FPGA Implementation of Soft Output Viterbi Algorithm Using Memoryless Hybrid Register Exchange Method

International Journal of VLSI design & Communication Systems (VLSICS) Vol.2, No.3, September 2011
DOI : 10.5121/vlsic.2011.2304 51
FPGA IMPLEMENTATION OF SOFT OUTPUT
VITERBI ALGORITHM USING MEMORYLESS HYBRID
REGISTER EXCHANGE METHOD
R .D. Kadam1
and S. L. Haridas2
1
Department of Electronics and Telecomm. BDCE, Sevagram, RTM Nagpur University,
India
rdk_arvi@rediffmail.com
2Department of Electronics and Telecomm. BDCE, Sevagram, RTM Nagpur University,
India
slhlec@rediffmail.com
ABSTRACT
The importance of convolutional codes is well established. They are widely used to encode digital data
before transmission through noisy or error-prone communication channels to reduce occurrence of errors
and memory. This paper presents novel decoding technique, memoryless Hybrid Register Exchange with
simulation and FPGA implementation results. It requires single register as compared to Register Exchange
Method (REM) & Hybrid Register Exchange Method (HREM); therefore the data trans-fer operations and
ultimately the switching activity will get reduced.
KEYWORDS
Traceback method, Register Exchange Method, Hybrid Register Exchange Method, Memoryless HREM.
1. INTRODUCTION
The task facing the designer of a digital communication system is that of providing transmitting
information from one end of the system at a rate and a level of reliability and quality that are
acceptable to the user at the other end. Convolutional encoding with Viterbi decoding is a
technique that is particularly suited to a channel in which the transmitted signal is corrupted
mainly by additive white Gaussian noise (AWGN). Viterbi decoding was developed by Andrew
J. Viterbi in 1967 [1]. As it was recognized by early 1970’s, the algorithm was a maximum
likelihood decision device for any symbol sequence that could be modeled as a state diagram [2].
Then in 1971 viterbi published another paper [3] focusing on convolutional codes, it begins with
an elementary presentation of fundamental properties and structure of convolutional codes and
proceeds with the development of maximum likelihood decoder. Since then, other researchers
have expanded on his work by finding good convolutional codes, exploring the performance
limits of the technique [4] and varying decoder design parameters to optimize the implementation
of the technique for hardware and software. They are widely used to encode digital data before
transmission through noisy or error prone communication channels to reduce occurrence of error.
The traceback (other is register exchange) method is used as a data decoding technique offers
reduced hardware complexity with longer latency as the trades off cyper et al [5] first present a
algorithm for implementing a traceback survivor memory unit (SMU), which turned to be a
generalization of the implementation is used in [6]. The hardware complexity of a viterbi decoder
is proportional to the number of states in the trellis [7]. In order to make the Viterbi algorithm a
practical decoding techniques certain refinement were made on the basis algorithm. The viterbi
algorithm uses the trellis diagram to decode an input sequence [8] is a dynamic programming

52
algorithm for finding the shortest path through a trellis. In case of soft decision technique,
variation of the signal at the output of the demodulator are sampled and quantized, soft output
viterbi algorithm accepts and delivers soft sample values and can be regarded as a device for
improving the SNR [9], BMU finds the equivalent branch metric using correction between the
received sample values and the expected symbols and add compared select (ACS) unit perform
cumulative addition of branch metric to generate equivalent state metric, which multiple
operation results in complex control logic. The two decoding methods called modified traceback
and Hybrid register exchange method was proposed in [10]. A short overview of a soft input
viterbi decoder implementation in field programmable gate array (FPGA) for code division
multiple access (CDMA) wireless communication system is presented in [11]. It is well known
that data transmission over wireless channels are affected by attenuation, distortion, interference
and noise which affect the receiver’s ability to receive correct information. Particularly for
terrestrial cellular telephony, the interference suppression feature of CDMA can result in a many-
fold increase in capacity with multiple access techniques were not as analog and even competing
digital techniques [12].
In the following sections the coder used, system component, various decoding techniques and
finally the proposed new decoding technique with simulation and FPGA results will be described.
2. CONVOLUTION CODES
Convolution codes are well described in the literature [1], [3]. They are commonly specified by
three Parameters; (n, k, m), where: n is the number of output bits, k is the number of input bits,
and m is the number of shift register stages of the coder. The constraint length K of the code
represents the number of bits in the encoder memory that affect the generation of the n output bits
and is defined as K = m + 1. The code rate r of the code is a measure of the code efficiency and is
defined by r = k/n.
2.1. Structure of the convolutional code:
Figure 1 show the convolutional encoder structure (3, 1, 2) used in this paper and is built from its
parameters. It consists of 2 (m=2) shift register stages and two modulo-3 adders (n = 3) giving the
outputs of the encoder. The rate of the code is r = 1/3. The minimum distance of the code is dmin =
8. The outputs of the adders are sampled sequentially yielding the code symbols. The total
number p of bit symbols is given by p = n (b + m) where b is the total number of bits of
information [8].
The outputs Y0, Y1 and Y2 of the adders are governed by the following generator polynomials.
The generator polynomial for the output Y0 and Y1 is given by g0(x) = g1(x) = 1 + x + x2
the
generator polynomial for the output Y2 is given by g2(x) = 1 + x2
. g0(x), g1(x) and g2 (x) select the
shift register stages bits to be added to give the outputs of the encoder which are, for the case of
(3, 1, 2) encoder as follow. The polynomials give the code its unique error protection quality.
Y2 = U0 U2 ----------------1
Y1 = U0 U1 U2 ----------------2
Y0 = U0 U1 U2 ----------------3

