
discussed in detail in Chapter 3 including the encoder and decoder architecture, the
interleaver design and decoding algorithms. Chapter 4 concentrates on the turbo code
for UMTS standard. The specifications of the UMTS turbo encoder, channel and
decoder are discussed. The MATLAB implementation details and the simulation results
are given in Chapter 5. Chapter 6 concludes and summarizes the thesis.


6



2. CHANNEL CODING
Over the years, there has been a tremendous increase in the trends of digital
communication especially in the fields of cellular, satellite and computer
communications. In the digital communication system, the information is represented as
a sequence of bits. The processing is done in the digital domain. The binary data is then
modulated on an analog waveform and transmitted over the communication channel.
The signal is corrupted by the noise and interference introduced by the communication
channel. At the receiver end, the corrupted signal is demodulated and mapped back to
the binary bits. Due to channel induced impairments, the received signal may not be a
true replica of the transmitted signal; rather it is just an estimate of the transmitted
binary information. The bit error rate (BER) of the received signal depends on the noise
and interference of the communication channel. In a digital transmission system, error
control is achieved by the use of channel coding schemes. Channel coding schemes
protect the signal from the effects of channel noise and interference and ensure that the
received information is as close as possible to the transmitted information. They help to
reduce the BER and improve reliability of information transmission.
2.1 Introduction
Channel coding schemes involve the insertion of redundant bits into the data stream that
help to detect and correct bit errors in the received data stream. Due to the addition of
the redundant bits, there is a decrease in data rate. Thus the price paid for using channel
coding to reduce bit error rate is a reduction in data rate or an expansion in bandwidth.
2.2 Types of channel codes
There are two main types of channel codes, block codes and convolutional codes [7].
Block codes accept a block of k information bits, perform finite field arithmetic or
complex algebra, and produce a block of n code bits. These codes are represented as (n,
k) codes. The encoder for a block code is memory less, which means that the n digits in
each codeword depend only on each other and are independent of any information
contained in previous codeword. Some of the common block codes are Hamming codes
and Reed Solomon (RS) codes. RS codes are non-binary cyclic error correcting codes
that could detect and correct multiple random symbol errors. A Hamming code is a
linear error-correcting code which can detect up to two simultaneous bit errors, and
correct single-bit errors. For multiple error corrections, a generalization of Hamming
codes known as Bose Chaudhuri Hocquenghem (BCH) codes is used.
7



Block codes can either detect or correct errors. On the other hand, convolutional codes
are designed for real-time error correction. The code converts the entire input stream
into one single codeword. The encoded bit depends not only on the current bit but also
on the previous bit information. The decoding is traditionally done using the Viterbi
algorithm [8].
2.3 Need for better codes
The design of a channel code is always a trade-off between energy efficiency and
bandwidth efficiency [9]. Low rate codes having more redundant bits can usually
correct more errors. That means that the communication system can operate at lower
transmit power, tolerate more interference and noise and transmit at higher data rate.
Thus the code becomes more energy efficient. However, low rate codes also have a
large overhead and have more bandwidth consumption. Also, the decoding complexity
of the code also grows exponentially with code length. Thus, low rate codes set high
computational requirements to the conventional decoders.
There is a theoretical upper limit on the data transmission rate for a given bandwidth,
channel type, signal power and received noise power such that the data transmission is
error-free. The limit is called the channel capacity or the Shannon capacity. The formula
for additive white Gaussian noise (AWGN) channel is

(2.1)
Where W is the bandwidth, S is signal power, N is received noise power and R is the
data transmission rate. In practical transmission, no such thing as an ideal error free
channel exists. Instead, the bit error rate is brought to an arbitrarily small constant often
chosen at

or

. Shannon capacity sets a limit on the energy efficiency of the

code as it defines the lower bound for the amount of energy that be expended to convey
one bit of information given a fixed transmission rate, bandwidth and noise power.
Shannon gave his theory in 1940s but practical codes were unable to operate even close
to the theoretical bound. Until the beginning of 1990s, the gap between these theoretical
bounds and practical implementations was still at best about 3dB, i.e., the practical
codes required about twice as much energy as the theoretical predicted minimum.
Efforts were made to discover new codes that allow easier decoding. That led to the
introduction of concatenated codes. Simple codes were combined in parallel fashion so
that each part of the code can be decoded separately, thus decreasing decoder
complexity. Also, the decoders can exchange information with each other to increase
reliability. This lead to the introduction of near Shannon capacity error correcting code
known as turbo code.
8




Figure 2.1 Bit error rate for iterations (1,218) of the rate turbo code presented in
1993 [10] .
The first turbo code, based on convolutional encoding, was introduced in 1993 where a
scheme was described that uses a rate code over an AWGN channel and achieves a
bit error probability of

using Binary Phase Shift Keying (BPSK) modulation at an

of 0.7 dB [9]. The BER curves for different iterations of rate r=1/2 turbo code
are shown in Figure 2.1 [10].
The turbo codes consist of two or more component encoders separated by interleavers
so that each encoder uses an interleaved version of the same information sequence.
Contrary to conventional decoders which use hard decision to decode the bits,
concatenated codes like turbo codes do not limit themselves by passing hard decision
among the decoders. They rather exchange soft decision from the output of one decoder
to the input of the other decoder. The decoding is done in an iterative fashion to increase
reliability of the decision.
9



3. TURBO CODES
The generic form of a turbo encoder consists of two encoders separated by the
interleaver. The two encoders used are normally identical and the code is systematic,
i.e., the output contains the input bits as well. Turbo codes are linear codes. Linear
codes are codes for which the modulo sum of two valid code words is also a valid
codeword. A good linear code is one that has mostly high-weight code words. The
weight or Hamming weight of a codeword is the number of ones that it contains, e.g.,
the Hamming weight of codeword 000 and 101 is 0 and 2 respectively. High-weight
code words are desirable because it means that they are more distinct, and thus the
decoder will have an easier time distinguishing among them. While a few low-weight
code words can be tolerated, the relative frequency of their occurrence should be
minimized.
The choice of the interleaver is crucial in the code design. Interleaver is used to
scramble bits before being input to the second encoder. This makes the output of one
encoder different from the other encoder. Thus, even if one of the encoders occasionally
produces a low-weight, the probability of both the encoders producing a low-weight
output is extremely small. This improvement is known as interleaver gain. Another
purpose of interleaving is to make the outputs of the two encoders uncorrelated from
each other. Thus, the exchange of information between the two decoders while decoding
yields more reliability. There are different types of interleavers, e.g., row column
interleaver, helical interleaver, odd-even interleaver, etc.
To summarize, it can be said that turbo codes make use of three simple ideas:
Parallel concatenation of codes to allow simpler decoding
Interleaving to provide better weight distribution
Soft decoding to enhance decoder decisions and maximize the gain from
decoder interaction.
3.1 Turbo code encoder
The fundamental turbo code encoder is built using two identical recursive systematic
convolutional (RSC) encoders concatenated in parallel [9]. Convolutional codes are
usually described using two parameters: the code rate r and the constraint length K. The
code rate k/n, is expressed as a ratio of the number of bits into the convolutional encoder
k to the number of channel symbols output by the convolutional encoder n in a given
encoder cycle. The constraint length parameter K denotes the length of the
10



convolutional encoder, i.e., the maximum number of input bits that either output can
depend on. Constraint length K is given as
(3.1)
Where m is the maximum number of stages (memory size) in any shift register. The
shift registers store the state information of the convolutional encoder and the constraint
length relates the number of bits upon which the output depends.
In turbo code encoder, both the RSC encoders are of short constraint length in order to
avoid excessive decoding complexity. An RSC encoder is typically of rate r = 1/2 and
is termed a component encoder. The two component encoders are separated by an
interleaver. The output of the turbo encoder consists of the systematic input data and the
parity outputs from two constituent RSC encoders. The systematic outputs from the two
RSC encoders are not needed because they are identical to each other (although ordered
differently) and to the turbo code input. Thus the overall code rate becomes r = 1/3.
Figure 3.1 shows the fundamental turbo code encoder [9].
The first RSC encoder outputs the systematic output c
1
and recursive convolutional
encoded output sequence c
2
whereas the second RSC encoder discards its systematic
sequence and only outputs the recursive convolutional encoded sequence c
3
.
3.1.1 Recursive Systematic Convolutional (RSC) encoder
The RSC encoder is obtained from the conventional non-recursive non-systematic
convolutional encoder by feeding back one of its encoded outputs to its input. Figure
3.2 shows a conventional rate r = 1/3 convolutional encoder with constraint length K=3.


RSC Encoder 1
Interleaver
RSC Encoder 2
Input Systematic Output c
1

Output c
2

Output c
3

Figure 3.1 Fundamental turbo code encoder.
11




Figure 3.2 Conventional convolutional encoder with r =1/2 and K =3.
For each adder in the convolutional encoder, a generator polynomial is defined. It shows
the hardware connections of the shift register taps to the modulo-2 adders. A 1
represents a connection and a 0 represents no connection. The generator polynomials
for the above conventional convolution encoder are given as g
1
= [111] and g
2
= [101]
where the subscripts 1 and 2 denote the corresponding output terminals. The generator
matrix of the convolutional encoder is a k-by-n matrix. The element in the i
th
row and j
th

column indicates how the i
th
input contributes to the j
th
output. The generator matrix of
the above convolutional encoder is given in equation (3.2).

(3.2)
The conventional encoder can be transformed into an RSC encoder by feeding back the
first output to the input. The generator matrix of the encoder then becomes

(3.3)
Where 1 denotes the systematic output, g
2
denotes the feed forward output, and g
1
is
the feedback to the input of the RSC encoder. Figure 3.3 shows the resulting RSC
encoder.


D D
+
+
Output 1
Output 2
Input
12




Figure 3.3 RSC encoder obtained from the conventional convolution encoder with
r =1/2 and K =3.

It was suggested in [11] that good codes can be obtained by setting the feedback of the
RSC encoder to a primitive polynomial, because the primitive polynomial generates
maximum-length sequences which adds randomness to the turbo code.
3.1.2 Representation of turbo codes
Turbo codes are often viewed as finite state machines and are represented using state
diagrams and trellis diagrams. The contents in the memory elements of a coder
represent its state. For the rate r =1/2 convolutional encoder of Figure 3.2 with
constraint length K=3, the number of memory elements is L=K-1 =2. The input bit can
be either one or zero so the size M of the input alphabet 2. The number of possible states
of the coder is then M
L
= 2
2
=4 [12].
The operation of a convolutional coder can be represented by a state diagram which
consists of a set of nodes S
j
representing the possible states of the encoder where j {
0M
L
}. The nodes are connected by branches and are labelled by the input symbol and
the corresponding output symbol. Figure 3.4 shows the state diagram of the rate r = 1/2
convolutional encoder of Figure 3.2. The state diagram has four states S = { 00, 01, 10,
11}. The arrows represent the transition from one state to other and the labelling on
arrow gives the input bit (1 or 0) and the output of the encoder.
D D
Output 1
Output 2
Input
+
+
+
13




Figure 3.4 State diagram of rate r =1/2 constraint length K =3 convolutional
encoder
The state diagram can be expanded into the trellis diagram. All state transitions at each
time step are explicitly shown in the diagram to retain the time dimension. The trellis
diagram is very convenient for describing the behaviour of the corresponding decoder.
Figure 3.5 shows the trellis diagram for the encoder in Figure 3.2 [13].
The four possible states of the encoder are depicted as four rows of horizontal dots.
There is one column of four dots for the initial state of the encoder and one for each
time instant during the message. The solid lines connecting dots in the diagram
represent state transitions when the input bit is a one whereas the dotted lines represent
state transitions when the input bit is a zero. The labels on the branch represent the
output bits.

Figure 3.5 Trellis diagram for the encoder in Figure 3.2 [13].
14




Figure 3.6 Trellis termination for the encoder in Figure 3.2 [14].
Consider the trellis diagram for encoding 15 input bits as shown in Figure 3.6 [14]. The
trellis is in state 00 at the beginning. i.e., t=0. The trellis shows the next two states at
time t=1 with input bit 1 and 0. At the end of the 15 bit input, the encoder can be in any
of the states. However, for better performance of the decoder, both the initial and final
states of the encoder should be known. Thus, it is desirable to bring the encoders to a
known state after the entire input has been encoded. For this purpose, trellis termination
is performed by passing m=k-1 tail bits from the constituent encoders bringing them to
all zeros state. This brings the trellis to the initial all zero state as well. This is known as
trellis termination.
3.1.3 Trellis termination
Unlike conventional convolutional codes which always use a stream of zeros as tail bits,
the tail bits of a RSC depend on the state of the encoder when all the data bits have been
encoded [15]. Also because of the presence of interleaver between the two encoders, the
final states of the two component encoders will be different. Thus, the trellis termination
bits for the two encoders will also be different and an RSC cannot be brought to an all
zero state simply by passing a stream of zeros through it. However, this can be done by
using the feedback bit as the encoder input. This is done by using a switch at the input
as shown in Figure 3.7.

Figure 3.7 The trellis termination strategy for RSC encoder.
+
Input

Output 1
D D
Output 2
+
+
A
B
15



The switch is in position A while encoding the input sequence and is switched to
position B at the end of the input bit sequence for termination of trellis. The XOR of the
bit with itself will be zero (output of left most XOR) and thus the encoder will return to
all zero state after m=k-1 clock cycles by inputting the m=k-1 feedback bits generated
immediately after all the data bits have been encoded. Thus, a feedback is used for
terminating the trellis bringing the encoder to the all zero state.
3.1.4 Recursive convolutional encoders vs. Non-recursive encoder
The recursive convolutional encoders are better suited for turbo codes as compared to
non-recursive encoders because they tend to produce higher weight code words.
Consider a rate r = 1/2 constraint length K = 2 non-recursive convolutional encoder
with generator polynomial g
1
= [11] and g
2
= [10] as shown in Figure 3.8. The
corresponding RSC encoder with generator matrix G = [1, g
2
/ g
1
] is shown in Figure
3.9.

Figure 3.8 Non-recursive r=1/2 and K=2 convolutional encoder with input and output
sequences.

Figure 3.9 Recursive r =1/2 and K =2 convolutional encoder of Figure 3.8 with
input and output sequences.


Output 1
D
Input

+
Output 2
Output 2
D
Input

+
Output 1
16



Table 3.1 I nput and output sequences for convolutional encoders of Figure 3.8 and
Figure 3.9.
Encoder Type Input Output 1
c
1

Output 2
c
2

Weight
Non- RSC 1000 1000 1100 3
RSC 1000 1000 1111 5
Table 3.1 shows the output sequences corresponding to the same input sequence given
to the two encoders. The non-recursive convolutional encoder produces output c
1
=
[1100] and c
2
= [1000], thus it has a weight of 3. On the other hand, the recursive
convolution encoder outputs c
1
= [1000] and c
2
= [1111] which has a weight of 5. Thus, a
recursive convolutional encoder tends to produce higher weight code words as
compared to non-recursive encoder, resulting in better error performance. For turbo
codes, the main purpose of implementing RSC encoders as component encoders is to
utilize the recursive nature of the encoders and not the fact that the encoders are
systematic [16].
Figure 3.10 shows the state diagram of the non-recursive and recursive encoders.
Clearly, the state diagrams of the encoders are very similar. Also, the two encoders have
the same minimum free distance and can be described by the same trellis structure [9].
However, these two codes have different BERs as the BER depends on the input-output
correspondence of the encoders [17]. It has been shown that the BER for a recursive
convolutional code is lower than that of the corresponding non-recursive convolutional
code at low signal-to-noise ratios Eb/No [9][17].

(a) State diagram of the non-recursive encoder in Figure 3.8.

(b) State diagram of recursive encoder in Figure 3.9.
Figure 3.10 The state diagram of recursive and non recursive encoders.
17



3.1.5 Concatenation of codes
In coding theory, concatenated codes form a class of error correcting codes obtained by
combining an inner code with an outer code. The main purpose of concatenated codes is
to have larger block length codes with exponentially decreasing error probability [18].
There are two types of concatenation, namely serial and parallel concatenations. In
serial concatenation, the blocks are connected serially such that the output of one block
is input to the next block and so on.
An interleaver is often used between the encoders in both serial and parallel
concatenated codes to improve burst error correction capacity and increase the
randomness of the code. In case of serial concatenation, the output of encoder 1 is
interleaved before being input to encoder 2 whereas in parallel concatenation, the input
data is interleaved before being input to the second encoder. The serial concatenated
coding scheme is shown in Figure 3.11.
The total code rate for serial concatenation is the product of the code rates of the
constituent encoders and is given as [18].

