Received November 9, 2018, accepted December 2, 2018, date of publication January 9, 2019, date of current version January 29,

2019.
Digital Object Identifier 10.1109/ACCESS.2019.2891597

Toward High-Performance Implementation

of 5G SCMA Algorithms
ALIREZA GHAFFARI 1 , MATHIEU LÉONARDON 2 , ADRIEN CASSAGNE 2,3 ,

CAMILLE LEROUX 2 , AND YVON SAVARIA 1 , (Fellow, IEEE)

1 École Polytechnique de Montréal, Montreal, QC H3T 1J4, Canada
2 CNRS IMS Laboratory, Bordeaux INP, University of Bordeaux, 33400 Bordeaux, France
3 Inria, Bordeaux Institute of Technology, LaBRI/CNRS, 33405 Bordeaux, France

Corresponding author: Alireza Ghaffari ([email protected])

This work was supported in part by the Natural Sciences and Engineering Research Council of Canada, in part by the Huawei Canada, and
in part by the Prompt Quebec.

ABSTRACT The recent evolution of mobile communication systems toward a 5G network is associated
with the search for new types of non-orthogonal modulations such as sparse code multiple access (SCMA).
Such modulations are proposed in response to demands for increasing the number of connected users.
SCMA is a non-orthogonal multiple access technique that offers improved bit error rate performance and
higher spectral efficiency than other comparable techniques, but these improvements come at the cost of
complex decoders. There are many challenges in designing near-optimum high throughput SCMA decoders.
This paper explores means to enhance the performance of SCMA decoders. To achieve this goal, various
improvements to the MPA algorithms are proposed. They notably aim at adapting SCMA decoding to
the single instruction multiple data paradigm. Approximate modeling of noise is performed to reduce the
complexity of floating-point calculations. The effects of forwarding error corrections such as polar, turbo,
and LDPC codes, as well as different ways of accessing memory and improving power efficiency of modified
MPAs are investigated. The results show that the throughput of an SCMA decoder can be increased by
3.1 to 21 times when compared to the original MPA on different computing platforms using the suggested
improvements.

INDEX TERMS 5G, BER, exponential estimations, intel advanced vector extensions (AVX), iterative
multi-user detection, knights corner instruction (KNCI), log-MPA, maximum likelihood (ML), message
passing algorithm (MPA), power efficiency, SCMA, single instruction multiple data (SIMD), streaming
SIMD extension (SSE).

I. INTRODUCTION connectivity for Augmented Reality (AR) and reliable com-

Non-orthogonal Multiple Access (NOMA) mechanisms are munication for intelligent industrial controls, Internet of
investigated as means to improve the fifth-generation mobile Things (IoT) or Internet of Mission Critical Things (IoMCT).
communication systems (5G) [1] to realize massive connec- Unfortunately, massive connectivity and spectral efficiency
tivity and to reduce bit error rates. Sparse Code Multiple of SCMA come at the cost of high complexity in the decoder,
Access (SCMA) is a NOMA mechanism that offers better bit making the design of high throughput and low complexity
error rate performance and higher spectral efficiency, while decoders a challenge for systems exploiting SCMA [5].
the sparsity of the codebooks ensures lower complexity of Exploiting sparsity of the codebooks, Belief Propagation
decoding compared to other non-orthogonal modulations [2]. (BP) or Message Passing Algorithm (MPA) decoders were
SCMA is a promising candidate for 5G communication introduced to achieve near Maximum Likelihood perfor-
systems since it provides up to 300% more connectivity mance with lower complexity [6]. Substantial research works
by spreading information of each user’s codebook over were conducted on improving SCMA decoders to satisfy
sets of shared OFDM frequency tones [3]. According to the uplink requirements of 5G. Indeed, MPA is populated
the NGMN white paper [4], 5G is seriously considered to with many exponential computations to calculate the extrinsic
fulfill more diverse scenarios compared to 4G. Applica- information and probabilities of the received signal. This
tions can be broadband support in dense areas, low latency is based on modeling the channel noise with a Gaussian
2169-3536
2019 IEEE. Translations and content mining are permitted for academic research only.
10402 Personal use is also permitted, but republication/redistribution requires IEEE permission. VOLUME 7, 2019
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

probability density function (PDF). A classical improvement vectorized instructions. Therefore, complicated and control
to this bottleneck is the computation of extrinsic information heavy algorithms such as MPA have to be adapted for
in the logarithm domain, which led to develop the log-MPA efficient execution on heterogeneous architectures and their
decoder. In [7], fixed point and floating-point implemen- exploitable parallelism must be expressed at every level of
tations of the MPA and log-MPA on FPGA are studied. the code, whether in arithmetic or memory access instruc-
The bit error rate performance and complexity of the MPA tions. Particularly, various Single Instruction Multiple Data
and log-MPA are compared and it is concluded that using (SIMD) extensions and thread-level parallelism are used
log-MPA with 4 message passing iterations achieves a good to increase the throughput of MPA decoding on various
tradeoff between performance and complexity. In [8], several platforms.
complexity reduction techniques are proposed to increase the This paper reports on two contributions that can be useful
system throughput. These techniques are 1) SCMA code- for any variation of the aforementioned MPA. First, MPA
book design with minimum number of projections, 2) clus- and log-MPA have been adapted to use SIMD extensions
tered MPA (CMPA) which defines sub-graphs in MPA and leveraging the available data-level parallelism. The algo-
runs MPA on them, and 3) selected exponential computa- rithms are revised to have aligned and contiguous access to
tions. In [9] an adaptive Gaussian approximation is used memory, which is crucial to achieve high memory through-
to unselect the edges of the graph with smaller modulus. put. Various SIMD instruction set architectures (ISAs) such
In addition, mean and variance feedbacks are employed to as Advanced Vector Extensions (AVX), Streaming SIMD
compensate information loss caused by unselected edges. Extension (SSE), Knights Corner Instruction (KNCI) and
User’s codebooks play an important role for fast conver- ARM NEON are used to enhance the performance of var-
gence of the MPA or log-MPA. As investigated in [10]–[12], ious parts of the algorithm. Multi-threaded programming
revisiting codebook design can help to reduce the number technique and power efficiency are also studied in this
of iterations needed for MPA decoding of SCMA. In [13], paper.
an improved MPA is proposed which eliminates determined Second, efforts have been made to reduce the high dynamic
user codewords after certain number of iterations and con- ranges and high storage burden that are induced by the numer-
tinue the iterations for undetermined user’s codewords. Sim- ous calculations of the exponential function embedded in
ilarly, in [14], a belief threshold is set to choose the most MPA, which is one of its main points of congestion. To elim-
reliable edge probabilities and continue the iterations for inate calculations of the exponentials in the MPA, this paper
the others. A Shuffled MPA (S-MPA) is introduced in [15]. uses approximate modeling of noise. Indeed, a Gaussian
S-MPA is based on shuffling the messages between func- Probability Density Function (PDF) is estimated with sub-
tion nodes and variable nodes. As a result, the convergence optimal, bell shaped, polynomial PDFs. Using polynomial
rate is accelerated. A Monte Carlo Markov Chain Method PDFs enables a significant throughput improvement with a
is proposed in [16] to decode SCMA signals and sphere very small degradation on the bit error rate performance. In
decoding is also explored in [17] and [18] for SCMA receiver addition, this technique enables to use vectorized instructions
design. for the calculation of the probabilities, as opposed to log-
The main difference between this work and previously MPA. Details will be explained in the sequel. The impacts
cited works is that the present paper combines an analytic of turbo codes [19], polar codes [20] and LDPC codes [21]
view of MPA complexity with hardware and resource aware are investigated.
programming, exploiting hardware features available in gen- In this paper, symbols B, N, Z, R and C denote binary,
eral purpose processors (GPPs). The SCMA decoding algo- natural, integer, real and complex numbers. Scalar, vector and
rithms are revised in light of the needs of Cloud Radio matrix are presented as x, x, X respectively. The n’th element
Access Networks (C-RANs) and to take full advantage of of a vector denoted by xn and X n,m is the element of n’th row
key hardware features available in GPPs such as their SIMD and m’th column of matrix X. Notation diag(x) shows a diag-
engine. In the early 2000s, the performance of many pro- onal matrix where its n’th diagonal element is xn . In addition,
cessors improved significantly due to clock rate increases. the transpose of a matrix is expressed as X T . The paper is
This increase of performance needed very minimal if any organized as follows, Section II introduces the SCMA algo-
programming effort, however the drawbacks of high clock rithm. Maximum Likelihood, MPA and log-MPA decoding
rate was more power and energy consumption, overheating methods are explained in this section as a background to this
of processors, leakage currents and signal integrity problems. research. Section III elaborates on proposed improvements
These disadvantages led designers to follow new paradigms such as vectorizing the algorithm, exponential estimations,
such as thread-level and data-level parallelisms that pro- contiguous access to memory and other hardware oriented
vide good performance at lower clock speeds. Another chal- techniques. Section IV explores the bit error rate performance
lenge was data access efficiency in cache and RAM for as well as the throughput, the latency, the power consumption,
performance critical algorithms. Higher performance also and the energy efficiency of the proposed MPA and log-
came from improved cache access efficiency of heteroge- MPA implementations. Some profiling metrics are given to
neous processors and parallel access to the L1 cache through better understand the results. Section V is dedicated to study

