2524 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO.

9, SEPTEMBER 2009

A Class of Quasi-Cyclic LDPC Codes over GF(2𝑚 )

Christian Spagnol and William Marnane

Abstract—Low Density Parity Check (LDPC) codes over II. LDPC OVER GF(2𝑚 )
GF(2𝑚 ) are an extension of binary LDPC codes. Performances
of GF(2𝑚 ) LDPC codes have been shown to be higher than In this paper we focus on GF(2𝑚 ) LDPC codes applied to
binary LDPC codes, but the complexity of the encoders/decoders the binary-input AWGN channel.
increases. Hence there is a substantial lack of implementations A major drawback of LDPC over GF(2𝑚 ) codes is their
for LDPC codes over GF(2𝑚 ) codes. This paper presents a decoding complexity that increases with the order of the field.
class of quasi-cyclic LDPC codes over GF(2𝑚 ). These codes can A straightforward implementation of the non-binary belief
alleviate the encoding/decoding complexity without excessive loss
of performances. It is shown how, from a performance point of propagation decoder has a very large decoding complexity [2],
view, such codes are better than 𝑚 times bigger binary codes but recent implementations of the algorithm greatly reduced
and as good as 2𝑚 longer binary codes. such complexity [4], [5]. From a hardware point of view the
Index Terms—Low density parity check codes, quasi-cyclic main problem of implementing LDPC decoders is not the
codes, permutational codes, FPGA. computational complexity but the routing complexity [6]. In
general, for the GF(2𝑚 ) case the messages passing along the
edges are probability weight vectors. Hence, every message
I. I NTRODUCTION passed between nodes is a vector of 2𝑚 elements that must be
represented with a finite number of bits. This means a drastic

I NTEREST in non-binary LDPC codes started with the

work of Gallager [1] where arbitrary-alphabet LDPC codes
using modulo-q arithmetic were introduced. A generalization
increase in the size of buses connecting the computation units,
thus increasing the difficulty of optimally routing these buses
while keeping the area consumption to a minimum.
of binary LDPC codes and their decoding algorithm to finite Due to this, using codes with some regularity is even more
fields (Galois Fields,) of GF(2𝑚 ) type have been presented important and carries more benefits in the case of LDPC
in [2]. For this class of LDPC codes the non-zero entries in over GF(2𝑚 ) than in the binary case. For this reason quasi-
the parity check matrix H are elements in GF(2𝑚 ). It has cyclic codes [7] are considered. The idea of using quasi-
been shown how, with properly chosen values for the entries, cyclic codes over GF(2𝑚 ) is also considered in [8].
random LDPC codes over GF(2𝑚 ) out-perform binary LDPC
codes and turbo codes [2]. It is also known how increasing
the order of the Galois field increases the performances of III. Q UASI - CYCLIC LDPC C ODES
the codes. A detailed analysis under iterative decoding of A quasi-cyclic code of index 𝑡 is a linear block code C
coset LDPC codes over GF(q) when used over arbitrary in which a cyclic shift of any codeword in C by 𝑡 positions
discrete-memoryless channel ( in particular for non-binary is also a codeword. The generator matrix G for these codes
and asymmetric channels) has been presented by Bennetan et can be decomposed in 𝑡 circulant matrices each of dimension
al. [3]. [𝑝 × 𝑝]. Each of these circulant matrices can be represented
This letter introduces an extension over GF(2𝑚 ) of a class by a polynomial in the field 𝐺𝐹 (2𝑚 )[𝑥]/(𝑥𝑝 + 1). These are
of quasi-cyclic binary codes. The suitability of these codes called the generator polynomials. The parity check matrix
with respect to encoding and decoding is discussed. The associated with a quasi-cyclic code C is also composed of
encoding benefits, especially, are described and the building circulant matrices. Hence a parity check matrix H for a quasi-
block for encoding circuits is presented. A discussion on the cyclic code, of dimension 𝑀 × 𝑁 can be constructed by
performance advantages of quasi-cyclic over GF(2𝑚 ) LDPC assembly circulant matrices, each of dimension 𝑝 × 𝑝 with
code when compared with longer binary codes is presented. 𝑡 × 𝑝 = 𝑁 and stacking vertically 𝛼 circulant matrix with
Three different codes, from a very small code (𝑁 = 160) 𝛼 × 𝑝 = 𝑀.
to a medium size one (𝑁 = 2408) are considered. We
show how, even if the complexity grows, the use of quasi- A. Implementation aspects
cyclic GF(2𝑚 ) LDPC codes is of interest due to the error
correction capabilities of these class of codes, and the fact Implementation of binary quasi-cyclic LDPC codes has
that they are comparable to binary codes which have much been studied in many previous works. Reduced complexity
higher routing complexity. encoding scheme that takes advantage of the representation of
such codes has been presented in [9]. Decoding architectures
Paper approved by T.-K. Truong, the Editor for Coding Theory and Tech- that capitalize on the structure of the H matrix have been
niques of the IEEE Communications Society. Manuscript received December presented in [10], [11]. The advantages associated with binary
17, 2007; revised August 20, 2008 and January 6, 2009 quasi-cyclic LDPC codes can be carried over in the case of
The authors are with the Dept. of Electrical and Electronic Engineering,
University College Cork, Ireland (e-mail: {c.spagnol, l.marnane}@ue.ucc.ie). the GF(2𝑚 ) quasi-cyclic codes since the structure of the H
Digital Object Identifier 10.1109/TCOMM.2009.09.070644 and G matrices is unchanged. In the case of GF(2𝑚 ) codes
0090-6778/09$25.00 ⃝
c 2009 IEEE
SPAGNOL and MARNANE: A CLASS OF QUASI-CYCLIC LDPC CODES OVER GF(2𝑀 ) 2525