(3.4)
For parallel concatenation, the blocks are connected in parallel. The parallel
concatenation scheme is shown in Figure 3.12. The total code rate for parallel
concatenation is calculated as

(3.5)

Figure 3.11 Serial concatenated code.
Encoder 1
r
1
= k
1
/ n
1

Encoder 2
r
2
= k
2
/ n
2

Modulator
Channel
Decoder 1 Decoder 2

Demodulator
Interleaver
Deinterleaver
18




Figure 3.12 Parallel concatenated code.
Turbo codes use the parallel concatenated encoding scheme in which the two RSC
encoders are connected in parallel separated by an interleaver. However, the turbo code
decoder is based on the serial concatenated decoding scheme. The performance of serial
concatenated decoders is better than the parallel concatenated decoding scheme as the
serial concatenation scheme has the ability to share information between the
concatenated decoders. On the other hand, the decoders for the parallel concatenation
scheme are primarily decoding independently.
3.1.6 Interleaver design
Turbo codes use an interleaver between two component encoders. The purpose of using
the interleaver is to provide randomness to the input sequences and increase the weight
of the code words.
Consider a constraint length K = 2 rate r = 1/2 convolutional encoder as shown earlier
in Figure 3.9. The input sequence x
i
produces output sequences c
1i
and c
2i
respectively.
The input sequences x
1
and x
2
are different permuted sequences of x
0
. Table 3.2 shows
the resulting code words and weights.
Table 3.2 I nput and Output Sequences for Encoder in Figure 3.9.
Input
Sequence
Output Sequence
c
1i

Output Sequence
c
2i

Codeword
Weight
x
0
1100 1100 1000 3
x
1
1010 1010 1100 4
x
2
1001 1001 1110 5
Modulator
Channel
Demultiplexer

Demodulator
Encoder 1
r
1
= k
1
/ n
1

Multiplexer
Interleaver
Encoder 2
r
2
= k
2
/ n
2

General
Concatenated
Decoding
(Depends on
Decoder Structure)
19



The values in Table 3.2 show that the codeword weight can be increased by permuting
the input sequence using an interleaver. The interleaver affects the performance of turbo
codes by directly affecting the distance properties of the code [19]. The BER of a turbo
code is improved significantly by avoiding low-weight code words. The choice of the
interleaver is important in the turbo code design. The optimal interleaver is the one that
produces fewest low weight coded sequences. The algorithm shows the basic interleaver
design concept:
Generate a random interleaver.
Generate all possible input information sequences.
Determine the resulting code words for all possible input information
sequences. Determine the weight of the code words to find the weight
distribution of the code.
From the collected data, determine the minimum codeword weight and the
number of code words with that weight.
Repeat the algorithm for a reasonable number of times. By comparison of the data, the
interleaver with the largest minimum codeword weight and lowest number of code
words with that weight is selected.
A number of interleavers are used in turbo codes [20]. A few of them are discussed
below in detail.
3.1.6.1 Matrix interleaver
The matrix interleaver is the most commonly used interleaver in communication
systems. It writes data in a matrix columnwise from top to bottom and left to right
without repeating or omitting any of the data bits. The data is then read out rowwise
from left to right and top to bottom. For example, if the interleaver uses a 2-by-3 matrix
to do its internal computations, then for an input of [A B C D E F], the interleaver
matrix is

The interleaved output is [A D B E C F]. At the deinterleaver the data is written in
columnwise fashion and read rowwise to obtain the original data sequence.
3.1.6.2 Random (Pseudo-Random) interleaver
The random interleaver uses a fixed random permutation and maps the input sequence
according to the permutation order. Consider an input sequence of length L=8 and the
random permutation pattern be [2 3 5 6 7 4 8 1]. If the input sequence is [A B C D E F
G H], the interleaved sequence will be [B C E F G D H A]. At the deinterleaver, the
reverse permutation pattern is used to deinterleave the data.
20




Figure 3.13 The interleaving using random interleaver.
3.1.6.3 Circular-shifting interleaver
The permutation p of the circular-shifting interleaver is defined by
(3.7)
Where i is the index, b is the step size, s is the offset and L is the size of the interleaver.
The step size b should be less than L and b should be relatively prime to L [21]. Figure
3.14 illustrates a circular-shifting interleaver with L=8, b=3, and s=0.

Figure 3.14 The interleaving using circular shifting interleaver.
It can be seen that the adjacent bit separation is either 3 or 5. This type of interleaver is
suited for permuting weight-2 input sequences with low codeword weights into weight-
2 input sequences with high codeword weights. However, because of the regularity of
adjacent bit separation (3 or 5 in the above case), it is difficult to permute input
sequences with weight higher than 2 using this interleaver [21].
3.1.6.4 Odd-Even interleaver design
The code rate of the turbo code can be changed by puncturing the output sequence.
Puncturing is the process of removing some of the parity bits after encoding. This has
the same effect as encoding with an error-correction code with a higher rate, or less
21



redundancy. However, with puncturing the same decoder can be used regardless of how
many bits have been punctured, thus puncturing considerably increases the flexibility of
the system without significantly increasing its complexity. A pre-defined pattern of
puncturing is used at the encoder end. The inverse operation, known as depuncturing, is
implemented by the decoder. By puncturing the two coded output sequences of a rate
r = 1/3 turbo code, a rate r = 1/2 turbo code can be obtained. However, by simple
puncturing the two coded output sequences there is a possibility that both the coded bits
corresponding to a systematic information bits be punctured. Due to this, the
information bit will have none of its coded bits in the sequence and vice versa. As a
result, the performance of the turbo decoder degrades in case of error for an unprotected
information bit. This can be avoided by using an odd-even interleaver design. First, the
bits are left uninterleaved and encoded, but only the odd-positioned coded bits are
stored. Then, the bits are scrambled and encoded, but now only the even-positioned
coded bits are stored. As a result, each information bit will now have exactly one of its
coded bits [22]. This guarantees an even protection of each bit of information, a uniform
distribution of the error correction capability and better performance of the code.
3.2 Turbo code decoder
In traditional decoding approach, demodulation is based on hard decision of the
received symbol. This discrete value is then passed on to the error control decoder. This
approach has some disadvantages. The decoder cannot make use of the certainty of
information available to it while decoding [9].
A similar but better rule is to take into account the a priori probabilities of the input. If
the +1 symbol has a probability of 0.9 and if the symbol falls in negative decision range,
the Maximum Likelihood Decoder (MLD) will decide it as -1. It does not take into
account the 0.9 probability of symbol being +1. A detection method that does take this
conditional probability into account is the Maximum a posteriori probability (MAP)
algorithm. It is also known as the minimum error rule [23].
Another algorithm used for turbo decoding is the soft output Viterbi algorithm (SOVA).
It uses Viterbi algorithm but with soft outputs instead of hard.
3.2.1 Map decoder
The process of MAP decoding includes the formation of a posteriori probabilities
(APP) of each information bit followed by choosing the data bit value corresponding to
MAP probability for that data bit. While decoding, the decoder receives as input a soft
(i.e. real) value of the signal. The decoder then outputs for each data bit an estimate
expressing the probability that the transmitted data bit was equal to one indicating the
reliability of the decision. In the case of turbo codes, there are two decoders for outputs
from both encoders. Both decoders provide estimates of the same set of data bits, but in
22



a different order due to the presence of interleaver. Information exchange is iterated a
number of times to enhance performance. During each iteration, the estimates are re-
evaluated by the decoders using information from the other decoder. This allows the
decoder to find the likelihood as to what information bit was transmitted at each time
instant. In the final stage, hard decisions will be made, i.e., each bit is assigned the value
1 or 0 [24].
The APPs are used to calculate the likelihood ratio by taking their ratio. The logarithmic
form of likelihood ratio is the log likelihood ratio (LLR).

(3.8)

(3.9)
Where

is the likelihood ratio,

is the LLR. The term

is described as
the joint probability that data

and state

, observed from time k =1 to N

and conditioned on the received corrupted binary sequence

which has been

transmitted through the channel, demodulated and presented to the decoder in soft
decision form.

(3.10)
The sequence

can be written as

(3.11)
Substituting equation (3.11) in equation (3.10), we get

(3.12)
The Bayess theorem gives the conditional probabilities as

(3.13)
Applying the results of equation (3.13) to equation (3.12) yield,

23

(3.14)
Equation (3.14) will now be explained in terms of the forward state metric, reverse state
metric and the branch metric.
3.2.1.1 The state metrics and the branch metric
FORWARD STATE METRIC
Consider the first term on the right side of equation (3.14)

(3.15)
For i =0,1 and time k and state

implies that the events before time k are not

influenced by events after time k, i.e., the future does not affect the past. Thus the two
terms

and

are irrelevant and

is independent of these terms.

However, since the encoder has memory, the state

is based on the past so this

term is relevant. This simplifies equation 3.15 as

(3.16)
The equation represents the forward state metric at time k as being a probability of the
past sequence that is dependent only on the current state

induced by the
sequence.
REVERSE STATE METRIC
The second term of equation (3.14) represents the reverse state metric

at time k and
state

(3.17)
Where f (i,m) is the next state given an input i and state m and

is the reverse
state metric at time k+1 and state f(i,m). Equation (3.17) represents the reverse state
metric

at future time k+1 as being a probability of the future sequence which

depends on the state at future time k+1. The future state f (i,m) in turn is a function of
the input bit and the state at current time k.



24



BRANCH METRIC
The third term in equation (3.14) is defined as the branch metric

at time k and state

m.

(3.18)
Substituting equation (3.16), (3.17), (3.18) in equation (3.14) yields

(3.19)
Equation (3.8) and (3.9) can be expressed in terms of equation (3.19) as

(3.20)

(3.21)
Equations (3.20) and (3.21) represent the likelihood ratio

and the LLR

of
the k
th
data bit respectively.
3.2.1.2 Calculating the forward state metric
The forward state metric in equation (3.16) can be expressed as the summation of all
possible transition probabilities from time k-1, as follows [24]

(3.22)
Applying Bayess theorem yields

(3.23)
As the knowledge about state m and the input j at time k-1 completely defines the path
resulting in state

, so the equation can be simplified to:

25




Figure 3.15 The graphical representation for calculating the forward state metric.

(3.24)
Where b(j, m) is the state going backward in time from state m, via the previous branch
corresponding to input j. Equation (3.24) can be simplified using equations (3.16) and
(3.18) .

(3.25)
Figure 3.15 represents equation (3.25) in terms of transitions from state k-1 to state k.
3.2.1.3 Calculating the reverse state metric
The reverse state metric was given in equation (3.17) as

Using this equation, the reverse state metric for state m at time k is given as:

(3.26)
Equation (3.26) can be expressed in terms of summation of all possible transition
probabilities to time k+1 as follows [24]

(3.27)

j=0
j=1

k-1

k


26






Using Bayess theorem

(3.28)
In the first term of equation (3.28), the terms

and

define the path given

an input j and state m resulting in state

. Thus, the term

can
be replaced with

in the second term of equation (3.28).

(3.29)
Using equation (3.17) and (3.18), equation (3.29) becomes

(3.30)
The equation represents the reverse state metric at time k as the weighted sum of state
metrics from time k+1 where the weighting consists of branch metrics associated with
transitions corresponding to data bits 0 and 1. Figure 3.16 shows the graphical
representation of calculating the reverse state metric.
3.2.1.4 Calculating the branch metric
The branch metric was given in equation (3.18) as

j=0
j=1

k k+1
Figure 3.16 The graphical representation for calculating reverse state metric.

27



The equation can be solved using Bayess rule as

(3.31)
Where

is the received noisy signal and it consists of both noisy data bits

and
noisy parity bits

.The noise affects data and parity bits independently so the current
state is independent of the current input and can therefore be any of the

states where
is the number of memory elements in the convolutional code system. Thus the
probability is given as

(3.32)
Solving equation (3.31) using equation (3.32) yields

(3.33)
The term

is the a priori probability of

and is given as

, thus equation
(3.33) becomes

(3.34)
Now, the probability density function (pdf)

of a random variable X
k
is related
to the probability of that random variable X
k
taking on the value x
k
as

(3.35)
For an AWGN channel with zero mean and variance

, the probability terms in

equation (3.34) are replaced with the pdf equivalents of equation (3.35) yielding

(3.36)
Where u
k
and v
k
represent the transmitted data bits and parity bits respectively. The
parameter

represents data that has no dependence on state m whereas the parameter

represents parity which depends on the state m since the code has memory.
Simplifying equation (3.36) gives [24]:

(3.37)
Substituting equation (3.37) in equation (3.8), we obtain:
28

(3.38a)

(3.38b)
Where

is the input a prior probability ratio

at time k and

is the output
extrinsic likelihood. The term

can be considered as the correction term that changes

the input priori knowledge about a data bit due to coding and is passed from one
decoder to the other during iterations. This improves the likelihood ratio for each data
bit and minimizes the decoding errors. Taking log of equation (3.38b) yields the final
LLR term.

(3.39)
The final soft number

is made up of three LLR terms; the a priori LLR

,
the channel measurement LLR

, and the extrinsic LLR

.
Implementing the MAP algorithm in terms of the likelihood ratios is very complex
because of the multiplication operations that are required. However, by implementing it
in the logarithmic domain, the complexity is greatly reduced.


29



4. THE UMTS TURBO CODE
Turbo codes are discussed in detail in chapter 3. This chapter will focus on the
implementation of Turbo encoder and decoder for UMTS systems. The architecture of
the encoder will be discussed in section 4.1. It explains the basic convolutional encoders
and interleaver used in UMTS turbo code in detail. After coding, the systematic bits and
code bits are arranged in a sequence, modulated using BPSK modulation and
transmitted over the channel. The channel models and the decoder architecture are
discussed in detail in section 4.2 and section 4.3 respectively.
4.1 UMTS turbo encoder
The UMTS turbo coder scheme is Parallel Concatenated Convolution Code (PCCC). It
comprises of two constraint length K = 4 (8 state) RSC encoders concatenated in
parallel. The overall code rate is approximately r = 1/3. Figure 4.1 shows a UMTS
turbo encoder [26].
The two convolutional encoders used in the Turbo code are identical with generator
polynomials [25].
g
0
(D)=1+D
2
+D
3
(4.1)
g
1
(D)=1+D+D
3
(4.2)

Figure 4.1 The UMTS turbo encoder [26].
30



Where g
0
and g
1
are the feedback and feed forward generator polynomials respectively.
The transfer function of each constituent convolutional encoder is:

(4.3)
The data bits are transmitted together with the parity bits generated by two
constituent convolutional encoders. Prior to encoding, both the convolutional encoders
are set to all zero state, i.e., each shift register is filled with zeros. The turbo encoder
consists of an internal interleaver which interleaves the input data bits X
1
, X
2
.X
K
to
X
1
, X
2
. X
K
which are then input to the second constituent encoder. Thus, the data
is encoded by the first encoder in the natural order and by the second encoder after
being interleaved. The systematic output of the second encoder is not used and thus the
output of the turbo coder is serialized combination of the systematic bits X
k
, parity bits
from the first (upper) encoder Z
k
and parity bits from the second encoder Z
k
for k =
1,2,K.

(4.4)
The size of the input data word may range from as few as 40 to as many as 5114
bits. If the interleaver size is equal to the input data size K the data is scrambled
according to the interleaving algorithm, otherwise dummy bits are added before
scrambling. After all the data bits K have been encoded, trellis termination is performed
by passing tail bits from the constituent encoders bringing them to all zeros state. To
achieve this, the switches in Figure 4.1 are moved in the down position. The input in
this case is shown by dashed lines (input=feedback bit). Because of the interleaver, the
states of both the constituent encoders will usually be different, so the tail bits will also
be different and need to be dealt separately.
As constraint length K=4 constituent convolutional encoders are used, so the
transmitted bit stream includes not only the tail bits {X
k+1
, X
k+2
, X
k+3
} corresponding to
the upper encoder but also tail bits corresponding to the lower encoder {X
k+1
, X
k+2
,
X
k+3
}. In addition to these six tail bits, six corresponding parity bits {Z
k+1
, Z
k+2
, Z
k+3
}
and {Z
k+1
, Z
k+2
, Z
k+3
} for the upper and lower encoder respectively are also
transmitted. First, the switch in the upper (first) encoder is brought to lower (flushing)
position and then the switch in the lower (second) encoder. The tail bits are then
transmitted at the end of the encoded data frame. The tail bits sequence is:

(4.5)
The total length of the encoded bit sequence now becomes 3K+12, 3K being the
coded data bits and 12 being the tail bits. The code rate of the encoder is thus
31



r = K / (3K+12). However, for large size of input K, the fractional loss in code rate due
to tail bits in negligible and thus, the code rate is approximated at 1/3.
4.1.1 Interleaver
Turbo code uses matrix interleaver [25]; [26]. Bits are input to a rectangular matrix in a
rowwise fashion, inter-row and intra-row permutations are performed and the bits are
then read out columnwise. The interleaver matrix can have 5, 10 or 20 rows and
columns between 8 and 256 (inclusive) depending upon the input data size. Zeros are
padded if the number of data bits is less than the interleaver size and are later pruned
after interleaving.
The input to the interleaver is the input data bit sequence X
1
, X
2
.. .X
K
where K
is the number of data bits. Interleaving is a complicated process and it comprises of a
number of steps which are described below.
4.1.1.1 Bits input to the rectangular matrix
1. Determine number of rows R :
The number of rows can be either 5, 10 or 20 depending on the size of input data. If the
input data size lies between 40 and 159, the number of rows R is five. If the input data
size K is either between 160 and 200 or 481 and 530, the number of rows of the
interleaver matrix R is ten. The interleaver matrix has twenty rows for size of input data
K lying anywhere outside the range specified.

