TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4393 ~ 4
3
9
9
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.547
3
4393
Re
cei
v
ed
De
cem
ber 2
1
, 2013; Re
vi
sed
Febr
uary 10,
2014; Accept
ed Feb
r
ua
ry
22, 2014
Real-time Implementation of a xAVS
Video Decoder
Qing Cha
ng, Xin Liu,
YaoLi Wang*
Coll
eg
e of Information En
gi
ne
erin
g, T
a
iyuan
Univers
i
t
y
of
T
e
chn
o
lo
g
y
,
T
a
iyua
n, 030
0
24, Chi
na, Ph./F
ax: +
008
6-35
160
14
052/3
5
1
601
40
52
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:
w
i
lli
ng
w
a
n
@
gmail.c
o
m
A
b
st
r
a
ct
AVS vid
e
o
dec
oder
cons
u
m
e
s
a
hu
ge
nu
mb
er of c
o
m
put
ati
on, so
re
al-ti
m
e
i
m
ple
m
entati
on of
a
n
AVS decoder has some challenging on
x86 c
o
mputing platform
. This arti
cle describes a x
AVS open sour
c
e
proj
ect to solve this prob
le
m.
F
i
rst of all, th
e reaso
n
for the low
efficienc
y of t
he code
of the existin
g
AV
S
vide
o deco
der
ope
n sourc
e
re
ference softw
are RM52J
_r
1 i
s
analy
z
e
d
acc
o
rdi
ng to the d
e
scripti
on of AVS
key techn
o
lo
gi
es an
d dec
od
i
ng pr
inci
ples
i
n
the offici
al
docu
m
entati
o
n
.
T
hen accor
d
ing to th
e
ma
in
problem
s
of the referenc
e software, re-design the
optim
i
z
ed xAVS
dec
oder arc
h
itectur
e
, and real-time
prop
erty sig
n
ifi
c
antly i
m
prove
d
w
i
th C co
de.
F
i
nal
ly,
use
the x8
6 p
l
atfor
m
multi
m
e
d
ia
i
n
struction s
e
ts
to
further opti
m
i
z
e xAVS seman
t
ic proce
ssor. The exper
i
m
en
tal results sho
w
that, under the prec
on
ditio
n
o
f
ensur
ing the quality of dec
oding, the dec
odi
ng speed of the xAVS decoder fo
r D1 has increas
ed
m
o
re than
14 times, to full
y meet the n
e
e
d
s of real-ti
m
e
deco
d
in
g
.
Ke
y
w
ords
:
xA
VS, decod
er, real-ti
m
e
deco
d
i
n
g
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
AVS is independently developed
by Chinese as the
second
generation
source codi
ng
standard. AVS video codec performance is 2 to 3 ti
mes higher than MPEG-2, and quite well with
H.264 [1
-4],
but the imple
m
entation
co
mplexity is
lo
wer t
han the
H.264. In o
r
der to ve
rify the
c
o
rrec
tness
of Algorithm,
AVS work
ing group
s
u
bmitted the reference
s
o
ftware RM52J
_
r1,
but
RM52
J_
r1
was
not optimi
z
ed
for
a pla
tform. The
d
e
co
ding
sp
ee
d of RM52
J_
r1 i
s
le
ss th
an
satisfa
c
to
ry, and the
de
co
ding
rate i
s
o
n
ly 3 to 5
per se
con
d
in th
e cu
rrent hi
g
h
-en
d
x86
ke
rnel,
so there i
s
a
big gap for practical appli
c
ations
. The task
of implem
enting AVS algorithm
s on t
he
x86 platform
was propo
sed by AVS Workgroup, that is
xAVS open source pr
oject. Our task
force a
s
sum
e
d the optimization of xAVS deco
der.
This paper first introduces
the structure and characteristi
cs
of AVS, then according to
the defic
ienc
ies
of AVS referenc
e
s
o
ft
ware, pr
oposes
the
s
o
lution of xAVS
decoder, whic
h
desi
g
n
s
a
ne
w d
e
codin
g
p
r
ocess and
p
r
ogra
m
stru
ctu
r
e, an
d a
c
hi
e
v
es the
go
al
with
C
cod
e
.
By
optimizatio
n
of the multimedia in
stru
cti
on set,
the D1 decodin
g
speed i
s
more
than 50 fra
m
es
per second, so it meets the
real-time
req
u
irem
ents.
2.
Deco
ding Principles of AVS
AVS dec
oding algorithm s
t
ruc
t
ure is
s
hown in Figure
1.
