TELKOM
NIKA
, Vol. 11, No. 5, May 2013, pp. 2699 ~
2709
ISSN: 2302-4
046
2699
Re
cei
v
ed
Jan
uary 19, 201
3
;
Revi
sed Ma
rch 1
8
, 2013;
Acce
pted Ma
rch 2
5
, 2013
Low Cost Quasi Binary Weighting Addition Log-SPA
LDPC Decoders
Po-Hui Yang
*, Ming-Yu Lin
Dep
a
rtment of Electron
ic Engi
neer
ing, Nati
on
al
Yun
lin U
n
iv
e
r
sit
y
of Scie
nce
and T
e
chnol
o
g
y
123 U
n
ivers
i
t
y
Roa
d
, Section
3, Doul
iou, Yu
nlin
640
02, T
a
iw
a
n
, telp:+
88
6
-
5-53
426
01-
43
40
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: ph
yan
g
@
y
u
n
t
ech.edu.t
w
A
b
st
r
a
ct
T
h
is pap
er pro
poses a
new
additi
on
meth
od
, w
h
ic
h can be
app
lie
d to red
u
c
e the har
dw
are cost of
LDPC
dec
od
er
s usi
n
g
lo
g-SP
A, w
h
ich tra
d
iti
ona
lly
has
the
best l
o
w
e
st bit
error r
a
te (BE
R
) b
u
t the
h
i
gh
est
hardw
are r
e
q
u
i
r
ement. T
he pr
opos
ed
quas
i-
bin
a
ry w
e
ight
i
n
g ad
ditio
n
ca
n
be si
mply i
m
pl
emente
d
by O
R
gates. With auxiliary
pseu
do-carry circuit, the BER pe
rform
a
nce can reac
h a fair
level.
In t
he l
og-SPA
mess
ag
e-pass
i
ng
path, n
u
m
eric transf
o
rm reducti
on
ar
chitecture
is
prop
osed
for
further h
a
rdw
a
r
e
reducti
on. Synt
hesi
z
e
d
and
n
u
merica
l res
u
lt
s show
that
th
e new
pr
op
ose
d
arch
itecture
achi
eved
an
u
p
t
o
32% a
nd 1
8
%
total hardw
ar
e reducti
on, c
o
mpar
ed w
i
th
traditio
nal l
og-
SPA deco
ders
,
and the si
mp
lest
sign-
mi
n arch
itecture, respect
i
vely, w
i
th fair BER perfor
m
a
n
ce.
Ke
y
w
ords
: LD
PC deco
der, fo
rw
ard error correction co
de, lo
g-SPA
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Ne
w VLSI te
chn
o
logy all
o
ws th
e lo
w-d
ensity pa
rity-che
ck (LDP
C) code [1,
2] to be
reali
z
ed
in
ha
rdware. The
excelle
nce of
the e
r
ro
r correctio
n
p
e
rfo
r
mance of
the
LDP
C
cod
e
has
introdu
ce
d ex
tensive
discu
ssi
on
s a
nd
st
udie
s
[3-6]. T
he o
r
igin
al L
D
PC
de
cod
e
r ado
pting
a
sum-
produc
t algorithm (SPA) require
d such large
hardware that m
any new methods were proposed
for hardware
reduction [7-15]. Logarithm field tr
ansf
e
r, called log-SPA [16], is
a known met
hod
for repl
aci
n
g
multipliers
with add
ers
and havin
g
the be
st bit erro
r rate
(BE
R
) p
e
rfo
r
man
c
e.
Although a
d
o
p
ting loga
rith
mic op
eratio
n
s
to turn th
e origin
al multi
p
licatio
n into
addition, a la
rge
numbe
r
of a
ddition
devices
have
al
so
be
com
e
h
a
r
d
wa
re
bu
rde
n
s. M
any
h
a
r
d
wa
re
-sim
plify
i
ng
methodol
ogie
s
have be
en
propo
se
d to deal with
mass additio
n
operation
s
. For exampl
e,
parall
e
l archi
t
ecture [9, 1
2
] and reorg
anized ad
de
r trees
with re-ma
p
s
skills [11] have b
een
prop
osed to
i
m
prove
effectively the co
mplexity
of the h
a
rd
wa
re.
