Analysis of 5G LDPC Codes Rate-matching Design

Fatemeh Hamidi-Sepehr, Ajit Nimbalker, Gregory Ermolaev*

Next Generation and Standards

Intel Corporation
Santa Clara, USA *Nizhny Novogorod Russia
Email: {fatemeh.hamidi-sepehr, gregory.ermolaev}@intel.com

Abstract— This paper analyzes the rate-matching algorithms This article first discusses 5G NR LDPC code design,
for LDPC codes in 5G New Radio (NR) framework. Key features structure, and also related work from literature. Then, an
of NR LDPC rate-matching design includes flexible bit-selection investigation of fundamental aspects with respect to 5G LDPC
based on circular buffer operation, enhanced decoding latency rate-matching design and bit-selection algorithms is provided,
via code construction based on a single-parity-check extension
which is the main focus of this paper. The study is followed
from a high rate code, systematic bit puncturing, and support for
limited buffer rate-matching (LBRM). IR-HARQ is supported and concluded based on performance evaluations and analysis.
via several redundancy versions (RVs) defined non-uniformly While convolutional and Turbo coding schemes have been
over the transmit circular buffer, enabling possibility of multiple previously standardized (in 3GPP/IEEE) and are well-
self-decodable RVs. This paper, discusses NR LDPC structure, recognized [1]-[3], deep understanding of flexible 5G NR
design, and its key features, followed by an in-depth investigation LDPC codes supporting IR-HARQ, still benefits from detailed
of IR-HARQ functionality, rate-matching, and (re-)transmissions analysis and examination, as presented in this paper.
via different RV ordering selections. Performance of single
transmissions, as well as re-transmissions using different schemes II. DATA CHANNEL CODING CHAIN
are studied and evaluated using throughput and BLER metrics.
This study attests the flexibility and robustness of NR LDPC
This section discusses some of the key functionalities
design in addition to efficient support of IR-HARQ operation. supported by the NR data channel encoding chain, where
LDPC is applied for forward error correction (FEC). The
Keywords—5G, Channel coding; LDPC; Circular Buffer; processing entails an error detection/correction channel coding
Rate-matching; HARQ, Redundancy versions. scheme, rate-matching, interleaving, and transport channel or
mapping onto (de-mapping from) the physical channels. Data
I. INTRODUCTION AND BACKGORUND channel coding chain in 5G, includes the following steps:
3GPP is working on a New Radio design to develop 5G air • CRC attachment: For Transport Blocks (TBs) ≤ 3824
interface, expected to support multiple use cases including bits, a smaller 16-bit TB-level CRC is applied to enable
EMBB (Enhanced Mobile Broad-Band, supporting multi- additional error detection LDPC parity-check offers. For
TBs greater than 3824 bits, 24-bit CRC is applied. In case
Gbps throughputs for cloud technologies, virtual reality, etc.),
URLLC (Ultra-Reliable Low Latency communication), and of segmentation, 24-bit code block CRC is also applied.
MMTC (Massive Machine Type Communication for IoT). • Code block segmentation: The maximum LDPC code
Channel coding is a fundamental component being thoroughly block (CB) size is 8448 bits for base graph 1 and 3840
investigated to ensure the design can efficiently meet the bits for base graph 2 (defined in Section III). For TB sizes
stringent latency and/or ultra-high throughput requirements. larger than the maximum CB size, the TB is segmented
EMBB data channel - Channel coding is one of the essential into multiple CBs, each having a CB-level CRC attached.
building blocks of any communication system. Discussions in • Filler bits insertion: Filler bits (or zero-padding) are
3GPP NR design revolved comparison among three primary appended to each CB before LDPC encoding, to match
channel coding candidates (Turbo codes, Low-Density Parity the information block-size to the encoder input size.
Check (LDPC) codes, and Polar codes), to be employed for • LDPC encoding: Discussed in Section III.