53
Figure 1. The 1/3 Convolutional Encoder
3. BLOCK DIAGRAM OF VITERBI DECODER
In this paper, the viterbi decoder is designed and implemented aims for 3GPP standard. The
diagram of proposed architecture is shown in figure 2. It is composed of branch metric unit
(BMU), add compare select unit (ACSU), survival memory unit (SMU), and a decoding unit.
3.1. Branch Metric Unit (BMU)
BMU works for calculating the branch metrics according to received sequence. For three 3-bit
soft decision input bits (i0, i1, i2) each ranging from -3 to +3, eight 5-bit branch metrics are
generated. The decision bits are represented in the two’s complement representation. The BMU
perform simple add; subtract operations on the input bits to generate the output. For example the
branch metric for the state transition which produces the binary output (010) is i0 - i1 + i2. The
BMU performs the computations, as represented in table 1. The output of the BMU is still in a
two’s complement format. The bit serial format of the branch metrics is generated by the parallel
to serial module at the output of the BMU, as shown in Figure 3. The bit serial format of the BMs
is then fed into the ACSU. In the VA for decoding convolutional codes, the squared Euclidean
distance is the optimum branch metric for decoding sequences that are transmitted in an AWGN
ACSU
BMU SMU
Decoding Unit
In X FF1 FF0
Y0
Y1
Y2
+
+
+
+
+
+
U0 U1 U2
Figure 2. Block diagram of viterbi decoder

54
channel. Multiplication operations or look up tables are required for the Viterbi algorithm to
compute the squared distances to obtain the branch metrics. However, for binary convolutional
codes, it is proven that linear distances (Hamming distances) can be used as the optimum branch
metrics. This is true for convolutional codes.
Table 1. Branch Metric
3.2. Add compare select (ACS)
The ACSU work in serial architecture style with 2 ACS as shown in figure 4, for having small
implementation area, each trellis state S at time t is a state metric SMp and SMq which is the
accumulated metric along the shortest path leading to that state.
BMU000 (i0+i1+i2) = (-3-3-3) -9
BMU001 (i0+i1-i2) = (-3-3+3) +3
BMU010 (i0-i1+i2) = (-3+3-3) -3
BMU011 (i0-i1-i2) = (-3+3+3) +3
BMU100 (-i0+i1+i2) = (+3-3-3) -3
BMU101 (-i0+i1-i2) = (+3-3+3) +3
BMU110 (-i0-i1+i2) = (+3+3-3) +3
BMU111 (-i0-i1-i2) = (+3+3+3) +9
BMU
000
BMU
001
BMU
010
BMU
011
+
+
BMU
011
BMU
101
BMU
100
+
+
+ +
BMU
111
+
+
i0
i1
i2
Figure 3. Branch Metric Unit

55
The state metrics at time can be recursively calculated in terms of the state metrics of the previous
iteration as follows:
For state metric SMt
SMp = min (SMi+BMi1, SMj+BMj1),
SMq = min (SMi+BMi0, SMj+BMj0)
To demonstrate the functionality of VA, a sample input to the encoder is traced until the input is
decoded. The encoder has an input sequence, 11011000 and generates the code stream, (111, 100,
100, 000, 100, 100, 111 and 000). The VA which uses soft decision formats to decode survivor
path.
4. DECODING TECHNIQUES
4.1. Register Exchange method
In the register exchange method [2], a register assigned to each state contains information bits for
the survivor path from the initial state to the current state. In fact, the register keeps the partially
decoded output sequence along the path, as illustrated in figure 5. The register of state S1 at t = 3
contain ‘110’, which is the decoded output sequence along the hold path from the initial state.
The resister exchange approach does require the copying of all the registers at each stage. The
need to traceback is eliminated since the register of the final state contain the decoded output
sequence. However, this approach results in complex hardware due to the need to copy the
contents of all the register in a stage to the next stage. At last stage, the decoded output sequence
is the one that is stored in the survivor path register S0, the register assigned to the state with the
minimum Path Metric. Since the Register Exchange method does not need tracing back, it is
faster. However, the RE method does require the copying of all the registers at each stage.
Figure 5. Register Exchange Approach
S2
SM
SMj
SMi SM
S1
S0 S0
Figure 4. ACSU Butterfly