Figure 1. Struc
t
ure of AVS Dec
o
der
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4393 – 4
399
4394
The
de
codi
n
g
p
r
o
c
e
s
s is:
the
re
ceive
d
bit
stre
am
whi
c
h
i
s
co
mpre
ssively
encode
d
enters the in
put terminal
of t
he deco
d
e
r, and a
c
cording to the e
n
tropy de
codi
ng it can obt
ain a
seri
es of
para
m
eters. On
th
e on
e h
and, t
he q
uanti
z
ed
coeffici
ents can b
e
o
b
taine
d
through
a
n
ti-
scanni
ng th
e
pa
ramete
rs, after
de
-qu
a
n
tization
an
d
inverse
tra
n
s
form
ation it
ca
n
obtain
the
resi
dual coeff
i
cient
s.
On the othe
r hand, a
c
cording to the rece
ive
d
predi
ction mo
de a
nd data the
decode
r
obtain
s
the predi
cted valu
es of intra predictio
n or
inter predictio
n, the re
co
nstructed value
s
are
the sum
of th
ese
p
r
edi
cted
value
s
and
the
re
si
dual
coefficient. Fi
n
a
lly, though
l
oop filte
r
the
s
e
recon
s
tru
c
ted
values can g
e
t the final image value
s
.
3.
The De
sign of xAVS Dec
oder
The de
codi
n
g
referen
c
e
softwa
r
e RM
52J_r1
whi
c
h
was relea
s
e
d
by the official have
alrea
d
y reali
z
ed.
All syntax elements of the
AVS standard and ba
si
c
semantics, but
the efficiency of th
e
code is
still
very low. Analysis
of the reasons are [5-7]: the st
ructure of
ref
e
rence code
is
irratio
nal, so there i
s
a
larg
e numb
e
r
of unne
ce
ss
ary
crite
r
ia a
nd ju
mp; the refe
rence code
exists
compl
e
x multi
-
cy
cle, a
nd th
ere
are m
u
lti-level ne
sted
call
s b
e
twe
e
n
functio
n
; vari
able
definitio
ns
are n
o
t unif
o
rm, too ma
ny duplicate
definition;
the dynami
c
memory all
o
cation
wa
s
use
d
repe
atedly, the CPU
resou
r
ce
s wa
s
con
s
umed u
n
re
as
onably, so th
e efficien
cy of the memory
is
low. For the
main problem
of RM52
J_r1
, this
paper p
r
ese
n
ts a ne
w des
ig
n of xAVS decod
er.
Table 1. DT
C-CSF System
Table 2. Indu
ction Ma
chin
e Paramete
rs
Parameter
value
Proportional gai
n, Kp
36
Integral gain, Ki
12643
Flux h
y
s
t
eresis band
0.01 Wb
Sampling frequ
enc
y
40kHz
Sw
itching freq
u
enc
y
4kHz
Parameter
value
Stator resistance
5.5
Ω
Rotor r
e
sistance
4.51
Ω
Stator self inductance
306.5 mH
Rotor self inductance
306.5 mH
Mutual inductance
291.9 mH
Momen of inertia
0.01 kg.m
2
Number of
poles
4
Rated speed
1410 rpm
DC-link voltage
565 V
3.1. xAVS
Proce
s
s
Figure 2. The
Decodin
g
Proce
s
s of xAVS
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Real
-tim
e Implem
entation of a
xAVS Vi
deo Decoder (Qing
Chang)
4395
The
decoding process of
x
AVS is
shown i
n
Fi
gure 2. The im
age data is divi
ded i
n
to
some f
r
ame
s
.
One fra
m
e is a slice. Every slice h
a
s
a
different sta
r
t
cod
e
. First the sta
r
t co
de
o
f
image h
eade
r inform
ation
is analy
s
ed, t
he sli
c
e ty
pe
can b
e
got by the start
cod
e
. Then t
h
e
decode
r choo
se
s different
decodin
g
pro
c
e
ss
by the
slice type. Wh
en pa
rsing
o
u
t the sta
r
t code
is I, P, or B picture h
ead
er,
we de
cod
e
them re
sp
ecti
vely.
3.2. Module
Op
timization
AVS us
es
block
-
based hy
brid coding framewor
k
.
Its
c
o
re tec
h
nologies
inc
l
ude [8-10]
entropy de
coding, inverse quanti
z
atio
n and inve
rs
e tra
n
sfo
r
m
a
tion, intra
predi
ction, in
ter
predi
ction, an
d loop filter.