EMBB data/control channels. Several aspects such as • Rate-matching and bit-selection: The main focus of this
performance, flexibility (in terms of the code-rates and block- paper, elaborated in Section IV.
sizes), support of HARQ including Incremental Redundancy • Bit-interleaving: A post rate-matching bit-interleaver is
(IR), implementation complexity (e.g. at multi-Gbps applied and it is limited to each code block individually.
throughputs), latency (e.g. faster turnaround than LTE), were A row-column interleaver with number of rows equal to
meticulously investigated. In particular, flexible LDPC was the modulation order, and row-wise write and column-
recognized as providing the best trade-off overall and agreed wise read is included at the output of the rate-matching.
as the coding scheme for NR EMBB data channel, while Polar This can improve the performance by enabling systematic
code was agreed for NR EMBB control channels. bit priority ordering for the first redundancy version.

978-1-5386-6355-4/18/$31.00 ©2018 IEEE

 III. LDPC FEATURES AND DESIGN the shift coefficient matrix for the corresponding set
LDPC codes [4,5] provide comparable or better performance (described later in this Section).
than turbo codes with lower decoding complexity [6], and Figure 1 presents the general structure of NR LDPC base
have been successfully incorporated into wireless standards graph. Sub-matrix A corresponds to the systematic bits. The
including WiFi (802.11n/11ac), 802.3an, WiMax (802.16e), as encoding process for the high-rate portion of the base graph
well as hard-disk industry [7-9]. LDPC code construction (the top left submatrix) is performed based on the dual-
provides large degrees of freedom, allowing matrix design to diagonal structure so that most parity bits can be obtained via
be tailored based on requirements, e.g. taking into account block-wise back-substitution (similar to the 802.16e/802.11n
throughput, latency, and performance trade-offs. Structured LDPC [7][8]), as shown by sub-matrix B in Figure 1. Sub-
LDPC codes based on submatrices (formed from shifted block B corresponds to the first set of parity bits; its 1st
Identity matrices and all zero-matrix) lend themselves to very column has weight 3, while its other columns have a dual
high throughput and low latency with simplified hardware diagonal structure. Sub-block C is an all-zero matrix.
implementations, based on efficient block-wise encoding and For efficient support of IR-HARQ, single parity-check (SPC)
decoding. This is specially achieved by parallelized operations based extension is used to support lower rates, as shown in
(check node update and variable node update), and use of Figure 1. Sub-block D corresponds to SPC rows, and E is an
layered belief propagation decoding to reduce the number of identity matrix corresponding to the second set of parity bits,
iteration, and Min-Sum algorithm (with a correction factor i.e., SPC extension. This is similar to the Raptor-like extension
based on offset or scaling)[10] (and references therein). as described in [16] (and reference therein).

Structured (also known as quasi-cyclic [11]) LDPC can

support flexible block sizes through support of suitable
submatrix sizes. Additionally, similar to LTE, shortening/zero-
padding can be applied for added flexibility. LDPC codes can
also offer additional benefits, e.g., CRC overhead reduction
since the parity-check criterion can be used for early stop of
decoding, and channel interleaver may not be strictly required Fig. 1. 5G NR LDPC base-graph structure.
as the LDPC encoding structure based on submatrices can
provide built-in interleaving. The LDPC codes for NR, support flexible information block-
size (maximum of 8448 bits), coarse shift sizes (Z) for
With respect to the IR-HARQ support, rate-compatible LDPC expansion ( Z max = 384 ), and filler bits (or zero-padding). In
codes can support flexible code-rates based on puncturing or
order to improve performance, built-in puncturing of the first
parity-check matrix extension. Rate-compatible codes refer to
2ˑZ systematic bits is also adopted.
a family of nested codes where the codeword bits for the
NR LDPC code is targeted to support at least 20 Gbps at code-
higher rates are embedded in the codewords of lower-rate.
rate 8/9, and to support good throughputs at other code-rates.