Ho
weve
r, a
large
nu
mbe
r
of
adde
rs and
b
i
g loo
k
-u
p tab
l
es
(L
UT)
are
still the h
eavy hard
w
a
r
e lo
ading
s. The
prop
osed
sig
n
-
min algo
rith
m (SMA) a
r
chitecture [8] has
no ad
ditions in
ch
eck node o
perat
ions to
save
on
hard
w
a
r
e at
the expense of BER perform
an
ce.
Mixed-sign
al circuits hav
e been u
s
e
d
by
resea
r
chers
to
redu
ce
the
po
we
r and
hard
w
a
r
e req
u
irem
ents [1
7, 18]. Howe
ver, this solu
tion
hinde
rs the
realization
of tech
nolo
g
y mi
gration,
e
s
pe
cially in
all
di
gital processes. F
o
r SMA to
have good B
E
R perfo
rma
n
ce an
d to re
quire a
ddition
al comp
en
sat
i
on circuit
s
[7, 13], hardwa
r
e
and BER are treated as tradeoffs. Hav
i
ng log-
SPA
with naturally good BER
performance,
a
simplified
ci
rcuit is utili
zed
to repl
ace th
e bina
ry
ad
d
e
rs,
whi
c
h
be
come
the m
a
jor fo
cu
s of t
h
is
study
2. LDPC Dec
oder Algori
t
hm
Con
s
id
erin
g an LDP
C
cod
e
visuali
z
ed
by a Tanne
r
grap
h con
s
ist
i
ng of
M
ch
e
ck n
ode
s
and
N
varia
b
l
e node
s, a
set of the ch
eck no
de
s
conne
cted to
variable
nod
es i
s
den
ote
d
as
and
the
set
of the va
ria
b
le n
ode
s
conne
cted
to
che
c
k n
ode
s is
den
oted
as
.
rep
r
e
s
ent
s
the set
wit
h
the
che
c
k no
de
s ex
cluded
and
rep
r
e
s
ent
s th
e set
with
variabl
e no
de
s exclud
ed. During
the d
e
coding
pro
c
e
s
se
s of
the LDPC
co
de, the SPA update
s
the
soft info
rma
t
ion iteratively between
check no
de
s and
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2699 – 270
9
2700
variable
nod
e
s
. Fo
r notatio
n sim
p
lificatio
n,
is d
e
fined
the soft information
sent f
r
om
the variable n
ode
to the check nod
e
as
(1)
whe
r
e the
q
uantities
are
the pro
babili
ty informatio
n se
nt from
the variabl
e
node
to the ch
eck no
d
e
along
a
con
n
e
c
ting e
dge of a T
a
nner
graph,
indicating
. Als
o
,
is d
e
fined a
s
the
so
ft informat
ion sent
from
the variable nod
e
to
the che
c
k no
de
as
(2)
whe
r
e th
e qu
antities
are the probability informatio
n
sent from th
e check
node
to the
varia
b
l
e no
de
alo
n
g
a
co
nne
ctin
g ed
ge
of a
Tanne
r
gra
p
h
,
indicating
.
Next, the iteration pro
c
e
d
u
r
es of the SP
A us
ed in L
D
PC cod
e
s
ca
n be sum
m
ari
z
ed a
s
follo
ws.
Step 1.
Initia
lization. Fo
r the dom
ain of
log-li
keliho
o
d
, the log-li
kelihoo
d ratio
(LL
R
) i
s
obtaine
d for the ca
se of the transmitted bit
= 0 and
= 1 given the received bit
whi
c
h may be
corrupte
d
by the cha
nnel n
o
ise, like
(3)
The valves of
and
are initialized to the values
(4)
and
(5)
r
e
spec
tively.
Step 2.
Che
c
k node to variable nod
e (u
pdate ch
eck
node
s). Acco
rding to the standard
SPA, a c
h
eck
node
gat
hers all th
e
incomi
ng
LL
R me
ssag
es and
evaluat
es the
LL
R
messag
e se
n
t
to the variable node
, which can b
e
expressed a
s
(6)
whe
r
e
(7)
and
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Low
Cost Quasi Binary
We
ighting Addit
i
on Log-SPA LDPC
Decoders (Po-Hui
Y
ang)
2701
Step 3.
