Documente Academic
Documente Profesional
Documente Cultură
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2016.2537812, IEEE
Communications Letters
Abstract—A modified bit-reliability based decoding algorithm ISRB (IISRB) algorithm and its re-selection scheme (RS-
is presented based on a recent work by Huang et al.. For the IISRB) [9][10].
presented algorithm, only one Galois field element is passed and
exchanged along the edges of the Tanner graph. At variable Recently, Huang et al. proposed a weighted bit-reliability
nodes, full messages rather than extrinsic messages are processed based (wBRB) decoding algorithm [11][12] for NB-LDPC
to further reduce the computational complexity. At check nodes, codes. Different from the existing NB-LDPC decoders, the
only hard reliability is considered and the main operation
is to compute the check-sum and send one extrinsic symbol
wBRB algorithm updates the reliability in terms of bits rather
back to variable nodes. Simulation results show that, even with than field elements. Combined with distance-based coeffi-
lower complexity and less memory consumption, the presented cients, the wBRB algorithm can achieve good error perfor-
algorithm still can perform very closely to the original wBRB mance for NB-LDPC codes with various column weights, but
algorithm with low quantization bits (3 ∼ 4bits) when decoding the complexity and memory consumption are greatly reduced
the majority-logic decodable non-binary LDPC codes. For codes
constructed in high order fields, the presented algorithm can even
compared with the existing MP and MLgD algorithms.
outperform the original wBRB algorithm. This letter presents a modified algorithm, referred to as the
Index Terms—Hard information, LDPC decoding, majority- hard-information bit-reliability based decoding algorithm (HI-
logic decoding, non-binary LDPC codes, reliability-based. BRB) based on the work in [11]. Similarly, the presented algo-
rithm updates reliability messages bit-by-bit in the iterations.
I. I NTRODUCTION Our work however can be differentiated from that of [11] based
on the following facts:
First introduced by Gallager [1], and then investigated by
Davey and Mackay [2], non-binary low density parity check
• Full messages rather than extrinsic messages are passed
(NB-LDPC) codes are shown to outperform their binary coun-
from variable nodes to check nodes, which can save
terparts when decoded by message-passing (MP) decoding
integer operations introducing by the extrinsic informa-
algorithms, such as the q-ary sum-product algorithm (QSPA)
tion computing and thus can further reduce the decoding
and the FFT-QSPA [3]. However, the performance gain is
complexity;
obtained at the expense of extremely high decoding complexity
• At variable nodes, the reliability message for each bit
and thus it is difficult for hardware implementations in prac-
within a symbol is modified to its binary constellation
tice. In [4], Declercq et al. proposed an extended min-sum
signal, which actually corresponds to the bit-based hard
(EMS) algorithm, which can reduce both the computational
reliability. Such modification can save memory loads and
and the storage complexities. Several variations of the EMS
integer operations since no soft reliability is involved;
algorithm were proposed [5][6]. Besides the MP algorithms,
• At check nodes, the bit-reliability of the extrinsic infor-
there exists an important class of reduced-complexity decoding
mation is also changed to its hard counterpart and hence
algorithms for NB-LDPC codes, known as the majority-logic
can save much integer comparisons introducing by the
decoding (MLgD) algorithms, which is specially designed
min operations;
for the codes with heavy column weights. Related works
• Finally, an efficient non-uniform quantization scheme is
include the iterative MlgD algorithm [7] and the iterative
combined with the presented algorithm to further reduce
soft reliability-based (ISRB) algorithm [8] and the improved
the memory consumption.
Manuscript received September 15, 2015. The associate editor coordinating
the review of this paper and approving it for publication was H. Saeedi. This Simulation results show that the presented HI-BRB algo-
work was supported by NSF of China (No.61102090,61261023,61362010)
and NSF of Guangxi (No.2015GXNSFAA139290, 2014GXNSFBA118276, rithm can perform very closely to the original wBRB algorithm
2012GXNSFAA053217). with low quantization bits (3 ∼ 4bits) when decoding the NB-
X. Li is with the School of Electronic and Information Engineering, South LDPC codes with heavy column weights. More interestingly,
China University of Technology, Guangzhou, 510641, China. T. Qin, H. Chen,
Y. Sun and Q. Liang are with the School of Computer, Electronics and for the NB-LDPC codes constructed in high order fields (say,
Information, Guangxi University, Nanning, 530004, China. L. Luo is with larger than 64), the HI-BRB algorithm can even outperform
the School of Information Science and Engineering in Guangxi University the wBRB algorithm with a more detailed optimized distance-
for Nationalities, China.
(Corresponding author: Haiqiang Chen, e-mail: haiqiang@gxu.edu.cn) based parameter. Therefore, the presented algorithm can make
an efficient tradeoff between performance and complexity.
1089-7798 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2016.2537812, IEEE
Communications Letters
1089-7798 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2016.2537812, IEEE
Communications Letters
TABLE I
COMPUTATIONAL COMPLEXITIES REQUIRED FOR VARIOUS DECODING ALGORITHMS, WHERE δ = nγ = mρ
P (k)
for 0 ≤ j ≤ n−1 and 0 ≤ t ≤ ℓ−1. This modification can save 2) for 0 ≤ i ≤ m − 1, compute si = zj hi,j ;
integer operations introducing by the extrinsic information j∈Ni
3) if s = (s0 , s1 , ..., sm−1 ) = 0, exit the iteration;
computing thus it can further reduce the decoding complexity.