This enables encoding and decoding process using a single
Two base graphs (BG1, and BG2) are designed for better
encoder/decoder pair, and also achieves high throughput
performance optimization and improved decoding latency for
efficiency with a high degree of flexibility [12-14]. There is a
different range of block-lengths and code-rates:
rich trove of literature on rate-compatible LDPC design based
on density evolution and EXIT analysis (including [15,16] and • BG1 (dimension 46 × 68, K b = 22 systematic columns) is
references therein). Particularly, for various code-rates, rate- mainly designed for block-sizes within approximately 300
compatible LDPC codes of different block-sizes can use a to 8448 bits, code-rates between approximately 8/9 and
uniform hardware structure. Rate-matching and HARQ for NR 1/3, and further lower rates via repetition.
LDPC codes, is discussed in following sections. • BG2 (dimension 42 × 52, K b ∈ {6,8,9,10} systematic
columns) is mainly designed for block-sizes within 40 to
A. 5G NR LDPC Characteristics 3840 bits, code-rates between approximately 2/3 and 1/5,
An LDPC code is defined by a parity check matrix H and further lower rates via repetition.
( H .x T = 0 ), where x is the codeword. The parity check matrix The exact base graph usage (including block-size/code-rate),
for NR structured LDPC codes can be defined by the base as well as the definition of the two matrices are detailed in the
graph (BG), shift size Z, and shift coefficients ( Pij ), as follows NR standard specification TS 38.212 [17].
Shift size (Z) determination: NR LDPC codes support eight
(similar to 802.11n standard): 1s and 0s in BG are replaced by sets of shift sizes Z, and eight shift coefficient designs (one per
a circularly-shifted Identity matrix and zero matrix of size Z x
set, j = [1,...,8] ), defined by Z = a j × 2 K j , where
Z, respectively. The shift coefficients Pij are right cyclic shifts
applied to the identity matrix for (i,j)-th element in BG. The a j = [ 2, 3, 5, 7, 9, 11, 13, 15], K j = 0, 1, ... , K j , max ,
shift values are calculated using modulo K j ,max = [7, 7, 6, 5, 5, 5, 4, 4].
operation: Pij = mod (Vij , Z ) , where Vij is the (i,j)-th element of These provide hardware friendly implementation of shift
network, while providing good performance for the range of
code block-sizes and code-rates with both base graphs. Each
BG supports native block-sizes given by K b ⋅ Z , as well as the
code block-sizes in between via filler bits (zero-padding).
For BG1 and a given information block-size K, Z value is
selected as the smallest value such that K b ⋅ Z ≥ K . The
corresponding shift coefficient matrix is then chosen
according to the Z set. For BG2, Kb is first selected based on
the given information block-size K, from the defined set
above. Then, similar steps as for BG1 are followed for
determining shift size Z and the shift coefficient matrix.
Fig. 2. NR LDPC circular buffer and RV placements, for BG1.
IV. RATE-MATCHING
Following the overview of NR rate-compatible LDPC code Supporting more than one self-decodable RV (e.g., RV3 in
structure which enables efficient support of IR-HARQ, next addition to RV0), can provide benefits in certain
we discuss the NR rate-matching design and performance. retransmission scenarios. In some cases, the gNB (or 5G Base
Rate-matching for the NR LDPC codes is applied per coded station) may have ambiguity between a NACK (negative
block and consists of bit selection from the transmit-side acknowledgement1) and a DTX (discontinuous transmission2)
circular buffer depending on a redundancy version (RV) feedback from the user equipment (UE). Then, a good choice
index. LTE-like circular buffer-based rate-matching [2] for for gNB is to use a second self-decodable RV that can provide
IR-HARQ is supported for NR LDPC codes. Particularly, better HARQ combining gains (if initial transmission to the
circular buffer is filled with an ordered sequence of systematic UE was successful) or provide successful (i.e. self-decodable)
bits and parity bits. Each RV is assigned a starting bit location second transmission (if initial transmission to UE was lost). In
in the circular buffer. The RV enables the receiver to know case of a NACK feedback, it is preferred to retransmit a
where the received packet started in circular buffer and allows redundancy version with newer parity bits to achieve IR gains.