VOLUME 7, 2019 10403

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

FIGURE 1. a) SCMA encoder with 6 users (layers) and 4 physical resources, b) SCMA uplink chain with
channel coding, c) Factor graph representation of a decoder, d) Message Passing Algorithm based on
Bayesian factor graph: (I) Resource to user message, (II) Guess swap at each user and user to resource
message, (III) Final guess at each user.

the effects of suggested techniques on block error rate after where xj = (x1 , . . . , xKj )T and hj = (h1 , . . . , hKj )T are
channel coding. Finally, the main findings of this research are respectively codeword and channel coefficients of layer j.
summarized in Section VI.
B. SCMA DETECTION SCHEMES
1) MAXIMUM LIKELIHOOD
II. BACKGROUND For an arbitrary codeword, the optimum decision, i.e. the
A. OVERVIEW OF THE SCMA SYSTEM MODEL one that minimizes the likelihood of transmission errors after
An SCMA encoder with J users (layers) and K physical decoding, is the one resulting from the Maximum Likeli-
resources is a function that maps a binary stream of data hood (ML) estimation, which can be described as:
to K -dimensional complex constellations f : Blog2 (M ) →
X, x = f (b) where X ⊂ Ck . The K -dimensional complex ˆ = arg min ||y − c||2 ,
xML (2)
codeword x is a sparse vector with N < K non-zero entries. c∈X
Each layer j = 1, . . . , J has its own codebook to generate
the desired codeword according to the binary input stream. given the received codeword. In (2), the soft outputs x̂ are also
Fig. 1 shows SCMA uplink chain with J = 6, K = 4 called Log-Likelihood Ratios (LLRs) that can be calculated
and N = 2. SCMA codewords are spread over K physical with the following equation:
resources, such as OFDMA tones. Fig. 1a shows that in the P
C∈L0x P(y|c)

multiplexed scheme of SCMA, all chosen codewords of the J
LLRx = ln P , (3)
layers are added together after being multiplied by the chan- C∈L1x P(y|c)
nel coefficient hj . Then, the entire uplink chain is shown
in Fig. 1b. The output of the SCMA encoder is affected by where LLRx is the log likelihood ratio of bit x obtained from
the white additive noise n. codeword x̂. This codeword comes from L1x the set of code-
J
X words in which bit x is 1 and L0x the set of codewords in which
y= diag(hj )xj + n, (1) bit x is 0. The probability function P(y|c) can be expressed
j=1 as in (4) when a signal is transmitted over an additive white

10404 VOLUME 7, 2019

A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

Gaussian channel with σ 2 variance: As shown in Fig. 1d(II) there are only two resources con-
nected to each user. A message from a user to a resource is a
||y − c||2
 
1
P(y|c) = √ exp − . (4) normalized guess swap at the user node:
2π σ 2σ 2
µRES3→UE3 (i)
µUE3→RES1 (i) = P , (9)
Although the ML method provides the best guess for O xML , i µRES3→UE3 (i)
performing the computation with this method requires unac-
message passing between users and resources (see (8)
ceptable complexity in real applications. In the case of six
and (9)) will be repeated three to eight times to reach
users and codebooks matrices size 4 × 4 as in Fig. 1a,
the desired decoding performance. The final belief at each
the calculation of the soft output for each bit in (4) needs
user B (i) is the multiplication of all incoming messages as
4096 exponential function computations, which is unaccept-
illustrated in Fig. 1d(III) and (10) for UE4 and codeword i.
able. Nevertheless, in this article the result of this method is
Finally, (11) is used to calculate soft outputs for bit bx :
used to compare with practical methods to characterize the
BER performance degradation of MPA and log-MPA. B3 (i) = µRES1→UE3 (i) × µRES3→UE3 (i), (10)
 
P(y|bx = 0)
2) MESSAGE PASSING ALGORITHM (MPA)
LLRx = ln
P(y|bx = 1)
Fig. 1c shows a Bayesian factor graph representation of an P
Bm (i) when bx =0

MPA decoder with six users and four physical resources. = ln Pm . (11)
Thanks to sparsity of the codebooks, exactly three users m Bm (i) when bx =1
collide in each physical resource. There are four possible 3) LOG-MAP
codewords for each of the three connected user’s codebooks
Since calculation of exponentials in (7) requires relatively
which gives 64 possible combined codewords in each phys-
high computational effort, changing the algorithm to log
ical resource. In the first step of the MPA, the 64 distances
domain using the Jacobi formula (12) is a classical improve-
between each possible combined codewords and the actual
ment of MPA:
received codeword are calculated. X N 
ln exp(fi ) ≈ max{f1 , f2 , . . . , fN } (12)
X
dRESβ (m, H) = ||yβ − hl,mu xl,mu ||, (5)
l⊂ζ,mu ∈{1,...,K } i−1

using Jacobi formula, (8) can be reduced to:

ζ is the set of users connected to resource β and the con-
d 2 (i, j, k, H)
 
sidered codeword is denoted as m. For instance, (5) can be
µRES1→UE5 (i) = max − RES1 2
re-written for resource 4 as: 2σ
j,k=1,...,4
dRES4 (m2 , m4 , m6 , h2 , h4 , h5 ) + µUE2→RES1 (j) + µUE3→RES1 (k), (13)


= ||y4 − h2 x2 (m2 ) + (h4 x4 (m4 ) + (h5 x5 (m5 ) ||. (6)
m2,4,6 =1,2,3,4
due to elimination of exponential’s high dynamic ranges,
there is no need to normalize the guess swap and (9) will be:
In which m2 , m4 , m5 indicate the different codewords for
µUE3→RES1 (i) = µRES3→UE3 (i). (14)
users 5, 4, and 2 in (6). Assuming perfect channel estima-
tion and Gaussian noise, these Euclidean distances can be The rest of the algorithm can be expressed as follows:
expressed as probabilities using (7):
B3 (i) = µRES3→UE3 (i) + µRES1→UE3 (i), (15)
 2
dRESβ

LLRx = max(Bm (i) when bx =0 ) − max(Bm (i) when bx =1 ).
9(dRESβ ) = exp − . (7) i i
2σ 2 (16)
After calculating the residual probability of each codeword
with (7), iterative MPA starts exchanging beliefs (probabil- III. PROPOSED IMPROVEMENTS
ities) on possible received codewords among the users and Besides methodical improvements of the MPA such as log-
resources nodes of the factor-graph. According to Fig. 1d(I), MPA, hardware oriented improvements are important to take
a message from resources to users has been defined to con- full benefit of C-RAN servers capabilities. Since MPA and
tain extrinsic information of two other connected users. For log-MPA are control heavy algorithms, mishandling of data
instance, a message from resource 4 to user 2 containing the can induce huge performance losses. This section explores
probability information of codeword i can be expressed as: how MPA can be reformulated: 1) to improve data locality
in cache and to reduce cache misses and branch mispredic-
4 X
X 4 tions 2) to reorder the data paths in order to help exploiting
µRES4→UE2 (i) = 9 dRES4 (i, j, k, H)


data-level parallelism at each step of the MPA and log-MPA
j=1 k=1 algorithms and 3) to exploit approximated modeling of addi-
×µUE4→RES4 (j) × µUE5→RES4 (k). (8) tive white Gaussian noise in order to eliminate exponential

VOLUME 7, 2019 10405

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

calculations and to drastically reduce the number of instruc- in (5) and (6), 2) calculate the exponentials in (7), 3) perform
tions for SSE, NEON, AVX and KNCI ISAs. users to resources messaging and final guesses at each user.

A. FLATTENING MATRICES TO REDUCE CACHE 1) SIMD COMPUTATION OF COMPLEX NORMS

MISSES AND BRANCH MISSES Equations (5) and (6) use a complex norm function ||.||,
Considering (6) and (7), there are 64 calculations of distances it can be optimized by using SIMD instructions. There
and probabilities for each resource (256 for all resources). are two ways to perform this computation: Fig. 2a depicts
Using a multidimensional array (4×4×4) should be avoided, how to implement the norm function using an Array of
because it typically causes bad data locality, which leads to an Structures (AoS) for complex numbers. In this method,
increased number of cache misses. These misses negatively the complex numbers are represented as two consecutive
affect the throughput, and this is significant, since this process floating-point numbers. The implementation with AoS uses
must be repeated in the decoder for each received 12-bit
block of data. Flattening a d-dimensional array to a vector
using (17) is appropriate to prevent cache misses and improve
the spatial locality of data. This is done with the help of an
index defined as:
X d  Y d 
index = Nj ni . (17)
i=1 j=i+1

jth
where Nj is the size of the dimension of the array and ni is
the location of a target element in that dimension. Improving
data locality with a stride of a single floating-point number in
each element makes it easier for the processor to have aligned
and contiguous accesses to the memory through SIMD ISA.
Utilizing SIMD instructions helps to reduce the total num-
ber of mispredicted branches in the algorithm. Contiguous
accesses to the L1 cache are performed by chunks of 128-bit,
256-bit or 512-bit. It reduces the number of iterations in
the for-loops and consequently it reduces the number of
branches. On the other hand, for a vector of sixty four
32-bit floating-point numbers, 64 iterations are needed in the
scalar mode, while only 16, 8 or 4 iterations are required
in the vectorized modes using respectively SSE (or NEON),
AVX or KNCI ISAs.

B. ADAPTING THE ALGORITHMS TO IMPROVE

DATA-LEVEL PARALLELISM
SSE, NEON, AVX and KNCI ISAs handle SIMD opera-
tions [22]. KNCI and AVX use 512-bit and 256-bit registers,
while SSE and NEON use 128-bit registers. For instance,
an AVX operation can process eight 32-bit floating-point
numbers simultaneously. The AVX instructions also provide
high-performance loads and stores to the cache memory
due to data vectorization. Flattening matrices to vectors is
a prerequisite to enable AVX contiguous accesses to the
memory. Vectorized instructions such as AVX are accessi-
ble in C++ through intrinsic functions. An intrinsic is a
function that directly maps to an assembly instructions (for
some rare exceptions it can be more than one instruction).
Nowadays, AVX units use sixteen 256-bit YMM registers
and a 32-bit MXCSR control register to handle vectors of
eight 32-bit or four 64-bit floating-point numbers. The AVX
ISA allows to perform SSE instructions using the lower
128-bit lane of the YMM registers. For MPA, the SIMD FIGURE 2. Complex norm AVX algorithm using a) Array of Structures
instructions are used to 1) compute the complex norm ||.|| (AoS), b) Structure of Arrays (SoA).

10406 VOLUME 7, 2019

A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

six intrinsic functions: one load (_mm256_loadu_ps),

one store (_mm256_storeu_ps), one multiplication
(_mm256_mul_ps), one permutation of the lanes
(_mm256_permute2f128_ps), one horizontal addition
(_mm256_hadd_ps) and one extraction of the highest
lane of the AVX register (_mm256_extractf128_ps).
Fig. 2b sketches the computation of the complex norm
using a Structure of Array (SoA) data layout. This imple-
mentation also uses six intrinsic functions: two loads
(_mm256_loadu_ps), one store (_mm256_storeu_ps),
two standard multiplications (_mm256_mul_ps), one addi-
tion (_mm256_add_ps).
Our experiments demonstrated that these two methods
have similar performances, however we used the Structure of
Arrays (SoA) since it is 1) easier to port for the ISAs that lack
from shuffle instructions and 2) trivial to extend for different
register lengths.