trary number of generators is a combination of the architecture

presented here.

B. A class of binary Quasi-cyclic codes

In this section a particular family of quasi-cyclic matri-
ces [12] is presented. These codes have 𝑡 = 2𝛼 fixing the
rate at 1/2. The structure of such codes is:
⎡ ⎤
𝐻11 0 ⋅⋅⋅ 0 𝐼 𝐻12 𝐼 0 ⋅⋅⋅ 0
⎢ ⎥
⎢ . ⎥
⎢
⎢ 𝐼 𝐻21 0 ⋅⋅⋅ 0 0 𝐻22 𝐼 .. 0 ⎥
⎥
⎢ .. .. .. .. .. .. .. ⎥
H =⎢
⎢ 0 𝐼 . . . . 0 . . . ⎥ (1)
⎥
⎢ ⎥
⎢ .. .. .. .. .. .. .. ⎥
⎢ . . . . 0 0 . . . 𝐼 ⎥
⎣ ⎦
0 ⋅⋅⋅ 0 𝐼 𝐻𝛼1 𝐼 0 ⋅ ⋅ ⋅ 0 𝐻𝛼2
Where 0 indicates the 𝑝 × 𝑝 zero matrix and 𝐼 the 𝑝 × 𝑝
identity matrix. It is evident that these two particular matrices
are circulant matrices. The matrices 𝐻𝑖𝑐 with 𝑐 ∈ {1, 2} and
𝑖 ∈ {1, ⋅ ⋅ ⋅ , 𝛼} are circulant matrices generated by a binomial
𝑐 𝑐
𝑝𝑐𝑖 = 𝑥𝑎𝑖 + 𝑥𝑏𝑖 . Where 𝑎𝑐𝑖 and 𝑏𝑐𝑖 represent the position of the
𝑐
hh ones in the 𝐻𝑖 matrix. The choice of using a binomial instead
of a generic polynomial, fixes the code to be a (3,6) regular
Fig. 1. Design schematic of a 1-generator quasi-cyclic encoder, 𝑡 = 2. The
message word u is loaded serially.
LDPC code. Moreover fixing the value of 𝑝 the dimension of
the code can be changed by the value of 𝛼 and vice-versa.
This family of H matrices has been chosen because if the
polynomials representing the circulant matrices satisfy certain
the decoder will have the same structure but the computational conditions then the Tanner graph does not contain any cycle of
blocks will be different and the encoder will be required to length less than eight, as demonstrated in [12], thus performing
compute operations in GF(2𝑚 ). We present how the scheme better than other quasi-cyclic codes with smaller girth. It
of the encoder for quasi-cyclic LDPC codes can be easily is of particular interest that these conditions can be easily
modified to work for quasi-cyclic codes over GF(2𝑚 ). The met even for small codes. Hence it is possible to construct
hardware implementation of 1-generator quasi-cyclic codes, short codes suitable for implementation on devices where the
with rate 1/2, will be discussed in this section. resources are restricted (e.g. wireless sensors) that achieve
The hardware encoder of a rate 1/2 1-generator quasi- higher performances than random codes of the same size. This
cyclic code with 𝑡 = 2 is shown in Fig. 1, where 𝑒𝑖 represents particular type of codes is of interest because they provide
the 𝑖-th entry value of the first row of the circulant matrix good performance and are hardware friendly.
that forms the G matrix. In this case the storage requirement
is one 𝐾-stage shift register (for a rate 1/2 1-generator quasi- IV. P ERFORMANCE OF QUASI - CYCLIC LDPC OVER