it to perform appropriate soft-combining operations. In case of a lost grant or an erased initial transmission, a self-
There are several ways to define redundancy versions. Since decodable retransmission is preferable. If the initial
LDPC code structure is aligned closely with the shift size Z, transmission is received at the UE, but the NACK feedback is
for simplicity of RV addressing, aligning the starting point in turned to a DTX, then self-decodable retransmission can
the circular buffer based on the base matrix dimension (i.e. achieve performance gains due to soft-combining. Evaluation
multiples of shift size Z) associated with the parity-check results showing self-decodability of different RVs for BG1
matrix can be beneficial. The RV starting locations are then and BG2, are presented in Section V.
multiples of Z, where the first RV address is always 0 i.e.
starting at the beginning of the circular buffer. Note that the B. Limited Buffer Rate-Matching (LBRM)
first 2ˑZ systematic bits are not entered into the circular buffer.
Typically, LDPC decoding latency depends on the number of
In NR, four RVs are specified at fixed locations in the circular
edges in the base graph used for decoding. Thus, transmissions
buffer and the coded bits are read out sequentially from the
at higher code-rates (e.g. 1st transmission or higher MCS) can
circular buffer, starting with the corresponding bit location
be decoded faster compared to the transmissions at lower rates
with wraparound if the end of circular buffer is reached.
(e.g. soft combined 1st and 2nd transmission or lower MCS).
A. Redundancy versions for LDPC Thus, the decoding latency for large packets can be reduced by
In NR, the fixed locations of RVs {0,1,2,3} are defined at limiting (via LBRM) the lowest rate supported by such
{0,17,33,56} x Z for BG1 (as depicted in Figure 2), and packets. This allows very high decoding throughput and lower
{0,13,25,43} x Z, for BG2, where the RVs (especially RV3) latency. LBRM is thus an important factor that not only affects
are placed at non-uniform locations. If the RVs were defined the UE complexity from the soft buffer perspective (or HARQ
uniformly spaced over the circular buffer, the start position for storage), but also facilitates decoding latency reduction.
RV3, would have been 50ˑZ for BG1 and 48ˑZ for BG2. LBRM can be handled by limiting the circular buffer size
However, defining the starting bit location of RV3 to be closer corresponding to the code blocks that belong to a large
to the end of the LDPC codeword, allows more overlap with transport block. This can be done by applying a limitation for
the systematic portion. Then, RV3 contains a part of the rate-matching on the circular buffer based on e.g. a reference
systematic bits as well as some extra parity bits which are not minimum coding rate (e.g. 2/3) for the largest transport block-
included in RV0. This results in RV3 being self-decodable up size schedulable for the UE [17]-[19]. When LBRM is
to a higher code-rate compared to the case it is defined applied, the start position of the RVs are scaled from the full
uniformly. Self-decodability refers to the property where the
decoder is able to recover the original information based on a
single transmission, i.e., without soft-combining with other 1
The TX knows the RX has received the data, but decoding error happened.
RV transmissions [18]. RV0 is self-decodable for the largest 2
Scheduling message (PDCCH) is missed at the receiver, and the transmitter
range of code-rates, compared to all other RVs. is not sure whether the receiver has received the data or not.
buffer positions (described in Section IV-A), while
maintaining alignment to integer multiples of Z.
C. RV order for special cases
Since adaptive HARQ is supported in NR for both uplink and
downlink, the redundancy version information will be signaled
along with the resource allocation (RA) information within the
Downlink Control Information (DCI) on the PDCCH. In this
sub-section, we discuss the RV order for special cases where
the RV index may not be explicitly signaled.
In some cases (e.g. uplink (UL) grant free transmission3), the
RV order for packet transmissions and retransmissions may
need to be predefined or configured by the higher layers (with
a possible default ordering sequence) [19].