2) SIMD COMPUTATION OF EXPONENTIAL

To speedup the computational time of the exponentials used
in (7), the MIPP wrapper [23] has been used. MIPP proposes
a vectorized implementation of the exponential based on
a series expansion. Many intrinsic functions are encapsu-
lated to compute the exponential. MIPP also allows to write
portable intrinsic codes. A single SIMD code is written for
multiple ISAs such as SSE, NEON, AVX, AVX512 and KNCI
thanks to the meta-programming techniques.
The flattened complex and normalized numbers are cal-
culated as shown in Fig. 2a and Fig. 2b to produce the
preliminary values used to compute the probabilities. Fig. 3a FIGURE 3. a) Vectorized Exponentials (N0 = 2σ 2 ), b) Vectorized
illustrates the full process on a vector of eight floating-point calculation of final guess at user 4.
numbers. First the values are loaded into the YMM registers,
then they are multiplied by −1/2σ 2 and finally the exponen- leading to marginal bit error rate degradation after the MPA
tial function is performed according to (7). decoding. However, these estimated probabilities cause small
degradations of the block error rate (BLER) performance
3) SIMD MESSAGE PASSING after the channel decoding (cf. Section V). The proposed PDF
Some remaining parts of the MPA can be vectorized too. must satisfy two conditions to be valid: 1) it must be positive
Especially, the guess swaps and the computation of the final and lower bounded at zero, 2) its integral over (−∞, ∞) must
guesses at each user node can be vectorized using SSE be equal to 1. The following function is suggested to estimate
instructions. Fig. 3b shows the computation of final guesses the exponentials:
for user 4. There are four messages from a resource to a
2/π
user containing the probabilities of four different codewords, 9d0 RESβ = 4
. (18)
which are the elements of the SSE vectors. According to 2σ 2 + 4dRESβ
Fig. 3b these vectors of probabilities are loaded into SSE,
The computation of 9 0 is faster than the original 9 [24].
NEON or the lowest lane of the AVX registers.
The probabilities produced using (7) and (18) are normalized
C. ESTIMATED-MPA (E-MPA)
according to (9). Furthermore, the numerator 2/π does not
play an important role in MPA and can be uniformly elimi-
Computation of the exponentials in (7) is one of the most
nated from all calculations to reduce the computational effort.
important bottlenecks of the MPA algorithm. It is possible
Thus,
to further accelerate the computation by using proper esti-
mations. The exact exponential computation is not essential 1
to produce a satisfying estimation in the MPA algorithms. 9d0 RESβ ≈ 4
, (19)
2σ 2 + 4dRESβ
Considering that (7) represents a Gaussian PDF, it can be
replaced by sub-optimal bell-shaped polynomial distributions can be used as a systematic replacement to the vectorized
to model the noise. It will be shown in Section IV-B that using exponential MIPP function used in Fig. 3a. It reduces the
a polynomial estimation can increase the throughput while overall number of instructions to three intrinsic functions:

VOLUME 7, 2019 10407

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

two multiplications (_mm256_mul_ps) and one addition

(_mm256_add_ps).

D. ACCURACY OF FLOATING-POINT COMPUTATIONS

The finite precision of floating-point calculations induces
losses in the results. Thus, technical standards such as IEEE
754 define rounding rules, precision of calculations, excep-
tion handling and underflow behavior. However, the MPA
delivers approaching bit error rate results with less precise
floating-point models. For instance, in the GNU compiler,
-Ofast is a high-level compiler option which includes
fast math libraries to handle floating-point calculations
(-ffast-math). The compiler uses various mathematical
simplifications as explained in [25] and uses approximated
libraries for the division and the square root functions. The
compiler also forces the value to zero in the case of an
underflow. Using -Ofast can improve the throughput of the
MPA algorithm as will be shown in Section IV.
In this work, other well-known optimization techniques,
such as loops unrolling, using references instead of pointers,
avoiding type conversions, preferring prefixed operators, and
functions inlining have been used to enhance the throughput
of the various message passing algorithms.

IV. PERFORMANCE ANALYSIS

In this section, the effects of the various optimizations consid-
ered in Section III are investigated. A key concern is to ensure
that the decoding error performance is not affected by the FIGURE 4. Performance of MPA compared with E-MPA. (a) BER
execution time improvements, particularly when approxima- performance comparison of ML, MPA and E-MPA for 5 iterations.
(b) Convergence behavior of E-MPA and MPA.
tions are involved. Energy efficiency and power consumption,
throughput, memory access efficiency, hardware complexity
analysis are all important aspects that must be considered. throughput and energy consumption is a main goal of sys-
tem designers. Energy optimization can reduce the cost of
A. EVALUATION OF ERROR PERFORMANCE cloud services significantly while it can contribute to decrease
Fig. 4a shows the performance comparison of a maximum the emission of greenhouse gases. Power utilization is also
likelihood (ML) decoder, an MPA decoder performing 5 iter- important because improved performance per Watt is useful
ations and an estimated-MPA (E-MPA) decoder as explained to limit power demands. This section explores the power,
in Section III also performing 5 iterations. There are very energy efficiency and throughput of the various message
small differences in the bit error rate performance of the three passing algorithms suggested in this work. Tests have been
decoders (less than 0.10 dB). Although both MPA and E-MPA conducted on three platforms running the Ubuntu Linux oper-
show their optimum behavior with 5 iterations, the conver- ating system. The three systems are : 1) an IntelTM Core-i7
gence behavior of the two methods are different as illustrated 6700HQ processor with AVX instructions (256-bit SIMD)
in Fig. 4b. E-MPA has a slower convergence rate for less than and four physical cores using 2-way Simultaneous Multi-
three iterations. This phenomenon is expected as the probabil- Threading (SMT or Intel Hyper-ThreadingTM technology)
ity functions produced by bell-shaped polynomial PDF do not running at nominal frequency of 2.6 GHz, 2) an ARMTM
have the quality of probabilities produced by exponentials. Cortex-A57 with NEON instructions (128-bit SIMD) and
However, the convergence behavior is almost identical for four cores (no SMT) running at 2.0 GHz and 3) an IntelTM
more than 4 iterations. The other optimizations like loops Xeon-Phi Knight-Corner 7120P with KNCI instructions
unrolling, fast math libraries and vectorization were not found (512-bit SIMD) and 61 cores using 4-way SMT and running
to degrade the BER performance or the convergence rates. at 1.2 GHz.
Table 1 shows the comparison of throughput, latency,
B. CHARACTERIZING THROUGHPUT GAINS, ENERGY power consumption and energy of different decoding algo-
EFFICIENCY AND POWER CONSUMPTION rithms that are executed on the three platforms to decode
Energy efficiency is of interest in the design of C-RAN 768 Million bits. The average power and energy consump-
servers. It is determined by the rate of computation that tion measured on the Core-i7 processor were obtained
can be delivered by a processor. Joint optimization of the with the turbostat software [26] which exploits the Intel

10408 VOLUME 7, 2019

A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

TABLE 1. Throughput, latency, power and energy characteristics. throughput and latency are improved by much larger factors.
It means that the overall energy consumption have been
decreased with AVX.

