New Trends of High-Throughput Forward Error Correction: Polar Coding

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/328062453
New Trends of High-Throughput Forward Error Correction: Polar

Coding
Presentation · September 2018
CITATIONS READS
0 197
1 author:
Toshiaki Koike-Akino
Mitsubishi Electric Research Laboratories (MERL)
410 PUBLICATIONS 2,396 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Applied Deep Learning View project
Brain-Machine Interface View project
All content following this page was uploaded by Toshiaki Koike-Akino on 04 October 2018.
The user has requested enhancement of the downloaded file.

MITSUBISHI ELECTRIC RESEARCH LABORATORIES
Cambridge, Massachusetts
New Trends of High-Throughput Forward Error Correction:

Polar Coding
Toshiaki Koike-Akino
Mitsubishi Electric Research Laboratories (MERL), Cambridge, MA 02139, USA
© MERL
Outline
• Polar Coding
– Successive cancellation list (SCL) decoding + cyclic-redundancy check (CRC)
– Polar codes vs. low-density parity-check (LDPC) codes
– Computational complexity u x1
1 + + +
• Polar design for bit-interleaved coded modulation (BICM) u2 + + x2
– Interleaver design for quadrature-amplitude modulation (QAM) u3 + + x3
– EXIT-based frozen design u4 + x4
– Polarization rate analysis + +
– Non-uniform shaped QAM +
+
• Polar-based turbo product codes (TPC) uN xN
– Highly-parallel and pipelining processing
– Related works based on BP decoding Stage 1st 2nd mth
– SCL-based soft-in soft-output decoding
• Irregular polar coding
– Pruning polarization units
– Complexity & latency reduction
Information Row
• Summary k kxk Parity
– Irregular polar-TPC, capable of highly parallel decoding
– Lower complexity; 1/4 complexity
– Lower latency; 1/10 latency
– Higher throughput; 256-times faster Parity
Column
of
Parity
Parity
© MERL 9/19/18: T. Koike-Akino: Polar Coding 2

DCN Technology Evolution
• Higher speed: 400Gb/s and beyond
> 400G
[Zhou-Liu ECOC17]
Forward Error Correction (FEC) Codes
• Hard-decision (HD): RS, BCH, …
• Soft-decision (SD): Spatially-coupled, TPC (A), LDPC-C (B,C), Polar code
• Rateless: LT, Raptor, S-K feedback [ECOC2018]
(C)

Why Polar Codes?
• Arikan proposed in 2008:
– Capacity-achieving code in arbitrary discrete memoryless channels with proof
– Low-complexity encoding and decoding; Cooley-Tukey-like butterfly architecture
– Flexible in code rates with frozen bit selection
– 5G standard
Frozen bits u0 + + + +
c0
u1 + + +
c1
u2 + + + c2
u3 + +
c3
u4 + + +
c4
u5 + +
c5
u6 + +
c6
u7 +
c7
u8 + + + c8
u9 + + c9
u10 + + c10
u11 + c11
u12 + + c12
u13 + c13
u14 + c14
u15 c15
Polarization Phenomenon
• Polar kernel polarizes messages into ‘bad’ and ‘good’ sub-channels
• Proportion of good sub-channels approaches channel capacity
u0 0.9375
+
0.75
+
0.5 c0
u1 0.4375 + 0.25 0.5 c1
u2 0.5625 0.75
+
0.5
c2
u3 0.0625 0.25 0.5

c3

Successive Cancellation (SC) Decoding
• Log-linear decoding complexity: N log2(N)
• Capacity achieving in long codes
• Disappointing performance compared to state-of-the-art LDPC codes
– Error propagation
– Very long codes are required: Decoding latency issue
Polar
Long
Polar
LDPC

Successive Cancelation List (SCL) Decoding + CRC
• Recent breakthrough to be competitive against LDPC codes [Tal-Vardy 2015]
– List decoding to prevent error propagation
– Cyclic-redundancy check (CRC) to validate codeword in the list

Polar vs. LDPC Codes (4QAM, List-32, Ite-32)
• Systematic polar+CRC vs. Pareto-optimal LDPC codes [KoikeAkino OFC 2017]
32 ite. 32 list
256
1024
LDPC
4096
16384 Polar