Varia
b
le nod
e to check no
de (u
pdate vari
abl
e node
). Accordin
g to the
stand
ard
SPA, a variable node
passe
s the
LL
R
messag
e to
a
ll the
con
n
e
c
ted
che
c
k n
o
d
e
s,
whi
c
h
ca
n
be expre
s
sed
as
(9)
Step 4.
Che
c
k
stop
criterio
n. The
overall
LL
R me
ssag
e of a
vari
abl
e no
de
indi
cating
the proba
bility to de
cod
e
t
he vari
able
n
ode
to 1
or
0 can
be o
b
tained
by ad
d
i
ng u
p
all th
e
incomi
ng LL
R message
s to
the variable
node
as
(10
)
After each ite
r
ation, a h
a
rd
deci
s
ion
on t
he varia
b
le n
ode
is mad
e
, i.e., the variable
is de
code
d a
s
“o
ne”
whe
n
, and de
code
d as “ze
r
o
”
ot
herwise. If all the deci
s
ion
bits con
s
titute a valid co
de
word, the alg
o
rithm
stop
s
and o
u
tputs t
he de
co
ding
result. Othe
rwi
s
e,
the algorith
m
repe
ats Step
s 2-4.
3.
Log-SPA LDPC Deco
d
e
r
Hard
w
a
r
e
The hardware structur
es of traditional l
og-SPA for the ch
eck node is sh
own in
Figure 1
and the data
path of the variable n
ode i
s
sho
w
n in
Fig
u
re 2, re
sp
ectively. Focusi
ng on the che
ck
node a
r
chite
c
ture is im
po
rtant be
cau
s
e t
he ha
rd
wa
re
and
comp
utation co
mplexiti
es of the
che
ck
node a
r
e hig
h
. Generally, the che
ck n
ode can
be i
m
pleme
n
ted
by signin
g
a
nd addi
ng two
indep
ende
nt operational p
a
rts [9].
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
1
lo
g
1
x
x
x
e
e
Figure 1. The traditional Log-SPA check node architecture.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2699 – 270
9
2702
Figure 2. Dat
apath of tradit
i
onal SPA variable node.
The q
uanti
z
e
r
s
re
stri
ct the
input
ran
ge f
o
r in
co
ming
LLR value
s
, roundi
ng off t
he inp
u
t
values to the
nearest value
,
the Q-LL
R, in a set
ofqua
ntization level
s
. Table 1 i
s
an example o
f
a
4-bit
sign
ed
quanti
z
ation t
able. While t
he me
ssag
e
is up
dated
in
the che
ck
n
ode by
(6
), the
input valu
es
have to
be
m
appe
d into
a
hyperb
o
lic ta
ngent
(
tanh
) functio
n
. Th
en
, the
cal
c
ulat
ed
che
c
k nod
e is converte
d ba
ck to LL
R by (8) fo
r variabl
e node me
ssage up
dating.
Table 1. Qua
n
tization
Ran
g
e
LLR
Q-
LLR
000.101
(
0.625)
000 (0)
000.110-
-001.1
0
1
(0.75-
-1.62
5
)
001 (1)
001.110-
-010.1
0
1
(1.75-
-2.62
5
)
010 (2)
010.110-
-011.1
0
1
(2.75-
-3.62
5
)
011 (3)
011.110-
-100.1
0
1
(3.75-
-4.62
5
)
100 (4)
100.110-
-101.1
0
1
(4.75-
-5.62
5
)
101 (5)
101.101
(
5.625)
110 (6)
The functio
n
of
tanh
is usually implem
ented by an LU
T. Base
d on the cha
r
a
c
teri
stic
curve
of
(7
),
whi
c
h i
s
sh
o
w
n i
n
Fi
gure
3, the i
nput
a
nd the
out
put
map
p
ing
tabl
es
co
uld
be
b
u
ilt
as Ta
ble 2 (LUT 3
-
7
)
an
d Table 3
(L
UT 7-3),
respectively. No
ticeably, the
bit length of the
quanti
z
ed L
L
R
value is a trade-off betwe
en har
dware compl
e
xity and BER perfo
rmance.
