I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
8
,
No
.
1
,
Feb
r
u
ar
y
201
8
,
p
p
.
70
~
75
I
SS
N:
2088
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v
8
i
1
.
p
p
7
0
-
75
70
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e
.
co
m/
jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
I
JE
C
E
O
pti
m
i
z
ing
App
le
Lo
ss
less
Audio C
o
dec
A
lg
o
rith
m
u
sing
NVIDI
A
CUDA
A
rchite
ctur
e
Ra
f
id Ah
m
ed,
M
d.
Sa
zz
a
du
l Is
la
m
,
J
ia
Uddi
n
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter S
c
ien
c
e
&
En
g
in
e
e
rin
g
,
BRA
C
Un
iv
e
rsity
,
Ba
n
g
lad
e
sh
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
4
,
2
0
1
7
R
ev
i
s
ed
Oct
3
,
2
0
1
7
A
cc
ep
ted
Oct
1
7
,
2
0
1
7
A
s
m
a
jo
rit
y
o
f
th
e
c
o
m
p
re
ss
io
n
a
lg
o
rit
h
m
s
a
re
i
m
p
le
m
e
n
tatio
n
s
f
o
r
CP
U
a
rc
h
it
e
c
tu
re
,
th
e
p
rim
a
r
y
f
o
c
u
s
o
f
o
u
r
w
o
rk
w
a
s
to
e
x
p
lo
it
th
e
o
p
p
o
rt
u
n
it
ies
o
f
G
P
U
p
a
ra
ll
e
li
sm
in
a
u
d
io
c
o
m
p
re
ss
io
n
.
T
h
is
p
a
p
e
r
p
r
e
se
n
ts
a
n
im
p
le
m
e
n
tatio
n
o
f
A
p
p
le
L
o
ss
les
s
A
u
d
io
Co
d
e
c
(A
LAC)
a
lg
o
rit
h
m
b
y
u
sin
g
NV
IDIA
G
P
Us
Co
m
p
u
te
Un
if
ied
De
v
ice
A
r
c
h
it
e
c
tu
re
(CUD
A
)
F
ra
m
e
w
o
rk
.
T
h
e
c
o
re
id
e
a
w
a
s
to
id
e
n
ti
f
y
t
h
e
a
re
a
s
w
h
e
re
d
a
ta
p
a
ra
ll
e
li
s
m
c
o
u
l
d
b
e
a
p
p
li
e
d
a
n
d
p
a
ra
ll
e
l
p
ro
g
ra
m
m
in
g
m
o
d
e
l
CUD
A
c
o
u
ld
b
e
u
se
d
to
e
x
e
c
u
te
th
e
id
e
n
ti
f
ied
p
a
ra
ll
e
l
c
o
m
p
o
n
e
n
ts
o
n
S
in
g
le
In
str
u
c
ti
o
n
M
u
lt
ip
le
T
h
re
a
d
(S
IM
T
)
m
o
d
e
l
o
f
CUD
A
.
T
h
e
d
a
tas
e
t
w
a
s
re
tri
e
v
e
d
f
ro
m
Eu
ro
p
e
a
n
B
ro
a
d
c
a
stin
g
Un
io
n
,
S
o
u
n
d
Qu
a
li
ty
A
s
se
ss
m
e
n
t
M
a
teria
l
(S
QA
M
).
F
a
ste
r
e
x
e
c
u
ti
o
n
o
f
th
e
a
lg
o
rit
h
m
led
to
e
x
e
c
u
ti
o
n
ti
m
e
re
d
u
c
ti
o
n
w
h
e
n
a
p
p
li
e
d
to
a
u
d
i
o
c
o
d
in
g
f
o
r
larg
e
a
u
d
io
s.
T
h
is
p
a
p
e
r
a
lso
p
re
se
n
ts
th
e
re
d
u
c
ti
o
n
o
f
p
o
w
e
r
u
sa
g
e
d
u
e
to
ru
n
n
in
g
th
e
p
a
ra
ll
e
l
c
o
m
p
o
n
e
n
ts
o
n
G
P
U.
E
x
p
e
rim
e
n
tal
re
su
lt
s re
v
e
a
l
th
a
t
w
e
a
c
h
iev
e
a
b
o
u
t
8
0
-
9
0
%
sp
e
e
d
u
p
th
ro
u
g
h
CUD
A
o
n
th
e
id
e
n
ti
f
ied
c
o
m
p
o
n
e
n
ts
o
v
e
r
it
s C
P
U i
m
p
lem
e
n
tatio
n
w
h
i
le sa
v
in
g
CP
U p
o
w
e
r
c
o
n
su
m
p
ti
o
n
.
K
ey
w
o
r
d
:
AL
AC
Au
d
io
d
ec
o
d
in
g
Au
d
io
e
n
co
d
in
g
C
UD
A
P
C
M
SIM
T
Co
p
y
rig
h
t
©
2
0
1
8
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts
re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
J
ia
Ud
d
in
,
Dep
ar
t
m
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
&
E
n
g
in
ee
r
i
n
g
,
B
R
AC
U
n
iv
er
s
it
y
.
6
6
Mo
h
ak
h
ali,
D
h
ak
a
1
2
1
2
,
B
an
g
lad
e
s
h
.
E
m
ail: j
ia.
u
d
d
in
@
b
r
ac
u
.
ac
.
b
d
1.
I
NT
RO
D
UCT
I
O
N
B
e
s
t
u
s
e
o
f
e
xp
e
ns
i
ve
c
o
m
p
u
ti
ng
r
e
s
o
ur
c
e
s
s
uc
h
a
s
m
e
m
or
y,
ne
t
w
or
k
ba
n
d
w
i
dt
h
or
pr
oc
e
s
s
i
ng
un
i
t
s
i
s
gr
o
w
i
n
g
da
y
b
y
da
y.
A
s
a
r
e
s
ul
t
,
c
o
ns
u
m
pt
i
on
o
f
t
ho
s
e
r
e
s
o
ur
c
e
s
ne
e
ds
t
o
be
c
a
r
e
f
ul
l
y
pl
a
nn
e
d
i
n
or
de
r
t
o
a
c
hi
e
ve
m
a
xi
m
u
m
pe
r
f
or
m
a
nc
e
.
D
a
t
a
c
o
m
pr
e
s
s
i
o
n
he
l
ps
t
o
ut
il
iz
e
s
pa
c
e
li
m
it
e
d
r
e
s
o
ur
c
e
s
m
or
e
e
f
f
e
ct
i
ve
l
y.
