scieee Science in your language
[en] (orig)

Comparison of matroid intersection algorithms for large circuit analysis

Abstract

This paper presents two approaches to symbolic analysis of large analog integrated circuits via simplification during the generation of the symbolic expressions. Both techniques are examined from the point of view of matroid theory. Finally, a new approach which combines the positive features of both approaches is introduced.

Read accessible full text

Comparison of matroid intersection algorithms for large circuit analysis

Author: Galán, Mariano; Fernández Fernández, Francisco Vidal; Rodríguez Vázquez, Ángel Benito
Publisher: Institute of Electrical and Electronics Engineers
Year: 1997
DOI: 10.1109/ISCAS.1997.621491
Source: https://idus.us.es/bitstreams/f17ee266-cc22-421f-85cb-3ccd419297b1/download
1997
IEEE
In ema ional Symposium
on
Ci cui s and Sys ems,
June
9-12,1997,
Hong
Kong
Compa ison
o
Ma oid In e sec ion Algo i hms o L,a g : Ci cui Analysis*
Ma iano Galin, F ancisco
V.
Fe nindez a id Angel Rod iguez-Vizquez
Dep . o Analog and Mixed-Signal In eg a ed Ci cui Design,
IMSE,-CNM
Edi . CICA, A da. Reina Me cedes s/n, 13-41012 Se illa,
SPAlN
Tel.: +34
5
4239923, FAX: +34
5
4231832, E-mail: paco @cnm
us.es
Abs ac
This pape p esen s wo app oaches
o
symbolic
analysis o la ge analog in eg a ed ci cui s ia simpli ica-
ion du ing he gene a ion o he symbolic exp essions.
Bo h echniques a e examined om he poin
o
iew o
ma oid heo y. Finally, a new app oach which combines
he posi i e ea u es o bo h app oaches is in oduced.
1.
In oduc ion
Symbolic analyze s a e CAD ools which calcu-
la e ne wo k unc ions o analog ci cui s, wi h he com-
plex equency
s
and he ci cui pa ame e s kep as
symbols. Those ne wo k unc ions a e ypically gi en as
a
cancella ion- ee sum o p oduc s:
2
N
2
M
(1)
,(x)
+. ,(.,
+. ,(.,
+...+s
gg(x)
+.gl
(x)
+s
g*(.)
+
...
+s
g,(x)
H(s,x)
=
T
in
which
x
={xl,
x2,
. .
.,
xe}
is he ec o o symbolic
pa ame e s, and and
gi
a e sums o p oduc s. Many
po en ial applica ions ha e been epo ed o symbolic
analyse s, mos o which can be ound
in
[I].
One
o
he majo d awbacks ha has p e en ed
symbolic analyze s om being widely accep ed by he
analog design communi y has been ha he size
o
he ci -
cui s ha hey can analyze is s ill much smalle han o
nume ical simula o s. This is due o he exponen ial
g ow h o he symbolic o mula complexi y wi h he ci -
cui size. This no only limi s he maximum analysable
ci cui size bu also makes mo e di icul o mula in e -
p e a ion and i s use in design au oma ion applica ions,
whe e exp essions should be he smalles possible o
inc eased e alua ion e iciency.
Expe ience wi h he applica ion o symbolic ana-
lyze s
o
eal analog in eg a ed ci cui s shows ha , usu-
ally, a e y small pa o he symbolic exp ession con ains
mos ele an in o ma ion.
So
he app oxima ion o sym-
bolic exp essions, de ined as he educ ion o o mula
complexi y while main aining he accu acy as high as
possible, has become
a
mus . The gene al opic o exp es-
sion simpli ica ion will be desc ibed in Sec ion 2.
*
This
wo k
has
been pe o med
in
he
amewo k
o
he
AMADEUS
P ojec
o
he ESPRIT
IV
P og am
o
he
CEC.
0-7803-3583-X/97
$10.00
01997
IEEE
1784
]Howe e , o mula app oxima ion a e hle com-
ple e exac exp ession has been gene a ed
only
eases
in e p e a ion and manipula ion bu ci cui size limi a-
ions emain.
A
majo b eak h ough has been p ioduced
wi h lhe in oduc ion
o
simpli ica ion du ing gene a ion
algo i hms [2],[3],[41. Sec iion
3
will explain hsow he
app oxima ion du ing genwa ion p oblem can be unde -
s ood in e ms o ma oids,
a
gene al ma hema ical he-
o y ha gua an ees ha i I he ma oid p oblem has a
solu ion, he pa icula p oblem a hand can
also
be
sol ed, and a leas as e icien ly as he ma oid p oblem.