Generally, RV0 is well-suited for the first transmission and
has good self-decodability (though at lower rates, all RVs may
have similar performance). Hence, RV0 can be used for
transmitting data such as paging or system information (SI) Fig. 3. Throughput (Mbps) vs SNR (dB) for different RV orders, BG1.
that are either used for 1-shot transmission or generally coded
down to the lower code-rates. For SI transmission requiring
soft-combining, an implicit RV order may be applied. In such
cases, implicit RV indicator (or RV order indicator) allows for
reduced DCI overhead and improved PDCCH performance. If
an explicit RV indicator (RV order indicator) is present in all
DCIs, then the gNB scheduler can indicate the RV in the
corresponding DCI.
Overall, when a single fixed RV value is needed for special
cases with no explicit RV index signaling, RV0 is the best
option for both BG1 and BG2. Regarding the predefined RV
order, evaluations in Section V reveal the best choice.
V. EVALUATION RESULTS
Performance evaluations can be assessed and presented
through BLER or throughput versus SNR curves. In this
Section, we present the results using these performance
metrics. QPSK modulation, with 50 iterations of Sum Product
Algorithm for LDPC decoding with flooding schedule, and Fig. 4. Throughput (Mbps) vs SNR (dB) for different RV orders, BG2.
parity check based early termination is used.
A. HARQ performance B. RVs performance and self-decodibility
Figure 3 compares the IR-HARQ throughput performance BG1 Evaluations: Figure 5 presents BLER results for one
with a maximum of four equal-sized transmissions using transmission using BG1 with different RVs, for block-size
different RV orders: [0231], [0213], [0123], and [0321], for 4224 bits. The BLER curves for the code-rates at which the
block-size 4224 bits, and a few code-rates, with BG1. Based RV is not self-decodable, are all ones. In order to show the
on these evaluations, the RV orders [0231] and [0213] show effect of RV spacing over the circular buffer, results with both
performance gains up to about 1.8dB (depending on the code- RV3 as defined in Section IV-A, and uniformly spaced RV3
rate) in the 2nd transmission SNR regime, compared to the are presented. Figure 6 summarizes the range of code-rates for
other two RV orders. Figure 4 presents similar results for self-decodability of different RVs for BG1, based on the
block-size 2560 bits, with BG2. RV orders [0231] and [0213] evaluated block-sizes – we note the values are approximate as
show up to about 1dB performance gain for medium rates in the rate at which self-decodability vanishes can depend on the
the 2nd transmission regime, compared to other two RV order, block-size. Note that the range of the rates indicates that RV0
for BG2 as well [19]. Thus, for both BGs, [0231] provides a and RV3 provide better self-decodability relative to RV2 and
good choice for RV order that can be used for special cases RV1.
described in Section IV-C. BG2 Evaluations: Figure 7 shows BLER results for different
RVs as defined in Section IV-A., for BG2, block-size of 2560
bits. Figure 8 summarizes the range of code-rates for self-
3
UL grant-free transmission reduces control signaling overhead of scheduling decodability of different RVs for BG2, based on the evaluated
procedure. For efficient radio resource utilization, higher layers indicate all block-sizes - we note the values are approximate as the rate at
the scheduling parameters, e.g., modulation and coding (MCS), RA, etc.
which self-decodability vanishes can depend on the block- VI. CONCLUSION
size. Note that the range of the code-rates indicates that RV0 In this paper, we analyzed and evaluated the circular buffer
and RV3 provide better self-decodability relative to RV2 and rate-matching and bit-selection algorithms for the LDPC codes
RV1. in the 5G air interface. Particularly, NR LDPC structure,
BlockSize = 4224
design, as well as its key features were first discussed.
100
r0.91, RV3 Nonuniform
Performance of HARQ re-transmissions was then studied and
r0.92, RV3 Nonuniform compared, using RVs defined over the circular buffer, spaced
r0.87 RV3 Uniform
r0.88 RV3 Uniform uniformly, and non-uniformly. Self-decodability of different
r0.55, RV2
r0.56, RV2 RVs, as well as predefined RV orders for the special cases
10-1
r0.43, RV1
r0.44, RV1
where RV index is not explicitly signaled, were also discussed
r0.97, RV0 and evaluated. The results show that the NR LDPC design is
r0.98, RV0
flexible, robust and supports efficient IR-HARQ operation.