T
he
r
e
a
r
e
s
e
ve
r
a
l
a
l
go
r
i
t
h
m
s
o
n
l
os
s
l
e
s
s
a
u
di
o
c
o
de
c
s
;
F
r
e
e
L
os
s
le
s
s
A
ud
i
o
C
o
de
c
(
F
L
A
C
)
[
1]
,
S
h
or
t
e
n
[
2]
,
W
i
nd
o
w
s
Me
di
a
A
u
di
o
(
W
M
A
)
[
3]
,
MP
E
G
-
4
[
4]
,
W
a
v
P
a
c
k
,
Mo
n
ke
y’
s
A
u
di
o
[
5]
a
nd
ot
he
r
s
be
i
n
g
us
e
d
by
pr
o
gr
a
m
s
t
o
a
ll
e
vi
at
e
s
pa
c
e
us
a
ge
.
L
o
s
s
les
s
au
d
io
co
m
p
r
ess
io
n
d
etec
ts
an
d
d
is
ca
r
d
s
s
tatis
tica
l
r
ed
u
n
d
an
c
y
in
a
u
d
io
f
ile
i
n
o
r
d
er
to
g
et
r
ed
u
ce
d
b
it
-
r
ate
an
d
th
is
p
r
o
ce
s
s
is
g
e
n
er
all
y
o
b
tain
ed
b
y
i
m
p
le
m
en
t
in
g
li
n
ea
r
p
r
ed
ictio
n
f
o
r
r
ed
u
n
d
an
c
y
r
em
o
v
al
an
d
t
h
e
en
tr
o
p
y
e
n
c
o
d
er
f
o
r
co
d
in
g
th
e
p
r
ed
icto
r
o
u
tp
u
t
[
6
]
.
T
he
r
e
a
r
e
al
s
o
s
o
m
e
t
r
a
de
of
f
s
o
n
t
he
de
ci
s
i
on
of
us
i
n
g
c
o
m
pr
e
s
s
i
o
n.
O
ne
of
t
he
m
a
i
n
i
s
s
ue
s
i
s
i
nc
r
e
a
s
e
i
n
e
n
c
od
i
n
g/
de
c
od
i
n
g
t
i
m
e
a
s
w
e
l
l
a
s
gr
o
w
t
h
o
f
p
o
w
e
r
c
on
s
u
m
pt
i
o
n.
A
p
p
le
L
o
s
s
less
Au
d
io
C
o
d
ec
(
AL
AC
)
i
s
an
au
d
io
co
d
in
g
f
o
r
m
at
d
e
v
elo
p
ed
b
y
A
p
p
le
I
n
c.
f
o
r
lo
s
s
les
s
d
ata
co
m
p
r
ess
io
n
o
f
d
ig
ital
m
u
s
ic.
Fil
en
a
m
e
ex
ten
s
io
n
.
m
4
a
is
u
s
ed
f
o
r
A
p
p
le
L
o
s
s
les
s
d
ata
to
s
to
r
e
w
it
h
i
n
a
n
MP
4
co
n
tain
er
.
T
h
is
ex
te
n
s
io
n
i
s
als
o
u
s
ed
b
y
A
p
p
le
f
o
r
lo
s
s
y
AAC
a
u
d
io
d
ata
w
h
ic
h
is
n
o
t
a
v
ar
ian
t
o
f
AL
A
C
[
7
]
.
An
iT
u
n
es
.
m
4
p
f
ile
i
s
a
DR
M
-
e
n
cr
y
p
ted
M4
A
w
h
ich
i
s
n
’
t
i
n
ter
o
p
er
ab
le
o
u
ts
id
e
o
f
t
h
e
A
p
p
le
ec
o
s
y
s
te
m
.
A
s
a
r
es
u
lt,
o
th
er
lo
s
s
les
s
co
d
ec
s
,
s
u
ch
a
s
FLAC a
n
d
Sh
o
r
te
n
,
ar
e
n
o
t n
a
tiv
el
y
s
u
p
p
o
r
ted
b
y
A
p
p
le
'
s
iT
u
n
es
s
o
f
t
w
ar
e.
So
,
u
s
er
s
o
f
iT
u
n
es
s
o
f
t
w
ar
e
w
h
o
w
an
t
to
u
s
e
a
lo
s
s
les
s
f
o
r
m
at
h
a
v
e
to
u
s
e
AL
AC
[
8
]
.
A
p
p
le
L
o
s
s
le
s
s
s
u
p
p
o
r
ts
b
it
d
ep
th
s
o
f
1
6
,
2
0
,
2
4
an
d
3
2
b
its
a
n
d
an
y
ar
b
itra
r
y
i
n
te
g
er
s
a
m
p
le
r
ate
f
r
o
m
1
to
3
8
4
,
0
0
0
Hz.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
Op
timiz
in
g
A
p
p
le
Lo
s
s
less
A
u
d
io
C
o
d
ec
a
lg
o
r
ith
m
u
s
in
g
N
V
I
DI
A
C
UDA
a
r
ch
itectu
r
e
(
R
a
f
id
A
h
med
)
71
I
n
co
m
p
ar
i
s
o
n
w
it
h
th
e
tr
ad
it
io
n
al
GP
GP
U
tech
n
iq
u
e
s
,
C
UD
A
o
f
f
er
s
s
ev
er
al
ad
v
a
n
ta
g
es,
s
u
c
h
as
s
h
ar
ed
m
e
m
o
r
y
,
f
a
s
ter
d
o
w
n
l
o
ad
s
,
f
u
ll
y
s
u
p
p
o
r
t
f
o
r
in
teg
e
r
an
d
b
it
w
is
e
o
p
er
atio
n
s
a
n
d
also
C
UD
A
co
d
e
is
ea
s
il
y
u
n
d
er
s
tan
d
ab
le
th
a
n
Op
en
C
L
[
9
]
.
T
h
e
p
r
o
ce
s
s
f
lo
w
o
f
C
UD
A
p
r
o
g
r
a
m
m
i
n
g
i
s
d
escr
ib
ed
b
elo
w
[
1
0
]
:
a.
Data
is
co
p
ied
f
r
o
m
th
e
m
ai
n
m
e
m
o
r
y
to
th
e
GP
U
g
lo
b
al
m
e
m
o
r
y
b.
C
P
U
s
en
d
s
t
h
e
p
r
o
ce
s
s
in
g
i
n
s
t
r
u
ctio
n
s
to
t
h
e
GP
U
c.
E
ac
h
co
r
e
at
th
e
GP
U
m
e
m
o
r
y
p
ar
allell
y
ex
ec
u
tes t
h
e
d
ata
d.
R
es
u
lt
f
r
o
m
GP
U
g
lo
b
al
m
e
m
o
r
y
i
s
co
p
ied
b
ac
k
to
m
ai
n
m
e
m
o
ry
I
n
th
is
p
ap
er
,
w
e
p
r
o
p
o
s
e
a
n
i
m
p
le
m
en
tatio
n
o
f
AL
AC
au
d
io
co
m
p
r
ess
io
n
o
n
NVI
D
I
A
GP
Us.