56
4.2. Hybrid Register Exchange Method
The main drawback of register exchange method is its frequent switching activity and long
constraint length. One of the promising solutions to reduce the switching activity can be achieved
by combining the REM and TB techniques [4]. The initial state can be first traced back through
an m cycle, and then transfer the content of initial state to the current state and the next m bits of
the register is the m bits of current state itself.
Figure 6. Hybrid Register Exchange Approach
4.3. Memoryless Register Exchange Method
The RE approach generates the decoded bits in the correct order. The decoded bits are produced,
and then read out from the decoder. Thus, a memory free viterbi decoder can be implemented by
solely resetting the encoder contents for each L bits that are encoded. The new VD
implementation is called the memoryless viterbi decoder (MLVD). Since the MLVD needs to
track only one row, the MLVD requires only one pointer to track the current position of the
decoder in the trellis. If the initial state is zero, then only the first row of memory is needed. In
other words, the storage of the decoded bits is necessary in order to choose only one row of
memory at the end to represent the actual decoded bits. If the required row of memory is
predetermined, and then there is no need for the storage of the other rows as shown in figure 7.
Figure 7. Memoryless REM
4.4. Memoryless Hybrid Register Exchange method
The MLVD keeps track of the current state position of the decoder in the memory unit as shown
in Figure 8. It makes use of the fact that the bit appended to each row of memory is exactly the bit
that is shifted into the pointer to form the new pointer to that row of memory [13]. To show the
00 01 11 10 01 11 10 00 00
t=0
1
t=1
11
t=2
110
t=3
1101
t=4
11011
t=5
110110
t=6
1101100
t=7
11011000
t=8

57
functionality, K = 3 a (four states) and rate = 1/3 convolutional encoder (g2 = 101, g1 = 111, g0 =
111) is employed to encode the input sequence of (11011000). The code stream (111, 100, 100,
000, 100, 100, 111 and 000) is generated and transmitted over a channel. The noisy code stream
(111, 110, 101, 000, 100, 100, 111 and 000) for example is received at the decoder. The
underlined bits are incorrect because of the noise encountered during transmission. Applying the
MLVD method result in the successive values for the pointer and row of memory for the decoded
data over time. The pointer contains the current state of the decoder (m bits). The data at memory
is the pointer value at that instant. Every time the digit get decode. It is reset to zero (the initial
state of the encoder) after decoded last bits. After every ‘m’ cycles the pointer content and
memory contents are upend, so the switching will get reduced.
Figure 8. Memoryless Hybrid Register Exchange Method
5. SIMULATION RESULT FOR MEMORYLESS HREM
Figure 9 shows the simulation result for memoryless HREM, using pointer implementation.
Encoding the data of 8 bit as adding 2 bit noise error. We get the result after 14 clock pulse unit
as shown in Figure. It is same decoding input data to decode the serially output data. It shows the
noise of two bits can be corrected by the decoder since it has free distance of 8.
Fig.9. Simulation Result for memoryless hybrid register exchange method
6. FPGA Implementation
To prepare the MLVD using Hybrid Register Exchange processing VHDL design that is
implemented on the FPGA, the design is synthesized by using Synopsys tool. Then, the design is
imported into Xilinx tools for the mapping routing, then a VHDL file with timing information is
11 01 10 00
11
t=2
1101
t=4
110110
t=6
11011000
t=8