x
01
2345
67
x
0.0
0.5
1.0
1.5
2.0
2.5
3.0
x
)
=log((e
x
+1)/
(e
x
-1))
LU
T
3
-
7
LU
T
7
-
3
Figure 3. Qua
n
tization
curv
e for lookup
-table
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Low
Cost Quasi Binary
We
ighting Addit
i
on Log-SPA LDPC
Decoders (Po-Hui
Y
ang)
2703
Table 2. 3-bit to 7-bit LUT
Q-
LLR
000 (0)
1.111111
(1.984
375)
001 (1)
0.110001
(0.765
625)
010 (2)
0.010001
(0.265
625)
011 (3)
0.000111
(0.109
375)
100 (4)
0.000010
(0.031
250)
101 (5)
0.000001
(0.015
625)
110 (6)
0.000000
(0.000
000)
111 (7)
0.000000
(0.000
000)
Table 3. A 7-bit to 3-bit LUT Example for Tra
d
itional
Archite
c
ture
Q-
LLR
1.111111-
-1.10
0
000 (1.98
4375-
-1
.500000)
000 (0)
1.011111-
-0.10
0
000 (1.48
4375-
-0
.500000)
001 (1)
0.011111-
-0.00
1
010 (0.48
4375-
-0
.156250)
010 (2)
0.001001-
-0.00
0
011 (0.14
0625-
-0
.046875)
011 (3)
0.000010
(
0
.031250)
100 (4)
0.000001
(
0
.015625)
101 (5)
0.000000
(
0
.000000)
110 (6)
The lo
ga
rith
m dom
ain
o
peratio
n, a
pra
c
tical
sol
u
tion, re
pla
c
es th
e multi
p
licative
operation
by
additiveo
peration to
avoi
d inten
s
iv
e
multiplicative
cal
c
ul
ation
s
on
conventi
onal
SPA. consequently, the speed bottlen
eck and the hardware
cost are shifted from multipliers
to
adde
rs, a
nd
become
signi
ficant on l
a
rg
e LDP
C
de
coders. the m
i
n-sum al
gori
t
hm [10] is t
h
e
main
L
D
PC hard
w
a
r
e architecture whi
c
h red
u
ces
h
a
rd
wa
re
co
st effectively. howeve
r
, its B
E
R
performance is unsatis
factory. based on the origi
nal
Log-SPA and the min-
sum hardware
redu
ction e
ssences, some
simple lo
gic
gates a
r
e p
r
o
posed to repl
ace the tradit
i
onal ad
ders i
n
the che
c
k no
de. In the next section, the
prop
osed ide
a
is explain
e
d
in detail.
4. The Propo
sed Scheme
4.1. Quasi Bi
nar
y
Weighting Addition
Chec
k No
de
As sh
own in
Figure
1, when the in
pu
t LLR
value
s
are
sent int
o
the ch
eck
node fo
r
addition
op
eration, they
are first ma
ppe
d into
t
he
de
sired
qu
antize
d
value
sho
w
n in
Tabl
e 2.
The
corre
s
p
ondin
g
value i
n
T
able 2
is the
tanh
fun
c
tio
n
after
rea
s
o
nable
qua
ntization. Based
on
easi
e
r
circuit
s
to re
place the traditio
nal
adde
r, the value cha
r
a
c
te
ristics of the
t
anh
func
tion
in
Table
2 a
r
e f
u
rthe
r ob
se
rv
ed. The
characteri
stic
s
of
the su
mmed
value could
b
e
re
pla
c
ed
by the
OR gate in the addition o
peratio
n of the che
ck n
o
d
e
(few value
errors are fu
rther discu
s
se
d).
OR
ope
ration
is
perfo
rme
d
dire
ctly, with
bits of th
e O
R
g
a
te u
nde
r
the same
bin
a
ry weightin
g
.