AL
AC
alg
o
r
it
h
m
f
o
llo
w
s
t
h
e
s
a
m
e
b
asic
p
r
in
cip
le
o
f
o
th
er
lo
s
s
les
s
au
d
io
co
m
p
r
ess
io
n
il
lu
s
tr
ated
in
Fig
u
r
e
1
[
1
1
]
.
I
t
is
a
h
i
g
h
l
y
s
er
ialized
alg
o
r
it
h
m
w
h
ic
h
i
s
n
o
t
e
f
f
icie
n
t
e
n
o
u
g
h
to
b
e
u
s
ed
o
n
GP
U.
Fo
r
ex
a
m
p
le,
a
s
t
h
e
i
n
p
u
t
au
d
io
f
i
le
is
s
ep
ar
ated
in
to
p
ac
k
et
s
,
t
h
e
en
co
d
in
g
o
f
n
e
x
t
p
ac
k
et
d
ep
en
d
s
o
n
th
e
r
esu
lts
o
f
p
r
ev
io
u
s
e
n
co
d
ed
p
ac
k
et.
Ou
r
r
ed
esig
n
ed
i
m
p
le
m
en
ta
ti
o
n
f
o
r
th
e
C
UD
A
f
r
a
m
e
w
o
r
k
ai
m
s
to
p
ar
allell
y
i
m
p
le
m
en
t
t
h
e
p
ar
ts
o
f
th
e
alg
o
r
ith
m
w
h
er
e
t
h
e
co
m
p
u
tat
io
n
o
f
a
m
et
h
o
d
is
n
o
t
s
er
ialized
b
y
p
r
ev
io
u
s
co
m
p
u
tatio
n
s
.
T
h
e
p
ap
er
is
p
r
esen
ted
as
f
o
llo
w
s
.
Sectio
n
2
d
escr
ib
es
th
e
i
m
p
le
m
e
n
tatio
n
o
f
C
UD
A
m
o
d
el
o
n
AL
AC
E
n
co
d
in
g
an
d
Dec
o
d
in
g
p
r
o
ce
s
s
.
T
h
e
ex
p
er
im
e
n
ta
l
r
esu
l
t
w
it
h
h
ar
d
w
ar
e
an
d
s
o
f
t
w
ar
e
s
etu
p
s
ar
e
d
is
cu
s
s
ed
in
s
ec
tio
n
3
an
d
f
i
n
a
ll
y
co
n
cl
u
s
io
n
s
w
i
ll b
e
m
ad
e
i
n
s
ec
tio
n
4
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
2
.
1
.
SI
M
T
M
o
del
T
h
e
GP
U
co
r
e
co
n
tain
s
a
n
u
m
b
er
o
f
Stre
a
m
i
n
g
M
u
ltip
r
o
ce
s
s
o
r
s
(
SMs).
On
ea
ch
SM,
ex
ec
u
t
io
n
o
f
ea
ch
in
s
tr
u
c
tio
n
f
o
llo
w
s
a
m
o
d
el
lik
e
SIM
D
w
h
ic
h
is
ca
lle
d
SIM
T
r
ef
er
r
ed
b
y
N
v
id
ia.
I
n
SIM
T
,
th
e
s
a
m
e
in
s
tr
u
ctio
n
is
a
s
s
ig
n
ed
to
all
th
e
t
h
r
ea
d
s
i
n
t
h
e
c
h
o
s
e
n
war
p
.
A
ll
t
h
r
ea
d
s
i
n
a
w
ar
p
a
r
e
is
s
u
ed
th
e
s
a
m
e
in
s
tr
u
ctio
n
,
alt
h
o
u
g
h
n
o
t
e
v
e
r
y
t
h
r
ea
d
n
ee
d
s
to
ex
ec
u
te
t
h
at
in
s
tr
u
c
tio
n
.
A
s
a
r
es
u
lt,
th
r
ea
d
s
in
a
w
ar
p
d
iv
er
g
i
n
g
ac
r
o
s
s
d
if
f
er
en
t
p
ath
s
in
a
b
r
an
ch
r
es
u
lts
i
n
a
lo
s
s
o
f
p
ar
allelis
m
[
1
2
]
.
I
n
o
u
r
i
m
p
le
m
e
n
tat
io
n
,
t
h
e
b
r
an
ch
i
n
g
f
ac
to
r
is
h
a
n
d
led
in
C
P
U
b
ef
o
r
e
ca
llin
g
t
h
e
k
er
n
e
l
to
g
ain
to
tal
p
ar
allelis
m
.
T
h
e
k
er
n
el
b
lo
ck
s
ize
m
u
s
t
b
e
c
h
o
s
e
n
les
s
t
h
an
o
r
eq
u
al
to
tile
s
ize
s
u
c
h
t
h
at
o
n
e
o
r
m
o
r
e
ele
m
e
n
ts
o
f
a
t
ile
i
s
lo
ad
ed
in
to
s
h
ar
ed
m
e
m
o
r
y
b
y
ea
c
h
t
h
r
ea
d
in
a
b
lo
ck
.
R
es
u
lta
n
tl
y
,
t
h
e
d
ev
i
ce
p
er
f
o
r
m
s
o
n
e
in
s
tr
u
ct
io
n
f
etch
f
o
r
a
b
lo
ck
o
f
th
r
ea
d
s
w
h
ic
h
is
i
n
SIM
T
m
an
n
er
.
T
h
is
s
h
o
r
ten
s
in
s
tr
u
c
tio
n
f
etc
h
an
d
p
r
o
ce
s
s
in
g
o
v
er
h
ea
d
o
f
lo
ad
in
s
tr
u
ctio
n
[
1
3]
.
I
n
th
e
test
s
,
w
e
d
eter
m
i
n
e
th
at
5
1
2
t
h
r
ea
d
s
p
er
b
lo
ck
c
o
n
f
i
g
u
r
atio
n
i
s
g
i
v
in
g
t
h
e
b
e
s
t
p
er
f
o
r
m
a
n
ce
.
I
n
t
h
e
C
UD
A
i
m
p
le
m
e
n
tatio
n
o
f
AL
AC
al
g
o
r
ith
m
,
w
e
h
a
v
e
d
ec
id
ed
to
ex
p
lo
it
t
h
e
f
r
a
m
i
n
g
a
n
d
m
ix
in
g
p
h
ase
o
f
en
co
d
in
g
.
T
h
e
d
ec
o
d
in
g
o
f
an
AL
A
C
a
u
d
io
to
p
cm
(
p
u
ls
e
c
o
d
e
m
o
d
u
latio
n
)
d
ata
ar
e
d
o
n
e
b
y
f
o
llo
w
i
n
g
th
e
s
tep
s
o
f
Fi
g
u
r
e
1
r
e
v
er
s
el
y
.
So
,
f
o
r
th
e
d
ec
o
d
in
g
s
ec
tio
n
,
w
e
atte
m
p
ted
to
p
ar
allelize
th
e
u
n
-
m
i
x
i
n
g
a
n
d
co
n
ca
ten
ati
n
g
p
h
ase.