Polar vs. LDPC Codes (4QAM, List-4, Ite-4)
• Polar codes can outperform LDPC codes for lower complexity and latency regimes
4 ite. 4 list
256
1024
LDPC
Polar
16384 4096

Outline
• Polar Coding
1 + + +
– Interleaver design for high-order quadrature-amplitude modulation (QAM) u3 + + x3
– Shaped QAM + +
– Polarization rate analysis +
+
Information Row
– Fading performance in wireless OFDM k kxk Parity
• Summary
– Lower latency; 1/5 latency Parity
Column
of
Parity

Polar-Coded BICM for High-Order QAM
• High-order modulation has non-uniform reliability for different bit-plane significance
• Interleaver is used for BICM
– Interleaver does not always work for polar codes because polarization speed is affected
– Appropriate interleaver design can improve polarization speed
[KoikeAkino OFC17]
Polar Coding Interleaver

u0 + + + + c0 LSB0
u1 + + + c1 MSB0
u2 + + + c2 LSB1
u3 + + c3 MSB1
u4 + + + c4 LSB2
u5 + + c5 MSB2
Data & Frozen Bits
u6 + + c6 LSB3
u7 + c7 MSB3
u8 + + + c8 LSB4
u9 + + c9 MSB4
u10 + + c10 LSB5
u11 + c11 MSB5
u12 + + c12 LSB6
u13 MSB6 256QAM
+ c13
u14 + c14 LSB7 (16PAM)
u15 c15 MSB7
Interleaver Design
• Rectangular block interleaver
– All combinations for the power-of-two numbers of columns and rows to design
• Quadratic permutation polynomial (QPP) interleaver

– Used for turbo codes in wireless communications standard
– Maximum contention free
– Few number of parameters to design
Write
Read
Block interleaver

Polar-Coded 256QAM
>0.7 dB loss
Optimized QPP
>0.5dB gain

Non-Uniform Multi-Level QAM
• Uniform QAM has shaping loss from optimal Gaussian signal distribution
• We propose geometric shaping for QAM with super-Gaussian non-uniformity
Gaussian distribution

Shaping Gain of Non-Uniform QAM
Shannon
Shaped
256QAM
256QAM

Polar-Coded 256QAM with Constellation Shaping
Uniform QPP
Shaped QPP 0.5dB gain
0.9dB gain

Design via EXIT (Extrinsic Information Transfer)
• Frozen bit location can be designed by density evolution (DE), while DE is
cumbersome for non-uniform bit reliability
• Gaussian approximation (GA) can simplify DE, assuming input and output
messages as Gaussian
• EXIT chart does not impose Gaussian assumption, thus more accurate than GA
I0 I0
I4 I4

Polarization Rate Boosting (256QAM)
• Joint interleaver and frozen bit location design can boost polarization speed
Random Interleave
Interleave
Interleave + Frozen

Outline
• Polar Coding
1 + + +
– Shaped QAM +
+
Information Row
Column
of
Parity
Parity

Computational Complexity
• Polar decoding requires nonlinear complexity: L N log2(N)/2
• LDPC BP decoding has linear complexity: 2 I dv N
dv=5
LDPC
dv=3
dv=2
Polar 4k 1M
We achieve:
- Longer polar performance
- Shorter polar complexity
- Parallel hardware implementation

Turbo Product Codes (TPC)
• Resolve the issue of parallelism in SCL decoding by polar product codes [OFC18]
• Highly parallel encoding and decoding are enabled; 256-times faster throughput
• Polar-TPC outperforms conventional BCH-TPC by 0.5dB
Column Constituent Code

Row Constituent Code Polar (N, k)
k
u1 + + x1
+
u2 + x2
+
Information Row u3 + + x3
k kxk Parity u4 x4
+
N
+ +
+
Parity +
Column
of uN
Parity xN
Parity
N Stage 1st 2nd mth
TPC (N, k)2 (N + k)th Polar