1) Entropy de
codi
ng
Acco
rdi
ng to
the
syntax
element [1
1-14],
entropy
decodin
g
can
gen
erate
Le
vel array
and
Ru
n a
r
ra
y. Level a
rra
y contai
ns th
e am
plitude
of
the non
-ze
r
o qua
ntized
coeffici
ents, run
array contai
ns the n
u
m
ber of
con
s
ecutiv
e
zero
before th
e
current n
o
n
-zero qua
ntized
coeffici
ents.
The p
r
o
c
e
s
s
of entro
py de
codi
ng i
s
that
, First
determ
i
ne if the
en
coded
data
of
the
c
u
rrent 8x8
block
exis
ts
or not, If it e
x
is
ts
,
initialize the ma
ppi
ng table
and
jump thresh
old,
otherwise exi
t. Then tran
s_co
efficient i
s
pa
rsed,
if tran
s_
coeffici
ent is le
ss th
an 59, qu
anti
z
ed
coeffici
ents
a
nd ru
n
can b
e
get by loo
k
up Curr
entVL
CTabl
e, So L
e
vel array an
d Ru
n array
can
be got .If trans_coeffici
ent i
s
bi
gger than
or e
qual
to 59, escape_level_diff shoul
d
be
parsed
out,
Level arrays
and Ru
n arra
y can be cal
c
ulated by
escape_l
evel_dif
f. If trans_co
e
fficient equa
l to
EOB, then th
e end of the block co
effici
ent deco
d
ing.
Finally, acco
rding to ab
sL
evel to update
the numbe
r o
f
the next stopwat
ch.
From the pro
c
e
ss of entro
py decodi
ng, if tr
ans_
c
oeffi
cient less tha
n
59, Level and Run
are
u
s
ed
in t
he
Cu
rre
ntVLCTabl
e a
n
d
e
s
cape
_level
_
d
iff isn’t u
s
e
d
.
If trans_coe
fficient is big
ger
than o
r
equ
al
to 5
9
, only
escap
e_level
_diff is u
s
ed.
The
case
of less than
59
is u
s
ed
in
the
majority of the ent
ropy
decode
d; in
most
ca
se
s escap
e_level
_diff
of Cu
rrentVLCT
able
is
superfluous.
Thus two st
ructures are
defined
when we design
code
table in
xAVS decoder.
Structu
r
e con
t
ains Level
a
nd Ru
n will
be u
s
ed in
t
he case
of less than 5
9
,anothe
r st
ru
cture
whi
c
h only
contain
s
esca
pe_level_
d
iff will be u
s
e
d
in other
ca
se
s. Two ta
ble
s
are u
s
ed f
o
r
storing a VLC code tabl
e in xAVS decoder.
Whe
n
a
c
cord
ing ab
sL
evel
value to up
da
te
the serial
n
u
mbe
r
of a
st
opwatch,
RM
52J_r1
uses a seri
es
of
judgment statement, so
increa
sing the running tim
e
of
the code.
xAVS decoder
use
s
the
met
hod of lo
oki
n
g-up t
able in
stead
of
co
n
d
itional ju
dg
ment. The
switchi
ng rule
s of
cod
e
table are sho
w
n in T
able 3 to Tabl
e 5.
Table 3. xAVS Intra Luma
Table Swit
c
h
ing Rules
absLevel
0
1 2 3 4 5 6
7 8 9 10 >10
table_num
0
1 2 3 3 4 4
4 5 5 5
6
Table 4. xAVS Inter Luma
Table Swit
c
h
ing Rules
absLevel
0 1
2 3 4 5 6 7
8 9 >9
table_num
0 1
2 3 4 4 4 5
5 5 6
Table 5. xAVS Intra Luma
Table Swit
c
h
ing Rules
absLevel
0 1 2
3 4 >4
table_num
0 1 2
3 3 4
2) Inverse qu
antizatio
n an
d Inverse tran
sform
a
tion
The intege
r inverse qua
ntization an
d inverse
tran
sf
ormatio
n
of 8x8 block a
r
e
use
d
in
AVS. Dimens
ional integer invers
e trans
f
orma
tion
c
an
be dec
o
mposed into horiz
ontal
and
vertical one-dimensi
onal integer
inverse transformati
on. xAVS
and RM52J_r1
are
basi
cally the
s
a
me, but xAVS has
all
z
e
ro block
judgment befor
e t
he invers
e trans
f
ormation.