In
the OR ope
ration sho
w
n i
n
the middle of Figure
4, the output of the bottom right OR gate
is
, in which
is missin
g be
cau
s
e the
u
p
load of
would not contain the value of
. The
prop
osed ne
w circuit
s
do
not really
execute the
addition in traditional bin
a
ry adde
rs
. Beside
s, in the appli
c
atio
n of the LDPC
decode
r, havi
ng the O
R
g
a
te clo
s
e to t
he bina
ry ad
dition ope
rati
on is p
o
ssibl
e
be
cau
s
e of
the
value
cha
r
a
c
teristi
c
s in th
e
mappi
ng ta
bl
e. The
propo
sed
ne
w m
e
thod i
s
the
r
ef
ore
call
ed
qu
asi
binary
weight
ing additio
n
. In terms of
overall
op
era
t
ion, the 7-bi
t updated va
lues
sho
u
ld
be
mappe
d to 3
-
bit values
bef
ore th
e qu
asi
bina
ry wei
g
h
t
ing additio
n
che
c
k no
de v
a
lue i
s
up
dat
ed
to a va
riabl
e
nod
e. Ta
ble
4, mo
dified
from the
tra
d
itional o
ne
(e.g., Table
3
)
, is doi
ng
the
mappin
g
fo
r
update
d
valu
es,
whe
r
e
”x” stan
ds for “d
on't
care”.
Fu
rther,
all val
u
es i
n
th
e
che
c
k
node
are th
e
estimated
p
r
o
babilitie
s that
do
not
req
u
ire p
r
e
c
ise
cal
c
ulatio
n a
s
l
o
ng a
s
ite
r
atio
ns
can b
e
don
e
with conve
r
gen
ce. As a
result, the
che
c
k nod
e hard
w
a
r
e
co
mplexity can
b
e
redu
ce
d with
few comp
utational erro
rs. The use
of
logical O
R
-gates, sho
w
n
in Figure 4,
is
prop
osed to repla
c
e the bi
nary add
ers i
n
t
he che
c
k n
ode with
simp
ler logi
c ci
rcui
ts.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2699 – 270
9
2704
Table 4. A Modified 7-bit to 3-bit LUT
Q-
LLR
1xxxxxx 000
(0)
01xx
xxx 001
(1)
001xxxx 010
(2)
0001xxx 011
(3)
00001xx 100
(4)
000001
x 101
(5)
0000001
110
(6)
0000000
111
(7)
Figure 4.The
pse
udo
-carry circuit.
The pe
rform
ance of the
prop
osed q
u
a
si bi
nary
weighting
addit
i
on for p
e
rfo
r
ming the
SPA has been evaluated by numeri
c
al
experiments. T
he code
used for sim
u
lati
on is
(1008, 3,
6)
regul
ar LDP
C
code [19]. For compari
s
ons, two
al
gorithms, namel
y
, the original SPA, LUT 3-7
[9], and the SMA [7], are u
s
ed to
ob
serv
e any impr
ovement on th
e
perfo
rman
ce
of the pro
p
o
s
ed
scheme. Sim
u
lation
re
sults a
r
e
sho
w
n
in Figu
re
9, whe
r
e th
e
BER pe
rform
ance versu
s
the
signal-to-noi
se ratio (S
NR) is
compared am
ong the SPA, LUT 3-7 [9], the SMA [7], and the
prop
osed
qu
asi
bina
ry we
ighting
additi
on. Th
e maxi
mum ite
r
ation
numb
e
r is se
t to eighty in
t
h
e
simulatio
n
. In Figu
re 9,
BER perfo
rm
ance of the
prop
osed q
u
a
si bi
nary
weighting
addi
tion
outperfo
rmin
g the SMA does n
o
t lose too mu
ch. As
a re
sult of the new
ch
eck
node a
r
chite
c
ture
that uses
si
mple a
ddition
and
a
regul
ar L
U
T,
the
OR
ope
ration
archite
c
ture
ha
s ab
out 1
5
%
hard
w
a
r
e
re
ductio
n
whe
n
comp
are
d
with
the t
r
a
d
itional
LUT
3-7 a
r
chite
c
ture
ba
sed
on a
0.18
m stand
ard
cell library. In addition, the pro
posed qu
asi
binary weig
hting additio
n
architectu
re
can
wo
rk up
to 100 M
H
z de
co
ding
spe
ed a
s
th
e propo
sed
architectu
re
use
s
con
c
i
s
e logi
cal gates to pe
rform a
ddition
s in the LDP
C
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Low
Cost Quasi Binary
We
ighting Addit
i
on Log-SPA LDPC
Decoders (Po-Hui
Y
ang)
2705
4.2. Quasi Bi
nar
y
Weighting
Addition With
Pseu
do
-Ca
r
r
y
Without the
carry sig
n
in addition, ou
r new
a
ddition
method in check no
de
s lead
s to
some
num
eri
c
al e
r
rors
an
d bad
BER
p
e
rform
a
n
c
e.