2.
Symbolic exp ession app oxima ion
Con en ionally, epo ed symbolic analyze s ha e
inco po a ed he app oxima ion ea u e by i s ciilcula -
ing he exac ne wo k unc ion and hen simpli ying i as
a pos p ocessing s ep
[5],[6],[7].
Fo his eason
'we
will
call i
Simpli ica ion
A e Gene a ion
(SAG). I
is
based
on he elimina ion
o
he leas signi ican e ms o subex-
p essions while some e o c i e ion is sa is ied. Rela i e
signi icance is e alua ed based on nume ical es ima es o
he symbolic ci cui pa ame e s. Assume
T
h,
(x:l
=
h,,
(x)
+
h,,
(x)
+
.
.
.
+
h,,
(x)
=
c
h,,
(x)
(2)
/=
I
ep exn s ei he k(x) o
gk(x)
in
(1).
The
P
leas signi i-
can e ms in (2) a e elimina ed, one by one and begin-
ning wi h he smalles , while he sum o he eli nina ed
e ms keeps below
some
gi len h eshold,
P
T
(3)
whe e
xo
ep esen s he design poin o he ci cui pa am-
e e s and
ek
is a disc imina ion h eshold.
The p incipal d awback o SAG echniques is ha
simpli ica ion is pe o med al e he exac ne wo k unc-
ion has been gene a ed.
So
he exac exp ession, which
usually is no used any mo e a e he app oxiima ion,
limi s he analysable ci cui size o a ound
10
an,sis o s.
In o de o ex end
he
capabili ies o con en ional
symbolic analyze s, he new concep o
Simpl@ca ion
Du ing Gene a ion
(SDG) has ecen ly been p oposed.
SDG is based on he ollowing idea: e ms a e gene a ed
in s ic ly dec easing o de o magni ude, s a ing wi h
he la ges , o each powe o
s,
un il hey ep esen a
gi en ac ion o he o al magni ude o he coe icien .
SDG has wo main ad an ages. Fi s , he analysis
is as e , as no ime is was ed gene a ing e ms ha would
be neglec ed la e on. Second, i s smalle memo y needs
make possible ha much la ge ci cui s can be analyzed.
This is clea ly illus a ed in Fig. l(b), which shows he
compu a ion ime o app oxima ed exp essions wi h
E+
=
0.25
e sus he numbe o ci cui nodes
in
Fig. l(a).
I
can be seen ha he new SDG app oach is
o de s o magni ude as e han he con en ional one
o
ge he same app oxima ed exp ession. Mo eo e , he
maximum analysable ci cui size inc eases conside ably.
Ml*
2
...*n
1
o5
I
---
SAG app oach
SDG app oach
-
~
I
G
io4
I
'i;
102-
/
V
2
103:
/
/
SAG app oach
SDG app oach
---
1
o5
I
-
I
G
io4
I
'i;
102-
,/
/
V
2
103:
/
%---5
'
'io
'
'is'
' '
'io'
35
.
'
4d
(b)
Numbe
o
ladde s ages
Figu e
1:
(a)
Ladde ne wo k;
(b)
CPU
ime compa i.
son o he calcula ion
o
i s ol age gain.
2.1.
Te m gene a ion wi h he wo-g aph me hod
Te m gene a ion by he wo-g aph ee enume a-
ion me hod has been shown o be he mos e icien ech-
nique o implemen he SDG idea. The wo-g aph
me hod makes
use
o
a ol age g aph
G
and a cu en
g aph
GI,
bo h
easily
buil om he o iginal ne wo k
[8].
Fig.
2
illus a es he cons uc ion o bo h g aphs o a
simple ci cui . Each alid e m co esponds o he admi -
ance p oduc o e e y b anch in a spanning ee common
o bo h g aphs.
The e o e, he e m gene a ion p oblem o each
powe o
s
(say he k- h powe ) educes o he ollowing
g aph p oblem: enume a e all spanning ees common o
bo h, he ol age and he cu en g aph, in dec easing
o de o weigh and con aining
k
capaci ance b anches.
Repo ed echniques [2],[3],[4] s a ed om he
algo i hm in
[9]
o he enume a ion o spanning ees o
a one-colo ed g aph in dec easing o de o weigh . As he
alid e ms a e gi en by common spanning ees o ol -
Figu e
2:
Example ci cui and i s wo-g aph ep esen a ion
age and cu en g aphs, his algo i hm is applied o he
ol age g aph and o each gene a ed spanning ee
i
is
checked i i is also a spanning ee in he cu en g aph.