Fig
u
r
e
1
.
T
h
e
b
asic o
p
e
r
atio
n
o
f
AL
AC
e
n
co
d
er
2
.
1
.
1
.
SI
M
T
M
o
del in En
co
din
g
I
n
s
er
ial
C
P
U
i
m
p
le
m
en
tatio
n
o
f
AL
AC
,
t
h
e
in
p
u
t
d
ata
i
s
d
i
v
id
ed
to
s
e
v
er
al
f
r
a
m
e
s
b
y
t
h
e
en
co
d
in
g
p
h
ase,
w
h
er
e
a
f
r
a
m
e
is
s
p
lit
in
to
ev
e
n
s
m
aller
p
iece
s
to
ca
r
r
y
o
u
t
m
i
x
i
n
g
o
p
er
atio
n
.
As
s
h
o
w
n
i
n
Fi
g
u
r
e
2
,
th
e
p
o
s
s
ib
le
w
a
y
to
u
tili
ze
th
i
s
s
er
ializatio
n
o
f
f
r
a
m
i
n
g
a
n
d
m
i
x
in
g
o
f
en
co
d
i
n
g
a
s
ter
eo
a
u
d
io
is
to
b
atch
all
th
e
i
n
p
u
t
p
ac
k
et
s
i
n
to
C
UD
A
g
lo
b
al
m
e
m
o
r
y
f
o
r
a
s
i
n
g
le
p
ar
allel
o
p
er
atio
n
o
f
m
i
x
in
g
t
h
e
d
ata
as
s
h
o
w
n
i
n
Fig
u
r
e
3
tak
i
n
g
ad
v
a
n
ta
g
e
o
f
SIM
T
n
atu
r
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
1
,
Feb
r
u
ar
y
201
8
:
70
–
75
72
Fig
u
r
e
2
.
C
P
U
i
m
p
le
m
e
n
tatio
n
o
f
Fra
m
i
n
g
a
n
d
Mix
i
n
g
p
h
ase
Fig
u
r
e
3
.
GP
U
im
p
le
m
e
n
tatio
n
o
f
Fra
m
i
n
g
a
n
d
Mix
i
n
g
p
h
ase
2
.
1
.
2
.
SI
M
T
M
o
del in D
ec
o
din
g
T
h
e
d
ec
o
d
in
g
p
r
o
ce
s
s
is
r
e
v
er
s
e
o
f
t
h
e
e
n
co
d
in
g
p
r
o
ce
s
s
.
F
i
r
s
t
AL
A
C
au
d
io
d
ata
i
s
d
ec
o
m
p
r
e
s
s
ed
.
T
h
en
th
e
p
r
ed
icto
r
s
ar
e
r
u
n
o
v
er
th
e
d
ata
to
co
n
v
er
t
it
to
p
c
m
d
ata.
Fi
n
all
y
,
f
o
r
s
t
er
eo
au
d
io
,
u
n
-
m
i
x
f
u
n
ctio
n
is
ca
r
r
ied
o
u
t
to
co
n
ca
ten
ate
th
e
2
ch
a
n
n
el
s
in
to
a
s
in
g
le
o
u
tp
u
t
b
u
f
f
er
.
A
s
t
h
e
s
a
m
e
i
n
d
ep
en
d
en
t
b
eh
a
v
io
r
ex
is
t
i
n
t
h
e
d
ec
o
d
in
g
p
r
o
ce
s
s
t
o
m
a
k
e
u
s
e
o
f
t
h
e
d
ata
p
ar
alle
lis
m
i
n
C
UD
A
,
w
e
d
is
tr
ib
u
te
t
h
e
w
o
r
k
o
f
th
e
en
d
o
f
th
e
d
ec
o
d
in
g
p
r
o
ce
s
s
ac
r
o
s
s
th
e
GP
U.
2
.
2
.
M
em
o
ry
Co
a
lescin
g
A
cc
o
r
d
in
g
to
J
an
g
B
,
Sch
aa
D,
et
al,
GP
U
m
e
m
o
r
y
s
u
b
s
y
s
te
m
s
ar
e
d
esig
n
ed
to
d
eliv
er
h
i
g
h
b
an
d
w
id
t
h
v
er
s
u
s
lo
w
-
late
n
c
y
ac
ce
s
s
.
T
o
g
ain
h
i
g
h
est
th
r
o
u
g
h
p
u
t,
a
lar
g
e
n
u
m
b
er
o
f
s
m
a
l
l
m
e
m
o
r
y
ac
ce
s
s
e
s
s
h
o
u
ld
b
e
b
u
f
f
er
ed
,
r
eo
r
d
er
ed
,
an
d
co
alesced
in
to
a
s
m
al
l n
u
m
b
e
r
o
f
lar
g
e
r
eq
u
est
s
[
1
4]
.
Fo
r
s
tr
id
ed
g
lo
b
al
m
e
m
o
r
y
ac
ce
s
s
,
th
e
e
f
f
ec
ti
v
e
b
an
d
w
id
t
h
i
s
al
w
a
y
s
p
o
o
r
.
W
h
en
co
n
cu
r
r
e
n
t
t
h
r
ea
d
s
s
i
m
u
lta
n
eo
u
s
l
y
ac
ce
s
s
m
e
m
o
r
y
ad
d
r
ess
e
s
th
a
t
ar
e
lo
ca
ted
f
ar
ap
ar
t
in
p
h
y
s
ical
m
e
m
o
r
y
,
th
er
e
is
n
o
p
o
s
s
ib
ilit
y
f
o
r
co
alescin
g
t
h
e
m
e
m
o
r
y
ac
ce
s
s
[
1
5
]
.
T
h
u
s
,
w
e
r
es
tr
u
ct
u
r
e
th
e
b
u
f
f
er
in
s
u
c
h
a
w
a
y
t
h
at
g
lo
b
al
m
e
m
o
r
y
lo
ad
s
is
s
u
ed
b
y
t
h
r
ea
d
s
o
f
w
ar
p
ar
e
co
alesced
in
en
co
d
in
g
p
r
o
ce
s
s
as s
h
o
w
in
F
ig
u
r
e
4
.
Fig
u
r
e
2
.
Un
co
alesced
an
d
co
alesced
m
e
m
o
r
y
ac
ce
s
s
2
.
3
.
P
inn
ed
M
e
m
o
ry
B
y
d
e
f
a
u
lt,
C
P
U
d
ata
allo
ca
tio
n
s
ar
e
p
a
g
ea
b
le.
A
s
C
UD
A
d
r
i
v
er
’
s
u
s
es
p
ag
e
-
lo
ck
ed
o
r
p
in
n
ed
m
e
m
o
r
y
,
GP
U
ca
n
n
o
t
d
ir
ec
tl
y
ac
ce
s
s
d
ata
p
r
o
m
p
ag
ea
b
le
h
o
s
t
m
e
m
o
r
y
.