When the
c
u
rrent
8x8 sub
-
blo
c
k i
s
th
e all
-
ze
ro
blo
ck, thi
s
block
doe
sn’t
exist inve
rse t
r
an
sform
a
tion
, becau
se
after
inverse transf
ormation these bloc
k are st
ill the a
ll-zero block.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4393 – 4
399
4396
3) Intra p
r
edi
ction
Intra pre
d
icti
on incl
ude
s
determi
ning t
he pre
d
ictio
n
mode of e
v
ery 8x8 su
b-blo
c
k
,
getting refere
nce
sa
mple v
a
lue a
nd
cal
c
ulating p
r
edi
cted value
s
. In RM5
2
J_r1, I
t
firstly obtain
s
and
save
s th
e refe
ren
c
e
pixel value,
and the
n
d
e
termin
es
a p
r
edictio
n mo
d
e
to p
r
edi
ct. The
processes
of xAVS is different
from RM
52J_r1.
Fi
rst it
determi
nes
the prediction mode of
each
8x8 su
b-block, then
rea
d
s the corre
s
po
nding
pixel
value
s
a
c
cordi
ng to the
pre
d
iction
mod
e
l, so
many unne
ce
ssary refe
ren
c
e pixel value
s
ca
n be avoi
ded.
4) Inter predic
t
ion
In AVS, P or
B frame
has
t
w
o
referenc
e frames
at
mos
t, P frame
c
a
n refer to t
he mos
t
forwa
r
d
ne
arest d
e
code
d
I frame
or P f
r
ame
,B fr
am
e can
refe
r to
a front
and
rear of the
mo
st
nearest
decoded I frame
or P frame. I
n
xAVS t
he
macrobl
ock
decoding of P
and B f
r
ame
are
simila
r to I-frame, al
so a
c
cording
to th
e ty
pe of the
macrobl
ock
to decode. F
i
rst, obtai
n th
e
corre
s
p
ondin
g
refere
nce index of 8x8 sub-blo
ck. Next, based
on the motion vector of
the
adja
c
ent blo
c
k to obtain a predi
ction mo
tion vector
of
the current b
l
ock. recon
s
truct them to get
motion vecto
r
of the cu
rren
t bl
ock. The
correspon
ding
referen
c
e
sa
mples
ca
n be
got throu
gh t
he
motion ve
cto
r
a
nd th
e
ref
e
ren
c
e
ind
e
x. If the reference
sam
p
le
are n
o
t in
the inte
ge
r pi
xel
locatio
n
, so these refe
ren
c
e sampl
e
s
sh
ould be inte
rp
olated to obta
i
n the predi
ct
ed value
s
.
In order to op
timize the efficien
cy of cod
e
execution, the inter pr
edi
ction of xAVS has a
different de
si
gn idea
s from
RM52
J_
r1.
Firstly, in
RM52J_r1, cal
l
oc fu
nctio
n
i
s
ca
lle
d at
the
start
of a
n
ima
ge
de
coding
to
dynamically allocate mem
o
ry. After the end of
an im
age de
co
ding
relea
s
e the
memory, so the
effic
i
enc
y
of
this
memory
is
very low.
In
xAVS dec
oder, the dynamic
memory alloc
a
tion of
referen
c
e fra
m
e are p
u
t before
de
cod
i
ng the entire
video seq
u
e
n
ce, an
d the
s
e mem
o
ry a
r
e
relea
s
e
d
afte
r the en
d of the entire vide
o se
quen
ce.
Secon
d
ly, in the de
codi
ng
pro
c
e
ss
of ea
ch
frame, re
sid
u
a
l data, the
predi
ctive val
ue and
th
e reco
nstructe
d
value are st
ored
by a on
e
-
dimen
s
ion
a
l
array, the di
mensi
on of t
he array a
r
e
redu
ce
d. Third, the P fram
e or B frame
can
have up to t
w
o referenc
e frames
, therefore x
AVS ass
i
gns
three frame buffers
to
s
t
ore
the
informatio
n o
f
the currentl
y
decod
ed frame a
s
well
as the two re
feren
c
e fra
m
es. And expa
nds
boun
dary of frame b
u
ffer, thus it’s b
e
tter for interpol
ation optimization.
Duri
ng the interpol
ation proce
s
s, if the r
e
fere
n
c
e inte
ger sample i
s
out of the referen
c
e
image, inste
ad it with the nearest int
eger
sam
p
le
(edge
or co
rne
r
of the sample
) whi
c
h
is
nearest
to th
e refere
nce
sample t
hat i
s
motion
vecto
r
can
poi
nt to
sa
mple
s
whi
c
h
are o
u
t of
a
referen
c
e im
age [15-22]. So during the
interpolati
o
n
process, eve
r
y refere
nce pixels sh
ould
be
judge
d wheth
e
r is out
side
the bound
ary
of the re
fere
nce ima
ge. The max and min function
are
use
d
in refe
rence software, so
that ea
ch 8x8
sub
-
b
l
ock nee
ds
6
4
judgm
ents,
the efficien
cy o
f
the program is
reduc
ed.