2) ARMTM CORTEX-A57
On this platform [27], the throughput difference caused by
the fast math libraries of the GNU compiler is still visible for
MPA and log-MPA algorithms. With level three optimization
(-O3), MPA and log-MPA run at 1.60 Mbps and 3.01 Mbps
respectively. When using fast math libraries (-Ofast) the
throughputs increased to 4.07 and 4.70 Mbps. It should be
noted that the four physical cores of the ARM platform were
utilized for those tests. Power consumption and energy used
per decoded bit is lower on the ARM platform than on the
Intel processors. The low power consumption of the ARM
platform notably comes at the cost of less powerful floating-
point arithmetic units (cf. MPA+NEON and E-MPA+NEON
in Table 1). Eliminating the exponential computations almost
doubled the performance in E-MPA (15.30 Mbps) as com-
pared to MPA+NEON (8.40 Mbps), which shows the limits of
low power processors when calculating many exponentials.
Nevertheless, by using E-MPA, the ARM low power proces-
sors can be a good candidate for implementation of SCMA
decoders on C-RAN servers as it allows significant energy
savings.

performance counters in Machine Specific Registers (MSRs) 3) INTELTM XEON-PHI 7120P

to monitor CPU and RAM utilizations. However, in the The Xeon-Phi Knights Corner [28] benefits from the ability
case of ARM and Xeon Phi platforms, external cur- to execute four hardware threads per core, while having
rent sensors were used to measure the energy and power 61 cores and 512-bit SIMD registers. In this case, 244 threads
consumptions. can be run to handle the MPA decoding task. Despite these
benefits, the Xeon-Phi Knight Corners suffers from two
1) INTELTM CORE-i7 6700HQ main disadvantages: 1) the KNC instructions diversity is
The baseline implementation of MPA with level 3 (-O3) reduced compared to AVX or AVX-512 ISAs and 2) the
optimization of the GNU compiler reaches 3.51 Mbps uti- cores frequency is relatively low in order to keep reasonable
lizing all four physical cores of the processor (SMT on). power consumption and limits the heat dissipation. As an
Log-MPA improves the performance to 6.37 Mbps bene- example of missing instruction, the KNCI ISA does not
fiting from elimination of the exponential calculations, still offer coalesced division (_mm512_div_ps) for floating-
in -O3. However, using the fast math libraries (-Ofast) point numbers. Beside those limitations, the E-MPA+KNCI
and the loop optimizations from Section III-D increases the exhibits the highest throughput among the three mentioned
throughput to 14.85 Mbps for MPA and to 10.31 Mbps platforms (up to 114.60 Mbps). However, it consumes almost
for log-MPA. It is important to observe that MPA outper- three times more energy per bit compared to the ARM-based
forms the log-MPA with the fast math libraries and more implementations. The MPA decoding algorithm exhibits its
aggressive optimizations, without compromising on the bit best performance on this platform when cross compiled using
error rate performance. This is because log-MPA induces -O2 -mmic flags by an Intel icpc compiler. Using fast
inefficient data accesses due to the messages passed from math options such as -no-prec-div -no-prec-sqrt
resources to users. This phenomenon will be investigated -fp-speculation=fast -fp-model-fast=2 do
further in Section IV. Using the AVX and SSE SIMD ISAs not change the results significantly with the Intel compiler.
reduces the branch mispredictions and the cache misses Fig. 5 focuses on the energy consumed per decoded bit
(cf. Section III-A). Consequently, the throughput is increased (also mentioned in Table 1). In summary, the SIMD algo-
to 67.83 Mbps in MPA and to 75.46 Mbps for the E-MPA rithms have a higher energy efficiency per decoded bit. The
where the 9 0 estimated exponentials from (19) are performed. processor resources are well stressed and the power does not
These results confirm significant throughput gains for the increase too much. Among the obtained results, the Xeon-Phi
proposed implementations, while the energy consumption is obtains the best throughput while the Cortex-A57 has the
reduced. Utilizing AVX increases the average power con- lowest energy consumption. In the case where the number
sumption of MPA and log-MPA from 35 to 40 Watts but of users in the cloud is increased, the results presented in

VOLUME 7, 2019 10409

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

dynamic range of branch mispredictions shows that reducing

the total number of branches and branch mispredictions have
more impact on increasing throughput of the MPA algorithm
in comparison to reducing cache misses. This phenomenon
also shows that using optimization methods such as log-MPA
which produces large number of branches due to the max(.)
function is not ideal for multi-processor servers in C-RAN.
These reported significant improvements have been brought
by SIMD instructions. Improving data locality, contiguous
access to memory and parallelizing loops are the main rea-
sons that made SIMD algorithms exhibit better performance
when it comes to cache interface.
FIGURE 5. Graphical comparison of the energy consumed per decoded
bit for three different platforms. Table 2 also reports the number of Instructions per
Cycle (IPC) of each implementation. It is obvious that the
number of IPC was reduced in MPA -O3 and log-MPA due to
this section are scalable up to the number of processing units poorer memory access efficiency. This reduces the through-
dedicated to them. put of those algorithms. On the other hand, without using
contiguous access to memory, the processor spends more time
C. MEMORY (CACHE) ACCESS EFFICIENCY for scalar load and stores. This can cause a bottleneck in
Apart from SIMD operations and parallelization, cache interfacing memory while other resources of the processor
access efficiency plays an important role in the high- are waiting to receive data and consequently it decreases the
performance implementation of algorithms on GPP. Table 2 IPC. By contrast, in the case of contiguous access to memory
shows the performance characterization of different MPA (or cache) the processor can fetch sufficient data all at once
algorithms on the Core-i7 6700HQ processor for decoding to support sustained processing thus reducing the memory
768 Million bits. As reported in Table 2, contiguous accesses bottleneck and improving internal processing as reflected by
to the memory using AVX instructions reduces the total better IPC indices.
number of branches and references to the cache. Reducing
the number of branches and references to the cache increases D. PROFILING AND HARDWARE COMPLEXITY
the throughput of the algorithm. Previous sections explored how processor parallel resources,
efficient and contiguous memory access, and compiler opti-
TABLE 2. Cache performance characterization.
mizations play an important role in getting efficient imple-
mentation of the SCMA algorithms. In [6], [7], [13], and [15],
computational complexity, measured as operation counts,
was used to represent the complexity of the MPA. Opera-
tion counts can be misleading metrics when characterizing
algorithmic complexity of algorithms executing on general
purpose processors. Indeed, it misses significant factors such
as cache misses, memory efficiency and precision of floating-
point calculations. In this section, the time complexity of the
various forms of SCMA decoders are investigated using the
Intel Vtune StudioTM profiler [22].
Fig. 6 reports profiling results obtained with different
SCMA decoders variations when applied to the decoding
of 768 Million bits. Results were organized to show the exis-
According to Table 2, MPA+AVX shows almost ten tence of five bottlenecks i.e. logarithms in (11), exponentials
times fewer branches (12845 Million) versus MPA -Ofast in (7), complex norm and complex subtraction in (5) and
(126578 Million) and consequently it offers better perfor- messages passed from resources to users in (8).
mance. For MPA+AVX, 401 Million branches have been Observing MPA and MPA (-Ofast) reveals the overhead
mispredicted by the processors, compared to 7093 Millions of exponentials and complex norms in the algorithm. For
for MPA. For cache misses MPA+AVX produced two example, the decoder spent more than 62 percent of its time
Millions fewer cache misses when compared to MPA and (32.35 seconds) to calculate exponentials and norms in MPA
the total number of cache references are also significantly (-Ofast). This led us to explore SIMD calculation of these
(122 Millions) less than with MPA. The total number of cache two steps. Comparing E-MPA+SIMD and MPA+SIMD
misses for various algorithms in Table 2 are between 70 to implementations to others such as MPA (-O3 or -Ofast)
79 Millions, while the total number of branch mispredictions shows a clear gain in throughput for calculation of the expo-
varies between 422 Millions to 6454 Millions. This high nentials and norms. In more details, E-MPA+SIMD spends