C
o
n
s
eq
u
e
n
tl
y
,
t
h
e
GP
U
f
ir
s
t
allo
ca
tes
a
te
m
p
o
r
ar
y
p
i
n
n
ed
m
e
m
o
r
y
,
co
p
ies
h
o
s
t
d
ata
in
to
it
an
d
t
h
en
tr
an
s
f
er
s
d
ata
to
t
h
e
d
ev
ice
m
e
m
o
r
y
.
So
,
it
is
b
etter
to
u
s
e
p
in
n
ed
m
e
m
o
r
y
b
ef
o
r
eh
a
n
d
r
ath
er
t
h
a
n
u
s
i
n
g
p
a
g
ed
m
e
m
o
r
y
i
n
o
r
d
er
to
r
ed
u
ce
co
p
y
i
n
g
co
s
t [
1
6
]
.
A
s
w
e
h
a
v
e
to
tr
a
n
s
f
er
c
h
an
n
el
b
u
f
f
er
s
f
r
o
m
GP
U
to
C
P
U
f
o
r
ea
c
h
i
ter
atio
n
w
h
i
l
e
en
co
d
in
g
f
r
a
m
es,
w
e
k
ep
t th
e
c
h
an
n
el
b
u
f
f
er
o
f
C
P
U
as p
in
n
ed
m
e
m
o
r
y
f
o
r
f
a
s
t tr
an
s
f
er
o
f
d
ata.
3.
RE
SU
L
T
S
A
ND
AN
AL
Y
SI
S
T
o
an
aly
ze
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
C
UD
A
i
m
p
le
m
e
n
tatio
n
,
we
u
s
ed
a
GeFo
r
ce
GT
X
9
6
0
ca
r
d
w
it
h
C
UD
A
v
er
s
io
n
7
.
5
i
n
s
talled
o
n
a
m
ac
h
i
n
e
w
i
th
I
n
tel(
R
)
C
o
r
e(
T
M)
i5
-
4
5
9
0
C
P
U
r
u
n
n
in
g
at
3
.
3
0
GHz
.
T
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
Op
timiz
in
g
A
p
p
le
Lo
s
s
less
A
u
d
io
C
o
d
ec
a
lg
o
r
ith
m
u
s
in
g
N
V
I
DI
A
C
UDA
a
r
ch
itectu
r
e
(
R
a
f
id
A
h
med
)
73
C
P
U
i
m
p
le
m
e
n
tatio
n
o
f
AL
AC
i
s
al
s
o
tes
ted
o
n
th
e
s
a
m
e
t
estb
ed
.
T
o
co
m
p
ar
e
t
h
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
GP
U
i
m
p
le
m
en
ta
tio
n
o
f
AL
A
C
alg
o
r
ith
m
,
w
e
s
elec
ted
r
ep
r
esen
t
ativ
e
o
f
f
i
v
e
a
u
d
io
f
ile
s
f
o
r
test
in
g
.
A
ll
th
e
s
e
f
ile
s
h
av
e
t
h
e
C
D
f
o
r
m
a
t,
i.e
.
4
4
.
1
k
Hz,
1
6
b
its
.
a.
T
r
ac
k
5
o
f
[
1
7
]
(
2
8
s
)
: E
lectr
o
n
ic
Go
n
g
5
k
Hz
(
Mo
n
o
)
b.
T
r
ac
k
2
0
o
f
[
1
7
]
(
3
9
s
)
: Sax
o
p
h
o
n
e
c.
T
r
ac
k
5
0
o
f
[
1
7
]
(
2
2
s
)
: M
ale
s
p
ee
ch
(
E
n
g
lis
h
)
d.
T
r
ac
k
6
8
o
f
[
1
7
]
(
2
m
in
4
4
s
)
: O
r
ch
estra
e.
T
r
ac
k
7
0
o
f
[
1
7
]
(
2
1
s
)
: So
n
g
b
y
E
d
d
ie
R
ab
b
itt
T
h
e
au
d
io
f
iles
ar
e
co
llected
f
r
o
m
So
u
n
d
Q
u
ali
t
y
Ass
e
s
s
m
en
t
Ma
ter
ial
(
S
Q
A
M)
r
ec
o
r
d
in
g
s
f
o
r
s
u
b
j
ec
tiv
e
test
s
9
.
T
h
e
f
lie
s
w
er
e
co
n
v
er
t
ed
to
W
A
V
f
r
o
m
F
L
AC
to
w
o
r
k
w
it
h
p
c
m
d
ata.
T
o
m
ea
s
u
r
e
th
e
r
esu
lts
,
w
e
r
a
n
o
u
r
test
1
0
ti
m
es
o
n
ea
c
h
d
atas
et
f
o
r
ea
ch
r
ea
d
in
g
a
n
d
s
h
o
w
e
d
th
e
av
er
ag
e
r
u
n
n
i
n
g
r
es
u
lt
s
i
n
th
i
s
s
ec
tio
n
.
3
.
1
.
E
nco
din
g
Re
s
ults
3
.
1
.
1.
E
x
ec
utio
n Speed
G
a
in
T
h
e
r
esu
lts
o
n
T
ab
le
1
s
h
o
w
s
th
e
av
er
ag
e
r
u
n
n
in
g
s
p
ee
d
r
esu
lt
s
o
f
b
o
th
m
o
n
o
an
d
s
te
r
eo
au
d
io
(
T
r
ac
k
6
8
)
f
o
r
C
P
U
an
d
GPU
s
ep
ar
atel
y
.
Fo
r
1
6
b
it
au
d
io
w
e
ac
h
ie
v
e
s
p
ee
d
g
ai
n
ab
o
u
t
6
7
%
f
o
r
m
o
n
o
an
d
8
5
%
f
o
r
s
ter
eo
au
d
io
ty
p
e.
Fo
r
2
4
b
it
au
d
io
th
e
s
p
ee
d
u
p
is
a
r
o
u
n
d
9
0
%
f
o
r
b
o
th
au
d
io
t
y
p
es.
Fo
r
3
2
b
it
au
d
io
,
th
e
s
p
ee
d
in
cr
ea
s
e
is
ar
o
u
n
d
8
1
%.
Her
e
w
e
ca
n
in
f
er
th
at
2
4
b
it
au
d
io
co
n
v
er
s
io
n
r
es
u
lt
s
in
f
as
ter
en
co
d
in
g
s
p
ee
d
f
o
r
GP
U
w
h
er
e
f
o
r
1
6
b
it a
u
d
io
it is
s
lo
w
er
th
a
n
t
h
e
o
th
er
s
.