To comp
ensate for th
e su
mmation e
r
ro
r in
the O
R
op
eration a
r
chitecture, a
con
c
i
s
e
circuit
i
s
develop
ed in
place of the
tradition
al a
dde
r
carry-o
u
t ci
rcuits. A pse
u
d
o
-carry ci
rcuit is t
hen d
e
si
gned to d
eal
with loss
ca
rry-o
ut proble
m
.
The p
s
eu
do
-carry ci
rcuit can be
simply
impleme
n
t
ed
by combi
nati
onal lo
gic
circuits. Fi
gure
5 is
a de
sign exa
m
ple. Wh
en
the qua
si bin
a
ry wei
ghting
adde
r re
ceiv
es the m
app
ed value
s
, the
pse
udo
-carry ci
rcuits
dete
c
t the
l
ogi
cal
high fo
r
each
sa
me
bina
ry wei
ghting
co
lumn of
bits. I
f
a
colum
n
of bits ha
s more than on
e “1” ,
a pseu
do-ca
rry-o
ut circuit
is sent to the highe
r-o
rd
e
r-bit
colum
n
of the succe
edin
g
OR op
eratio
n
.
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
..
.
.
.
.
..
.
..
.
..
Figure 5.Prop
ose
d
qua
si bi
nary wei
ghtin
g addition
with pse
udo
-carry circuits.
Numeri
cal si
mulation results, shown
in Fi
gure 9,
illustrate th
at the quasi
binary
weig
hting a
d
d
ition, LDP
C
decode
r, eq
uippe
d wi
th
pse
udo
ca
rry
, has
only 0.
08 dB BER l
o
ss
when com
p
ared
with the log-SPA paral
l
el archit
ecture [9]. On account
of the added pseudo-
carry
circuit, t
he total
LDP
C
h
a
rd
wa
re
h
a
s
abo
ut 6%
overhe
ad
co
mpared
with t
he o
r
igin
al q
u
asi
binary weight
ing addition,
yet the BER
perfo
rman
ce
improve
s
to about 0.2 dB over the quasi
binary weigh
t
ing addition
archite
c
tu
re
. Although
extra pse
u
d
o
-carry ci
rcu
i
ts increa
se
the
hardware, the total hardware
co
st is still less than adder-bas
ed log-SPA. Further, the new
architectu
re,
whi
c
h ad
opts
prop
er L
U
T, require
s le
ss hard
w
a
r
e tha
n
sele
ctor-b
a
s
ed
sign
-min.
5. Design fo
r
Hard
w
a
re Reduc
tions
A unified TC-based archite
c
ture i
s
pro
p
o
se
d
to simp
lify the numeric tra
n
sfe
r
b
e
twee
n
TC
and si
gn magnitude i
n
traditional
l
og-SPA,
and to reduce the hard
ware without affect
ing
BER pe
rform
ance lo
ss. F
u
rthe
r,
the bit
-
width
of L
U
Ts i
s
examin
ed to come
out with o
p
timum
bits for the da
ta path.
5.1.