10410 VOLUME 7, 2019

A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

embed the SCMA encoder and decoder in a full communica-

tion chain. To this purpose, we used the AFF3CT1 software
which is an ideal tool that provides the necessary simulation
models and allows performing the desired verifications.
AFF3CT is Open-source and specifically designed to
offer an efficient environment to the communication systems
designers. Monte-Carlo simulations can be run to mea-
sure various metrics such as the BER and BLER perfor-
mance, or the throughputs and latencies of each module,
e.g. FEC encoders and decoders, modulation and demodu-
lation blocks, or different channel models.

FIGURE 6. Profiling results of different MPA algorithms using Intel Vtune

ProfilerTM on Core-i7 6700HQ platform for decoding 768 Million bits.

1.68 seconds computing exponentials and norms which is

more than 19 times faster than the initial computation of
norms and exponentials in MPA (-Ofast). On the other
hand, exponentials and norms computations are performing
as fast as complex subtract. This profiling results show the
efficiency of the proposed SIMD implementation methods.
By contrast, log-MPA has not shown good performance using
fast math library when compared to MPA. Inefficient memory
access, cache misses and high number of branches are among
the reasons that made log-MPA exhibits lower throughput
than expected. Those phenomena are induced by comparison
operations embedded in the max(.) function in (13). Never-
theless, without using fast math libraries, log-MPA still offers
performance gains over MPA. FIGURE 7. BLER evaluation of SCMA MPA and E-MPA decoders combined
with LDPC, polar and turbo codes. (a) Code rate R = 1 = 3. (b) Code rate
R = 1 = 2.
V. CHANNEL CODING
A. COMPLETE SIMULATION CHAIN
According to the latest 3GPP report [29], in the 5G stan-
In the previous sections of this article, algorithmic improve-
dard, the two selected code families are the LDPC and polar
ments and implementation techniques have been proposed.
codes. Being implemented in the AFF3CT software, it is
These optimizations lead to drastic reductions of the pro-
possible to test our SCMA decoders in a complete commu-
cessing time and to an increase of the processing power
nication chain, in conjunction with state-of-the art LDPC,
efficiency. This is done with approximately no degradation of
polar and even turbo decoders that were used in the LTE
the BER performance after SCMA decoding. Nevertheless,
standard [30]. Fig. 7 shows the BLER performances of MPA
in a full communication chain, multiple access algorithms
and E-MPA decoders when combined with different channel
are closely linked to the Forward Error Correction (FEC)
codes. For a matter of reproducibility, the full parameters of
modules. Indeed, the input of the FEC decoder consists in
the FEC used are reported in the next section. This research
the outputs of the SCMA decoder.
In order to claim that the proposed improvements do not 1 AFF3CT is an Open-source software (MIT license) for fast forward error
degrade the overall error performance, it is necessary to correction simulations, see http://aff3ct.github.io

VOLUME 7, 2019 10411

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

does not claim any novelty in channel coding, however, two curves are plotted: one for the E-MPA and the other for
we found crucial to validate our proposed SCMA optimiza- the MPA. Only 0.2 to 0.4 dB separate the two versions of the
tions in a sufficiently complete communication chain. algorithm for all the considered channel codes. These results
show the extent to which uncertainty of estimations affects
B. CHANNEL CODING CONFIGURATIONS channel coding. The decoding speed improvement brought
1) TURBO CODES by the E-MPA algorithm has a cost in terms of decoding
In a first validation, the turbo code from the LTE stan- performance. This trade-off should be considered in order to
dard is used. In the decoder, 6 iterations are done. The two meet the system constraints.
sub-decoders implement the max-log Maximum A Posteri-
ori algorithm (max-log-MAP) [31] with a 0.75 scaling fac- VI. CONCLUSIONS
tor [32]. In Fig. 7a, the rate is R ≈ 1/3, no puncturing is In this paper, in consideration of the potential of Cloud-RAN
performed, the number of information bits K is 1024 and that would support 5G communication, we focused on
the codeword length N is 3084. In Fig. 7b, R ≈ 1/2 with improving the efficiency of 5G SCMA receivers on the type
the puncturing of half of the parity bits, K = 2048, and of multiprocessors that can be found in such servers. We pro-
N = 4108. vided test results using different platforms such as ARM
Cortex, Xeon-Phi and Core-i7. The benefits of using SIMD
and various algorithmic simplifications have been studied and
2) LDPC CODES
test results were presented. Among the platforms, the ARM
In a second set of validations, the LDPC codes used in this
Cortex-A57 was shown to offer the lowest energy consump-
paper are based on MacKay matrices that have been taken
tion per decoded bit, while many-core platforms such as
from [33]. In Fig. 7a, the matrix used is (K = 272, N = 816),
Xeon-Phi Knight’s Corner 7120P had the best throughput.
and in Fig. 7b the matrix is (K = 2000, N = 4000). In both
In addition, an estimation of conditional probabilities using
figures, the decoder used is a Belief Propagation (BP) decoder
polynomial distributions instead of Gaussian distribution was
with an Horizontal Layered scheduling [34]. For the update
proposed to increase throughput. This estimation has shown
rules, the Sum-Product Algorithm (SPA) has been used [35].
to offer throughput improvements of 15 to 90 percent depend-
The number of iterations is 100.
ing on the platform used, while it causes a very small degrada-
tion of BLER after channel decoding. To support this claim,
3) POLAR CODES the error performance of telecommunication chains combin-
In the final validation, polar codes are built by suitably ing MPA and E-MPA with channel coding with LDPC, polar
selecting the frozen bits. We used the Gaussian Approxima- codes and turbo codes with code rates R = 1/3 and R = 1/2
tion (GA) technique of [36]. The input SNR for the code were tested.
construction with the GA is 1 dB, which apparently is very
low considering that the SNR are 4 to 5 dB in the convergence REFERENCES
zone. This is motivated by the fact that the GA algorithm is [1] S. M. R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, ‘‘Power-domain
designed to work with the BPSK modulation. Using SCMA non-orthogonal multiple access (NOMA) in 5G systems: Potentials and
completely modifies the histogram of the LLR values for a challenges,’’ IEEE Commun. Surveys Tuts., vol. 19, no. 2, pp. 721–742,
2nd Quart., 2017.
given SNR. Therefore, a shift on the input SNR of the GA [2] H. Nikopour and H. Baligh, ‘‘Sparse code multiple access,’’ in Proc. IEEE
algorithm must be applied in order to efficiently select the Int. Symp. Pers. Indoor Mobile Radio Commun., London, U.K., Sep. 2013,
frozen bits. If this shift is not applied, the decoding perfor- pp. 332–336.
[3] 5G Algorithm Innovation Competition. (2015). Altera Innovate Asia
mances of the polar code degrades drastically. The number FPGA Design Contest. [Online]. Available: http://www.innovateasia.com/
of information bits and the codeword length are (K = 682, 5g/en/gp2.html
N = 2048) in Fig. 7a and (K = 2048, N = 4096) in Fig. 7b. [4] NGMN Alliance, ‘‘5G white paper,’’ Next Gener. Mobile Netw., Frankfurt,
Germany, White Paper, 2015, pp. 1–125.
The decoder is a Successive Cancellation List (SCL) decoder [5] L. Lei, C. Yan, G. Wenting, Y. Huilian, W. Yiqun, and X. Shuangshuang,
with L = 32 and a 32-bit GZIP CRC that was proposed ‘‘Prototype for 5G new air interface technology SCMA and performance
in [37]. evaluation,’’ China Commun., vol. 12, no. 9, pp. 38–48, Sep. 2015.
[6] S. Zhang, X. Xu, L. Lu, Y. Wu, G. He, and Y. Chen, ‘‘Sparse code multiple
access: An energy efficient uplink approach for 5G wireless systems,’’ in
C. EFFECTS OF E-MPA ON ERROR CORRECTION Proc. IEEE Global Commun. Conf., Dec. 2014, pp. 4782–4787.
In Fig. 7, the number of iterations of the SCMA demodulator [7] J. Liu, G. Wu, S. Li, and O. Tirkkonen, ‘‘On fixed-point implementation of
Log-MPA for SCMA signals,’’ IEEE Wireless Commun. Lett., vol. 5, no. 3,
is 5. The objective of simulating multiple channel codes is pp. 324–327, Jun. 2016.
not to compare them with each other. A fair comparison of [8] A. Bayesteh, H. Nikopour, M. Taherzadeh, H. Baligh, and J. Ma, ‘‘Low
the different channel codes would indeed impose using the complexity techniques for SCMA detection,’’ in Proc. IEEE Globecom
Workshops, San Diego, CA, USA, Dec. 2015, pp. 1–6.
same code lengths and more importantly their computational [9] Y. Du, B. Dong, Z. Chen, J. Fang, P. Gao, and Z. Liu, ‘‘Low-complexity
complexity should be compared, which is not the case here. detector in sparse code multiple access systems,’’ IEEE Commun. Lett.,
Our goal here is to study the impact of using E-MPA on vol. 20, no. 9, pp. 1812–1815, Sep. 2016.
[10] M. Taherzadeh, H. Nikopour, A. Bayesteh, and H. Baligh, ‘‘SCMA code-
the BER and FER performances when the channel codes are book design,’’ in Proc. IEEE Veh. Technol. Conf., Las Vegas, NV, USA,
included in the communication chain. For each channel code, Sep. 2014, pp. 1–5.