(4.6)
The rows are numbered from 0 to R-1 from top to bottom.
2. Determine number of columns C :
For determining C, first a prime root p is calculated which is later used in the intra-row
permutations. If K lies in the range between 481 and 530 (inclusive), then both C and p
are assigned a value 53. Otherwise, the value of p is selected from Table 4.1 such that
the product of R and (p+1) is less than or equal to the number of input bits K. In this
case, the number of columns C can have any of the three values p-1, p and p+1
depending on the relation between number of data bits K, number of rows R and prime
root p. Like the rows, the number of columns is also numbered from 0 to C-1.

32

(4.7)
Table 4.1 The list of prime number p and associated primitive root v.
p v p v p v p v p v
7 3 47 5 101 2 157 5 223 3
11 2 53 2 103 5 163 2 227 2
13 2 59 2 107 2 167 5 229 6
17 3 61 2 109 6 173 2 233 3
19 2 67 2 113 3 179 2 239 7
23 5 71 7 127 3 181 2 241 7
29 2 73 5 131 2 191 19 251 6
31 3 79 3 137 3 193 5 257 3
37 2 83 2 139 2 197 2
41 6 89 3 149 2 199 3
43 3 97 5 151 6 211 2
3. Writing data in rectangular matrix:
The data is written in the interleaver matrix row wise. The size of the interleaver matrix
is R C. If the size of input stream is smaller than the interleaver size, dummy bits are
padded which are later pruned away after interleaving. The interleaver matrix is

Where K is the length of Input bit sequence and k =0, 1, 2..K.
4.1.1.2 Intra row and inter row permutations
The permutation process consists of seven steps which have been discussed in detail
below:
1. For the calculated value of prime root p, corresponding primitive toot v is selected
from Table 4.1.
33



2. The base sequence s for intra row permutation is constructed by using the primitive
root v, the prime root p and the previous value of base sequence. The value of base
sequence is initialized to one.

(4.8)
3. The prime number sequence q is found out using the value of prime root p which is
later used for the calculations of variables for intra-row permutation. The size of the
sequence is equal to the number of rows of the interleaver matrix. The value of q is
calculated using the greatest common divisor (gcd). The prime integer q
i
(i = 1, 2,
R-1) in the sequence is the least prime integer such that the gcd of q and (p-1)
equals one. The value of q should be greater than 6. Also, the next value of q should
be greater than the previous value in the sequence. The value of q is initialized to 1.

(4.9)
4. The inter-row permutation pattern is selected depending on the number of input bits
K and number of rows R of the interleaver matrix. The inter-row permutation pattern
simply changes the ordering of rows without changing the ordering of elements
within a row. If the number of rows is either five or ten, the inter-row permutation
pattern is simply the flipping of rows or in other words a reflection around the centre
row, e.g., if the number of rows is five, then the rows {0, 1, 2, 3, 4} are permuted to
rows {4, 3, 2, 1, 0} respectively. Same is the case when the number of rows is ten.
The rows {0, 1, 2, 3 .9} become rows {9, 8, 7, 6, 5, 4, 3, 2, 1, 0} respectively.
When the number of rows is twenty, there are two different cases based on the
number of input bits. The rows {0, 1, 2,19} become rows {20, 10, 15, 5, 1, 3, 6, 8,
13, 19, 17, 14, 18, 16, 4, 2, 7, 12, 9, 11} respectively, when the number of input bits
K satisfies either of the two conditions, i.e., 2281 K 2480 or 3161 K 3210.
For any other value of K, the permutation pattern is {20, 10, 15, 5, 1, 3, 6, 8, 13, 19,
11, 9, 14, 18, 4, 2, 17, 7, 16, 12}.
5. The prime integer sequence q
i
is permuted to get r
i
(for i = 0, 1, . R-1) which is
used in calculating the intra-row permutation patterns.

(4.10)
Where T(i) refers to the i
th
value of the inter-row permutation pattern calculated in
the previous step.
6. Next, the intra-row permutation is performed depending on the relation between
number of columns C and the prime root p. The intra-row permutation changes the
34



ordering of elements within a row without changing the ordering of the rows. The i
th

( i = 0, 1, .R-1) intra-row permutation is performed as:

(4.11)
Where

is the original bit position of the j

th
permuted bit of i
th
row. For
example, the values i=3, j=4 and

means that the bit at row 3 and column

4 now was originally at position 15 in the same row.
If the number of columns C is one less than the prime root p, then equation
(4.11) is used for permuting the sequence. If number of columns C is equal to the
prime root p, then the last column of i
th
row U
i
(p-1) is assigned a value zero.
If the number of columns C is one more than prime root p, then U
i
(p-1) is
assigned a value zero and U
i
(p) is assigned a value p. In this case, if the size of input
bits K is equal to the interleaver matrix size R C, then the first and last bits of the
last row are exchanged, i.e., U
R-1
(p) with U
R-1
(0).
7. The inter-row permutation is performed for the intra-row permuted rectangular
matrix based on the permutation pattern T introduced in step 4.
4.1.1.3 Bits output from the rectangular matrix with pruning
After intra-row and inter-row permutations, data is read out from the permuted R C
rectangular matrix in a columnwise fashion starting with the bit in row 0 of column 0
and ending with the bit at row R-1 of column C-1.

The output is pruned by deleting dummy bits that were added before interleaving.
For example, bits Y
k
that correspond to bits Y
k
for k > K were the dummy bits and are
removed from the output. The bit sequence obtained after interleaving and pruning is
X
k
for k K where X
1
corresponds to Y
k
with smallest index k after pruning. The
number of bits output from the turbo code interleaver is equal to the number of input
bits K and the number of pruned bits is (RC)-K.
4.2 Channel
After encoding, the entire n bit turbo codeword is assembled into a frame, modulated,
transmitted over the channel, and then decoded. The input to the modulator is assumed
to be U
k
where U
k
can be either systematic bit or parity bit and it can have a value either
35



0 or 1. The signal passing received after passing through the channel and demodulation
is Y
k
which can take on any value. Thus, U
k
is a hard value an Y
k
is a soft value.
The modulation scheme assumed for the proposed system was BPSK and the
channel model can be either AWGN or flat fading channel. For a channel with channel
gain a
k
and Gaussian noise n
k
, the output from the receivers matched filter is
Y
k
= a
k
S
k
+ n
k
where S
k
is the signal after passing through the matched filter. For
systematic data bits, S
k
= 2X
k
- 1. For upper and lower encoder parity bits S
k
= 2Z
k
- 1
and S
k
= 2Z
k
- 1 respectively. For an AWGN channel, channel gain a
k
equals one
whereas for a Rayleigh flat fading channel, it is a Rayleigh random variable. The
Gaussian noise n
k
has a variance
2
= 1/(2E
s
/ N
o
) where E
s
is the energy per code bit
and N
o
is the one sided noise power spectral density. For a coderate r = K/ (3K+12), the
noise variance in terms of energy per data bit E
b
becomes

(4.12)
The input to the decoder is in the log likelihood ratio LLR form to assure that the
effects of the channel .i.e. noise variance n
k
and channel gain a
k
have been properly
taken into account. The input to the decoder is in the form

(4.13)
By applying Bayess rule and assuming that P[S
k
=+1] = P[S
k
=-1]

(4.14)
where f
Y
(Y
k
|S
k
) is the conditional pdf of receiving Y
k
given code bit S
k
, which is
Gaussian with mean a
k
S
k
and variance
2
[27]. Thus, for decoding the bit, we need both
the code bit as well as knowledge of the statistics of the channel.
Substituting the expression for Gaussian pdf and simplifying yields:

(4.15)
36



The expression shows that the matched filter coefficients need to be scaled by a
factor 2a
k
/
2
in order to meet the input requirements of the decoder where R
k
shows the
received LLR for both data and parity bits. The notation R(X
k
) denotes the received LLR
corresponding to systematic data bits X
k
, R(Z
k
) is the LLR corresponding to the upper
encoder parity bit Z
k
and R(Z
k
) corresponds to the lower encoder parity bit Z
k
.
4.3 Decoder architecture
The architecture of the UMTS decoder is as shown in Figure 4.2. It operates in an
iterative manner as indicated by the presence of the feedback path [26] [27].
The decoder uses the received codeword along with the knowledge of the code
structure to compute the LLRs. Because of the presence of the interleaver at the encoder
end, the structure of the code becomes very complicated and it is not feasible to
compute the LLR by using a single probabilistic processor. It is rather feasible to break
the job of achieving a global LLR estimate in two iterations. As a result, each iteration
of the Turbo decoder consists of two half iterations, one for each constituent RSC
decoder. As indicated by the sequence of arrows, the timing of the decoder is such that
the RSC decoder 1 operates during the first half iteration and RSC decoder 2 operates
during the second half iteration.
Both the decoder structures compute LLR using a soft-input-soft-output processor
(SISO). Although the data bits used by the second decoder are interleaved, the two
SISO processors are producing LLR estimate of same set of data bits in different order.

Figure 4.2 The UMTS turbo decoder architecture [26].

)
-
-
RSC
Decoder
1
RSC
Decoder
2
I
n
t
e
r
l
e
a
v
e

D
e
i
n
t
e
r
l
e
a
v
e

+

+ +

37



The performance of the decoders can be greatly improved by sharing the LLR
estimates with each other. The first SISO decoder calculates the LLR and passes it to
the second decoder. The second decoder uses that value to calculate LLR which is fed
back to the first decoder. These back and forth exchange of information between the
processors make turbo decoder an iterative decoder. This also results in an improved
performance as after every iteration, the decoder is better able to estimate the data.
However as the number of iterations increase, the rate of performance improvements
decreases [27].
The value w(X
k
) for 1 k K, is the extrinsic information produced by RSC
decoder 2 and fed back to the input of RSC decoder 1. Prior to the first iterations, the
value of w(X
k
) is initialized to all zeros as decoder 2 has not yet decoded the data. After
each iteration, the value of w(X
k
) is updated to a new non zero value to reflect the
beliefs regarding the data and are propagated from decoder 2 to decoder 1 which uses
this information as the extrinsic information. As both RSC encoders at the transmitter
end are using independent tail bits so only the information regarding actual information
bits will be exchanged between the two encoders. For the value of k in range K+1 k
K+3, w(X
k
) is undefined or in other words zero after every iteration.
Decoder 1 must use the extrinsic information while decoding. Figure 4.2 gives a
detailed understanding of the steps being followed. The following sequence of steps is
followed:
1. The extrinsic information w(X
k
) is added to the received systematic LLR R(X
k
)
forming a new variable V
1
(X
k
).

(4.16)
As the extrinsic information w(X
k
) is non zero for 1 k K so the input to the RSC
decoder 1 is the received parity bits in LLR form R(Z
k
) and a combination of the
systematic data R(X
k
) and extrinsic information w(X
k
), i.e., V
1
(X
k
). For
K+1 k K+3 (tail bits) no extrinsic information is available, i.e., w(X
k
) equals
zero. In this case, the input to RSC decoder 1 is the received and scaled upper
encoders tail bits V
1
(X
k
) = R(X
k
) + 0 = R(X
k
), and the corresponding received and
scaled parity bits R(Z
k
).
2. The RSC decoder 1 uses this information to decode the data and outputs the LLR

1
(X
k
) for 1 k K as only the LLR of data bits is shared with the second encoder.

(4.17)
3. The extrinsic information w(X
k
) is subtracted from the first encoders LLR
1
(X
k
),
to obtain a new variable V
2
(X
k
) which is to be used by the second encoder. Similar to
V
1
(X
k
),

contains both the systematic channel LLR and the extrinsic

information produced by decoder 1.
38

(4.18)

4. The sequence

is interleaved to obtain

which is input to decoder 2.

(4.19)
This is done so that the sequence of bits in

matches that of the input to the

second decoder R

), the bits which were bits interleaved before convolutional

coding.
5. For 1 k K, the input to the decoder is

, i.e., the interleaved version of

and

, the channels LLR corresponding to second encoder parity bits.

The decoder uses this input to generate the LLR

(4.20)

6. The LLR

is deinterleaved to form

with the sequence of bits now

same as the original bits.

(4.21)
7. The sequence

is fed back to obtain the extrinsic information

by
subtracting

from it. This value of

will be used by decoder 1 during

the next iteration.

(4.22)
8. When the number of iterations are completed, a hard bit decision is taken using

to obtain the decoded bit

such that

equals one when

is
greater than zero and vice versa for 1 k K.

(4.23)
4.4 RSC decoder
The heart of the turbo decoder is the algorithm used to implement the RSC decoder. The
RSC decoders use a trellis diagram to represent all possible transitions between the
states along with their respective outputs. As the RSC encoder used by UMTS has three
shift registers, the number of distinct possible states is

, i.e., eight. The trellis diagram

39



not only shows the state for a particular bit k but also shows the permissible state
transitions leading to next state [26][27].
While decoding, each of the two RSC decoders sweep through the code trellis
twice, once in the forward and once in the backward direction. The sweep is
implemented using MAP algorithm which is a modified version of the Viterbi algorithm
to compute partial branch and path metrics. Although it is somewhat more complex than
the Viterbi algorithm but it has the advantage that it outputs the APPs of the information
and the channel [29]. The presented algorithm uses a version of the classical MAP
algorithm in the log domain.
4.4.1 Max* operator
The MAP algorithm poses technical difficulties because of a high number of additions
and multiplications. MAP algorithm if implemented in the log domain [30] [31]
significantly reduces the computational complexity. The log-MAP algorithm is based on
the Viterbi algorithm with two key modifications [31]:
Trellis should be swept in both forward and reverse directions.
Jacobi algorithm also known as max* operator should be used instead of the
add-compare-select (ACS) operation of the Viterbi algorithm.
The log-MAP algorithm is implemented twice during each half iteration; once in the
forward and once in reverse direction. It constitutes a dominant portion of the decoder
complexity and thus, the manner in which it is implemented is critical in determining
the performance and complexity of the decoder. Four versions based on four different
max* operations have been considered in the implementation [26].
4.4.1.1 Log MAP algorithm
The logarithm version of the MAP algorithm is the log MAP [31] algorithm. It has
reduced complexity as multiplication operations are transformed into additions.

(4.24)
Thus, max* operator in this case be calculated by taking the maximum of the two input
arguments x and y and then adding a correction function f
c
to it. The correction function
is simply a function of the absolute difference between the two arguments of the max*
operator. The correction function f
c
can be computed using log and exponential
functions or it can be pre-computed and stored in a look-up table to decrease
40




Figure 4.3 The correction function fc used by log-MAP, constant-log-MAP and
linear-log-MAP algorithm.
complexity. The log-MAP algorithm is the most complex of all the four algorithms but
it offers the best BER performance. The correction function f
c
used by log MAP
algorithm is illustrated in Figure 4.3 along with the correction functions used by
constant-log MAP and linear-log MAP algorithms.
4.4.1.2 Max-log MAP algorithm
In the max-log MAP algorithm, the log MAP algorithm is loosely approximated by
setting the correction function f
c
to zero in equation (4.24), i.e., it is not used at all [27].