I
n
Fi
g
u
r
e
5
,
w
e
s
ee
f
o
r
f
ir
s
t te
s
t
f
ile
(
T
r
ac
k
5
)
w
e
ac
h
ie
v
e
3
x
s
p
ee
d
u
p
,
f
o
r
s
ec
o
n
d
a
n
d
t
h
ir
d
test
f
ile
(
T
r
ac
k
2
0
&
5
0
)
w
e
ac
h
ie
v
e
7
x
s
p
ee
d
u
p
,
f
o
r
f
o
u
r
t
h
test
f
ile
(
T
r
ac
k
6
8
)
w
e
g
ai
n
6
.
5
x
s
p
ee
d
an
d
last
l
y
f
o
r
f
i
f
t
h
f
il
e
(
T
r
ac
k
7
0
)
w
e
g
et
5
x
s
p
ee
d
u
p
.
T
ab
le
1
.
Mix
in
g
/
C
o
n
v
er
ti
n
g
p
h
ase
e
x
ec
u
t
io
n
ti
m
e
co
m
p
ar
is
o
n
f
o
r
en
co
d
in
g
p
r
o
ce
s
s
B
i
t
D
e
p
t
h
C
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
mo
n
o
(
ms)
G
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
mo
n
o
(
ms)
C
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
st
e
r
e
o
(
ms)
G
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
st
e
r
e
o
(
ms)
16
2
.
8
5
0
.
9
4
2
0
.
3
3
.
0
2
5
9
24
2
5
.
6
8
1
.
6
2
6
3
.
5
5
.
2
4
6
32
9
.
3
2
1
.
7
1
38
4
.
8
2
6
6
Fig
u
r
e
3
.
Mix
i
n
g
/
C
o
n
v
er
ti
n
g
p
h
ase
s
p
ee
d
u
p
f
o
r
en
co
d
in
g
a
n
d
d
ec
o
d
in
g
p
r
o
ce
s
s
ag
ai
n
s
t
C
P
U
u
s
in
g
d
ataset
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
1
,
Feb
r
u
ar
y
201
8
:
70
–
75
74
3
.
1
.
2
.
P
o
w
er
C
o
ns
u
m
ptio
n Sa
v
ing
L
o
o
k
i
n
g
at
T
ab
le
2
,
th
e
p
o
w
e
r
co
n
s
u
m
p
tio
n
s
a
v
i
n
g
is
lea
s
t
f
o
r
1
6
b
it
au
d
io
f
o
r
b
o
th
m
o
n
o
an
d
s
ter
eo
au
d
io
f
iles
w
h
er
e
f
o
r
2
4
b
it a
u
d
io
p
o
w
er
s
a
v
in
g
is
t
h
e
h
ig
h
e
s
t.
T
ab
le
2
.
C
P
U
p
o
w
er
s
a
v
i
n
g
in
en
co
d
in
g
p
r
o
ce
s
s
B
i
t
D
e
p
t
h
P
o
w
e
r
c
o
n
su
mp
t
i
o
n
sav
i
n
g
f
o
r
mo
n
o
(
W
a
t
t
)
P
o
w
e
r
c
o
n
su
mp
t
i
o
n
sav
i
n
g
f
o
r
st
e
r
e
o
(
W
a
t
t
)
16
0
.
6
6
7
9
2
3
.
6
3
6
9
2
7
24
5
.
4
9
4
4
4
9
.
2
0
9
0
8
2
32
2
.
0
5
4
4
7
.
2
5
7
1
6
3
3.
2
.
Dec
o
din
g
Resul
t
s
3
.
2
.
1.
E
x
ec
utio
n Speed
G
a
in
Dec
o
d
in
g
r
e
s
u
l
ts
f
o
r
b
o
th
m
o
n
o
an
d
s
ter
eo
f
ile
(
T
r
ac
k
6
8
)
ar
e
s
h
o
w
n
i
n
T
ab
le
3
.
A
cc
o
r
d
in
g
to
t
h
e
r
esu
lt
s
,
w
e
g
et
ar
o
u
n
d
8
5
%
s
p
ee
d
in
cr
ea
s
e
f
o
r
1
6
b
it
m
o
n
o
au
d
io
an
d
9
0
%
f
o
r
1
6
b
it
s
ter
e
o
.
Fo
r
2
4
b
it
au
d
io
,
w
e
ac
h
ie
v
e
m
o
r
e
t
h
a
n
9
0
%
s
p
ee
d
u
p
.
L
astl
y
f
o
r
3
2
b
it
a
u
d
io
,
ar
o
u
n
d
8
9
%
s
p
ee
d
u
p
is
g
ai
n
ed
.
Fro
m
t
h
is
tab
le,
w
e
ca
n
also
s
tate
t
h
at
2
4
b
it
au
d
io
d
ec
o
d
in
g
is
f
aster
i
n
GP
U
th
a
n
th
e
o
t
h
er
s
w
h
er
e
1
6
b
it
au
d
io
co
n
v
er
s
io
n
i
s
s
lo
w
e
s
t.
T
ab
le
3
.
Un
-
m
i
x
i
n
g
/C
o
n
v
er
ti
n
g
p
h
ase
e
x
ec
u
tio
n
ti
m
e
co
m
p
a
r
is
o
n
f
o
r
d
ec
o
d
in
g
p
r
o
ce
s
s
B
i
t
D
e
p
t
h
C
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
mo
n
o
(
ms)
G
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
mo
n
o
(
ms)
C
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
st
e
r
e
o
(
ms)
G
P
U
e
x
e
c
u
t
i
o
n
t
i
me
f
o
r
st
e
r
e
o
(
ms)
16
4
.
0
2
0
.
5
8
9
2
1
0
.
2
5
1
.
0
7
6
4
24
1
0
.
8
8
4
0
.
8
3
5
2
2
8
.
0
3
2
1
.
8
1
9
2
32
7
.
0
9
3
0
.
8
6
4
2
0
.
3
5
4
1
.
7
1
5
3
.
3
.
2
.
P
o
w
er
C
o
ns
u
m
ptio
n Sa
v
ing
Fo
r
d
ec
o
d
in
g
s
tag
e,
w
e
g
et
les
s
p
o
w
er
s
av
in
g
t
h
an
t
h
at
o
f
e
n
co
d
in
g
s
ta
g
e.
Her
e
also
1
6
b
it
au
d
io
h
as
th
e
least p
o
w
er
s
av
in
g
w
h
er
e
2
4
b
it a
u
d
io
h
as
th
e
h
i
g
h
est
s
h
o
w
n
in
T
ab
le
4
.
T
ab
le
4
.
C
P
U
p
o
w
er
s
a
v
i
n
g
in
d
ec
o
d
in
g
p
r
o
ce
s
s
B
i
t
D
e
p
t
h
P
o
w
e
r
c
o
n
su
mp
t
i
o
n
sav
i
n
g
f
o
r
mo
n
o
(
W
a
t
t
)
P
o
w
e
r
c
o
n
su
mp
t
i
o
n
sav
i
n
g
f
o
r
st
e
r
e
o
(
W
a
t
t
)
16
0
.