cyclic code we have 𝑝 = 𝐾). In general, for 1-generator GF(2𝑚 )
quasi-cyclic code with 𝑡 > 2, (𝑡 − 1) 𝑝-stage shift registers NOTE: All simulations presented here have been obtained
are required. The generator matrix of the code is dense, so using a log likelihood domain decoder with a maximum limit
it can be expected that approximately half the entries in each of 15 repetitions. For every SNRdB point a minimum of 20
row of the parity part of G are non-zeros entries. Therefore, block errors has been found. Error bars are plotted on plots in
the logic requirement for the encoder is 𝑝/2 multipliers over this paper. The error associated with the Frame Error rate is
GF(2𝑚 ) and a block that compute the addition of 𝑝/2 GF(2𝑚 ) defined in [13]. We consider this an important aspect of the
values. In general, for 1-generator quasi-cyclic dual code with performances graph often overlooked in publications.
𝑡 > 2, (𝑡 − 1) 𝑝/2-input addition ports and 𝑝/2(𝑡 − 1) In this section performance results for some quasi-
multipliers are required. cyclic LDPC code over GF(2𝑚 ) are presented. Three different
In Fig. 1 the message word is read in serially, so it takes codes, from a very small code (𝑁 = 160) to a medium size
𝐾 clock cycles to fully load the register and another 𝐾 clock one (𝑁 = 2408) are investigated. We consider the class of
cycles to produce the parity part of the codeword. The data quasi-cyclic LDPC codes presented in previous section due
input rate of a serial loaded decoder is equal to half the data to their well known structure. A GF(2𝑚 ) code with the same
output rate. An encoder where the message word is loaded dimension of a quasi-cyclic binary code is obtained by main-
into the register in parallel requires just one clock cycle to taining the structure of the H matrix but changing the binary
load the message word and 𝐾 clock cycles to produce the coefficients of the polynomials that represent the circulant
rest of the codeword. matrices with elements in GF(2𝑚 ). In this contribution such
The hardware encoder for quasi-cyclic code with an arbi- elements are chosen randomly using a uniform distribution.
2526 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 57, NO. 9, SEPTEMBER 2009

Fig. 2. Performance comparison of a N=808 r=1/2 quasi-cyclic codes GF(8) Fig. 3. Performance comparison of a N=160 and N=480 quasi-cyclic codes
against a quasi-cyclic code of same length and two codes three times longer GF(8) against quasi-cyclic codes of same length and two codes three times
(𝑁 = 2424) one random and one quasi-cyclic longer