There is
boundary extens
ion of
frame buffer s
p
ace in xAVS dec
oder,
the ne
are
s
t i
m
age
bou
nd
ary pixel
will i
n
stea
d of th
e
pixel which i
s
b
e
yond th
e
image
bou
nd
ary.
So du
ring
the
interpolation
pro
c
e
ss,
the
referen
c
e
pixe
l doe
sn’t
judg
e whethe
r it
is o
u
t of
bou
n
d
s
or not.
5) loop filter
There a
r
e th
ree filterin
g m
ode
s: strong
filtering (BS
=
2), st
a
nda
rd
filtering (BS
=
1) a
nd
non-filter (BS=0). In this
article, the design idea of xAVS dec
oder is that, after the end of
one
image d
e
codi
ng, all the m
a
croblo
c
ks
of one fram
e
a
r
e filtered
wit
h
a ra
ster
scannin
g
meth
od
cycle. Th
e B
S
of every 8x8 macro
b
lock sho
u
ld be
g
o
t firstly, and
then filter the
luma an
d ch
roa
boun
dary. If the
curre
n
t de
codi
ng im
age
s i
s
I frame, t
hen BS i
s
eq
ual to
2,so
th
e calculation
of
the filter fun
c
t
i
on is omitted
.
For P a
nd
B frame,
the
r
e is
no ve
rtical bou
nda
ry filtering
of 16x
16
and 16x8 ma
cro
b
lo
cks, an
d no hori
z
ont
al bound
ary
filtering of 16
x16 and 8x1
6
macroblo
cks.
Whe
n
the
s
e
condition
s
are
met, the
corresp
ondi
ng B
S
don’t n
eed
to cal
c
ul
ate.
For filte
r
ed
pi
xels
whi
c
h
ch
rom
a
blo
c
k n
eed
s i
s
le
ss th
a
n
luma
bl
ock, so
the
ch
ro
ma filter fu
nction an
d
ch
ro
ma
filtering func
tion was
divided in
xAVS decoder. After s
u
c
h
optimi
z
a
tion, the loop filter module
saves a lot of calculation in xAVS decoder
, so the decoding
speed i
s
improved.
3.3. SIMD
Optimizatio
n
SIMD is a
CPU execution
mode
whi
c
h
is si
ngle
-
in
structio
n an
d e
x
ecute m
u
lti-cha
nnel
data for x86
platform, the
plurality of el
ements to
be
pro
c
e
s
sed in
a parallel ma
nner, to imp
r
ove
the speed of the pr
ogram. In this paper, xAVS decode
r was optimized with MMX and
SSE2.mainly optimize for inverse transform
ation,
interpol
ation, and loop filter module
whi
c
h
inclu
de a larg
e amount of calcul
ation.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Real
-tim
e Implem
entation of a
xAVS Vi
deo Decoder (Qing
Chang)
4397
H
o
r
i
z
o
n
t
a
l
in
ve
r
s
e
tr
a
n
s
f
o
r
ma
tio
n
o
f
in
ve
r
s
e
tr
a
n
s
f
o
r
ma
tio
n
c
a
n
be
o
p
t
imiz
ed
w
i
th
g
r
oup
s
u
m,
g
r
ou
p su
b
t
r
a
c
t
io
n as w
e
ll as
g
r
oup
sh
ift o
f
MM
X, the matrix
transpo
sition
is a
critical p
a
rt
[23-25]. MM
X instru
ction
s
can o
n
ly handl
e four
16-bit o
r
two
32-bit
data,
so the i
n
ve
rse
trans
formation module
was
ac
hieved with SSE2
[6]. SSE2 is
bas
ed
on
128-bit regis
t
ers
,
it’s
easi
e
r to impl
ement inverse transf
o
rmati
on and mat
r
ix transp
o
sitio
n
.