(4.25)
The max-log MAP algorithm is the least complex of all the four algorithms but it has
still twice the complexity of the Viterbi algorithm. It can be implemented using a pair of
Viterbi algorithm, one that sweeps through the trellis in the forward direction and one in
the reverse direction [31]. It offers the worst BER performance of all the four variants of
MAP algorithm. It has however the additional benefit of being intolerant of imperfect
noise variance estimates while operating in an AWGN channel.
4.4.1.3 Constant-log MAP algorithm
The constant-log MAP algorithm was first introduced by W. J. Gross and P. G. Gulak in
1998 [33]. It uses a lookup table with only two entries of the correction function, thus
decreasing the complexity.
0 0.5 1 1.5 2 2.5 3 3.5 4
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
|y-x|
f
c
|
y
-
x
|


log-MAP
linear-log-MAP
constant-log-MAP
41

(4.26)

The values of T and C used in the implementation are 1.5 and 0.5 respectively [26] [34].
The performance and complexity of constant-log MAP algorithm lies between log MAP
and max-log MAP however, it is more susceptible to noise variance estimation errors
than log MAP.
4.4.1.4 Linear log MAP algorithm
In the constant-log MAP algorithm, the correction function was approximated by a set
of lookup tables. This implies the need of a high speed memory. To avoid the need of
high-speed memory, a linear approximation can be used. The linear-log MAP algorithm
is also based on using a linear approximation of the Jacobi algorithm [35].

(4.27)
For a floating point processor, the parameters a and T used by the linear approximation
are chosen to minimize the total squared error between the exact correction function and
its linear approximation [26].

(4.28)
To minimize the function, partial derivates w.r.t a and T are set to zero. The partial
derivated of equation (4.26) w.r.t a is

42

(4.29)

Where K
1
and K
2
are constants :

Now taking partial derivative of of equation (4.28) with respect to T

(4.30)

Adding 3/2 times equation (4.29) and 2T times equation (4.30) yields

43

(4.31)

By solving g(T) which is momotonically increasing function of T, we arrive at the
optimal value of T. An iterative approach is used by first choosing T
1
and T
2
such that
g(T
1
) < 0 and g(T
2
) > 0. The mid point between T
1
and T
2
is T
0
. If g(T
0
)< 0, T
1
is set
equal to T
0
, otherwise T
2
is set equal to T
0
. The iterations are repeate until T
1
ans T
2
become very close. In the implementaion, the upper limit of the summation is set to 30
as with this value an error of 10
-10
can be achieved. By iteratively solving equation
(4.29) for g(T) and setting the upper value of summation to 30, the value of T and a
come out to be 2.50681640022001 and -0.24904181891710 respectively. For the
simulation, the value of T and a are approximated to 2.5068 and -0.24904 respectively.
The complexity and performance of linear-log-MAP algorithm lies between that of log-
MAP and constant log-MAP algorithms however it converges much faster than the
constant-log-MAP algorithm.
4.4.2 RSC decoder operation
As discussed earlier, each of the two RSC decoders sweep through the trellis twice,
once in forward and once in reverse direction. However, it does not matter in which
direction the sweep is performed first, i.e., one can sweep in either the forward or the
reverse direction first. Also, the partial path metrics for only the entire first sweep must
be stored in memory. The partial path metrics for the entire second sweep need not be
stored as the LLR values can be computed during the second sweep. Thus, partial path
metrics for only the current state and the previous state must be stored during the second
sweep. It is recommended to perform the reverse sweep first and save partial path
metrices for each node in memory. Then the forward sweep is performed and LLR
estimates of data are produced. If the forward sweep is performed first, the LLR
estimates are produced during the reverse sweep and are thus in reverse order [26].
44




Figure 4.4 A trellis section for the RSC code used by the UMTS turbo code [26].
The trellis structure used by the RSC decoder is shown in Figure 4.4. Each state has two
branches leaving it, one corresponding to an input one and one for input zero. Solid
lines indicate data one and dotted lines indicate data zero. The branches indicate which
next state can be reached from a particular state. The branches are labelled with branch
metrics. Every distinct codeword follows a particular path in the trellis [27] [26]
The branch metric connecting state S
i
(previous state, on left) and state S
j
(present
state, on right) is denoted as

. The branch metric depends on the data bit as

well as the parity bit associated with the branch. The branch metric is given as:

(4.32)
The RSC encoder being rate r=1/2, only four distinct branch metrics are possible:

(4.33)
where

and

are the inputs to the RSC decoder. In Figure 4.2, for the first
(upper) RSC decoder,

equals

and for the second (lower) decoder,

equals

. Also for the lower decoder, the second input is the parity bits
corresponding to the second encoder, i.e.,

instead of

.
45



4.4.2.1 Backward recursion
The simulated decoder sweep through the trellis in the backward direction first. The
normalized partial path metrics at all the nodes in the trellis are saved in memory. These
normalized partial metrics denoted by are later used to calculate LLRs during the
forward recursion [26]. The partial path metrics at trellis stage k with 2 k K+3 and at
state S
i
for (0 i 7) is denoted as

. The backward recursion starts at stage K+2

and proceeds through the trellis in the backward direction until stage 2 is reached. The
recursion is initialized with

equal to zero and rest all states initialized to

negative infinity.

(4.34)
The partial path metrics for the current state k are found using the partial path metrics
for the next (previously calculated) k+1 state and the associated branch metrics

(4.35)
Where

indicates the unnormalized partial path metric. The states

and

indicate the two states at stage k+1 that are connected to state

in trellis at stage k. The

partial path metrics are normalized after the calculation of

to obtain normalized
partial path metrics.

(4.36)
The purpose of normalization is to reduce memory requirements. As after normalization

equals zero, so only the other seven normalized partial path metrics

for
1 i 7, need to be stored. As a result, there is a 12.5% saving in memory compared to
when no normalization was used.
4.4.2.2 Forward recursion and LLR calculation
Once the backward recursion is complete, the trellis is then swept in the forward
direction. Unlike the backward recursion where the partial path metrics for all states at
all stages need to be stored in memory, for the forward recursion only the partial path
metrics for two stages of the trellis need to be stored in memory, i.e., the current stage k
and the previous stage k-1. The forward partial path metric for state S
i
at trellis stage k is
denoted as

with the stages k ranging from 0 to K-1 and all possible states 0 i
7. The forward recursion is initialized by setting the value of forward partial path metric
for stage zero,
0
initialized to zero for state S
0
and negative infinity for all other seven
states.

(4.37)
46



The forward recursion begins with stage k=1 and the trellis is swept in the forward
direction until stage K is reached. The un-normalized partial path metrics

are
calculated as

(4.38)
Where

and

are the two states at stage k-1 that are connected to state

at stage k.
The partial path metrics

are normalized to

after the calculation of

(4.39)
Using these values of normalized partial path metrics

are computed for stage k

along with the LLR estimate for data bit X
k
. For LLR calculation, first the likelihood of
the branch connecting state S
i
at stage k-1 to state S
j
at stage k is calculated. It is denoted
by

(4.40)
The likelihood of data 1 or 0 is then the Jacobi algorithm of the likelihood of all the
branches corresponding to data 1 or zero.

(4.41)
The

operator is computed recursively over the over the likelihoods of all the data
one branches

or data zero branches

. Once

is calculated,

is no longer needed and may be discarded.

4.5 Stopping criteria for UMTS turbo decoders
The complexity of the decoder increases with the increase in the number of iterations
whereas the performance of the decoder improves. There is always a trade-off between
the performance of the decoder and its complexity. Standard turbo code
implementations use a fixed number of iterations specified before starting the coding
process. However, the computational complexity can be decreased by keeping the
number of iterations variable [37].
Different iteration control criterions are available for Turbo codes. The code keeps on
iterating until a certain rule or stopping condition is satisfied. The stopping condition is
computed based on the information available during decoding and it determines when
the iteration process will be terminated. At the time of termination, the data block is
either successfully decoded or it cannot be decoded at all. After every iteration, the
decoder checks the stopping condition. If true, the code terminates. Otherwise it
continues to the next iteration [37].
47



The main code termination criteria used in the simulation is the average mutual
information of LLR after each iteration. If the mutual information measurement has
worsened in the iteration as compared to the previous iteration, the code will terminate
early [38].
48



5. MATLAB IMPLEMENTATION AND ANALYSIS OF
RESULTS
The UMTS Turbo code was explained in detail in Chapter 4. For the thesis, the turbo
code was implemented in MATLAB.
5.1 MATLAB implementation
The input data frame size is between 40 and 5114 bits as is the size of the UMTS
interleaver. The turbo encoder consists of two main blocks, i.e., the recursive
convolutional encoder and the interleaver. The convolutional encoded data is arranged
in the sequence described in equations (4.4) and (4.5). The encoded data frame is BPSK
modulated and sent over the channel. The channel models used in the simulation are
AWGN and Rayleigh flat fading channel. After adding noise to the data, the LLR is
calculated as in equation (4.15). The control flow graph of the implemented code is
shown in Figure 5.1.
The decoder implementation is complex and computationally extensive. It includes
processing using a number of loops. A limit is set on the maximum number of bits to be
encoded and maximum allowable error for early termination of the code. The decoder
decodes iteratively checking the number of errors after every iteration. If the number of
errors is zero for an iteration, the code will not execute the next iteration to decrease
processing load. If the error exists, the average mutual information is checked. If there
is an improvement, the code goes into next iteration. If the mutual information average
fails to improve for a specified number of times, the decoder is terminated at that
iteration. It then goes back to the loop and checks if the number of data bits to be
encoded has reached the maximum limit or error is less than maximum allowable error.
If true, it repeats the steps described above. Otherwise, it increments the SNR and
performs the encoding and decoding [36] [37] [34].
49




Figure 5.1 Control flow chart of UMTS turbo code MATLAB implementation.

50



5.2 Turbo code error performance analysis
Turbo codes are capable of producing low error rates at astonishingly low SNRs if the
frame size K is kept large and permutations are selected randomly. Turbo codes are
iterative codes and the exchange of mutual information between the constituent
decoders results in better performance of the code. The more is the number of iterations;
more is the exchange of information between the constituent encoders. As a result, the
code can perform better. However, when the number of iterations is made too high,
there is not much improvement in performance. Thus, upper bounds need to be
specified for the number of iterations and SNR for optimal performance of the code.
As the turbo code decoding is computationally very intensive so most of the simulated
performance results are for small frame lengths.
5.2.1 BER performance for frame size K = 40
The turbo code program was simulated for frame size K = 40 over a Rayleigh fading
channel. To keep the simulation fast, the number of frames for each SNR was taken as
500. Thus, for a frame size 40, 20000 bits were sent at each SNR value to get the BER.
The SNR range was used from 0 to 5 dB. The number of decoder iterations was chosen
to be 10. The BER for the iterations is shown in Figure 5.2.

Figure 5.2 BER for frame size K =40 UMTS turbo code over Rayleigh channel.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=40 turbo code as a function of number of iterations over Rayleigh channel
SNR (in dB)
B
E
R


iter 1
iter 2
iter 3
iter 4
iter 5
iter 6
iter 7
iter 8
iter 9
iter 10
51



Table 5.1 BER for K =40 UMTS turbo code over Rayleigh channel.
Iteration BER Iteration BER
1 0.005 6 1.338 x 10
-4

2 4.886 x 10
-4
7 1.339 x 10
-4

3 2.285 x 10
-4
8 1.339 x 10
-4

4 1.655 x 10
-4
9 1.34 x 10
-4

5 1.418 x 10
-4
10 1.34 x 10
-4

The BER values at the end of each iteration are given in Table 5.1 BER for K = 40
UMTS turbo code over Rayleigh channel. The turbo code was able to achieve BER of a
0.5 x 10
-2
after first decoder iteration. The BER improved to 1.34 x 10
-4
after the 10
th

iteration. It can be seen that as the number of iteration increases, the BER performance
improves. However, the rate of improvement decreases. This is depicted by the
overlapping curves after 5
th
iterations. The BER after 5 iterations is 1.418 x 10
-4
. The
BER does not show significant improvement after 5
th
iteration. Thus, the number of
iterations should be kept such as to avoid extra computations.
The BER curve for frame size K = 40 over an AWGN channel is shown Figure 5.3 BER
for frame size K = 40 UMTS turbo code over an AWGN channel. The BER after the
first decoder iteration is 3.628 x 10
-4
. The BER decreases with the increase of iterations
and at the end of tenth iteration; the BER is 1.152 x 10
-5
.

Figure 5.3 BER for frame size K =40 UMTS turbo code over an AWGN channel.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=40 turbo code over AWGN channel using log MAP
SNR (in dB)
B
E
R


iter 1
iter 2
iter 3
iter 4
iter 5
iter 6
iter 7
iter 8
iter 9
iter 10
52



5.2.2 BER performance for frame size K = 100
Turbo code performance can be improved by increasing the frame size K. The code can
achieve higher BER with the increase of frame size. This is because the interleaver
permutes the data and the decoder is better able to decode the data. The simulation for
the turbo code was run with frame size K = 100 keeping the number of iterations 6. The
BER for the iterations is given in Table 5.2.
The BER curve for frame size K = 100 UMTS turbo code is shown in Figure 5.4. The
figure shows an improvement in the BER performance as compared to the frame size K
40. The BER decreases with the increase in the number of iterations. This behaviour is
similar to the case with frame size 40. However, the frame size K 100 is able to achieve
BER of almost 10
-4
at SNR of 2.5 dB. The frame size K 40 was able to achieve such
BER at SNR 5 dB. Thus, we can conclude that by increasing the frame size K, the code
can achieve the same BER at much lower SNR.
Table 5.2 BER for frame size K =100 UMTS turbo code.
Iteration BER Iteration BER
1 1.07 x 10
-2
4 1.472 x 10
-4

2 5.662 x 10
-4
5 1.245 x 10
-4

3 2.1518 x 10
-4
6 1.246 x 10
-4


Figure 5.4 BER for frame size K =100 UMTS turbo code after 6 decoder iterations.
0 0.5 1 1.5 2 2.5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=100 turbo code as a function of number of iterations
SNR (in dB)
B
E
R


iter 1
iter 2
iter 3
iter 4
iter 5
iter 6
53



5.2.3 BER as a function of frame size K (Latency)
The performance comparison of turbo code can be done by plotting the BER for
different frame sizes K as in Figure 5.5. The figure shows that by increasing the frame
size K, the BER performance of the code improves. As a result, lower BER can be
achieved by keeping the SNR constant. The BER values for frame size K 40, 100 and
320 is given in Table 5.3. It can be seen in the figure that for frame size K 320, BER of
1.75 x 10
-5
is achieved at SNR 1.5 dB. However, frame size K 100 and 40 achieve only
a BER of 0.76 x 10
-2
and 0.0212 respectively for the same SNR. Larger frame sizes
mean more latency as the encoding and decoding is done per frame. Thus, the
performance improvement is achieved at the cost of increased latency.
Table 5.3 BER for frame size K =40, 100 and 320.
Frame size K BER SNR (dB)
40 1.152 x 10
-5
4.5
100 7.745 x 10
-6
3
320 1.75 x 10
-5
1.5


Figure 5.5 Performance comparison of UMTS turbo code after 10 decoder iterations.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for UMTS Turbo code as a function of frame size K
SNR (in dB)
B
E
R


K=40
K=100
K=320
54



5.2.4 BER performance over AWGN channel for max* algorithm
The turbo code program was simulated for frame size K = 40 over an AWGN channel.
The number of frames for each SNR was taken as 500 and the code was executed for 10
decoder iterations. Decoding was done using all four variations of the max* algorithm.
i.e., log MAP, max-log MAP, constant-log MAP, linear-log MAP. For performance
comparison, the BER after 10 iterations is plotted in Figure 5.6. The BER values after
10 decoder iterations are given in Table 5.4.
Table 5.4 BER for frame size K =40 after 10 decoder iterations.
Algorithm BER
Log MAP 1.1519 x 10
-5