10412 VOLUME 7, 2019

A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

[11] J. Peng, W. Chen, B. Bai, X. Guo, and C. Sun, ‘‘Joint optimization of ALIREZA GHAFFARI received the B.Sc. degree
constellation with mapping matrix for SCMA codebook design,’’ IEEE from the Amirkabir University of Technology
Signal Process. Lett., vol. 24, no. 3, pp. 264–268, Mar. 2017. (Tehran Polytechnic), Iran, and the M.Sc. degree
[12] C. Yan, G. Kang, and N. Zhang, ‘‘A dimension distance-based SCMA from Laval University Canada. He has wide
codebook design,’’ IEEE Access, vol. 5, pp. 5471–5479, 2017. range of industrial and research experience in
[13] M. Jia, L. Wang, Q. Guo, X. Gu, and W. Xiang, ‘‘A low complexity firmware design, FPGA, hardware acceleration,
detection algorithm for fixed up-link SCMA system in mission critical and the IoT. Recently, he is more focused on
scenario,’’ IEEE Internet Things J., vol. 5, no. 5, pp. 3289–3297, Oct. 2018.
hardware/software acceleration in heterogeneous
[14] L. Yang, Y. Liu, and Y. Siu, ‘‘Low complexity message passing algo-
platforms and clouds. He is currently an Enthusiast
rithm for SCMA system,’’ IEEE Commun. Lett., vol. 20, no. 12,
pp. 2466–2469, Dec. 2016. in electrical engineering and a Research Asso-
[15] Y. Du, B. H. Dong, Z. Chen, J. Fang, and L. Yang, ‘‘Shuffled multiuser ciate with the Electrical Engineering Department, École Polytechnique
detection schemes for uplink sparse code multiple access systems,’’ IEEE de Montréal.
Commun. Lett., vol. 20, no. 6, pp. 1231–1234, Jun. 2016.
[16] J. Chen, Z. Zhang, S. He, J. Hu, and G. E. Sobelman, ‘‘Sparse code multiple
access decoding based on a Monte Carlo Markov chain method,’’ IEEE
Signal Process. Lett., vol. 23, no. 5, pp. 639–643, May 2016.
[17] L. Yang, X. Ma, and Y. Siu, ‘‘Low complexity MPA detector based
on sphere decoding for SCMA,’’ IEEE Commun. Lett., vol. 21, no. 8,
pp. 1855–1858, Aug. 2017.
[18] F. Wei and W. Chen, ‘‘Low complexity iterative receiver design for MATHIEU LÉONARDON received the degree
sparse code multiple access,’’ IEEE Trans. Commun., vol. 65, no. 2,
in engineering from Bordeaux INP, Bordeaux,
pp. 621–634, Feb. 2017.
[19] C. Berrou, A. Glavieux, and P. Thitimajshima, ‘‘Near Shannon limit error-
France, in 2015. He is currently pursuing the
correcting coding and decoding: Turbo-codes. 1,’’ in Proc. IEEE Int. Conf. Ph.D. degree with the École Polytechnique de
Commun. (ICC), vol. 2, Geneva, Switzerland, May 1993, pp. 1064–1070. Montréal and the University of Bordeaux under
[20] E. Arıkan, ‘‘Channel polarization: A method for constructing capacity- a co-directorship between both institutions. His
achieving codes for symmetric binary-input memoryless channels,’’ IEEE research interests include the design of efficient
Trans. Inf. Theory, vol. 55, no. 7, pp. 3051–3073, Jul. 2009. and flexible implementations, both hardware and
[21] R. G. Gallager, ‘‘Low-density parity-check codes,’’ IRE Trans. Inf. Theory, software, for decoding error-correcting codes,
vol. 8, no. 1, pp. 21–28, Jan. 1962. in particular polar codes.
[22] Intel. (2018). Intel C++ Compiler 18.0 Developer Guide and Refer-
ence. [Online]. Available: https://software.intel.com/en-us/cpp-compiler-
18.0-developer-guide-and-reference-fxsave64
[23] A. Cassagne, O. Aumage, D. Barthou, C. Leroux, and C. Jégo, ‘‘MIPP:
A portable C++ SIMD wrapper and its use for error correction coding in
5G standard,’’ in Proc. Workshop Program. Models SIMD/Vector Process.
(WPMVP), Vösendorf, Austria, Feb. 2018, pp. 1–8.
[24] A. Ghaffari, M. Léonardon, Y. Savaria, C. Jégo, and C. Leroux, ‘‘Improving
performance of SCMA MPA decoders using estimation of conditional ADRIEN CASSAGNE received the M.Sc. degree
probabilities,’’ in Proc. 15th IEEE Int. New Circuits Syst. Conf. (NEWCAS), in computer science from the University of
Jun. 2017, pp. 21–24. Bordeaux, France, in 2013, where he is currently
[25] GCC. (2018). Semantics of Floating Point Math in GCC. [Online]. Avail- pursuing the Ph.D. degree. His research inter-
able: https://gcc.gnu.org/wiki/FloatingPointMath ests include the design of efficient and flexible
[26] L. Torvalds. (2018). Turbostat. [Online]. Available: https://github.com/ software implementations for modern decoding
torvalds/linux/tree/master/tools/power/x86/turbostat error-correcting codes such as LDPC, turbo, and
[27] NVIDIA. (2018). Jetson TX1. [Online]. Available: https://www.nvidia. polar codes. More precisely, he looks at different
com/fr-fr/autonomous-machines/embedded-systems-dev-kits-modules/ aspects of parallelism such as multi-node, multi-
[28] G. Chrysos, ‘‘Intel Xeon Phi coprocessor (codename Knights Corner),’’ in threading, or vectorization.