A. Comparison with equivalent binary length codes

In the following figures the performances of the
GF(2𝑚 ) codes are compared with the binary codes from which
the codes have been obtained and with codes (quasi-cyclic and
randomly generated) with dimension 𝑚 times bigger. Such
decision is justified by the consideration that the channel
under test is binary and the codewords are transmitted using
BPSK modulation for both types of codes. The output of
the GF(2𝑚 ) encoder must be converted in binary before
modulation and this requires 𝑚 bits for each GF(2𝑚 ) coded
symbol. The total number of {−1, 1} symbols sent trough
the channel is 𝑚𝑁 . Hence from a channel point of view the
GF(2𝑚 ) code has equivalent binary length of a 𝑚-times longer
binary code.
Fig. 4. Performance comparison of a 𝑁 = 808, 𝑟 = 1/2 quasi-cyclic codes
Fig. 2 presents a comparison between a binary quasi- over GF(8) versus a quasi-cyclic binary code with 𝑁 = 6464, 𝑟 = 1/2 and
cyclic LDPC code with 𝑁 = 808 and the GF(23 ) code a randomly generated code with 𝑁 = 6464, 𝑟 = 1/2. The random code is
obtained from it. Fig. 2 also presents the comparison be- a regular (3, 6) obtained with the optimal permutation method. The quasi-
cyclic code is of the same structure of the smaller code.
tween the performance of the GF(23 ) with 𝑁 = 808 quasi-
cyclic LDPC code and the performances of two (3,6) binary
codes of dimension 𝑁 = 2424, one quasi-cyclic and the outperform both the binary code with same length and binary
other random like. The latter has been generated using the codes with three times bigger dimension.
optimal permutation method. We can see how the quasi- An interesting characteristic of the performances presented
cyclic GF(23 ) code performs better than the original binary in the previous figures is that the quasi-cyclic GF(23 ) codes
code and of the two binary codes of dimension three times don’t show any sign of the error floor that can be seen for the
bigger. binary quasi-cyclic codes.
To verify if the considerations presented for the previous
code are valid for different length codes and different size
fields, a really small code (𝑁 = 160) and a code of medium B. Comparison with 2𝑚 longer codes
length (𝑁 = 2408) are considered in Fig. 3. The code with From a computational point of view the decoding algorithm
𝑁 = 2408 is defined over GF(23 ) and the code with 𝑁 = 160 for a GF(2𝑚 ) codes is more than 2𝑚 times more demanding.
is defined over GF(24 ). The comparison of such GF(2𝑚 ) codes For this reason a comparison of the performances of the quasi-
with binary codes with the same length and with three/four cyclic code over GF(23 ) with 𝑁 = 808 versus a binary quasi-
times longer codes (random and quasi-cyclic) confirms the cyclic code of dimension 𝑁 = 6464 code obtained from the
observations made previously. same construction is presented in Fig. 4. In the same figure
Even with a small code really high gain can be achieved the performances of a randomly generated code of length
against the same length binary code. It can also be seen how 𝑁 = 6464 can also be seen. The graph suggests that when the
the quasi-cyclic code over GF(24 ) of dimension 𝑁 = 160 decoding complexity of the codes is similar the GF(23 ) code
performs close to a 𝑁 = 2408 binary quasi-cyclic code. still performs slightly better than the eight times longer binary
Finally, also for the code with length 𝑁 = 2408 a considerable codes. This consideration might not necessarily be true for
gain in dB can be achieved using a GF(23 ) code. The code other random codes generated with more advanced methods,
SPAGNOL and MARNANE: A CLASS OF QUASI-CYCLIC LDPC CODES OVER GF(2𝑀 ) 2527

but it is an important result that increases the interest on quasi- ACKNOWLEDGMENTS