Bu ,
each spanning ee mus con ain exac ly
k
capaci-
o s. The e o e, he o iginal algo i hm o one-colo ed
g aphs had
o
be ex ended o wo-colo ed g aphs (con-
aining wo b anch ypes: capaci o s and conduc ances).
3.
The enume a ion p oblem in e ms
o
ma
aids
Ma oids a e ma hema ical abs ac ions which
o e a model o many ma hema ical s uc u es and com-
bina o ial p oblems. I ou p oblem can be exp essed
in
e ms o ma oids, hen, gene ally, be e algo i hms
imp o ing he unning ime (exploi ing special s uc u es
o hose ma oids) may be ound. Some gene al concep s
abou ma oids a e de ined nex .
3.1.
Ma oid p elimina ies
A
ma oid
M=(E,J)
is a s uc u e
in
which
E
is a
ini e se o
elemen s
and Jis a amily o subse s o
E,
which sa is y some axioms
[lo].
A
subse
I
in
3
is an
independen
se
o he ma oid
M=(E,J).
A maximal inde-
penden se is a
base
o he ma oid. A ma oid
M=(E,J)
is
said o be he
g aphic ma oid
o he g aph G i
E
is
he
se o a cs o G and a subse
I
E
is
in
gi and only i
I
is a cycle- ee
subse
o
a cs. In
a
connec ed g aph, a base
o i s g aphic ma oid co esponds o a spanning ee.
Finally, le
K
be a pa i ion ha sepa a es
E
in o
m
disjoin blocks
B1,
...,
Bm,
and le
di,
(i=l,
...,
m)
be
m
non-nega i e in ege s. Then,
M=(E, i
is
a
pa i ion
ma oid
i
J
is he amily o subse s
I
ha sa is y
InB.
Id.,i
=
1
,...,
m).
I
II
1
3.2.
Applica ion
o
ma oid heo y
o
ou p oblem
The ee enume a ion p oblem can be unde s ood
as a 3-ma oid in e sec ion p oblem. These h ee
ma oids a e:
1785
A g aphic ma oid de ined on he ol age g aph
G
Ano he g aphic ma oid de ined analogously on
Gp
*
A pa i ion ma oid de ined
so
ha each se o
b anches wi h
k
capaci o s is
a
base o he ma oid.
so
ha each spanning ee in
G,
is
a
base.
The ee enume a ion p oblem co esponds o he
enume a ion o bases common o he h ee ma oids
in
dec easing o de o weigh (g id a ea
in
Fig.
3).
Un o u-
na ely, he in e sec ion o h ee ma oids is conside ed o
be,
in
gene al,
a
NP-ha d p oblem
[I I].
Howe e , he e
a e polynomial ime algo i hms o he weigh ed in e -
sec ion o wo ma oids
[
12],[ 111. Repo ed symbolic
simula o s ha inco po a e
SDG
ha e sol ed he p ob-
lem by inding i s he in e sec ion o wo ma oids and
checking which elemen s o ha in e sec ion a e also
a
base
in
he hi d ma oid. Two di e en possibili ies
a ise:
Fi s in e sec ioning a g aphic ma oid and he pa i-
Fi s in e sec ioning he wo g aphic ma oids ( e i-
ion ma oid (ho izon ally-s iped a ea in Fig.
3).
cally-s iped a ea in Fig.
3).
du ion
space
Figu e 3: Illus a ing he 3-ma oid
in e sec ion
p oblem.
The i s one has been implemen ed
in
ADAGIO
[2],[3] and RAINIER [4]
as
i has been desc ibed
in
Sec-
ion 2.1. This me hod has he disad an age ha many
spanning ees o he ol age g aph may be gene a ed ha
a e no spanning ees in he cu en g aph and, hence, do
no lead o alid e ms. The a io o gene a ed alid e ms
o e he numbe o spanning ees in he ol age g aph
end o dec ease when he ci cui size inc eases, impos-
ing
a
limi o he maximum analysable ci cui size.
I
is
no easy o compa e he wo implemen a ions
o his app oach: ha in [2],[3], on he one hand, and ha
in
[4],
on he o he . This is mainly due o he di e en
e o c i e ia applied, and he ac ha he ool in [4] pe -
o ms
a
simpli ica ion be o e gene a ion p ocedu e
which signi ican ly educes he ci cui
size.
Fo illus a ion's sake, he ool
in
[3]
p o ides
a
simpli ied exp ession o he ol age gain o he olded-
cascode opamp
in
Fig. 4 in 54.7
s.