REFERENCES
10-2 [1] 3GPP Technical Specifications 36.212, “Channel coding, multiplexing
and interleaving (Release 14),” 2017.
[2] Cheng, J. F., Nimbalker, A., Blankenship, Y., Classon, B., Blankenship,
T. K. “Analysis of circular buffer rate matching for LTE turbo code,”
In 68th Vehicular Technology Conference, 2008, IEEE, (pp. 1-5).
10-3 [3] Sesia, Stefania, Matthew Baker, and Issam Toufik. LTE-the UMTS long
2 3 4 5 6 7 8 9 10 11 12 13
SNR
term evolution: from theory to practice. John Wiley & Sons, 2011.
[4] R. G. Gallager, Low density parity check codes, (MIT press, Cambridge,
Fig. 5. BLER vs SNR (dB) for different RVs; BG1. MA, 1963).
[5] T. J. Richardson, A. Shokrollahi and R. Urbanke, “Design of capacity-
approaching irregular low-density parity-check codes,” IEEE Trans.
Inform. Theory, Feb. 2001, 619-637.
[6] R1-164182, “Comparison of coding schemes for NR” Intel Corporation,
3GPP TSG RAN WG1 Meeting 85, Nanjing, China, 23rd – 27th May
2016.
[7] IEEE Std 802.16-2004, “Part 16: Air Interface for Fixed and Mobile
Broadband Wireless Access Systems,” Oct. 2004.
Fig. 6. Self-decodability (approximate) range for different RVs; BG1.
[8] IEEE 802.11 Working Group (1999-07-15). IEEE 802.11-1999:
Wireless LAN MAC and PHY Specifications.
[9] Zhao, Kai, et al. "LDPC-in-SSD: making advanced error correction
codes work effectively in solid state drives." FAST. Vol. 13. 2013.
[10] Hocevar, Dale E. "A reduced complexity decoder architecture via
layered decoding of LDPC codes." In IEEE Workshop on Signal
Processing Systems, 2004. pp. 107-112.
[11] Tanner, R.M., Sridhara, D. and Fuja, T., “A class of group-structured
LDPC codes,” In Proc. ICSTA, 2001, pp.365-370.
[12] Mohammad R. Yazdani, Student Member, IEEE, and Amir H.
Banihashemi, “On Construction of Rate-Compatible Low-Density
Parity-Check Codes”, IEEE Communication letters, 2004.
[13] Ha, Jeongseok, and S. W. McLaughlin. "Optimal puncturing
distributions for rate-compatible low-density parity-check codes."
In Proceedings. IEEE International Symposium on Information Theory,
2003, pp. 233-233.
[14] Li, Jing, and Krishna R. Narayanan. "Rate-compatible Low Density
Parity Check Codes for Capacity-approaching ARQ Schemes in Packet
Data Communications." In Communications, Internet, and Information
Technology, 2002, pp. 201-206.
Fig. 7. BLER vs SNR (dB) for different RVs; BG2. [15] Liva, Gianluigi, and Marco Chiani. "Protograph LDPC codes design
based on EXIT analysis." In IEEE Global Telecommunications
Conference, 2007, pp. 3250-3254.
[16] Van Nguyen, Thuy, Aria Nosratinia, and Dariush Divsalar. "The design
of rate-compatible protograph LDPC codes." IEEE Transactions on
communications 60, no. 10 (2012): 2841-2850.
[17] 3GPP Technical Specifications 38.212, v1.0.1 “Channel Coding,” 2017.
[18] R1- 1716329, “Finalization of remaining details of rate-matching” Intel
Corporation, 3GPP TSG RAN WG1 Meeting NR #3, Nagoya, Japan,
18th – 21st Sep 2017.
Fig. 8. Self-decodability range (approximate) for different RVs; BG2. [19] R1-1717405, “RV order for special cases” Intel Corporation, 3GPP TSG
RAN WG1 Meeting 90bis, Prague, CZ, 9th – 13th, October 2017.