Polar-TPC Encoding and Decoding
• Massively parallel and pipeline processing
• Soft-decision output is generated by Chase N-Parallel
k-Parallel
Row Encoder
algorithm [Pyndiah TCOM1998], exploiting Column Encoder
survival lists in SCL decoder Polar 1
Polar 1
• N-times faster decoding throughput is possible Polar 2
Polar 2
– For (256, 239)2, we achieve 256-times speed-up
Polar k
Polar N
N-Parallel N-Parallel N-Parallel N-Parallel

Pipeline TPC Column
Row Decoder Decoder
Encoding Row Decoder Column Decoder
SCL 1 SCL 1 SCL 1 SCL 1
SCL 2 SCL 2 SCL 2 SCL 2

Turbo Turbo Turbo
Chase Chase Chase
a, b a, b a, b
Pipelining
SCL N SCL N SCL N SCL N
1st Iteration 2nd Iteration

Related Works: Concatenated Polar Codes
• With Hamming codes
– M. Seidl and J. B. Huber, “Improving successive cancellation decoding of polar codes by usage of
inner block codes,” in Proc. Int. Symp. Turbo Codes Iterative Inf. Process., pp. 103–106, Brest,
France, Sep. 2010.
• With LDPC codes
– J. Guo, M. Qin, A. Guillen i Fabregas, and P. H. Siegel, “Enhanced belief propagation decoding of
polar codes through concatenation,” in Proc. IEEE Int. Symp. Inf. Theory (ISIT), pp. 2987–2991,
Honolulu, HI, June 2014,
– A. Eslami and H. Pishro-Nik, “On finite-length performance of polar codes: Stopping sets, error floor,
and concatenated design,” IEEE Trans. Commun., vol. 61, no. 3, pp. 919–929, Mar. 2013.
– Y. X. Zhang and A. Liu, “Polar-LDPC concatenated coding for the AWGN wiretap channel,” IEEE
Commun. Lett., vol. 18, no. 10, pp. 1683–1686, Oct. 2014.
• With BCH codes
– Y. Wang, K. R. Narayanan, and Y.-C. Huang, “Interleaved concate- nations of polar codes with BCH
and convolutional codes,” IEEE J. Sel. Areas Commun., vol. 34, no. 2, pp. 267–277, Feb. 2016.
• With convolutional codes
– Q. Zhang, A. Liu, Y. Zhang, and X. Liang, “Practical design and decoding of parallel concatenated
structure for systematic polar codes,” IEEE Trans. Commun., vol. 64, no. 2, pp. 456–466, Feb. 2016.
• With Reed-Solomon codes
– H. Mahdavifar, M. El-Khamy, J. Lee, and I. Kang, “Performance limits and practical decoding of
interleaved Reed-Solomon polar concatenated codes,” IEEE Trans. Commun., vol. 62, no. 5, pp.
1406–1417, May 2014.
No Polar-TPC BP decoding based We propose SCL-based TPC

Polar-TPC Performance
• Polar-TPC(256, 239)2 vs BCH-TPC(256, 239)2 vs Polar(2562, 2392+16)+CRC16
Polar-TPC BCH-TPC
Polar+CRC

Error Floor Analysis with Importance Sampling (IS)
• Importance sampling (IS) can reduce the required number of simulation runs to
achieve high confidence compared to Monte-Carlo (MC) via weighted sample mean
• IS has been used for BCH-TPC analysis:
– M. Ferrari, S. Bellini, Importance sampling simulation of turbo product codes, ICC, 2001
BCH-TPC(31,21)2
MC
IS

BER Performance via IS
• Error floor was observed for polar-TPC(256, 239)2
BCH-TPC
Polar-TPC 2-ite
4-ite
2-ite
4-ite

BER Performance via IS: Error Floor Mitigation
• Error floor was removed with polar-TPC(256, 240)2 + BCH(2402, 2392)
256
239 16
2-ite
Polar-TPC + BCH Polar-TPC
4-ite
0.6dB

Outline
• Polar Coding
1 + + +
• EXIT chart design method u2 + + x2
– Non-uniform shaped QAM +
+
Information Row
Column
of
Parity
Parity

Irregular Polar Coding
• For LDPC codes, it is well-known that irregular degree distribution can
significantly improve performance over regular counterparts
• What happens
2 for polar coding with3 irregularity? CND
Parity check 1 1 1 1 0 0
4 5 Factor graph
H= 0 0 1 1 0 1
1 0 0 1 1 0 VND
EXIT Chart
VND
CND