Lo
w
Har
d
w
a
re Cos
t
LUT
Figure 6 depi
cts BER p
e
rf
orma
nce versus bit-wi
dth o
f
mapping ta
ble from 3
-
bit
to 7-bit
for
(x), usin
g the (10
08, 50
4) matrix [19]. The 3-
bit to 4-bit mappi
ng
table (LUT 3
-
4) and the 3-
bit to 7-bit (L
UT 3
-
7
)
map
p
ing tabl
es
h
a
ve onl
y 0.1
dB BER pe
rforma
nce
er
ro
r. Ho
wev
e
r, they
have a
three
bit differen
c
e influe
nci
n
g
the
su
ccee
d
i
ng g
r
e
a
t de
a
l
of a
dde
rs.
The
synthe
si
zed
result sho
w
s that adoptin
g LUT 3
-
4 h
a
s a
bout 1
6
%
total area
red
u
ctio
n in
the LUT 3
-
7. In
con
s
id
eratio
n
of h
a
rd
wa
re
co
st, L
U
T
3-4, sh
own
in
Table
6, i
s
a
dopted
rather than
u
s
ing
the
LUT 3
-
7 u
s
ed
in [9] .
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2699 – 270
9
2706
SNR (
d
B)
0.
0
0
.
5
1.0
1
.
5
2.
0
2
.
5
Bit Err
o
r Rate
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Si
gn-M
i
n
LU
T 3
-
3
LU
T 3
-
4
LU
T 3
-
5
LU
T 3
-
6
LU
T 3
-
7
[6
]
SP
A
Figure 6. BER perfo
rma
n
ce of different LUT fo
r
(100
8, 504) L
D
PC code
with eig
h
ty decodi
ng
iteration
s
.
Table 6. 3-bit to 4-bit LUT
Q-
LLR
000 (0)
1.111 (1.8
75)
001 (1)
0.110 (0.7
50)
010 (2)
0.010 (0.2
50)
011 (3)
0.001 (0.1
25)
100 (4)
0.000 (0.0
00)
101 (5)
0.000 (0.0
00)
110 (6)
0.000 (0.0
00)
111 (7)
0.000 (0.0
00)
Table 7. Ne
w 3-bit to 4-bit LUT
Q-
LLR
000 (0)
1.111 (1.8
75)
001 (1)
0.110 (0.7
50)
010 (2)
0.010 (0.2
50)
011 (3)
0.001 (0.1
25)
100 (4 ,
-4)
0.000 (0.0
00)
101 (-3
)
0.001 (0.1
25)
110 (-2
)
0.010 (0.2
50)
111 (-1
)
0.110 (0.7
50)
5.2. TC Ba
se
d Messa
ge Passing Archi
t
ec
ture
As illust
rated
in Figu
re 7
(a
), messa
ge p
a
ssing
betwe
en che
ck
nod
e and va
riabl
e nod
e
is
sign-magni
tude in the traditi
onal
log-SPA architect
u
re [9], wher
e the variable node
processes
the TC value
s
and the
ch
eck nod
e pro
c
e
s
ses the S
M
values, so
that the numeric tran
sform is
pro
c
e
s
sed twice du
rin
g
on
e updatin
g. If messag
e pa
ssi
ng is T
C
-b
ase
d
num
bers, as
sho
w
n i
n
Figure 7 (b
), the variable
node d
o
e
s
n
o
t need SM
values fo
r TC tran
sfer.
Consequ
ently, after
summ
ation p
r
oce
s
sing at t
he varia
b
le n
ode, the ca
lculated T
C
values a
r
e the
n
sent to the ch
eck
node di
re
ctly. The only n
u
meri
c tra
n
sf
er left for th
e TC a
r
chite
c
ture i
s
o
n
the ch
eck n
o
d
e
outputs, a
nd
the LUT i
s
m
odified a
s
sh
own i
n
T
abl
e
7, inclu
d
ing t
he inp
u
t LLR.
Figure 8 sho
w
s
the effective
ness
of ha
rdwa
re
re
du
ction of
T
C
messag
e-pa
ssing
archite
c
ture. Th
e n
e
w
prop
osed
architecture i
s
smaller than
traditional
LUT
3-7
[9] an
d L
U
T 3
-
4
ha
rd
ware
by 26%
a
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Low
Cost Quasi Binary
We
ighting Addit
i
on Log-SPA LDPC
Decoders (Po-Hui
Y
ang)
2707
12% re
sp
ecti
vely. The co
mpari
s
o
n
is
made by
the
(100
8, 504
) matrix [19] with 0.18
m
stand
ard
cell
libra
ry, and the decodin
g
iteration
set at eighty.