For exam
ple,
in orde
r to o
b
tain one
-hal
f of the b example, this p
aper
de
script
how to
use the
MM
X to achieve
the optimiza
t
ion of the
in
terpolatio
n m
odule. By filtering
aro
und
the
hori
z
ontal di
rection o
n
the
four intege
r
point On
e
-
h
a
lf of the example b obtain
s
an interm
edia
t
e
valueb1
=(-C+5D+5E-F
), p
r
edi
ctive value b=max(
0,
min(2
55,(b
1+4)>>3
)) i
s
o
b
tained by t
he
interme
d
iate
valueb
1.Th
e main
in
structio
ns a
r
e
movd, p
s
h
u
fw, pu
np
ckl
bw, jn
z, p
a
ddw,
packu
swb. T
he fun
c
tion o
f
packu
swb i
s
limiting
ca
l
c
ulate
d
predi
ctive value from 0 to 25
5.The
reali
z
ation of
the pro
c
e
ss i
s
sh
own in Figure 3.
Figure 3. Sch
e
matic Di
ag
ram of Calcula
t
ing the Half Interpol
ation
The optimiz
at
ion of the loop filter is
ac
hi
eved by SSE2.For different filer s
t
rength, the
hori
z
ontal filter and verti
c
al filtering are writt
en sep
a
rately. The array which is rea
d
into the
regi
ster 12
8 in the vertical
filtering is sho
w
n in
Figu
re
4.In the operation of
the re
gisters, the d
a
ta
of the re
giste
r
s
are
rig
h
t of Figure 4 .Th
e
refo
re, in
th
e vertical filter matrix tra
n
spose shoul
d
be
operated firstly, in order to ac
hi
eve a parallel data pro
c
e
ssi
ng.
The syste
m
is fully optimized in a
c
cord
ance
with the
above meth
ods, an
d the effect is
desi
r
abl
e.
xmm0: a0 a1 a2 a3
a0 b0 c0 d0
xmm1: b0 b1 b2 b3
a1 b1 c
1
d1
xmm2: c
0
c
1
c
2
c3
a2 b2 c
2
d2
xmm3: d0 d1 d2 d3
a3 b3 c3 d3
Figure 4. Reg
i
ster Array of Vertical Filte
r
4. Experimenta
l
Results
The test env
ironm
ent of
xAVS decoder i
s
Windows7, Pentium dual
-core processor,
1.8GHz, com
p
iled and debugging in Visual St
udio 2008.The fram
e order i
s
IBBPBBP.
AVS
test
strea
m
of
30
000 f
r
ame
in
cludi
ng th
e type of
QC
IF
(176x14
4),
CIF (3
52x
28
8) and D1
(7
20x
576)
are te
st, and
comp
ared
with RM
52J_r1
decode
r. The
te
st re
sult
s o
f
frame rate is sho
w
n i
n
Ta
ble
6, and the test results
of
PS
NR i
s
shown i
n
T
able
7. From
Tabl
e
6 and T
able
7, xAVS decoder
and
RM52J_r1 decoder have
same PS
NR, the
decodi
ng
speed
of
xAVS decoder is
greater than
RM52
J_
r1
d
e
co
der.
The
decodin
g
sp
eed of
Q
C
IF,
CIF
and
D1
is re
sp
ectiv
e
ly 22.14
tim
e
s,
16.26 times a
nd 14.29 time
s than RM52
J_r1 de
cod
e
r.
Table 6. Frame Rate Tes
t
Results
Cont
ra
s
t
of xAVS
Dec
o
der and RM52J
_
r1
Dec
o
der
Test Sequence
Frames
Decoding speed(
fps)
RM52J_r1
x
AVS
akiy
o. qcif
30000
57.91
1282.05
foreman. cif
30000
13.58
220.75
SOCCER.
D1
30000
3.54
50.59
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4393 – 4
399
4398
Table 7. PSNR Test Resu
lt
s Contrast of xAVS De
coder and
RM52J_r1 Decoder
Test Sequence
RM52J_r1
xAVS
PSNR_
Y PSNR_U
PSNR_V
PSNR_
Y
PSNR_U
PSNR_V
akiy
o.
qcif
38.02
42.24
43.21
38.02
42.24
43.21
foreman.
cif
34.75
41.30
42.93
34.75
41.30
42.93
SOCCER.
D1
34.53
42.21
44.03
34.53
42.21
44.03
5. Conclu
sion
In the fou
n
d
a
tion of
anal
yzing
and
re
sea
r
ching
RM52J_r1, thi
s
p
ape
r
puts forward
optimiz
ed improvement
programs
and realizes
a
new x
AVS decoder. In the premis
e o
f
maintainin
g t
he o
r
iginal
i
m
age
quality
,
deco
d
ing
s
peed
ha
s b
e
en g
r
eatly
i
m
prov
e
d
. x
AVS
decode
r meet
s the re
quire
ment
s of re
al-time deco
d
in
g.