6
9
0
4
1
4
1
.
5
7
7
8
1
4
24
1
.
8
5
3
9
6
4.
3
7
9
1
1
8
32
1
.
1
2
4
8
8
6
3
.
0
8
7
1
9
4
4.
CO
NCLU
SI
O
N
I
n
th
i
s
p
ap
er
,
w
e
an
a
l
y
ze
d
th
e
f
ea
s
ib
ilit
y
to
u
s
e
C
UD
A
f
r
a
m
e
w
o
r
k
f
o
r
A
p
p
le
L
o
s
s
les
s
A
u
d
io
C
o
d
ec
co
m
p
r
es
s
io
n
al
g
o
r
ith
m
.
O
u
r
p
r
im
ar
y
f
o
cu
s
w
a
s
o
n
o
u
tp
er
f
o
r
m
i
n
g
t
h
e
m
i
x
i
n
g
/u
n
-
m
i
x
in
g
s
p
ee
d
o
f
th
e
C
P
U
b
ased
AL
AC
i
m
p
le
m
e
n
tatio
n
b
y
u
s
i
n
g
NVI
DI
A
GP
Us
w
it
h
o
u
t
lo
s
i
n
g
a
n
y
co
m
p
r
ess
io
n
r
a
tio
.
W
e
test
ed
o
u
r
i
m
p
le
m
en
ta
tio
n
o
n
s
e
v
er
al
d
atasets
an
d
m
ad
e
co
m
p
ar
is
o
n
.
E
x
p
er
i
m
e
n
tal
r
esu
lts
d
e
m
o
n
s
tr
a
te
th
at
w
e
ac
h
iev
e
av
er
ag
e
o
f
8
0
-
9
5
% sp
ee
d
u
p
f
o
r
m
i
x
i
n
g
/u
n
-
m
ix
i
n
g
au
d
io
d
ata.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8708
Op
timiz
in
g
A
p
p
le
Lo
s
s
less
A
u
d
io
C
o
d
ec
a
lg
o
r
ith
m
u
s
in
g
N
V
I
DI
A
C
UDA
a
r
ch
itectu
r
e
(
R
a
f
id
A
h
med
)
75
RE
F
E
R
E
NC
E
S
[1
]
F
irm
a
n
sa
h
L
,
S
e
ti
a
w
a
n
EB,
“
Da
ta
a
u
d
io
c
o
mp
re
ss
io
n
lo
ss
les
s
F
L
AC
fo
rm
a
t
t
o
lo
ss
y
a
u
d
io
M
P
3
fo
rm
a
t
w
it
h
Hu
ff
ma
n
s
h
if
t
c
o
d
in
g
a
lg
o
rith
m
,
”
2
0
1
6
4
t
h
In
tern
a
ti
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
In
f
o
rm
a
ti
o
n
a
n
d
Co
m
m
u
n
ica
ti
o
n
T
e
c
h
n
o
lo
g
y
(ICo
ICT
).
2
0
1
6
M
a
y
.
[2
]
S
a
lo
m
o
n
D.
A
c
o
n
c
ise
in
tro
d
u
c
t
io
n
t
o
d
a
ta
c
o
m
p
re
ss
io
n
.
L
o
n
d
o
n
:
S
p
ri
n
g
e
r
L
o
n
d
o
n
;
2
0
0
8
Ja
n
1
4
.
2
2
7
–
4
4
p
.
IS
BN:
9
7
8
1
8
4
8
0
0
0
7
1
1
.
[3
]
W
a
g
g
o
n
e
r
B,
“
Co
m
p
re
ss
io
n
f
o
r
g
re
a
t
v
id
e
o
a
n
d
a
u
d
i
o
:
M
a
ste
r
ti
p
s
a
n
d
c
o
m
m
o
n
se
n
se
,
”
2
n
d
e
d
.
Ox
f
o
rd
,
Un
it
e
d
Kin
g
d
o
m
:
El
se
v
ier S
c
ien
c
e
;
2
0
0
9
No
v
2
3
.
3
4
1
–
7
p
.
IS
BN:
9
7
8
0
2
4
0
8
1
2
1
3
7
.
[4
]
Yu
R,
Ra
h
a
rd
ja
S
,
X
iao
L
,
Ko
CC,
“
A
f
in
e
g
ra
n
u
lar
sc
a
lab
le
to
lo
ss
les
s
a
u
d
io
c
o
d
e
r,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
A
u
d
io
,
S
p
e
e
c
h
a
n
d
L
a
n
g
u
a
g
e
P
ro
c
e
ss
in
g
.
2
0
0
6
Ju
l
;
1
4
(4
)
:1
3
5
2
–
6
3
.
[5
]
S
a
lo
m
o
n
D,
M
o
tt
a
G
,
Bry
a
n
t
D,
“
Da
ta
c
o
m
p
re
ss
io
n
:
T
h
e
c
o
m
p
lete
re
f
e
r
e
n
c
e
,
”
4t
h
e
d
.
L
o
n
d
o
n
:
S
p
r
in
g
e
r
L
o
n
d
o
n
;
2
0
0
6
De
c
1
9
.
7
7
3
–
8
3
p
.
I
S
BN:
9
7
8
1
8
4
6
2
8
6
0
2
5
.
[6
]
G
u
n
a
w
a
n
T
,
Zain
M
,
M
u
in
F
,
Ka
rti
w
i
M
,
“
In
v
e
stig
a
ti
o
n
o
f
L
o
ss
les
s
A
u
d
io
Co
m
p
re
ss
io
n
u
sin
g
IE
EE
1
8
5
7
.
2
A
d
v
a
n
c
e
d
A
u
d
io
C
o
d
i
n
g
,
”
In
d
o
n
e
sia
n
J
o
u
rn
a
l
o
f
El
e
c
trica
l
En
g
i
n
e
e
rin
g
a
n
d
C
o
mp
u
ter
S
c
i
e
n
c
e
,
2
0
1
7
;
6
:
4
2
2
–
4
3
0
[7
]
A
p
p
le
L
o
ss
les
s
[
In
tern
e
t
]
.
A
p
p
lel
o
ss
les
s.co
m
.
2
0
1
7
[
c
it
e
d
2
3
S
e
p
tem
b
e
r
2
0
1
7
]
.
A
v
a
il
a
b
le
f
ro
m
:
h
tt
p
:
//
ww
w
.
a
p
p
lelo
ss
les
s.co
m
/
[8
]
W
a
g
g
o
n
e
r
B,
“
Co
m
p
re
ss
io
n
f
o
r
g
re
a
t
v
id
e
o
a
n
d
a
u
d
i
o
:
M
a
ste
r
ti
p
s
a
n
d
c
o
m
m
o
n
se
n
se
,
”
2
n
d
e
d
.