The gene a ion o a
symbolic exp ession o he pA741
's
ans e unc ion a
low equencies wi h
a
magni ude e o o
0.11%
(110
symbolic e ms) equi es
38
s.
An app oxima ed exp es-
sion (57 e ms) o he pA741 is p o ided in [4]
in
19
s.
Figu e
4:
:
Folded-cascode opamp.
1
0
The second app oach, ha
is,
in e sec ioning he
wo g aphic ma oids has been implemen ed
in
[13].
The e, common spanning ees o he ol age and cu en
g aphs, in dec easing o de o ee admi ance p oduc ,
a e di ec ly gene a ed using he algo i hm
in
[12].
So,
no
ime
is
was ed in gene a ing spanning ees
o
one o
hese g aphs which a e wa ds a e no spanning ees in
he o lhe g aph. Bu no con ol can be pe o med on he
b anch ype (conduc ance
o
capaci o ), and, hence,
capaci o admi ances ( he b anch weigh ) mus be e al-
ua ed a a ixed equency:
CUC.
This app oach app oxima es he ne wo k unc ion
o e
a
equency ange and conside s
a
se o sample e-
quencies wi hin his ange. The algo i hm
in
[
121 o gen-
e a ion o common spanning ees (wi hou any
cons ain in he numbe o capaci ance b anches) is
applied a each sample equency un il some gi en e o
c i e ion is me . Since he e: is
no
cons ain
in
he num-
be o capaci ance b anches, e ms a e gene a ed wi h
di e en powe s o
s.
A
g'ene a ed e m is kep in he
app oxima ed exp ession i i appea s o
a
leas one o
he sample equencies
in
he equency ange.
Bu his me hod has wo impo an d awbacks:
on
he one hand, mos e ms a e gene a ed a mo e han one
sample equency, and his means
a
loss
o
e iciency.
On
he o he hand, la ge e o s can occu a equencies di -
e en om hose
o
he se o sample equencies as
no
con ol is pe o med on he (exp ession accu acy be ween
wo sample equencies. Imp o ing he accu acy equi es
aking mo e sample equencies bu his
is
done a he
expense o addi ional compu a ion ime, g ea ly de e io-
a ing he e iciency
o
he me hod.
The same exp ession han [4] o he pA.741 has
been epo ed o be ob ained in
20
s.
p ac ically he same
han wi h he p e ious echinique [13]. Howe e , a mo e
accu a e exp ession (1185 , e ms) is ob ained in 57.7
s,
less han hal he ime needed wi h p e ious echnique.
Fo his example 4 sample equencies we e used.
1786
3.3.
A
3-ma oid in e sec ion algo i hm
The i s o he e iewed app oaches has
in
i s
e i-
ciency o la ge ci cui s i s main d awback. This is pa -
ially o e come
in
he second app oach bu
a
he cos o
uncon olled accu acy o he esul ing exp essions. The
new algo i hm in oduced in his sec ion combines he
posi i e ea u es o bo h app oaches by di ec ly add ess-
ing he in e sec ion
o
ou h ee pa icula ma oids.
The algo i hm
in
[12]
allows he gene a ion
o
spanning ees common
o
he ol age and cu en g aphs
in dec easing o de
o
weigh : om each common span-
ning ee ( he maximum weigh common spanning ee is
no di icul
o
ind), he nex one is ob ained by ca ying
ou he b anch exchanges indica ed by he bes
p imi i e
bo de pa h
(PBP) o he
bo de g aph
(BG), buil om
he spanning ee and bo h he ol age and cu en
g aphs. This bes p imi i e bo de pa h is mo e e i-
cien ly ound
in
he
condensed bo de g aph
(CBG),
which is easily de i ed om he bo de g aph
[
121.
This
algo i hm makes use o he lexicog aphic Floyd-Wa -
shall algo i hm o ind he bes PBP
in
ei he he
BG
o
he CBG.
Bu
his way,
in
bo h he cons uc ion o he
CBG and
in
he Floyd-Wa shall algo i hm, we ha e no
con ol on he ype o he b anches as i only looks a hei
weigh s,
so
i
is unde e mined he inal numbe o
b anches o each ype (capaci ance o conduc ance), and
hus i is no adequa e o ou needs.
In o de o sol e he p oblem o he con ol on he
b anch ypes, we pe o m ex ensi e modi ica ions in he
cons uc ion
o
he
CBG,
leading
o
he
Ex ended Con-
densed Bo de G aph
(ECBG).