Ackn
o
w
l
e
dg
ements
It is a project su
ppo
rte
d
by Shanxi Nati
onal Natural Scie
n
c
e Fou
ndati
ons (No.
2013
0110
15
-1).
Referen
ces
[1]
Gao Wen, Wang Qiang.
Digital Audio Video Co
ding Standard
of AVS.
Z
T
E
Commu
n
i
c
ations
. 20
06;
12(3): 6-9,1
3
.
[2]
Gong
Xia
o
x
i
a
, Liu
Xi
ng
w
a
n
g
. An achi
eveme
n
t of t
he full-I-frame real- tim
e
deco
d
in
g on
PC platform.
Scienc
epa
per Onlin
e
. http://
w
w
w
.
paper.edu.
c
n /index
.
php/def
ault/ rele
ase
pap
er/conte
n
t/200
91
1-16
2.
[3]
Li Yan, Liu
Xiulan. Bottl
eneck analy
sis and optimizati
on of AVS sof
t
w
ar
e decoder
.
Electroni
c
Measur
e
m
ent T
e
chno
logy.
2
010; 33(
4): 128
-130.
[4]
Audi
o Vi
deo
codi
ng Sta
n
d
a
r
d W
o
rkgro
u
p
of Ch
ina.GB
/T
20090.2-
20
0
6
.
Information te
chnology-
Advanc
ed co
di
ng of au
dio a
n
d
vide
o-Part2: Vide
o
. 200
6.
[5]
H Malv
ar, A
Hall
ap
uro, M
Karcze
w
i
cz,
L
Kerofsk
y
. L
o
w
-
C
omp
l
e
x
it
y t
r
ansform
and
qua
ntizati
on
in
H.264/AVC.
IEEE Trans.Circu
its Syst.Video Techno
l
. 200
3; 13: 598-6
03.
[6]
W
e
i F
ang, L
i
Xu
emi
ng. SIMD Optimiza
ti
on of T
r
ansform and Ou
ant
izatio
n in
H.2
64.
Co
mpute
r
Engi
neer
in
g an
d Appl
icatio
ns
. 200
4; 17: 24-2
7
.
[7]
Martuza M
uha
mmad, W
a
h
i
d
Kha
n
A. Imp
l
ementati
o
n
of
a c
o
st-shar
e
d
transform
ar
chitecture
for
multipl
e
vid
eo
coder.
Jour
nal
of Real-T
i
m
e Ima
ge Proc
essi
ng
. 201
2; 1-12.
[8]
Lv Qia
n
, Cu
i L
i
gon
g. F
P
GA verificati
on for
the OR
120
0 s
u
bs
y
s
t
e
m in
AV
S-SoC.
Adv
a
nced Mat
e
rials
Rese
arch(ICE
M 2012)
. 20
13;
651: 80
7-81
1.
[9]
Hu Qian, Zhang K
e
, Yu Lu.
Decoder
arc
h
itecture and
hard
w
ar
e implementation for
AVS-video.
Journ
a
l of Z
hej
ian
g
Univ
ersity
(Engin
eeri
ng
Scienc
e)
. 200
6
;
40(12): 21
39-
214
3.
[10]
W
ang F
e
i,
Li
Yuan, Ji
a H
u
iz
hu, Xie
Xia
o
d
ong,
Ga
o
W
e
n
.
An efficient fractional
m
o
tion
estim
a
tion
architectur
e
for AVS real-tim
e full HD video
encoder
. Proc
eedings
of IEEE In
ternational Conferenc
e
on Imagi
ng S
ystems and T
e
chni
ques(I
ST
2012 - 201
2). 201
2; 279-2
84.
[11]
Ding
Da
n-d
a
n
.
Rese
arch
on R
e
co
nfig
urabl
e Vi
deo
Codi
ng.
J
ourn
a
l of Z
h
eji
a
n
g
Un
iversi
t
y
(Engi
neer
in
g Scienc
e)
, 201
1.
[12] Dan
dan
Di
ng,
Hon
gga
ng Qi,
Lu Yu, W
en G
ao. Rec
onfi
gur
abl
e Vid
eo
Co
din
g
F
r
am
w
o
rk
and
Deco
de
r
Reconfigur
able Inst
antiation of AVS.
Signal P
r
ocessi
ng: Ima
ge Co
mmun
. 2
009; 24(
4): 287
-299.