Secondly, hard reliability rather than soft reliability message 4) for 0 ≤ i ≤ m − 1, 0 ≤ t ≤ ℓ − 1 and j ∈ Ni , compute
(k) (k) the extrinsic information and its bit-based reliability
is used for each bit of ẽj→i . Let R̃ẽj→i,t denote the bit-
(k) (k) (k)
reliability of the ẽj→i,t Ṫhen we just set
X
σi→j = h−1 i,j hi,j ′ ẽj→i ,
(k) j ′ ∈Ni \j
R̃ẽj→i,t = φ(ẽj→i,t ), i ∈ Mj (9)
and
for 0 ≤ j ≤ n − 1 and 0 ≤ t ≤ ℓ − 1. Such modification can R̃σ(k)
(k)
= φ(σi→j,t ) ;
i→j,t
save temporary memory loads and some integer operations,
since no soft reliability is involved in the iterations. 5) for 0 ≤ j ≤ n − 1 and 0 ≤ t ≤ ℓ − 1, compute the
(k)
Thirdly, since each bit-based reliability of ẽj→i is set to be bit-based weighted extrinsic information
φ(ẽj→i,t ), the reliability with respect to the extrinsic informa-
(k)
ξ˜j,t =
X
(k)
tion σi→j is consequently changed to its hard counterparts. θd(σ(k) ,z(k) ) R̃σ(k)
i→j,t
, (11)
i,j j
(k) (k) (k) i∈Mj
Let (R̃σi→j,0 , R̃σi→j,1 , · · · , R̃σi→j,ℓ−1 ) denote the bit-based
(k) and update the reliability information bit-by-bit
hard reliability vector of σi→j , then each component can be
computed as follows: (k)
(k)
Rz(k+1)
j,t
= Rz(0)
j,t
+ ξ̃j,t ;
R̃σ(k)
i→j,t
= φ(σi→j,t ), j ∈ Ni (10)
6) increment k by 1.
for 0 ≤ i ≤ m−1 and 0 ≤ t ≤ ℓ−1. Compared to the original (k) (k) (k)
• Output (z0 , z1 , · · · , zn−1 ) as the decoded codeword.
wBRB algorithm, this modification can reduce much integer
comparison operations that are related to the min operations
C. Complexity and Memory Consumption
at check nodes. Moreover, the memory consumption is also
reduced since we do not require to store the soft reliability in At check nodes, we need m(ρ − 1) Galois field addi-
the iterations. tions (GA) and δ Galois field multiplications (GM) (δ =
Based on the description above, the presented HI-BRB mρ = nγ) to compute the m check-sums. To compute the
(k),
algorithm can be summarized as follows. extrinsic information σi,j s, it requires δ GAs and δ GMs.
Algorithm 1 (The Hard Information Bit-reliability Based At variable nodes, it requires δℓ integer comparisons (IC) to
Decoding Algorithm - HI-BRB ): compute the distance vector and δℓ real additions (RA) to
(k),
• Input compute the weighted extrinsic information ξ˜j,t s. To update
received vector (y0 , y1 , · · · , yn−1 ) , quantization parameters the reliabilities of n variable nodes, we need δℓ RAs. The total
r, b and maximum iteration number Imax . computational complexities are shown in Table I. We observe
• Initialization that the complexity of the presented algorithm is equivalent
1) quantize each component yj into the integer vector qj ; to that of IHRB, but is lower than that of other algorithms.
2) set the loop counter k = 0; Besides this, the HI-BRB algorithm also has low memory
(0)
3) for 0 ≤ j ≤ n − 1 and 0 ≤ t ≤ ℓ − 1 , set Rzj,t = qj,t . consumptions. Firstly, we do not need any memory to store the
• Iterations: while k < Imax , do the following steps. soft reliabilites in the iterations. Secondly, it is shown in the
(k) next section that the HI-BRB algorithm can work efficiently
1) for 0 ≤ j ≤ n − 1 and 0 ≤ t ≤ ℓ − 1, compute zj
where at very low quantization bits (3 ∼ 4bits), which is significant
(
(k) (k) for extremely memory-limited scenarios.
zj,t = 0, Rzj,t > 0 Note that the IISRB and the RS-IISRB algorithm have less
(k) (k) ;
zj,t = 1, Rzj,t ≤ 0 IA/IC/RA/RC operations than the presented algorithm at each
1089-7798 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2016.2537812, IEEE
Communications Letters
0 0
10 10
wBRB−H wBRB−H
2.6,1.5,0.3 2,1,0.5
−1 wBRB Un−12bits −1 wBRB Un−12bits
10 2.4,1.8,0.2 10 5,5,1/3
IHRB−MLgD IHRB−MLgD
−2 ISRB−MLgD Un−12bits −2 ISRB−MLgD Un−12bits
10 10
IISRB−MLgD Un−12bits IISRB−MLgD Un−12bits
RS−IISRB−MLgD Un−12bits RS−IISRB−MLgD Un−12bits
−3 −3
10 HI−BRB NonUn−3bits 10 HI−BRB NonUn−4bits
2,2,0.8,0.7,0.5 5,5,1.1,0.7,0.7
HI−BRB NonUn−3bits
3,3,0.6,0.4,0.4
BER
BER
−4 −4
10 10
−5 −5
10 10
−6 −6
10 10
−7 −7
10 10
−8 −8
10 10
4.5 5 5.5 6 6.5 4 4.5 5 5.5 6
E /N (dB) E /N (dB)
b 0 b 0
Fig. 1. Performances of the PEG-C32 [1800, 1200] with column weight 9. Fig. 2. Performances of the EG-C256 [255, 175] with column weight 16.
1089-7798 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.