Irregular Pruning of Polarization Units
• We proposed to inactivate polarization units in an irregular fashion [KoikeAkino
ECOC2017, GLOBECOM2017]
• We can reduce
– the computational complexity for both encoding and decoding; 30%~80%
– the decoding latency of SCL; 25%~95%
– the bit error rate (BER); a marginal gain
Factor Graph
Polarization Unit No computation
Degree-3
u0 + + c0 u0 + c0
u1 + c1 u1 + c1
u2 + c2 u2 + c2
u3 c3 u3 c3
Degree-2

Why Irregular Polarization Helps?
• Impact of inactivated polarization units
Irregular Polar Code: UB = 0.2344
Regular Polar Code: UB = 0.2432 Inactive
1st stage 2nd stage
u1 + + x1 u1 0.9375
+
0.75
+
0.5 x1
0.9375 0.75 0.5
u2 0.4375 +
0.25 0.5 x2 u2 0.25 +
0.25 0.5 x2
u3 0.5625 0.75
+
0.5
x3 u3 0.5625 0.75
+
0.5
x3
u4 0.0625 0.25 0.5

x4 u4 x4
0.25 0.25 0.5
Irregular Polar Code: UB = 0.2432 Irregular Polar Code: UB = 0.2383

Inactive Inactive
u1 0.75 +
0.75
+
0.5 x1 u1 + +
0.5 x1
0.875 0.5
u2 0.4375 +
0.25 0.5 x2 u2 0.625
+
0.5 0.5 x2
u3 0.75 0.75
+
0.5
x3 u3 0.375 0.75
+ x3
0.5
u4 0.0625 x4 u4 x4
0.25 0.5 0.125 0.25 0.5

Irregular Polarization Gain
• BER performance in BEC channels; better BER and lower complexity
1st stage 2nd stage
u1 + +
0.5 x1
0.9375 0.75
u2 0.4375 + 0.25 0.5

x2
Regular
u3 0.5625 0.5
+ x3
0.5 Inactive
u4 x4 u1 + + x1
0.0625 0.25 0.5 0.875 0.5 0.5
u2 + x2
0.625 0.5 0.5
Irregular u3 0.375 0.75

+ x3
0.5
u4 x4
0.125 0.25 0.5

Irregular Polar Code Design Method
• We proposed greedy design method using EXIT chart analysis to jointly design
frozen bit locations and inactive polarization units [GLOBECOM 2017]

Union Bound, Complexity, and Latency Analysis
• Polar (256, 239): Lower complexity, lower latency, & improved BER
Latency
Reduction
86% reduction
BER
Regular
54% reduction Uncoded
7dB SNR
TPC with Irregular Polar Codes
• Parallel SCL decoding:
– 256-times higher throughput
• Irregular pruning:
– 50~70% complexity reduction
– 80~90% decoding latency reduction
Polar decoding can be low-complex

Irregular Pruning
than most LDPC decoding

Irregular Polar-TPC Performance
• No performance loss with beyond 50% pruning, up to 72% reduction
~ 1/10 latency

Side Comment: Irregularity
• We can of course consider many other irregular structures not only pruning
– Grafting edges, mixed multi-kernel, edge looping, etc.
• Irregular polar codes: There are a lot of rooms to investigate this new family!

Summary
• We have investigated parallelizable polar codes for high-speed communications
• We briefly introduced recent advancement of polar coding, competitive to LDPC
– SCL+CRC decoding
• We evaluated joint interleaver and frozen design for BICM
– Interleaver design for shaped QAM achieves 0.9dB gain
• We proposed polar-TPC to achieve 256-times faster decoding throughput
– 0.5dB better than conventional BCH-TPC
– 0.2dB from optimal performance towards long polar+CRC code
• We further reduced complexity and latency by using irregular polar codes
– 72% complexity reduction
– 87% latency reduction
• ECOC2018: Polar-TPC for high-dimensional modulation, S-K feedback code, Lattice
precoding, Wyner-Ziv coding
MERL, Cambridge, Massachusetts, USA