Ox
f
o
rd
,
Un
it
e
d
Kin
g
d
o
m
:
El
se
v
ier S
c
ien
c
e
;
2
0
0
9
No
v
2
3
.
2
1
5
–
2
1
p
.
IS
BN:
9
7
8
0
2
4
0
8
1
2
1
3
7
.
[9
]
G
u
o
S
,
Ch
e
n
S
,
L
ian
g
Y,
“
A
No
v
e
l
A
rc
h
it
e
c
t
u
re
o
f
M
u
lt
i
-
G
P
U
Co
m
p
u
ti
n
g
Ca
rd
,
”
T
E
L
KOM
NIKA
(
T
e
lec
o
mm
u
n
ica
ti
o
n
Co
mp
u
ti
n
g
El
e
c
tro
n
ics
a
n
d
C
o
n
tro
l
)
,
2
0
1
3
;1
1
(8
)
[1
0
]
Na
sre
e
n
A
,
G
S
,
“
P
a
ra
ll
e
li
z
in
g
M
u
lt
i
-
f
e
a
tu
re
d
Co
n
ten
t
Ba
se
d
S
e
a
rc
h
a
n
d
Re
tri
e
v
a
l
o
f
V
id
e
o
s
t
h
r
o
u
g
h
Hig
h
P
e
rf
o
rm
a
n
c
e
Co
m
p
u
ti
n
g
,
”
In
d
o
n
e
sia
n
J
o
u
rn
a
l
o
f
El
e
c
trica
l
En
g
in
e
e
rin
g
a
n
d
Co
mp
u
ter
S
c
ien
c
e
,
2
0
1
7
;5
(
1
):2
1
4
[1
1
]
Ha
n
s
M
,
S
c
h
a
f
e
r
RW
.
L
o
ss
les
s
c
o
m
p
re
ss
io
n
o
f
d
ig
it
a
l
a
u
d
i
o
.
IEE
E
S
i
g
n
a
l
P
r
o
c
e
ss
in
g
M
a
g
a
z
in
e
.
2
0
0
1
,
Ju
l:
1
8
(4
)
:5
–
1
5
.
[1
2
]
S
u
tsu
i
S
,
Co
ll
e
t
P
,
e
d
it
o
rs,
“
M
a
ss
iv
e
l
y
p
a
ra
ll
e
l
e
v
o
lu
ti
o
n
a
ry
c
o
m
p
u
tatio
n
o
n
G
P
G
P
Us
,
”
Be
rli
n
:
S
p
r
in
g
e
r
-
V
e
rlag
Be
rli
n
a
n
d
He
id
e
lb
e
rg
Gm
b
H &
Co
.
K;
2
0
1
3
De
c
6
.
1
4
9
–
5
6
p
.
IS
BN:
9
7
8
3
6
4
2
3
7
9
5
8
1
.
[1
3
]
Al
-
M
o
u
h
a
m
e
d
M
,
Kh
a
n
A
u
l
H,
“
Ex
p
lo
ra
ti
o
n
o
f
a
u
to
m
a
ti
c
o
p
ti
m
iza
ti
o
n
f
o
r
C
UD
A
p
ro
g
ra
m
m
in
g
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
P
a
ra
ll
e
l,
Eme
rg
e
n
t
a
n
d
Distrib
u
ted
S
y
ste
ms
,
2
0
1
4
S
e
p
1
6
;
3
0
(
4
):3
0
9
–
2
4
.
[1
4
]
Ja
n
g
B,
S
c
h
a
a
D,
M
istry
P
,
Ka
e
li
D,
“
Ex
p
lo
it
in
g
m
e
m
o
r
y
a
c
c
e
ss
p
a
tt
e
rn
s
to
im
p
ro
v
e
m
e
m
o
r
y
p
e
rf
o
r
m
a
n
c
e
in
d
a
ta
-
p
a
ra
ll
e
l
a
rc
h
it
e
c
tu
re
s,”
IEE
E
T
ra
n
s
a
c
ti
o
n
s o
n
P
a
ra
ll
e
l
a
n
d
Dis
trib
u
ted
S
y
ste
ms
,
2
0
1
1
Ja
n
;
2
2
(
1
):
1
0
5
–
1
8
.
[1
5
]
Ha
rris
M
.
[
I
n
tern
e
t
]
:
P
a
ra
ll
e
l
F
o
r
a
ll
.
Ho
w
to
a
c
c
e
s
s
g
lo
b
a
l
m
e
m
o
ry
e
ff
icie
n
tl
y
in
CUD
A
C/C++
k
e
rn
e
ls;
2
0
1
3
Ja
n
7
[
c
it
e
d
2
0
1
6
No
v
1
4
]
.
Av
a
il
a
b
le
f
ro
m
:
h
tt
p
s://
d
e
v
b
lo
g
s.n
v
id
ia.co
m
/p
a
ra
ll
e
lf
o
ra
ll
/h
o
w
-
a
c
c
e
ss
-
g
lo
b
a
l
-
m
e
m
o
r
y
-
e
ff
icie
n
tl
y
-
c
u
d
a
-
c
-
k
e
rn
e
ls/
[1
6
]
Ha
rris
M
.
[
In
ter
n
e
t]:
P
a
ra
ll
e
l
F
o
r
a
ll
.
Ho
w
to
o
p
ti
m
ize
d
a
ta t
ra
n
sf
e
r
s in
CUD
A
C/C++
;
2
0
1
2
De
c
4
[
c
it
e
d
2
0
1
6
No
v
1
4
]
.
A
v
a
il
a
b
le
f
ro
m
:
h
tt
p
s:
//
d
e
v
b
lo
g
s.n
v
id
ia.co
m
/p
a
ra
ll
e
lf
o
ra
ll
/h
o
w
-
o
p
ti
m
ize
-
d
a
ta
-
tran
sf
e
rs
-
c
u
d
a
-
cc/
[1
7
]
S
o
u
n
d
q
u
a
li
ty
a
ss
e
ss
m
e
n
t
m
a
ter
ial:
re
c
o
rd
in
g
s
f
o
r
su
b
jec
ti
v
e
tes
ts;
u
se
r
’s
h
a
n
d
b
o
o
k
f
o
r
th
e
EBU
-
S
QA
M
c
o
m
p
a
c
t
d
isc
.
Ge
n
e
v
a
:
T
e
c
h
n
ica
l
Ce
n
tre;
2
0
0
8
.
[
c
it
e
d
2
0
1
6
No
v
1
4
]
.
A
v
a
i
lab
le
f
ro
m
:
h
tt
p
s:/
/t
e
c
h
.
e
b
u
.
c
h
/
d
o
c
s/tec
h
/t
e
c
h
3
2
5
3
.
p
d
f
.
Evaluation Warning : The document was created with Spire.PDF for Python.