(a)
(b)
Figure 7. Dat
apath of varia
b
le nod
e and
che
c
k nod
e messag
e pa
ssing, (a
) tradit
i
onal
architectu
re, (b) propo
se
d TC ba
sed a
r
chitecture.
Figure 8. Hardwa
re redu
cti
ons
with ne
w TC ba
sed m
e
ssage p
a
ssin
g architectu
re
.
5.3.
Lo
w
Har
d
w
a
re Cos
t
Design
s
For l
o
w ha
rd
ware
co
st AS
IC imple
m
ent
ation, not
onl
y wa
s the
qu
asi
bina
ry we
ighting
addition a
r
chi
t
ecture
ado
pted, but the o
p
timized
bi
t-width of L
U
T
s
wa
s al
so e
m
ployed for
better
hardware effic
i
enc
y
. The
new low-cos
t
log-SPA LDP
C
decoder,
compos
ed of
OR
pseudo-carry
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 5, May 2013 : 2699 – 270
9
2708
architectu
re, the
TC-ba
s
e
d
me
ssage p
a
ssing dat
a
path, and
the
LUT
3-4 ma
pping ta
ble,
are
impleme
n
ted.
Simulation
re
sults,
sho
w
n i
n
Figu
re 9,
d
e
mon
s
trate th
e ne
w de
co
d
e
r (OR p
s
eud
o-
carry) perfo
rming
a re
aso
nable
BE
R. Whe
n
comp
a
r
i
ng it
with th
e ori
g
inal SP
A, it only has 0.2
dB loss at
BER. Ha
rd
ware co
mpa
r
ison
s, usi
ng t
he
(1008, 5
04) m
a
trix [19] and
a 0.18
m
stand
ard
cell
libra
ry unde
r
the sam
e
logi
c synth
e
si
s condition
s, are
sho
w
n in Fi
g
u
re 1
0
. For th
e
lowe
st ha
rd
ware
co
st, the
qua
si bi
nary
weig
hting
a
ddition
with
TC-ba
s
ed
archite
c
ture
is
the
sug
g
e
s
ted co
mbination.
SNR
(dB)
0.
0
0
.5
1.
0
1
.5
2.
0
2
.5
Bit
Erro
r Ra
te
10
-5
10
-4
10
-3
10
-2
10
-1
10
0
Si
gn
-
M
in
OR
Op
er
a
t
i
o
n
O
R
O
p
e
r
atio
n
w
ith
P
s
eu
do
Ca
rry
L
U
T 3-4 [6]
L
U
T 3-7 [6]
SPA
Figure 9. BER com
p
a
r
iso
n
for (10
08, 5
04)
L
D
PC
co
de with eig
h
ty deco
d
ing iterations.
Figure 10. Area com
p
a
r
iso
n
of
low hard
w
are co
st architectu
res u
s
i
ng 0.18µm st
anda
rd cell
libra
ry
.
6. Conclusio
n
For e
rro
r
correctio
n
ha
rd
ware im
pleme
n
t
ation,
su
ch a
s
in a
n
L
D
PC decode
r, errors are
allowed in th
e nume
r
ical algorith
m
. According to
th
e cha
r
a
c
teri
stic, the value follows
ce
rtain
rule
s when t
he num
eri
c
al
rang
e of ma
pping ta
ble i
s
slightly adj
u
s
ted, an
d re
p
l
ace
s
tra
d
itio
nal
adde
rs with
si
mple O
R
g
a
t
e
to exe
c
ute
addition
s.
Nu
meri
cal
simul
a
tion results f
u
rthe
r p
r
ove t
hat
the propo
se
d
qua
si bi
nary
weig
hting
ad
dition coul
d
succe
ssfully p
r
oce
ed
with L
D
PC
de
codi
n
g
.
More
over, th
e modified
T
C
-b
ased m
e
ssage
pa
ssi
n
g
tech
niqu
es can
effectively simplify the
Evaluation Warning : The document was created with Spire.PDF for Python.