Max-log MAP 1.8967 x 10
-5

Constant-log MAP 0.7137 x 10
-5

Linear-log MAP 1.0358 x 10
-5



Figure 5.6 BER for frame size K=40 UMTS turbo code after 10 decoder iterations
over an AWGN channel.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=40 UMTS Turbo code after 10 decoder iterations over an AWGN channel
B
E
R
SNR (in dB)


log MAP
Max log MAP
clog MAP
Linear log MAP
55



The figure shows that the code was able to achieve BER as low as 10
-5
for the input bits.
The log MAP, linear log MAP and constant log MAP algorithms show indistinguishable
performance for the specified frame size K=40. However, in practice the performance
of the log MAP algorithm is the best of all. This also hold true in our simulation when
the frame size K is increased. However, the code has been simulated for K=40 as larger
frame sizes take much longer time for the simulations to run.
It can be seen that performance of the max-log MAP algorithm is significantly worst of
the four algorithms. However, it has the least complexity and it takes the smallest time
to simulate. Thus, there is a trade off between complexity and performance.
In the figure, it is observed that the performance of linear-log MAP and constant-log
MAP is sometimes slightly better than log MAP. This is an interesting and unexpected
phenomenon. The reason for this is that the two constituent decoders are optimal in
terms of minimizing the local BER but the overall turbo decoder does not guarantee
minimizing the global BER [26]. This is because of the random perturbations due to
approximation in computing the partial path metrics and LLRs in constant log MAP and
linear-log MAP algorithm. These perturbations are minor and the results of constant-log
MAP and linear-log MAP are very close to the log MAP algorithm.
5.2.5 BER performance over Rayleigh channel for max* algorithm
The turbo code program was simulated for frame size K=40 over a Rayleigh fading
channel. The specifications were kept similar to the previous simulation. The BER after
10 iterations is plotted as a function of SNR in Figure 5.7.
After 10 decoder iterations, the log MAP algorithm achieved BER of approximately
0.74 x 10
-3
at SNR 4.5 dB. In case of AWGN channel, the BER achieved was
approximately 10
-5
. As seen earlier, the performance of the turbo code over Rayleigh
channel for a specified SNR is worse as compared to AWGN channel. Also, the BER
for max-log MAP is the highest of all and that of log MAP is the lowest of all. Thus, the
log MAP algorithm shows the best performance at the cost of highest complexity and
max-log MAP shows the worst performance but it is also the least complex of all. The
complexity and performance of constant-log MAP and linear-log MAP are between the
log MAP and max-log MAP algorithms.
Table 5.5 BER for K =40 over a Rayleigh fading channel after 10 decoder iterations.
Algorithm Log MAP Max-log MAP C-log MAP Linear-log MAP
BER 0.74 x 10
-3
0.14 x 10
-2
0.82 x 10
-3
0.84 x 10
-3
56




Figure 5.7 BER for frame size K=40 UMTS turbo code over a Rayleigh fading
channel after 10 decoder iterations.
5.2.6 Performance comparison over Rayleigh and AWGN channel
For comparison, the BER curves for AWGN and Rayleigh fading channels are plotted
together. Figure 5.8 shows the comparison. The plots are shown for SNR range 0 to 4.5
dB. The code is able to achieve much lower BER at a specified SNR over an AWGN
channel as compared to Rayleigh fading channel. The BER of approximately 10
-2
was
achieved over Rayleigh fading channel at SNR 3 dB. For an AWGN channel, the same
BER was achieved at SNR 1.75 dB. Thus, higher SNR or more number of iterations is
required to achieve the same performance of AWGN channel over Rayleigh fading
channel.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-4
10
-3
10
-2
10
-1
10
0
BER of K=40 UMTS Turbo code after 10 decoder iterations over a Rayleigh fading channel
B
E
R
SNR (in dB)


log MAP
Max log MAP
clog MAP
Linear log MAP
57




Figure 5.8 BER for frame size K=40 UMTS turbo code after 10 decoder iterations.
5.2.7 BER as a function of number of iterations
The turbo code performance improves with the increase in the number of iterations.
However, after a specific number of iterations, the BER stays constant and does not
decrease any further. For achieving lower BER either the SNR or frame size K needs to
be changed. By increasing the SNR, the signal power increases. As a result the a priori
information of the data available improves and it results in better decoding of the data.
Thus, lower BER can be achieved. Similarly, when frame size K increases, the
interleaver adds more randomness to the data. This also results in better decoding and
lower BER.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER of K=40 UMTS Turbo code after 10 decoder iterations
B
E
R
SNR (in dB)


log MAP
Max log MAP
clog MAP
Linear log MAP
AWGN
channel
Rayleigh
channel
58




Figure 5.9 The evolution of BER as a function of number of iterations.
Figure 5.9 shows the evolution of BER as a function of number of iterations at SNR 0
dB. It can be seen that for frame size K 40, the BER is much higher as compared to the
frame size K 320. These plots show that the decoder is well implemented as the BER is
evolving after every iteration. Also the BER decreases exponentially and approach a
very low value as the number of iteration increases.
5.3 EXIT chart analysis
The BER chart for iterative decoding has three parts;
The region with negligible BER reduction.
The turbo cliff with persistent BER reduction over many iterations.
The BER floor where low BER can be reached after just few iterations.
Although bounding techniques are used to terminate the code and avoid extra
computations, still BER analysis is not enough for performance analysis of the code.
This is because the bounding techniques are not perfect and they have some limitations
[39].

1 2 3 4 5 6 7 8 9 10
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
BER as a function of number of iterations at SNR 0 dB using log-MAP algorithm
B
E
R
number of iterations


K=40
K=320
59



EXIT chart is used to visualize the flow of information transfer through the constituent
decoders. It is useful for performance analysis of turbo decoder in low BER regions
where the BER curve was staying constant. For doing so, the mutual information is
plotted between the extrinsic information at the output of the constituent decoder and
the input message with respect to the mutual information between the a priori
information at the input of the constituent decoder and the message.
To account for the iterative nature of the suboptimal decoding algorithm, both decoder
characteristics are plotted into a single figure. For doing so, the transfer characteristics
of the second decoder the axes are swapped. This diagram is referred to as EXIT chart.
The exchange of extrinsic information can be visualized as a decoding trajectory
between the two curves. The exit chart for frame size K 320 at SNR 0 dB is shown in
Figure 5.10. The simulation was run for 100 frames and then the mean extrinsic
information was plotted. The dotted lines show the standard deviation of the extrinsic
information about the mean.


Figure 5.10 EXI T chart showing mean and standard deviation of turbo code.

0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
EXIT chart for BPSK modulation in an AWGN channel
I
E
I
A


mean
standards deviation
60



5.3.1 Extrinsic information as a function of a priori information
Extrinsic information was plotted expressed as a function of a priori information for
different SNR values in Figure 5.11. It can be seen that as the SNR increases, the
extrinsic information IE increases. As a result, we obtain higher values of extrinsic
information IE even when IA is null. When the channel is good .i.e., SNR is high; the
data is less corrupted after passing through the channel. As the input to the decoder now
contains lesser errors, so even if the a priori information is 0, the decoder can still
output a message that makes sense. Thus, the decoder is able to perform better with
lower BER. On the contrary, when the channel is noisier, the decoder fails to estimate
the message without efficient a priori information.

Figure 5.11 Extrinsic information expressed as a function of a priori information.




0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
UMTS turbo code over an AWGN channel
I
E
IA


SNR -1 dB
SNR 0 dB
SNR 1 dB
SNR 2 dB
61




Figure 5.12 The Extrinsic information plot as a function of SNR.
The EXIT chart for SNR 0 and 1 dB is shown in Figure 5.12. The figure shows that for
SNR 1 dB the extrinsic information is higher as compared to a priori information. Thus,
as discussed earlier, decoder is better able to decode the data.
5.3.2 Decoder trajectory
The decoder trajectory gives the visualization of the exchange of extrinsic information
in the EXIT chart. The trajectory of the decoder for frame size K 40 turbo code is
shown in Figure 5.13. It can be seen that with 4 iterations, the trajectory of extrinsic
information and a priori information reaches the point approx (0.7, 0.65) on the graph.
As the number of iterations is increased from 4 to 10, the trajectory reaches the point
approx (0.85, 0.85). Increasing the decoder iteration further to 20, there is not much
improvement in the a priori and extrinsic information. An increase of 10 decoder
iterations result only in a very small improvement in the extrinsic information and the
trajectory is not moving any further to (1, 1) point. It is rather oscillating around the
same point. Thus, the iterations need to be chosen keeping in mind the trade-off
between complexity and system performance.
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
EXIT chart for frame size K=40 over AWGN channel
I
A
2
,

I
E
1
IA1, IE2


SNR 0 dB
SNR 1 dB
62




Figure 5.13 Decoder trajectory for frame size K=40 for different number of iterations.

Figure 5.14 Decoder trajectory for frame size K=40 turbo code after 10 iterations.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.4
0.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
Decoder trajectories for K=40 at for different iterations
IA1, IE2
I
A
2
,

I
E
1


4 iterations
10 iterations
20 iterations
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Decoder trajectories for K=40 at different SNRs
IA1, IE2
I
A
2
,

I
E
1


SNR 0 dB
SNR 1 dB
SNR 2 dB
SNR 3 dB
63



The decoder trajectories for frame size K 40 after 10 decoder iterations for different
SNRs is shown in Figure 5.14. As the SNR increases, the a priori information to
decoder 1 increases. This also results in increased extrinsic information. There is an
improvement in system performance shown by the improvement in extrinsic
information 2 at the end of the iterations. As the SNR increases from 0 to 2 dB, the
extrinsic information of decoder 2 at the end of the 10
th
iteration improves significantly
from 0.65 to 0.9. On further increase in SNR from 2 dB to 3 dB, the extrinsic
information shows very little improvement. The decoder trajectory for frame size K 320
after 20 decoder iteration shows similar results as shown in Figure 5.15.
It can be seen that with the increase in SNR from 0 to 2 dB, the extrinsic information
after 20 iterations improves from 0.65 to 0.85 approx. However, the later iterations
dont show much improvement in the performance. Thus, there is a trade-off between
energy efficiency, complexity and performance of the system.

Figure 5.15 Decoder trajectory for K=320 turbo code after 20 decoder iterations.


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Decoder trajectories for K=320 turbo code
IA1, IE2
I
A
2
,

I
E
1


SNR 0 dB
SNR 2 dB
64



5.4 Comparison with simulation results from MATLAB coded
modulation library (CML)
CML is an open source toolbox for the simulation of capacity approaching codes in
MATLAB. It is available at the iterative solutions website [40]. It mainly runs in
MATLAB but the computationally intensive parts are simulated in C. It thus uses c-mex
files for simulations. A large database of previously simulated results is also available at
the website in the form of MAT files. These results are the result of many hours of
simulation run time. New simulations can be done using the CmlSimulate command in
MATLAB. A number of scenarios are given in the UmtsScenarios.m specifying the
frame size, number of iterations, channel model, etc. For better performance
comparison, the CML library code and the MATLAB implementation done in the thesis
were run for similar input parameters. The results of the CML library are stored in MAT
file and are plotted using the CmlPlot command along with the scenario number.
5.4.1 Turbo code performance comparison for frame size K = 40
using log MAP algorithm
The turbo code is simulated for frame size K 40 using log MAP algorithm. The number
of iterations was set to 10. The SNR range was specified from 0 to 4.5 dB. The
maximum trials and the minimum BER was kept 10
6
and 10
-6
respectively for the CML
simulation. The turbo code implementation was also run for similar conditions. The
BER curves for 10 decoder iterations for CML simulation are shown in Figure 5.16.
The results from the implemented code are shown in Figure 5.17.

Figure 5.16 BER curves for frame size K 40 using log MAP obtained from CML
simulation.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Eb/No in dB
B
E
R


UMTS-TC, BPSK, AWGN, 40 bits, log-MAP
65




Figure 5.17 BER curves for frame size K 40 using log MAP obtained from simulated
Turbo code.
The CML simulation achieved a BER between 10
-3
and 10
-4
at 4.5 dB after first decoder
iteration. The BER decreases with the increase in number of iterations. At the fourth
iteration, the BER reaches 10
-5
but after that the BER change is very small for the next
iterations. After the tenth iteration, the BER is close to 10
-5
. The turbo code
implementation done in the thesis also shows similar results. Figure 5.17 shows that the
BER for turbo code after first iteration is 5 x 10
-3
at SNR 4.5 dB. The BER falls with the
increase in number of iterations and at the end of tenth iteration the BER is 1.34 x 10
-5
.
This is very close to the BER achieved by the CML simulation. Thus, the implemented
turbo code performance is very close to the CML simulation.
5.4.2 Turbo code performance comparison for frame size K = 40
using max-log MAP algorithm
The turbo code simulation was run for frame size K = 40 over an AWGN channel using
10 decoder iterations. The decoding algorithm used was max-log MAP. The simulation
results for the CML and the simulated code are shown in Figure 5.18 BER for frame
size K = 40 using max-log MAP obtained from CML simulation.Figure 5.18 and Figure
5.19 respectively. The CML simulation achieves a BER of 10
-5
at SNR 4.5 dB. The
turbo code simulation was able to achieve a BER of 0.1896 x 10
-4
. Once again, the
results of the simulation are very close to the CML simulated results.

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=40 turbo code over AWGN channel using log MAP
SNR (in dB)
B
E
R


iter 1
iter 2
iter 3
iter 4
iter 5
iter 6
iter 7
iter 8
iter 9
iter 10
66




Figure 5.18 BER for frame size K =40 using max-log MAP obtained from CML
simulation.

Figure 5.19 BER for frame size K =40 using max-log MAP obtained from simulated
Turbo code.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Eb/No in dB
B
E
R


UMTS-TC, BPSK, AWGN, 40 bits, max-log-MAP
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K = 40 turbo code over AWGN channel using max-log MAP algorithm
B
E
R
SNR (in dB)
67



5.4.3 Turbo code performance comparison for frame size K = 40
over Rayleigh channel
The CML simulation and MATLAB turbo code were run for frame size K = 40 over
Rayleigh fading channel. The decoding algorithm was log MAP and the number of
decoder iterations was chosen as 10. The results of the CML code and the turbo code
simulation are shown in Figure 5.20 and Figure 5.21 respectively. The BER for both the
simulations is approximately 10
-4
at SNR 5 dB.

Figure 5.20 BER for frame size K =40 over Rayleigh channel obtained from CML
simulation.

Figure 5.21 BER for frame size K =40 over Rayleigh channel obtained from
simulated turbo code.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Eb/No in dB
B
E
R


UMTS-TC, BPSK, Rayleigh, 40 bits, log-MAP
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K=40 UMTS Turbo code after 10 decoder iterations over a Rayleigh channel
B
E
R
SNR (in dB)
68



5.4.4 Turbo code performance comparison for frame size K = 100
The turbo code was simulated for frame size K = 100 over an AWGN channel using log
MAP algorithm. The number of decoder iterations was chosen as 10. The code was run
for maximum trials 10
6
and minimum BER 10
-6
. The results obtained from the CML
simulation are shown in Figure 5.22. The turbo code simulation was run with similar
parameters. The results are shown in Figure 5.23. The curves show similar trends and
the BER at different SNRs are also close enough. Some variations are due to the fact
that the input was generated randomly and was different for both the simulations.

Figure 5.22 BER for frame size K =100 using log MAP from CML simulation.

Figure 5.23 BER for frame size K =100 using log MAP from simulated turbo code.
0 0.5 1 1.5 2 2.5 3
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Eb/No in dB
B
E
R


UMTS-TC, BPSK, AWGN, 100 bits, log-MAP
0 0.5 1 1.5 2 2.5 3
10
-6
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
BER for K =100 turbo code using log MAP over AWGN channel
B
E
R
SNR (in dB)
69



5.4.5 Turbo code performance comparison for frame size K = 320
The turbo code was simulated for frame size K = 320 over an AWGN channel using 10
decoder iterations. The results obtained from the CML code and the simulated turbo
code are shown in Figure 5.24 and Figure 5.25 respectively. The BER for the CML
simulation was between 10
-4
and 10
-5
at SNR of 1.5 dB. The simulated turbo code
achieved a BER of 1.75 x 10
-5
. Thus, the performance of the simulated turbo code is
very close to the CML simulation.

Figure 5.24 BER for frame size K =320 over AWGN channel from CML simulation.

Figure 5.25 BER for frame size K =320 using simulated Turbo code.
0 0.5 1 1.5
10
-5
10
-4
10
-3
10
-2
10
-1
Eb/No in dB
B
E
R


UMTS-TC, BPSK, AWGN, 320 bits, log-MAP
0 0.5 1 1.5
10
-5
10
-4
10
-3
10
-2
10
-1
BER for K = 320 turbo code using log MAP over AWGN channel
B
E
R
SNR (in dB)
70



6. SUMMARY AND CONCLUSIONS
The thesis aims at studying and analysing the performance of turbo codes. In the thesis,
turbo codes used in the UMTS standard were implemented in MATLAB. This chapter
summarizes the work done and then draws the conclusion based on the obtained
simulation results.
6.1 Summary
Turbo codes are a class of high performance error correction codes with error
performance close to the Shannon limit. They make use of the parallel concatenated
recursive coding to provide enough error protection with reasonable complexity. The
basic turbo code encoder consists of two RSC encoders connected in parallel. The RSC
encoders produce higher weight code words as compared to the non RSC encoders and
are thus better suited for turbo codes. An interleaver is used between the two encoders
to introduce randomness as random codes achieve better channel capacity. The output
of the encoder consists of systematic bits as well as encoded bits from the two encoders.
The decoder takes into account the received codeword along with the knowledge of the
code structure to compute the LLRs. Decoding is performed in an iterative fashion with
the information exchanged among the constituent encoders for better performance. The
decoder first computes the branch metrics for all the nodes of the trellis. The next step is
to compute the reverse path metric for the received frame of data and the result is stored
in memory. After that, the forward path metrics are calculated. At the end the LLR of
the data for the whole frame is calculated and the information is passed to the other
decoder. After the end of the iterations, a hard decision is taken on the LLR to get the
received data bits.
The RSC decoder uses a modified Jacobi algorithm for sweeping through the trellis.
Four versions of the Jacobi algorithm .i.e., log MAP, max-log MAP, c-log MAP and
linear-log MAP, were considered in the implementation. Performance comparison was
also done based on the BER curves. The algorithm is implemented twice in every half
iteration, and thus constitutes a major portion of the decoding complexity.