58
re-simulated for the timing verification. Afterward, the MLVDs design is downloaded to a
xs2s200 Xilinx chip. The MLVDs consumes only 2% of the total slices of the xs2s200,
comprising a total of 2352 gates. The design consumes 16 I/O bits.
All the decoding techniques are implemented and compared on the same platform.
Table 2. FPGA Simulation Report
Device Utilization REM HREM MLREM MLHREM
Number of Slices 254outof2352
10%
220 out of 2352
9%
60 out of 2352
2%
63 out of 2352
2%
Number of Slice
Flip Flop
144outof4704
3%
142 out of 4704
3%
68 out of 4704
1%
68 out of 4704
1%
Number of 4 input
LUTs
456outof4704
9%
390 out of 4704
8%
85 out of 4704
1%
83 out of 4704
1%
Number of bounded
IOBs
16outof 144
11%
16 out of 144
11%
16outof 144
11%
16 out of 144
11%
Registers 87 85 47 47
Multiplexers 44 3 6 6
Adder/ Substractor 21 21 13 13
Comparators 9 9 3 3
7. CONCLUSION
The MLVD using HREM is a memoryless implementation of the VA, and successfully decodes
the continuous data encoded by a convolutional encoder. The latency is only 2 data bits. The new
implementation is realized by applying the pointer concept to the HREM implementation, and by
using trellis truncation for every bit encoded. It is found that the new proposed decoding
technique requires lesser hardware as compared to REM, HREM and memoryless REM
techniques. The maximum throughput is 56 Mbps which is still more than 2Mbps which is
requirement of wireless communication. Increasing the MLVDs performance by increasing the
constraint length is still to be investigated. So the memoryless hybrid register exchange approach
can be used in place of traceback, register exchange for decoding of data.
References
[1] A. J. Viterbi, “Error Bounds on Convolutional Codes and an Asymptotically Optimum Decoding
Algorithm” Information and Control, 25(3), pp 260-269, April 1967.
[2] “A Personal History of the Viterbi algorithm” IEEE Signal Processing Magazine, July 2006.
[3] “Convolutional codes and their performance in communication systems” Andrew Viterbi, IEEE
trans. On communication technology, vol. com-19, 1971.
[4] G. Forney, “Convolutional codes II. Maximum - likelihood decoding.” Information and Control,
25(3), pp 222-266, July 1974

59
[5] “Generalized Traceback Techniques for Survivor Memory Management in the Viterbi Algorithm”,
Robert Cypher, C. Bernard Shung, IEEE 1990.
[6] H.A. Bustamante, I. Kang, C. Nguyen and R.E. Peile,” Stanford Telecomm design of a convolutional
decoder,” In MILCOM 89,Boston MA, October 1080 pp. 171-178.
[7] “A Low power Viterbi Decoder Design for wireless communication Application”, Samirkumar
Ranpara, Dong Sam Ha, IEEE 1999.
[8] A. J Viterrbi, J.K Omura, “Principals of digital communication”, McGraw Hill, Inc, 1979.
[9] “A Viterbi Algorithm with soft decision outputs and its application”, Joachim Hagenauer, Peter
Hoeher, IEEE 1989.
[10] “Prof. S. L. Haridas, Dr. N. K. Choudhari,” Design of Viterbi Decoder with Modified Traceback
and Hybrid Register Exchange Processing ICAC3’09.
[11] “FPGA Implementation of soft Input Viterbi Decoder for CDMA2000 System” Milos Pilipovic,
Marija Tadic, Novi Sad, Telefore 2008.
[12] “On the capacity of a cellular CDMA system “Klein S. Gilhousen, Irwin M. Jacobs, Roberto Padovani,
Andrew Lindsay A. Weaver, Jr Charles E. Wheatley 1211, IEEE trans. On vehicular technology,
vol.49, May 19991. “The CDMA Digital Cellular System an ASIC Overview” Richard Kerr, Kelein
Gilhousen, Butch, Weaver, Robert Padovani, Houtan DSehesh, Qualcomm Inc, San Diego, IUEEE
1992.
[13] Memoryless Viterbi Decoder Dalia A. El-Dib, Member, IEEE and M. I. Elmasry, Fellow, EXPRESS
BRIEFS, VOL. 52, NO. 12, IEEE 2005
Authors
Ravindra D. Kadam is pursuing M.Tech in Electronics Engineering from
RTM Nagpur University, Nagpur, India. He received CDAC in Pune, India in
1999. He received his B.E. degree in Electronics Engineering from RTM
Nagpur University, Nagpur in year 1997. He is currently working as Assistant
Professor in the Department of Electronics and Telecomm. Engineering ,
BDCOE, Sevagram, Wardha, M.S., India. His research interests include
computer networks and VLSI systems.
Sanjay L. Haridas received B.E. degree in Electronics Engineering from
RTM Nagpur University, Nagpur in year 1988 and M.Tech in Electronics
Engineering from VNIT, Nagpur in 1995. He is Professor in the Department of
Electronics and Telecommunication Engineering, BDCOE, Sevagram,
Wardha, M.S., India. His research interests are VLSI circuit design and Digital
Communiation.

FPGA Implementation of Soft Output Viterbi Algorithm Using Memoryless Hybrid Register Exchange Method

More Related Content

Similar to FPGA Implementation of Soft Output Viterbi Algorithm Using Memoryless Hybrid Register Exchange Method (20)

Recently uploaded (20)

FPGA Implementation of Soft Output Viterbi Algorithm Using Memoryless Hybrid Register Exchange Method