cyclic over GF(2𝑚 ) codes for hardware implementations. In Special thanks to the I.R.C.S.E.T. that supported this
fact, even if the two codes have the similar computational project. The authors would like to thank the Boole Centre
requirement there are other aspects to consider. It has been for Research in Informatics and the Irish Centre for High-End
mentioned how one major problem of implementing LDPC Computing for the use of their computer clusters. Moreover
decoders is the complexity of routing the wires. This com- we would like to thank M. O’Sullivan for providing opinions
plexity increases with the number of edges in the graph. All and assistance during the work.
codes considered are regular (3,6) codes hence the number
of edges on the binary code is eight times bigger than the R EFERENCES
number of edges for GF(23 ) code since it has three times
[1] R. Gallager, “Low density parity check codes,” 1963, Ph.D. thesis, MIT
more variables nodes. From an implementation point of view Press, Cambridge, MA.
this means more complex routing, memory addressing, ROM [2] M. Davey and D. MacKay, “Low-density parity-check codes over
size etc. Similar reasoning regarding the memory requirement GF(q)," IEEE Commun. Lett., vol. 2, no. 6, pp. 165-167, 1998.
[3] A. Bennatan and D. Burshtein, “Design and analysis of nonbinary LDPC
by the encoder can be done. codes for arbitrary discrete-memoryless channels," IEEE Trans. Inform.
These considerations together with the good performances Theory, vol. 52, no. 2, pp. 549-583, Feb. 2006.
of quasi-cyclic GF(23 ) codes justify the interest in such codes [4] C. Spagnol, E. Popovici, and W. Marnane, “Hardware implementation
of GF(2𝑚 ) LDPC decoders," to appear in IEEE Trans. Circuit Systems
even if the computational complexity of the decoder for LDPC I.
over GF(23 ) codes is higher. [5] A. Voicila, D. Declercq, F. Verdier, M. Fossorier, and P. Urard, “Low-
complexity, low-memory EMS algorithm for non-binary LDPC codes,"
in Proc. IEEE International Conf. Commun. ICC, pp. 671-676, 2007.
[6] C. Blanksby and A. J. Howland, “A 690-mw 1-Gb/s 1024-b, rate-1/2
V. C ONCLUSION AND F UTURE W ORK low-density parity-check code decoder," IEEE J. Solid-State Circuits,
vol. 37, no. 3, pp. 404-412, 2002.
In this contribution a new family of GF(2𝑚 ) LDPC codes,
[7] R. Townsend and J. E. J. Weldon, “Self-orthogonal quasi-cyclic codes,"
obtained from binary quasi-cyclic LDPC, has been intro- IEEE Trans. Inform. Theory, vol. 13, no. 2, pp. 183-195, 1967.
duced. It has been showed how the quasi-cyclic LDPC over [8] L. Zeng, L. Lan, Y. Y. Tai, B. Zhou, S. Lin, and K. Abdel-Ghaffar,
GF(2𝑚 ) codes outperform binary codes up to 2𝑚 times “Construction of nonbinary cyclic, quasi-cyclic and regular LDPC
codes: a finite geometry approach," IEEE Trans. Commun., vol. 56,
longer. Moreover we presented how the encoder for binary no. 3, pp. 378-387, 2008.
quasi-cyclic LDPC codes can be modified to encode quasi- [9] L. C. Zongwang Li, L. Zeng, S. Lin, and W. H. Fong, “Efficient encoding
cyclic LDPC over GF(2𝑚 ) codes. In this study the entries of quasi-cyclic low density parity check codes," IEEE Trans. Commun.,
vol. 54, no. 1, pp. 71-81, 2006.
of the H matrix have been chosen randomly and the results [10] C. Spagnol, E. Popovici, and W. Marnane, “Reduced complexity, FPGA
obtained are encouraging. Of interest for future studies is to implementation of quasi-cyclic LDPC decoder," in Proc. IEEE European
investigates methods to optimize such choice. The possibility Conf. Circuit Theory Design (ECCTD), vol. 1, pp. 289-292, 2005.
[11] Y. Chen and K. K. Parhi, “Overlapped message passing for quasi-cyclic
of increasing the performances of the codes just by picking low-density parity check codes," IEEE Trans. Circuits Syst. I, vol. 51,
optimal values for the coefficients is of great interest because no. 6, pp. 1106-1113, 2004.
this will allow to obtain better codes without increasing the [12] M. Rossi and M. Sala, “On a class of quasi-cyclic codes," in Proc.
Effective Methods Algebraic Geometry Conf. (MEGA), 2005.
overall complexity. Methods that will be investigated are [13] D. MacKay, S. Wilson, and M. Davey, “Comparison of constructions of
optimal row-column weight, maximal entropy and application irregular Gallager codes," IEEE Trans. Commun., vol. 47, no. 10, pp.
of the density propagation algorithm. 1449-1454, 1998.