71



6.2 Conclusions
Turbo codes are being used in 3G and 4G mobile telephony, WiMAX and satellite
communication systems. They show extraordinary performance at low SNRs
approaching the Shannon limit. However at higher SNRs, the BER curve begins to
flatten and hinders the ability to achieve extremely small bit error rates. The simulation
results showed that the performance of the turbo code depends on a number of
parameters including the frame size K, number of decoder iterations, the SNR and the
choice of log MAP algorithm.
Turbo codes are iterative codes and the performance improves with the increase in the
number of iterations. Typically 5 to 10 iterations produce most of the improvement. In
the simulated results, the BER dropped significantly from the first to the fifth iteration.
However, there was not much improvement from the fifth to the tenth iteration. Thus,
we can conclude that as the number of iterations increases, the rate of performance
improvement decreases. For achieving lower BER either the SNR or frame size K needs
to be changed.
When the signal power is increased, the SNR increases. As a result the a priori
information of the data available improves. As the input to the decoder now contains
lesser errors, the decoder can still output a message that makes sense thus, lower BER can
be achieved.
The performance of the turbo code profoundly depends on the randomness introduced
by the interleaver. An increased interleaver size K results in higher probability of higher
weight code words, thus resulting in better decoder performance. This was shown in the
simulation as the BER after 10 decoder iterations dropped from 10
-2
to 10
-5
by
increasing the frame size K from 40 to 320. Thus, by increasing the frame size K, lower
BER can be achieved by keeping the SNR constant but at the cost of increased latency.
Also, the code is able to achieve much lower BER at a specified SNR over an AWGN
channel as compared to Rayleigh fading channel.
The four variations of the classical log MAP algorithm were implemented. It was seen that
the performance of the max-log MAP algorithm was significantly worst of the four
algorithms and the log MAP algorithm had the best performance. However, the max-log
MAP has the least complexity and it takes the smallest time to simulate. Thus, there is a
trade off between complexity and performance. Although the log MAP algorithm is more
complex but it can achieve similar BER performance as that of the max-log MAP algorithm
using less iteration.
Thus, we can conclude that in designing turbo codes there is a trade-off between energy
efficiency, bandwidth efficiency, latency, complexity and error performance.

72



REFERENCES
[1] Muhammad Imran Anwar and Seppo Virtanen and Jouni Isoaho: a software
defined approach for common baseband processing, Journal of Systems
Architecture, Jan. 2008.
[2] S. Weiss, A. Shligersky, S. Abendroth, J. Reeve, L. Moreau, T. E. Dodgson
and D. Babb: A software defined radio testbed implementation, IEE
Colloquium on DSP Enable Radio, Livingston, Scotland, 2003, pp. 268-274.
[3] Walter H.W. Tuttlebee, Software defined radio: Facets of a developing
technology, IEEE Personal Communication, 6(2):38-44, April 1999.
[4] Z. Salcic, C. F. Mecklenbrauker, Software radio - architectural requirements,
research and development challenges, 8th International Conference on
Communication Systems, Vol 2, Nov. 2002.
[5] Petri Isomki, Nastooh Avessta, An Overview of Software Defined Radio
Technologies, TUCS Technical Report No 652, Turku, Finland, Dec. 2004.
[6] Bernard Sklar, Digital Communications: Fundamentals and Applications, Second
Edition, Prentice Hall (2001).
[7] Fu-hua Huang, Evaluation of Soft Output Decoding for Turbo Codes,
Virginia Polytechnic Institute and State University, Masters thesis, 1997.
[8] Core technologies, Viterbi algorithm for decoding of convolutional codes,
http://www.1-core.com/library/comm/viterbi/, cited May, 2012.
[9] Emilia Ksper, Turbo Codes, www.tkk.fi/~pat/coding/essays/turbo.pdf, cited
15 June, 2012.
[10] C. Berrou, A. Glavieux, and P. Thitimajshima., Near Shannon limit error
correcting coding and decoding: Turbo codes, IEEE International Conference
on Communications, Geneva, Switzerland, May 2003.
[11] G. Battail, C. Berrou, and A. Glavieux, Psuedo-Random recursive
convolutional coding for near-capacity performance, Globecom 1993, pp. 23-
27, Dec. 1993.
[12] Multu Koca, Turbo space time equalization of broadband wireless channels,
University of California, Davis, 2001.
[13] R.V. Rajakumar, Saswat Chakrabarti, Channel coding,
http://nptel.iitm.ac.in/courses/Webcourse-
contents/IIT%20Kharagpur/Digi%20Comm/pdf-m-6/m6l35.pdf, cited 22 Feb.
2012.
[14] Shu Lin and Daniel J. Costello, Error Control Coding: Fundamentals and
Applications, Prentice Hall, 1983.
[15] D. Divsalar and F. Pollara, Turbo codes for deep-space communications,
JPL TDA Progress Report 42-120, Feb. 15, 1995
[16] D. Divsalar and F. Pollara, Turbo codes for PCS applications, Proceedings
of ICC 1995, Seattle, Washington, pp. 54-59, June 1995.
[17] S. Benedetto and G. Montorsi, Unveiling turbo codes: some results on parallel
concatenated coding schemes, IEEE Transactions on Information Theory,
Vol. 42, No. 2, pp. 409-428, March 1996.
[18] J. G. Proakis, Digital Communications, 3rd ed., New York, McGraw-Hill,
1995.
73



[19] P. Jung and M. Nasshan, Performance evaluation of turbo codes for short
frame transmission systems, Electronics Letters, Vol. 30, No. 2, pp. 111-113,
Jan. 20, 1994.
[20] Shobha Rekh, S. Subha Rani, A. Shanmugam, Optimal choice of interleaver
for turbo codes, Academic Open Internet Journal, volume 15, 2005.
[21] S. Dolinar and D. Divsalar, Weight distributions for turbo codes using
random and non-random permutations, JPL TDA Progress Report 42- 122,
Aug. 15, 1995.
[22] A. S. Barbulescu and S. S. Pietrobon, Interleaver design for turbo codes,
Electronics Letters, Vol. 30, No. 25, pp. 2107-2108, Dec. 8, 1994.
[23] Charan Langton, Turbo coding and MAP decoding, intuitive guide to
principles of communications.
http://www.complextoreal.com/chapters/turbo1.pdf
[24] Bernard Sklar, Maximum aposteriori Decoding of turbo codes, online paper.
[25] European Telecommunications Standards Institute, Universal mobile
telecommunications system (UMTS): multiplexing and channel coding (FDD),
3GPP TS 125.212 version 8.6.0, Release 8, pp. 18-23, Sept. 2009.
[26] M. C. Valenti and J. Sun, The UMTS turbo code and an efficient decoder
implementation suitable for software-dened radios, International Journal of
Wireless Information Networks, Vol. 8, no. 4, pp. 203215, Oct. 2001.
[27] Teodor Iliev, A study of turbo codes for UMTS third generation cellular
standard, International Conference on Computer Systems and Technologies -
CompSysTech07, VI. 6.
[28] Matthew C. Valenti and Jian Sun, Handbook of RF and wireless technologies,
Turbo codes, Chapter 12, 375399. F. Dowla Editor, Newnes Press, 2004.
[29] L. R. Bahl, J. Cocke, F. Jelinek, and J. Raviv, Optimal decoding of linear
codes for minimizing symbol error rate, IEEE Trans. Inform. Theory, Vol. 20,
pp. 284287, Mar. 1974.
[30] P. Robertson, P. Hoeher, and E. Villebrun, Optimal and sub-optimal
maximum a posteriori algorithms suitable for turbo decoding, European
Transactions on Telecommunications (ETT), Vol. 8, pp. 119125, 1997.
[31] A. J.Viterbi, An intuitive justification and simplified implementation of the
MAP decoder for convolutional codes, IEEE Journal on Selected Areas of
Communication, Feb. 1998.
[32] A. J. Viterbi, Error bounds for convolutional codes and an asymptotically
optimum decoding algorithm, IEEE Transactions on Information Theory,
Vol. 13, pp. 260269, Apr. 1967.
[33] W. J. Gross and P. G. Gulak, Simplified MAP algorithm suitable for
implementation of turbo decoders, Electronics Letters, Vol. 34, pp. 1577
1578, Aug. 6, 1998.
[34] Matthew C. Valenti, "An efficient software radio implementation of the UMTS
turbo codec", IEEE 12th International Symposium on Personal, Indoor, and
Mobile Radio Communications (PIMRC '01), vol. 2, pp. G108-G113, San
Diego, USA, Sep. 2001.
[35] J. F. Cheng and T. Ottosson, Linearly approximated log-MAP algorithms for
turbo coding, IEEE Vehicular Technology Conference. (VTC), Houston, TX,
May 2000.
[36] Liang Li, Analysis of low power implementational issues of turbo-like codes
in body area networks, Progress report for continuation towards PhD,
University of Southampton, Nov. 2009.
74



[37] F. Gilbert, F. Kienle, N. Wehn, Low complexity stopping criteria for UMTS
turbo-decoders, IEEE Vehicular Technology Conference (VTC), April 2003.
[38] Lajos Hanzo, Robert G. Maunder, Jin Wang and Lie-Liang Yang, Near-
Capacity Variable-Length Coding: Regular and EXIT chart aided irregular
designs, Wiley, 2010.
[39] Stephan ten Brink, Convergence behavior of iteratively decoded parallel
concatenated codes, IEEE Transactions on Communications, vol. 49, Oct.
2001.
[40] Coded modulation library by iterative solutions,
http://www.iterativesolutions.com, cited 26 Feb, 2013.
75



Appendix: MATLAB code

Main program for UMTS turbo code

% The main Program for generating the UMTS Turbo code in MATLAB
clear all
close all
clc

K=100; % Specify frame size between 40 and
5114 bits.
iter_cnt = 10; % Specify the maximum number of
decoding iterations to perform.
symbols_per_frame=K*3+12; % Symbols in frame after encoding.
rate=K/symbols_per_frame; % Code rate of the Turbo code (ratio
of number of input bits to output bits)
frame_count_max=500; % Maximum number of frames to be
passed to the code
bit_count_max=frame_count_max*K; % Maximum bits at a
given SNR
error_max=10; % maximum number of errors (to be
used in condition later)
count_SNR=1; % counter for number of SNR values
results=[]; % The ouput results matrix defined

channel_type=1; % channel_type can be 1 or 2(for AWGN
and Rayleigh fading respectively)
sterric=1 % sterric can be 1 to 4 (for LogMAP, max-
logMAP, constant-logMAP, linear-logMAP)

max_fail_count = 3; % Specift the number of
iterations that should fail to improve the decoding before the
iterations are stopped early
err_out=[]; % error count matrix

% specifying the SNR starting value, step size and the range.
SNR_start=0;
SNR_delta=0.5;
SNR_stop=4;

BER=1;
data=randint(1,K); % generate random data
results=[zeros(1,iter_cnt+1) 1];
for SNR=SNR_start:SNR_delta:SNR_stop % for loop for SNR

error_counts=zeros(1,iter_cnt); % initialize error count
matrix
bit_count=0;

% Keep running the code until enough errors observed or enough
bits encoded.
while bit_count < bit_count_max || error_counts(iter_cnt) <
error_max

% function call for turbo code
76



[turbo_encoded]=turbo_enc(data);

%function call to bpsk modulate the turbo coded signal
[mod_signal]=bpsk_mod(turbo_encoded);

% Function call to pass through the channel
% channel type needs to be passed as function arguments
[symbol_likelihood]= channel ( mod_signal , SNR
,channel_type,symbols_per_frame,rate);
w_xk=zeros(1,K+3); % feedback from second component encoder
to first component encoder. It is initialized to zero before first
iteration.

avg_IA = 0; % initializing the mutual information
averaging value to zero
iteration_index = 1; % iteration index is initialized to 1
fail_count = 0; % number of times the code has failed

% Run the simulation while the number of failed iterations
% is less than max_fail_count and iteration index is less than
% maximum ietartions
while fail_count < max_fail_count && iteration_index <=
iter_cnt
% function call to turbo decode the data

[tilda2_xk,xk_hat,errors,w_xk]=turbo_dec(symbol_likelihood,1,data,w_xk
,sterric);
% The turbo code gives the error after each iteration.
% if all the errors have been corrected in the
% iteration, no need to carry on
if errors == 0
error_iter = 0;
fail_count = max_fail_count; % set fail count to max
to terminate the iterations
else
% Calculate the mutual information average to check
the performance of code in the iteration.
mutual_info = avg_mutual_inf(tilda2_xk);

% If IA is improved assign best_errors as errors
% otherwise increment the number of chances
if mutual_info > avg_IA
avg_IA = mutual_info; % if improved, assign new
value
error_iter = errors;
else
fail_count = fail_count + 1; % increment the
number of failures
end
end % end of if-else loop

% Accumulate the number of errors and bits that have been
simulated so far.
error_counts(iteration_index) =
error_counts(iteration_index) + error_iter;

% Increment the iteration_index counter for next iteration
iteration_index = iteration_index + 1;
end % end of while loop for iterations

77



% If the code is terminated early, assign the value of
% error of last iteration before termination to the next
% all iterations
% if error(iter)=0, assign zero to err(iter+1)...
% if error(iter)=10, assign 10 to err(iter+1).....
while iteration_index <= iter_cnt
error_counts(iteration_index) =
error_counts(iteration_index) + error_iter;

iteration_index = iteration_index + 1;
end

bit_count = bit_count + length(data); % add the number of
bits encoded so far

% Store the SNR and BERs in a matrix and display it.
results(count_SNR,1) = SNR;
results(count_SNR,2) = bit_count;
results(count_SNR,(1:iter_cnt)+2) = error_counts
% BER=results(count_SNR,iter_cnt+2)./results(count_SNR,2); %
calculate the BER
end % end of while loop for erro and bit count

count_SNR = count_SNR + 1; % Increment the SNR counter
end % end of for loop for SNR
results

% plotting the results
figure
for iteration_index = 1:iter_cnt
semilogy(results(:,1),results(:,iteration_index+2)./results(:,2));
hold on
end


% saving data in mat file
filename =
['AWGN_',num2str(K),'_',num2str(SNR_start),'_',num2str(SNR_stop),'.mat
'];
save(filename, 'results', '-MAT');

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function for turbo encoder implementation

function [output]=turbo_enc(xin)

% Function to turbo encode the data
% xin is the input data
% output is the coded data containing both data and flush bits

num=length(xin); % size of input data

[enc_out,flush1]=conv_enc(xin,1); % function call for convolutional
encoder

% interleaving
output=[]; % defining output matrix
num_blocks=ceil(num/5114); % determining number of data blocks
78



out_inter=[]; % interleaver output matrix
xin2=xin;
if num_blocks>1
for cnt_block=1:num_blocks-1
% interlevaing data block by block
[y]=interl(xin ((((cnt_block-1) *5114) +1):cnt_block*5114));
% function call fot interleaver
[enc_out_inter,flush]=conv_enc(y,0);
% convolutional encoding the interleaved data
for cnt=1:length(y)
output=[output xin2(cnt) enc_out(cnt) enc_out_inter(cnt)];
% storing encoded data in output matrix Xk,Zk ,Zk'
end
xin2=xin2(cnt+1:end);
enc_out=enc_out(cnt+1:end);
end % end of cnt_block loop
end % end of num_block loop

% interleaving the last block of data
input_interl = xin((( (num_blocks-1) *5114) +1) :end); % input to
interleaver
[y,interleaved]=interl(input_interl); % function
call to interleaver
[enc_out_inter,flush2]=conv_enc(y,1); % convolution
encode the interleaved data

% storing encoded data in output matrix Xk, Zk, Zk'
for cnt=1:length(y)
output=[output xin2(cnt) enc_out(cnt) enc_out_inter(cnt)];
end % end of for