Deep Learning for FEC
• Recent surge of machine learning field brought some potentials in FEC research
– [Gruber et al. 2017]
– [Cammerer et al. 2017]
– [Xu et al. 2017]
– [Doan et al. 2018]
– [Kim et al. 2018]
© MERL 9/19/18: T. Koike-Akino: Polar Coding

View publication stats 42

New Trends of High-Throughput Forward Error Correction: Polar Coding

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

New Trends of High-Throughput Forward Error Correction: Polar Coding

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

New Trends of High-Throughput Forward Error Correction: Polar

Presentation · September 2018

Applied Deep Learning View project

Brain-Machine Interface View project

The user has requested enhancement of the downloaded file.

New Trends of High-Throughput Forward Error Correction:

Mitsubishi Electric Research Laboratories (MERL), Cambridge, MA 02139, USA

© MERL 9/19/18: T. Koike-Akino: Polar Coding 2

© MERL 9/19/18: T. Koike-Akino: Polar Coding 5

u1 0.4375 + 0.25 0.5 c1

u3 0.0625 0.25 0.5

© MERL 9/19/18: T. Koike-Akino: Polar Coding 7

© MERL 9/19/18: T. Koike-Akino: Polar Coding 8

© MERL 9/19/18: T. Koike-Akino: Polar Coding 9

© MERL 9/19/18: T. Koike-Akino: Polar Coding 10

© MERL 9/19/18: T. Koike-Akino: Polar Coding 11

© MERL 9/19/18: T. Koike-Akino: Polar Coding 12

Polar Coding Interleaver

• Quadratic permutation polynomial (QPP) interleaver

© MERL 9/19/18: T. Koike-Akino: Polar Coding 14

© MERL 9/19/18: T. Koike-Akino: Polar Coding 15

© MERL 9/19/18: T. Koike-Akino: Polar Coding 16

© MERL 9/19/18: T. Koike-Akino: Polar Coding 17

© MERL 9/19/18: T. Koike-Akino: Polar Coding 18

© MERL 9/19/18: T. Koike-Akino: Polar Coding 19

© MERL 9/19/18: T. Koike-Akino: Polar Coding 20

© MERL 9/19/18: T. Koike-Akino: Polar Coding 21

© MERL 9/19/18: T. Koike-Akino: Polar Coding 22

Column Constituent Code

TPC (N, k)2 (N + k)th Polar

© MERL 9/19/18: T. Koike-Akino: Polar Coding 23

N-Parallel N-Parallel N-Parallel N-Parallel

SCL 2 SCL 2 SCL 2 SCL 2

1st Iteration 2nd Iteration

No Polar-TPC BP decoding based We propose SCL-based TPC

© MERL 9/19/18: T. Koike-Akino: Polar Coding 26

© MERL 9/19/18: T. Koike-Akino: Polar Coding 27

© MERL 9/19/18: T. Koike-Akino: Polar Coding 28

Polar-TPC + BCH Polar-TPC

© MERL 9/19/18: T. Koike-Akino: Polar Coding 29

© MERL 9/19/18: T. Koike-Akino: Polar Coding 30

© MERL 9/19/18: T. Koike-Akino: Polar Coding 31

© MERL 9/19/18: T. Koike-Akino: Polar Coding 32

u4 0.0625 0.25 0.5

Irregular Polar Code: UB = 0.2432 Irregular Polar Code: UB = 0.2383

© MERL 9/19/18: T. Koike-Akino: Polar Coding 33

1st stage 2nd stage

u2 0.4375 + 0.25 0.5

Irregular u3 0.375 0.75

© MERL 9/19/18: T. Koike-Akino: Polar Coding 34

© MERL 9/19/18: T. Koike-Akino: Polar Coding 35

Polar decoding can be low-complex

© MERL 9/19/18: T. Koike-Akino: Polar Coding 37

© MERL 9/19/18: T. Koike-Akino: Polar Coding 38

© MERL 9/19/18: T. Koike-Akino: Polar Coding 39

© MERL 9/19/18: T. Koike-Akino: Polar Coding 40

© MERL 9/19/18: T. Koike-Akino: Polar Coding

You might also like