Proc. IEEE Hot Chips 24 Symp. (HCS), Cupertino, CA, USA, Sep. 2012,
pp. 1–31.
[29] Multiplexing and Channel Coding (Release 15), document TS 38.212,
3GPP, Sep. 2017.
[30] Multiplexing and Channel Coding (Release 11), document TS 136.212,
3GPP, Feb. 2013.
[31] P. Robertson, E. Villebrun, and P. Hoeher, ‘‘A comparison of optimal and
sub-optimal MAP decoding algorithms operating in the log domain,’’ in
Proc. IEEE Int. Conf. Commun. (ICC), vol. 2, Jun. 1995, pp. 1009–1013. CAMILLE LEROUX received the M.Sc. degree
[32] J. Vogt and A. Finger, ‘‘Improving the max-log-MAP turbo decoder,’’ in electronics engineering from the University
Electron. Lett., vol. 36, no. 23, pp. 1937–1939, Nov. 2000. of South Brittany, Lorient, France, in 2005, and
[33] D. J. MacKay. (2018). Encyclopedia of Sparse Graph Codes. [Online]. the Ph.D. degree in electronics engineering from
Available: http://www.inference.org.uk/mackay/codes/data.html TELECOM Bretagne, Brest, France, in 2008.
[34] E. Yeo, P. Pakzad, B. Nikolic, and V. Anantharam, ‘‘High He was a Visiting Student with the Electrical and
throughput low-density parity-check decoder architectures,’’ in Proc. Computer Engineering Department, Aalborg Uni-
IEEE Global Commun. Conf. (GLOBECOM), vol. 5, Nov. 2001,
versity, Denmark, in 2004, and also with the Uni-
pp. 3019–3024.
versity of Alberta, AB, Canada, in 2005. From
[35] D. J. C. MacKay, ‘‘Good error-correcting codes based on very sparse
matrices,’’ IEEE Trans. Inf. Theory, vol. 45, no. 2, pp. 399–431, Mar. 1999. 2008 to 2011, he was a Postdoctoral Research
[36] P. Trifonov, ‘‘Efficient design and decoding of polar codes,’’ IEEE Trans. Associate with the Electrical and Computer Engineering Department, McGill
Commun., vol. 60, no. 11, pp. 3221–3227, Nov. 2012. University, Montreal, QC, Canada. He has been an Associate Professor
[37] M. Léonardon, A. Cassagne, C. Leroux, C. Jégo, L. Hamelin, and with Bordeaux INP, since 2011. His research interests include encompass
Y. Savaria, ‘‘Fast and flexible software polar list decoders,’’ CoRR, algorithmic and architectural aspects of channel decoder implementation.
vol. abs/1710.08314, pp. 1–11, Oct. 2017. [Online]. Available: More generally, he is interested in the hardware and software implementation
http://arxiv.org/abs/1710.08314 of computationally intensive algorithms in a real-time environment.

VOLUME 7, 2019 10413

 A. Ghaffari et al.: Toward High-Performance Implementation of 5G SCMA Algorithms

YVON SAVARIA (S’77–M’86–F’08) received machine learning, computational efficiency, and application specific archi-
the B.Ing. and M.Sc.A. degrees from the École tecture design. He holds 16 patents. He has published 140 journal papers and
Polytechnique Montreal, in 1980 and 1982, 440 conference papers. He was the thesis advisor of 160 graduate students
respectively, and the Ph.D. degree from McGill who completed their studies.
University, in 1985, all in electrical engineering. He has been a Consultant or was sponsored for carrying research by
Since 1985, he has been with the École Polytech- Bombardier, Inc., CNRC, Design Workshop, DREO, Ericsson, Genesis,
nique de Montréal, where he is currently a Profes- Gennum, Huawei, Hyperchip, ISR, Kaloom, LTRIM, Miranda, MiroTech,
sor with the Department of Electrical Engineering. Nortel, Octasic, PMC-Sierra, Technocap, Thales, Tundra, and VXP. He is
He has carried work in several areas related to a Fellow of IEEE. He has been a member of CMC Microsystems Board,
microelectronic circuits and microsystems, such as since 1999. He is a member of the Regroupement Stratégique en Microélec-
testing, verification, validation, clocking methods, defect and fault toler- tronique du Québec and of the Ordre des Ingénieurs du Québec. He was a
ance, the effects of radiation on electronics, high-speed interconnects and Tier 1Canada Research Chair on design and architectures of advanced micro-
circuit design techniques, CAD methods, reconfigurable computing and the electronic systems, from 2001 to 2015. He received the Synergy Award of
applications of microelectronics to telecommunications, aerospace, image the Natural Sciences and Engineering Research Council of Canada, in 2006.
processing, video processing, radar signal processing, and digital signal pro- He was a Program Co-Chairman of ASAP’2006 and NEWCAS’2018, and
cessing acceleration. He is currently involved in several projects that relate the General Co-Chair of ASAP’2007. He was a Chairman of CMC Microsys-
to aircraft embedded systems, radiation effects on electronics, asynchronous tems Board, from 2008 to 2010.
circuits design and test, green IT, wireless sensor networks, virtual networks,

10414 VOLUME 7, 2019