% appending flush bits Xk+1, Zk+1, Xk+2, Zk+2, Xk+3, Zk+3,
% X'k+1,Z'k+1,X'k+2, Z'k+2, X'k+3, Z'k+3

out=[];
m=length(xin); % length of input data
% loop for writing flush bits from upper encoder Xk+1, Zk+1, Xk+2,
Zk+2, Xk+3, Zk+3,
for cnt1=1:length(flush1)
out=[out enc_out(m+cnt1) flush1(cnt1)];
end % end of for

% loop for writing flush bits from lower encoder X'k+1,Z'k+1,X'k+2,
Z'k+2, X'k+3, Z'k+3
for cnt1=1:length(flush2)
out=[out enc_out_inter(m+cnt1) flush2(cnt1)];
end % end of for

output=[output out];

end % end of fuction

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

79




Function for constituent convolutional encoder

function [enc_out,flush]=conv_enc(xin)

% constraint length 4 convolutional encoder
% Input data is xin
% output is encoded data enc_out
% Flush bits are flush

% intilaize the states to zero
data_in=[0 0 0];
s1=0;
s2=0;
s3=0;

enc_out=[]; % defining the output matrix of encoded data

for cnt=1:length(xin)
input=xin(cnt); % input bit to convolutional encoder

out_back=xor(s2,s3); % feedback data

% new states of the convolutional encoder
s3_new=s2;
s2_new=s1;
s1_new=xor(input,out_back);

% output encoded bit
out1=xor(xor(s1_new,s1),s3);

% storing data in output array
enc_out=[enc_out out1];

% update the states
s1=s1_new;
s2=s2_new;
s3=s3_new;
end % end of for loop


% trellis termination
% switch go in down position
flush=[]; % initializing the flush matrix
for cnt=1:3 % loop for number of flush bits
out_back=xor(s2,s3); % feedback data

% new states
s3_new=s2;
s2_new=s1;
s1_new=xor(out_back,out_back);

% output encoded bit

out1=xor(xor(s1_new,s1),s3); %output
enc_out=[enc_out out1]; % storing encoded bit in output array
flush=[flush out_back]; % flush bit

80



% update ths states
s1=s1_new;
s2=s2_new;
s3=s3_new;
end %end of foor loop for flush bits

end % end of function

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function for interleaver main

function[y,interleaved]=interl(Input_data) % input data
% 40 <=k<=5114
% function for interleaving the data
% Input is Input_data
% Output is interleaved data y and interleaving matrix interleaved

K =length(Input_data); % length of input data
[interleaved]=interleave(K); % function call for generating
interleaved matrix array of specific size accoring to size of input
data

if length(interleaved) > K
% If length of input data is smaller than interleaver matrix, pad
zeros
pad_bits=length(interleaved)-K;
Input_data=[Input_data zeros(1,pad_bits)];
end % end of if loop

% Interleaving the data
for cnt=1:length(interleaved)
y(cnt)=Input_data(interleaved(cnt));
end % end of for loop

y=y(1:K); % removing padded bits from the interleaved data
end %end of function

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function to generate the interleaver matrix

function [interleaved,R,U,C,T]=interleave(K)

% function for generating the interleaver matrix based on the size of
data
% input is size of the data K
% output is interleaver matrix interleaved
% number of rows of the matrix R
% number of columns of matrix C
% Factor determing the intra row permutation U
% Factor determining the inter row permutation T

% determine number of rows of the interleaver matrix

if (K >= 40 ) && ( K<= 159)
R=5;
81



elseif ((K >= 160 ) && ( K<= 200) ) || ((K >= 481) && ( K<= 530) )
R=10;
else
R=20;
end % end of if loop

% determine number to be used in intra permutation p and number of
columns C
prim_num=primes(260);
prim_tab=[3 2 2 3 2 5 2 3 2 6 3 5 2 2 2 2 7 5 3 2 3 5 2 5 2 6 3 3 2 3
2 2 6 5 2 5 2 2 2 19 5 2 3 2 3 2 6 3 7 7 6 3];
prim_num=prim_num(4:end);
cnt=1;

if ((K >= 481) && ( K<= 530) )
p=53;
C=p;
V=2;
else
p=prim_num(cnt);
while K > R *(p+1)
cnt=cnt+1;
p=prim_num(cnt);



end % enf of while loop
V=prim_tab(cnt);

if K<= R*(p-1)
C=p-1;
elseif ((R*(p-1)) < K ) && (K <=( R*p))
C=p;
elseif K > R*p
C=p+1;
end % end of inner ifelse loop

end % end of outer if else loop

% generating array of data from 1 to size of matrix
x=1:(R*C);

% writing X in matrix form row wise
X=[];
cnt_r=0;
while cnt_r < R
X=[X ; x((cnt_r*C)+1 :((cnt_r+1)*C))];
cnt_r=cnt_r+1;
end % end od while loop

% inter row and intrarow permutations
% step 1 primitive root v
% step 2 construct base sequence for intra row permutation
s=1;
for cnt_s=2:p-1
s=[s mod((V*s(cnt_s-1)),p)];
end

% step 3 assign q0
82



q=1; % qo=1
cnt_prim=1;
prim_gen=primes(1000);
prim_gen=prim_gen(4:end);
prim=prim_gen(cnt_prim);
cnt_q=2;
while cnt_q<=R
mid_q=gcd(prim,p-1);
if mid_q==1
q(cnt_q)=prim;
cnt_q=cnt_q+1;
end
cnt_prim=cnt_prim+1;
prim=prim_gen(cnt_prim);
end

% step 4 permute q to get r
if R==5
T=fliplr([1:5]);
elseif R==10
T=fliplr([1:10]);
elseif (R==20) && (((2281 <= K) && (K <= 2480)) || ((3161 <= K) && (K
<= 3210)))
T=[20 10 15 5 1 3 6 8 13 19 17 14 18 16 4 2 7 12 9 11];
else
T=[20 10 15 5 1 3 6 8 13 19 11 9 14 18 4 2 17 7
16 12];
end

r=zeros(1,R);
for cnt_i=1:R
r(T(cnt_i))=q(cnt_i);
end

% step 5 intra row permutation
U=zeros(R,C);
for cnt_i=1:R

if C == p
for cnt_j=0:p-2
U(cnt_i,cnt_j+1)= s( (mod((cnt_j*(r(cnt_i))) ,(p-1) ))+1);

end
end

if C== (p+1)
for cnt_j=0:p-2
U(cnt_i,cnt_j+1)= s(( mod((cnt_j*(r(cnt_i))) ,(p-1) ))+1);

end

U(cnt_i,p+1)=p;

% exchanging first and last row
if K== (R*C)
new_U=U;
U(R,C)=new_U(R,1);
U(R,1)=new_U(R,C);

83



end
end

if C== (p-1)
for cnt_j=0:p-2
U(cnt_i,cnt_j+1)= (s(( mod((cnt_j*(r(cnt_i))) ,(p-1)
))+1))-1;

end

end

end

U=U+1;

% performing intra row permutation
intra_row=zeros(R,C);
for cnt_row=1:R
for cnt_col=1:C
intra_row(cnt_row,cnt_col)=X(cnt_row,U(cnt_row,cnt_col));

end
end

% % (6) Perform the inter-row permutation for the rectangular matrix
based on the pattern ( ) i?{0,1, ,R?1}
inter_row=zeros(R,C);

for cnt=1:R
inter_row(cnt,:)=intra_row(T(cnt),:);
end

Y=[];
for cnt_y=1:C
Y=[Y (inter_row(:,cnt_y))'];
end
interleaved=Y;
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function for bpsk modulation

function [mod_signal]=bpsk_mod(turbo_encoded)
%bpsk modulate the data

mod_signal=zeros(1,length(turbo_encoded));
for cnt=1:length(turbo_encoded)
if turbo_encoded(cnt)==0
mod_signal(cnt)=-1;
else
mod_signal(cnt)=1;
end
end

end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
84



Function to simulate channel models

function [symbol_likelihood]= channel (
input,SNR,channel_type,symbols_per_frame,rate)

% converting from db to linear scale
EbNo = 10.^(SNR/10);
EsNo = rate.*EbNo; % energy of symbol to noise
variance = 1/(2*EsNo); % variance of noise
N0=2*variance; % noise spectral density

% create the fading coefficients ,generate noise
if (channel_type==1) % AWGN channel
a = ones(1,symbols_per_frame);

elseif (channel_type==2) % Rayleigh fading
a = sqrt(0.5)*( randn( 1,symbols_per_frame) + j*randn(
1,symbols_per_frame) );
% noise =
sqrt(variance)*(randn(1,symbols_per_frame)+i*randn(1,symbols_per_frame
));
end

% add noise to the signal
noise = sqrt(variance)*(randn(1,symbols_per_frame));
r = abs(a).*input + noise;
symbol_likelihood = -2*r.*abs(a)/variance; % This is the LLR

end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function to simulate the turbo decoder

function
[tilda2_xk,xk_hat,err,w_xk]=turbo_dec(rx_signal,iter,data,w_xk,sterric
)
% main function for the decoder
% rx signal =received LLR
% iter=number of turbo code iterations to perform
% data = original data bits that will be used for calculating errors
% w_xk = feedback from 2nd decoder to 1st decoder

% the received signal rx_signal is in the form
% X1, Z1, Z1', X2, Z2, Z2', . . . , XK, ZK, ZK',XK+1, ZK+1, XK+2,
ZK+2, XK+3, ZK+3, XK'+1, ZK'+1, XK'+2, ZK'+2, XK'+3, ZK'+3,

length_data= length(rx_signal); % checking length of the signal
received
data_bits=rx_signal(1:end-12); % Seperating data bits from
received signal X1, Z1, Z1', X2, Z2, Z2', . . . , XK, ZK, ZK'
parity_bits=rx_signal(end-11:end); % parity part XK+1, ZK+1, XK+2,
ZK+2, XK+3, ZK+3, XK'+1, ZK'+1, XK'+2, ZK'+2, XK'+3, ZK'+3
K= (length_data-12)/3; % length of the input signal

% separating data and encoded bits Xk, Zk, Zk'
r_xk=data_bits(1:3:length_data-12);
r_zk=data_bits(2:3:length_data-12);
85



r_zk_dash=data_bits(3:3:length_data-12);

% appending flush bits Xk+1, Zk+1, Xk+2, Zk+2, Xk+3, Zk+3,
r_zk=[r_zk parity_bits(1:2:6)];
parity1=parity_bits(2:2:6);
% appending flush bits X'k+1,Z'k+1,X'k+2, Z'k+2, X'k+3, Z'k+3
r_zk_dash=[r_zk_dash parity_bits(7:2:12)];
parity2=parity_bits(8:2:end);
r_xk1=[r_xk parity1];
out=[];
% loop for iteration

for k_loop=1:iter
% call for upper decoder
v1_xk=r_xk1+w_xk; % input 1
r_zk; % input 2
[tilda1_xk]=dec_recursion(v1_xk,r_zk,sterric); % output of 1st
decoder
tilda1_xk=tilda1_xk(1:length(r_xk)); % removing parity bits
before passing infor to 2nd decoder
w_xk=w_xk(1:length(r_xk)); %
v2_xk=tilda1_xk-w_xk;

% Interleaving
% dividing data into blocks for interleaving
len_bits=length(r_xk);
num_blocks=ceil(len_bits/5114); % determining number of data
blocks
data_in=v2_xk; % data to be interleaved
interl_out=[]; % interleaver output
if num_blocks>1
for cnt_block=1:num_blocks-1
input_interl=(data_in((((cnt_block-1) *5114)
+1):cnt_block*5114));
interl_out=[interl_out interl(input_interl)];
end
end
% inetrleaving the last block
input_interl = data_in( ( ( (num_blocks-1) *5114) +1)
:end);
[y]=interl(input_interl);
interl_out=[interl_out y]; % all data blocks interleaved
v2_xk_dash= [interl_out parity2]; % interleaved data +
parity

% call for 2nd decoder
[tilda2_xk_dash]=dec_recursion(v2_xk_dash,r_zk_dash,sterric); %
output of 2nd decoder
tilda2_xk_dash=tilda2_xk_dash(1:length(r_xk));

% deinterleave the output
% number of blocks=num_blocks
deinterl_in=tilda2_xk_dash; % data to be deinterleaved
deinterl_out=[]; % deinterleaver output
if num_blocks>1
for cnt_block=1:num_blocks-1
deinterl_in((((cnt_block-1) *5114) +1):cnt_block*5114);
mid_dein=deinterl(deinterl_in((((cnt_block-1) *5114)
+1):cnt_block*5114));
deinterl_out=[deinterl_out mid_dein]
86




end
end

% inetrleaving the last block
deinterl_last = deinterl_in( ( ( (num_blocks-1) *5114) +1)
:end);
[y]=deinterl(deinterl_last);
deinterl_out=[deinterl_out y]; % all data blocks deinterleaved
tilda2_xk=deinterl_out;
% updating w_xk for next iteration
w_xk=tilda2_xk-v2_xk;
w_xk=[w_xk zeros(1,3)];

end

tilda2_xk=tilda2_xk+tilda1_xk ;

% perform hard decision to check for errors
xk_hat=[]; % initializing output

for count_check=1:length(tilda2_xk)
if tilda2_xk(count_check) > 0
xk_hat(count_check)=1;
elseif tilda2_xk <=0
xk_hat(count_check)=0;
end

end
end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function to deinterleaver the data

function[output]=deinterl(xin) % input data
% fucntion call for deinterleaver

K=length(xin);
[interleaved,R,U,C,T]=interleave(K); % function calll for calculating
the required parameters

% writing data in matrix form
size_matrix=R*C;

if K< size_matrix % if data smaller than matrix size,padding
required
xin=[xin (K+1):(size_matrix)];
end

% writing X in matrix form
X_out=[];
mid_in=xin;
for cnt_c=1:C
X_out=[X_out (mid_in(1:R))'];
mid_in=mid_in(R+1:end);
end
% X is now in matrix form
87




% reverse of inter row permutation
Inter_row=zeros(R,C);
for cnt=1:R
Inter_row(T(cnt),:)=X_out(cnt,:);
end

%reverse of intra row permutation
Intra_row=zeros(R,C);
for cnt_row=1:R
for cnt_col=1:C
Intra_row(cnt_row,U(cnt_row,cnt_col)
)=Inter_row(cnt_row,cnt_col);
end
end

% writing data in array
out_array=[];
for cnt_y=1:R
out_array=[out_array (Intra_row(cnt_y,:))];
end

output=out_array(1:K);

end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function to implement the constituent RSC decoder

function [LLR,LLR_coded] = dec_recursion(input1,input2,sterric)
% Function call for the component decoder
% Input 1 is the uncoded LLR
% Input 2 is the coded LLR
% Sterric defines the type of jacobian algorithm used

% The outputs from the code are:
% LLR= extrinsic LLR of the data bits
% LLR_c= estrinsic LLR of the coded bits

% The trellis structure is described by the trans_matrix.
% It gives the initial state,the final state, parity bit Z(i,j)
generated
% and the data bit X(i,j) input

% FromState, ToState, ParityBit Z(i,j), DataBit
X(i,j)
trans_matrix = [1, 1, 0, 0;
2, 5, 0, 0;
3, 6, 0, 1;
4, 2, 0, 1;
5, 3, 0, 1;
6, 7, 0, 1;
7, 8, 0, 0;
8, 4, 0, 0;
1, 5, 1, 1;
2, 1, 1, 1;
3, 2, 1, 0;
4, 6, 1, 0;
88



5, 7, 1, 0;
6, 3, 1, 0;
7, 4, 1, 1;
8, 8, 1, 1];

% count the number of total states
state_cnt = max(trans_matrix(:,1));

data_size=length(input1); % size of the input data
K=data_size-3; % data is from k=1...K+3
num_transition= 2*state_cnt; % number of transitions

% calculating the state transition metrics or gammas
% there are four distinct state transition metrics as marked on
the
% trellsi diagram
% gamma0 = 0
% gamma1 = V(Xk) uncoded
% gamma2 = R(Zk) encoded
% gamma3 = V(Xk)+R(Zk)

% Two gamma matrices with V(Xk) and R(ZK) values and zeros

gamma1=zeros(size(trans_matrix,1),data_size); % matrix for V(Xk)
gamma2=zeros(size(trans_matrix,1),data_size); % matrix for R(Zk)
for bit_cnt = 1:data_size
for trans_cnt = 1:size(trans_matrix,1)
if trans_matrix(trans_cnt, 3)==0 % when Z(i,j)=0
then gamma=X(i,j)V(Xk)
gamma1(trans_cnt, bit_cnt) =input1(bit_cnt);
end % end of if
if trans_matrix(trans_cnt, 4)==0 % when X(i,j)=0
then gamma=Z(i,j)R(Zk)
gamma2(trans_cnt, bit_cnt) = input2(bit_cnt);
end % end of if
end % end of for for trans_cnt
gamma=gamma1+gamma2; % the state transition matrix gamma
?_ij=V(X_k )X(i,j)+R(Z_k )Z(i,j)
end % end of bit_cnt for loop