[13]
DING Da
n-da
n
,
Yu L
u
. Overv
i
e
w
of MPEG
Reco
nfigur
ab
le
Vid
eo
Co
din
g
.
Vid
eo E
n
g
i
ne
erin
g
. 20
09
;
33(7): 12-
15.
[14]
W
A
NG W
en-xi
ang, SHE
N
Ha
i-hu
a. A Reco
n
f
igura
b
le S
ub-
Pixel Inter
pol
ation Arc
h
itectur
e
Desi
gn f
o
r
Multi-stan
dard
Video D
e
cod
i
ng.
Journ
a
l of
Computer-Ai
d
ed Desi
gn &
Co
mp
uter Graphics
. 20
11;
23(9): 16
03-
16
13.
[15]
DU Jua
n
, DING Dan-da
n, YU
Lu. Desi
gn
method
olo
g
y
o
f
F
P
GA based reconfi
gura
b
l
e
vide
o enc
oder
.
Journ
a
l of Z
hej
ian
g
Univ
ersity
(Engin
eeri
ng
Scienc
e)
. 201
2
;
46(5): 905-9
1
1
.
[16]
ZHAO Jing, Z
H
OU Li, Y
U
Q
i
ng-
dong, et
al. Research
of audio
v
i
deo
st
and
ard (AVS)
HD decoding
base
d
on a rec
onfig
urab
le pr
o
c
essor.
Journ
a
l
of Harbin En
gi
neer
ing U
n
iv
ersity
. 2012; 33(
2): 226-2
33.
[17]
W
ang Jin, Yua
n
Bin, Z
hang
Gang. T
he design of I frame based o
n
the
x26
4
archit
ecture.
Ta
iyu
a
n
Univers
i
ty of Techn
o
lo
gy
. 201
0; 41(2): 13
9-1
41, 146.
[18]
Bai Yuting. Hard
w
a
r
e
implement
of the key
m
odu
les
in AVS enc
oder.
T
a
iyua
n
Univers
i
ty of
T
e
chno
logy
. 2
012.
[19] X
i
ang
Hongli.
FPGA
implem
entation of
predicti
on module for AVS
encoder.
Taiy
uan
Univers
i
ty
o
f
T
e
chno
logy
. 2
012.
[20]
LI F
u
ji
ang.
T
he res
earch
of
ke
y a
l
gor
ithms
for re
al tim
e
a
vs enc
ode
a
n
d
the ASIC
d
e
si
gn
base
d
on
clou
d on ch
ip.
T
a
iyua
n Univ
er
sity of T
e
chnol
ogy
. 201
1.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Real
-tim
e Implem
entation of a
xAVS Vi
deo Decoder (Qing
Chang)
4399
[21]
LI Fujian
g
, Zha
ng Gang. Quic
k AVS in
tra predicti
on mod
e
selecti
on al
gor
i
t
hm.
Comp
uter
Engin
eeri
n
g
and Ap
plic
atio
ns
. 2010; 4
6
(2)
:
1-3.
[22]
Jian
g T
ao, Z
h
ou P
e
ih
ai. MI
N of B
aha
dur
KC.
Intra pr
edi
ction
in
H.26
4
an
d AVS
du
a
l
-mode
vi
deo
deco
der h
a
rd
w
a
re des
ign
and
impleme
n
tatio
n
.
Applic
atio
n of Electronic T
e
chn
i
qu
e
. 200
7; (9): 41-44.
[23]
Zhao Lo
ng
hui,
Che
n
Xinh
ua, Ren Hu
ai L
u
. AVS-base
d
ent
rop
y
dec
oder d
e
sig
n
.
ITS AP
PLICATIONS
.
201
0; (20): 53-
55 ,58.
[24]
Yi Bo-
n
ia
n. T
u
rbo C
o
d
e
D
e
si
gn
and
Im
pl
em
entatio
n of
Hi
g
h
-Spe
ed Para
ll
el
D
e
co
der.
T
E
LKOMNIKA
Indon
esi
a
Jour
nal of Electric
al
Engin
eeri
n
g
. 2
013; 11(
4): 211
6-21
23.
[25]
Yu F
an,
Yan
g
Hu
iji
ng,
Li G
ang. A
H
i
gh
Perf
ormanc
e
Sigma-D
e
lta
A
DC for
Au
dio
Deco
der
Chi
p
.
T
E
LKOMNIKA Indon
esi
a
Jour
nal of Electric
al
Engin
eeri
n
g
. 2
013; 11(
11): 65
70-6
576.
Evaluation Warning : The document was created with Spire.PDF for Python.