% Backward recursion

betas=zeros(state_cnt,data_size);
betas=betas-inf; % initializing beta to - Inf
betas(1,data_size)=0; % B(k+3)(S0)=0

% starting from k= K+2 and going to k=1

for bit_cnt = data_size-1:-1:1
if (bit_cnt<=K)
app_in = input1(bit_cnt);
else
app_in = 0;
end

beta_mid=zeros(state_cnt,1);
for trans_cnt = 1:state_cnt
beta1=betas(trans_matrix(trans_cnt,2),bit_cnt+1)
+gamma(trans_cnt, bit_cnt+1); % zero state connected to state S
89



beta2=betas(trans_matrix(trans_cnt+8,2),bit_cnt+1)+
gamma(trans_cnt+8, bit_cnt+1); % one state connected to state S

beta_mid(trans_matrix(trans_cnt,1))=max_sterric(beta1,beta2,sterric);
% max_sterric of both the betas
betas(trans_matrix(trans_cnt,1),bit_cnt) =
max_sterric(beta1,beta2,sterric);
end % end of for loop for trans_cnt
beta_mid=beta_mid-((ones(state_cnt,1))*(beta_mid(1,1)));
nomalizing Beta(Si)-Beta(S0)
betas(:,bit_cnt)=beta_mid; %
saving values in beta matrix

end % end of for loop for bit_cnt

% Forward recursion

alphas=zeros(state_cnt,data_size);
alphas=alphas-inf;
alphas(1,1)=0; % alpha0(so)=0
for bit_cnt = 2:data_size % statring from state 2 and going
till the end
if bit_cnt <= K
app_in=input1(bit_cnt-1);
else
app_in=0;
end

for trans_cnt = 1:num_transition
if alphas(trans_matrix(trans_cnt,1),bit_cnt-1) ==0
app_in=0;
end
alpha1=alphas(trans_matrix(trans_cnt,1),bit_cnt-1)
+gamma(trans_cnt, bit_cnt-1);
%sterric=1;

alphas(trans_matrix(trans_cnt,2),bit_cnt)=max_sterric(alpha1,alphas(tr
ans_matrix(trans_cnt,2),bit_cnt),sterric);

end % end of for loop for trans_cnt
alpha_mid=alphas(:,bit_cnt); % normalizing alphas
alpha0=alpha_mid(1,1);
alpha_mid=alpha_mid-alpha0;
alphas(:,bit_cnt)=alpha_mid ;
end % end of for loop for the bit_cnt


% LLR estimate for data bit Xk.
lamdas=zeros(num_transition,data_size);
for bit_cnt = 1:data_size
for trans_cnt = 1:size(trans_matrix,1)
lamdas(trans_cnt, bit_cnt) =
alphas(trans_matrix(trans_cnt,1),bit_cnt) + gamma(trans_cnt,bit_cnt) +
betas(trans_matrix(trans_cnt,2),bit_cnt);
end
end

% Calculating likelihood of data 1
LLR = zeros(1,data_size);
for bit_cnt = 1:data_size
90



% initializing the probability variables of -Inf
lamda_0=-inf;
lamda_1=-inf;
for trans_cnt = 1:size(trans_matrix,1)
lamda_mid=lamdas(trans_cnt,bit_cnt);
if trans_matrix(trans_cnt,3)==0 % Z(i,j)=0 ---->
?_ij=V(X_k )X(i,j)
lamda_0 = max_sterric(lamda_0,lamda_mid ,sterric); %
max* of all lamdas for Xi=0
else
lamda_1 = max_sterric(lamda_1, lamda_mid,sterric); %
max* of all lamdas for Xi=1
end
end
LLR(bit_cnt) = lamda_0-lamda_1; % LLR of bit being zero
or one
end


% extrinsic LLR estimate for coded bits
% LLR estimate for code bits.
lamdas2=zeros(num_transition,data_size);
for bit_cnt = 1:data_size
for trans_cnt = 1:size(trans_matrix,1)
lamdas2(trans_cnt, bit_cnt) =
alphas(trans_matrix(trans_cnt,1),bit_cnt) + gamma2(trans_cnt,bit_cnt)
+ betas(trans_matrix(trans_cnt,2),bit_cnt);
end
end

% Calculating likelihood of data 1
LLR_coded = zeros(1,data_size);
for bit_cnt = 1:data_size
% initializing the probability variables of -Inf
lamda_0c=-inf;
lamda_1c=-inf;
for trans_cnt = 1:size(trans_matrix,1)
lamda_mid2=lamdas2(trans_cnt,bit_cnt);
if trans_matrix(trans_cnt,4)==0 % Z(i,j)=0 ---->
?_ij=V(X_k )X(i,j)
lamda_0c = max_sterric(lamda_0c,lamda_mid2 ,sterric);
% max* of all lamdas for Xi=0
else
lamda_1 = max_sterric(lamda_1c, lamda_mid2,sterric);
% max* of all lamdas for Xi=1
end
end
LLR_coded(bit_cnt) = lamda_0c-lamda_1c; % LLR of bit
being zero or one
end

end


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

91



Function to compute the max* algorithm

function out=max_sterric(x,y,sterric)
% function for calculaing the Jacobian algorithm
if (x== -Inf && y== -Inf)
out= -Inf;
elseif sterric==1 % compute log MAP
ab=abs(x-y);
fc=(log(1+exp(-ab)))/log(2);
max_s=max(x,y)+fc;
[out]=max_s;

elseif sterric==2 %comput max-log MAP
max_s=max(x,y);

elseif sterric==3 % comput constant-log MAP
ab=abs(y-x);
T=1.5; % values as specified in text
C=0.5;
if ab > T
fc=0;
elseif ab <= T
fc=C ;
end
max_s=max(x,y)+fc;
out=max_s;

elseif sterric==4 % compute linear-log MAP
ab=abs(y-x);
T=2.5068; % constant values as specified in text
alpha=-0.24904;
if ab > T
fc=0;
elseif ab <= T
fc=alpha*(ab-T) ;
end
max_s=max(x,y)+fc;
out=max_s;
end

end


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function for hard decision decoding

function [xk_hat]=hard_decision(tilda2_xk)
% input is the LLR sequenct tilda2_xk
% output is the hard decoded bit sequence xk_hat

xk_hat=[]; % initializing output

for count_check=1:length(tilda2_xk)
if tilda2_xk(count_check) > 0
xk_hat(count_check)=1;
elseif tilda2_xk <=0
xk_hat(count_check)=0;
92



end
end

end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function for calculating average mutual information


% llrs is a 1xK vector of LLRs
% mutual_information is a scalar in the range 0 to 1
function mutual_information = avg_mutual_inf(llrs)

P0 = exp(llrs)./(1+exp(llrs));
P1 = 1-P0;
entropies = -P0.*log2(P0)-P1.*log2(P1);
mutual_information = 1-
sum(entropies(~isnan(entropies)))/length(entropies);

end

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Main function for plotting the EXIT chart

% The main Program for generating exit chart for the UMTS Turbo code
clear all
close all
clc

K=40; % Specify frame size between 40 and
5114 bits.
SNR =0; % Choose the SNR
frame_count_max = 100; % Choose how many frames to simulate
IA_count = 10; % Choose how many points to plot in the
EXIT functions
channel_type=1; % AWGN channel
sterric=1;
symbols_per_frame=K*3+12; % Symbols in frame after encoding.
rate=K/symbols_per_frame; % Code rate of the Turbo code (ratio
of number of input bits to output bits)

% Calculate the MIs to use for the a priori LLRs
IAs = (0:(IA_count-1))/(IA_count-1);
IE_means = zeros(1,IA_count);
IE_stds = zeros(1,IA_count);
% Determine each point in the EXIT functions

for IA_index = 1:IA_count % loop for IA count
IEs = zeros(1,frame_count_max); % initializng IE to all zeros

% This runs the simulation long enough to produce smooth EXIT
functions.
for frame_index = 1:frame_count_max

data=randint(1,K); % generate random data

93



% function call for turbo code
[turbo_encoded]=turbo_enc(data);

%function call to bpsk modulate the turbo coded signal
[mod_signal]=bpsk_mod(turbo_encoded);

% Function call to pass through the channel, demodulate and
% generate LLRs

[symbol_likelihood]= channel ( mod_signal , SNR
,channel_type,symbols_per_frame,rate);
w_xk=generate_llrs(data, IAs(IA_index)); % generate random
LLRs to use as reference
rx_signal=symbol_likelihood;
data_bits=rx_signal(1:end-12) ; % data part
parity_bits=rx_signal(end-11:end); % parity part XK+1,
ZK+1, XK+2, ZK+2, XK+3, ZK+3, XK'+1, ZK'+1, XK'+2, ZK'+2, XK'+3,
ZK'+3,
length_data=length(data_bits);

% seperating the systematic bits and coded bits
r_xk=data_bits(1:3:length_data);
r_zk=data_bits(2:3:length_data);
r_zk_dash=data_bits(3:3:length_data);

% appending flush bits Xk+1, Zk+1, Xk+2, Zk+2, Xk+3, Zk+3,
% X'k+1,Z'k+1,X'k+2, Z'k+2, X'k+3, Z'k+3
r_zk=[r_zk parity_bits(1:2:6)];
parity1=parity_bits(2:2:6);
r_zk_dash=[r_zk_dash parity_bits(7:2:12)];
parity2=parity_bits(8:2:end);
r_xk1=[r_xk parity1];

% loop for iteration
w_xk=[w_xk 0 0 0]; % extrinsic information
v1_xk=r_xk1+w_xk; % apriori information

[tilda1_xk]=dec_recursion(v1_xk,r_zk,sterric); %
function call for RSC decoding
IEs(frame_index) = avg_mutual_inf(tilda1_xk); %
calculate the mutual info of the decoder output.This is the extrinsic
information.

end

% Store the mean and standard deviation of the results
IE_means(IA_index) = mean(IEs);
IE_stds(IA_index) = std(IEs);
end

figure;
axis square;
title('EXIT chart for BPSK modulation in an AWGN channel');
ylabel('I_E');
xlabel('I_A');
xlim([0,1]);
ylim([0,1]);

hold on;
94




% Plot the EXIT function for component decoder 1
plot(IAs,IE_means,'-');
plot(IAs,IE_means+IE_stds,'--');
plot(IAs,IE_means-IE_stds,'--');

% Plot the inverted EXIT function for component decoder 2
plot(IE_means,IAs,'-');
plot(IE_means+IE_stds,IAs,'--');
plot(IE_means-IE_stds,IAs,'--');

% saving data in mat file
filename =
['AWGN_',num2str(K),'_','logMAP_','exit_chart_',num2str(SNR),'.mat'];
save(filename, 'IAs','IE_means', '-MAT');

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function call for the decoder trajectory in EXIT chart

% The main Program for generating the UMTS Turbo code in MATLAB
clear all
close all
clc
figure
K=320; % Specify frame size between 40
and 5114 bits.
iter_cnt = 20; % Specify the maximum number of
decoding iterations to perform.
symbols_per_frame=K*3+12; % Symbols in frame after encoding.
rate=K/symbols_per_frame; % Code rate of the Turbo code
(ratio of number of input bits to output bits)
frame_count_max=10; % Maximum number of frames to be
passed to the code
bit_count_max=frame_count_max*K; % Maximum bits at a given SNR
error_max=10; % Maximum number of errors (to be
used in condition later)
count_SNR=0; % Counter for number of SNR values
results=[]; % The ouput results matrix defined
SNR=0;
channel_type=1; % channel_type can be 1 or 2(for
AWGN and Rayleigh fading respectively)
sterric=1 ; % sterric can be 1 to 4 (for
LogMAP, max-logMAP, constant-logMAP, linear-logMAP)
max_fail_count = 3; % Specift the number of iterations
that should fail to improve the decoding before the iterations are
stopped early

% calculating Lc
EbNo = 10.^(SNR/10);
EsNo = rate.*EbNo; % energy of symbol to noise
variance = 1/(2*EsNo); % variance of noise
Lc=2/variance;

% Initiazing the output arrays
IA1mean=zeros(1,iter_cnt);
IE1mean=zeros(1,iter_cnt);
IA2mean=zeros(1,iter_cnt);
IE2mean=zeros(1,iter_cnt);
95



figure;
axis square;
title('BPSK modulation in an AWGN channel');
ylabel('I_E');
xlabel('I_A');
xlim([0,1]);
ylim([0,1]);
hold on;

for frame_index = 1:frame_count_max
err_out=[]; % error count matrix
BER=1; % initializig BER to max i.e. 1
data=randint(1,K); % generate random data
results=[zeros(1,2)];
error_counts=zeros(1,iter_cnt); % initialize error count matrix
bit_count=0;

% function call for turbo code
[turbo_encoded]=turbo_enc(data);

%function call to bpsk modulate the turbo coded signal
[mod_signal]=bpsk_mod(turbo_encoded);

% Function call to pass through the channel

[symbol_likelihood]= channel ( mod_signal , SNR
,channel_type,symbols_per_frame,rate);
extrinsic2=zeros(1,K); % feedback from second copmponent
encoder to first component encoder.It is initialized to zero before
first iteration.

avg_IA = 1; % initializing the mutual information
averaging value to zero
iteration_index = 1; % iteration index is initialized to 1
fail_count = 0; % number of times the code has failed
IA1 = 0;
IE1 = 0;
IA2=0;
IE2=0;

% Run the simulation while the number of failed iterations
% is less than max_fail_count and iteration index is less than
% maximum ietartions

% seperating the data and parity bits from the received signal
rx_signal=symbol_likelihood;
length_data= length(rx_signal); % checking length of the
signal received
data_bits=rx_signal(1:end-12); % data part
parity_bits=rx_signal(end-11:end); % parity part XK+1, ZK+1,
XK+2, ZK+2, XK+3, ZK+3, XK'+1, ZK'+1, XK'+2, ZK'+2, XK'+3, ZK'+3,
K= (length_data-12)/3; % length of the input signal


% separating data and encoded bits Xk, Zk, Zk'
r_xk=data_bits(1:3:length_data-12);
r_zk=data_bits(2:3:length_data-12);
r_zk_dash=data_bits(3:3:length_data-12);

96



% appending flush bits Xk+1, Zk+1, Xk+2, Zk+2, Xk+3, Zk+3,
% X'k+1,Z'k+1,X'k+2, Z'k+2, X'k+3, Z'k+3

r_zk=[r_zk parity_bits(1:2:6)];
parity1=parity_bits(2:2:6);
r_zk_dash=[r_zk_dash parity_bits(7:2:12)];
parity2=parity_bits(8:2:end);


while iteration_index <= iter_cnt

w_xk=extrinsic2; % extrinsic information 2
apriori1=extrinsic2; % apriori information 1
IA1 = avg_mutual_inf(apriori1);
% function call to turbo decoder. Ouputs are extrinsic
information
% for decoder 1 and 2 and apriori information for decoder 2.

[extrinsic1,extrinsic2,apriori2]=turbo_dec_traj(r_xk,r_zk,r_zk_dash,Lc
, K, apriori1,sterric);

% calculate IAs and IEs
IE1 = avg_mutual_inf(extrinsic1);
IA2 = avg_mutual_inf(apriori2);
IE2 = avg_mutual_inf(extrinsic2);

% store the results in output arrays
IA1mean(1,iteration_index)=IA1mean(1,iteration_index)+IA1;
IE1mean(1,iteration_index)=IE1mean(1,iteration_index)+IE1;
IA2mean(1,iteration_index)=IA2mean(1,iteration_index)+IA2;
IE2mean(1,iteration_index)=IE2mean(1,iteration_index)+IE2;

iteration_index = iteration_index + 1; % increment the
iteration index

end % end of while loop for iterations

end % end of for loop for frame count

% calculating means of IAs and IEs

IA1mean=(IA1mean)./frame_count_max
IE1mean=(IE1mean)./frame_count_max
IA2mean=(IA2mean)./frame_count_max
IE2mean=(IE2mean)./frame_count_max

% plotting the trjectory
j=0;
figure
for i=1:length(IA1mean)
j=j+1;
traj(j)=IE1mean(i);
indices(j)=IA1mean(i);
j=j+1;
traj(j)=IA2mean(i);
indices(j)=IE2mean(i);
end
plot(indices,traj,'k.-')
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%