In
he ECBG, each
b anch can ha e up o wo weigh s, ins ead
o
one weigh
as
i occu s
in
he con en ional CBG. These wo weigh s
co espond o he capaci ance and conduc ance nodes o
he BG ha sa is y he condi ion needed
o
ha e a co e-
sponding b anch in he CBG, wi h he highes weigh o
all he capaci ance/conduc ance nodes o he BG.
I
e lec s he ac ha a conduc ance/capaci ance b anch
can be exchanged by, ei he
a
conduc ance o a capaci-
ance b anch. Each b anch weigh has an associa ed
b anch ype, which co esponds o he numbe
o
induced
capaci ance b anch exchanges.
To ind he bes
PBP
in his ECBG an algo i hm,
called
Mul ile el Sho es Pa hs
(MSP), has been de el-
oped. The MSP algo i hm inds all sho es pa hs
be ween each pai o nodes
in
he ECBG. These pa hs
co espond o e e y possible ne numbe
o
capaci ance
b anch exchanges wi h espec o he common spanning
ee om which he BG was buil . This is done conside -
ing he combina ions o he di e en b anches be ween
nodes, and aking in o accoun also hei ypes. Thus, cal-
cula ion o he ollowing common spanning ee in he
same coe icien in
(1)
educes o aking he bes PBP
wi h ze o ne capaci ance b anch exchanges.
Conclusions
Two algo i hms o app oxima ion du ing gene a-
ion o he symbolic analysis
o
la ge analog in eg a ed
ci cui s ha e been e iewed. The i s
one
has in i s accu-
acy i s s onge ea u e, dec easing i s e iciency wi h
he ci cui size. The second gi es p io i y
o
he e i-
ciency bu wi h
a
high isk o la ge inaccu acies. This has
mo i a ed he in oduc ion o
a
new app oach which
o e comes p e ious d awbacks and cons i u es he i s
polynomial ime algo i hm o he h ee ma oid in e sec-
ion p oblem.
Re e ences
A. Rod iguez-Vazquez, F.V. Femandez, J.L. Hue as and
G.
Gielen,
Symbolic Anulysis Techniques and Applica ions
o
Anulog
Design
Au oma ion,
IEEE P ess, 1996.
F.V. Fe nindez, P. Wambacq,
G.
Gielen, A. Rod iguez-Vazquez
and W. Sansen. "Symbolic Analysis o La ge Analog In eg a ed
Ci cui s by App oxima ion Du ing Exp ession Gene a ion,"
P oc.
IEEE
ln .
Symp.
on Ci cui s
und
Sys ems,
pp.25-28, 1994.
P. Wambacq, F.V. Femandez,
G.
Gielen, W. Sansen and A.
Rod iguez-Vazquez, "E icien Symbolic Compu a ion
o
App oxima ed Small-Signal Cha ac e is ics
o
Analog In eg a ed
Ci cui s,"
IEEE
JSSC,
Vol. 30,
No.
3, pp, 327-330, Ma ch 1995.
Q.
Yu
and C. Sechen, "App oxima e Symbolic Analysis o La ge
Analog In eg a ed Ci cui s,"
P oc. IEEE In .
Co $
on
Compu e -
Aided Design,
pp.
664-611,
1994.
F.V. Fe nindez, A. Rod iguez-Vizquez and J.L. Hue as,
"In e ac i e AC Modeling and Cha ac e iza ion o Analog
Ci cui s ia Symbolic Analysis,"
Anulog In eg u ed Ci cui
und
Signal P ocessing,
Vol
1,
pp. 183-208, Kluwe , No . 1991
G.
Gielen, H. Walscha s and W. Sansen, "ISAAC: A Symbolic
Simula o o Analog In eg a ed Ci cui s,"
IEEE
Jou nul
o
Solid
S u e Ci cui s,
Vol.
24,
pp. 1587-1597, Dec. 1989.
G.
M. Wie zba
e
al.,
"Sspice
-
A
Symbolic SPICE P og am o
Linea Ac i e Ci cui s,"
P oc.
32nd
Midwe.F
Symp.
on
Ci cu s
und
Sys ems,
pp. 1197-1201, 1989.
P.M. Lin,
Symbolic
Ne wo k
Analysis.
Else ie , 199
I.
H.N. Gabow, "Two Algo i hms o Gene a ing Weigh ed
Spanning T ees in O de ",
SIAM
J.
o
Compu ing,
Vol.
6,
No.
I,
pp.
139-150,
Ma ch 1977.
[IO]
E.L. Lawle ,
Combinu o iul Op imizu ion: Ne wo ks
und
[
1
I]
C.H. Papadimi iou and
K.
S eigli z,
Combinu o iul
Op imizu ion: Algo i hms
und
Complexi y,
Englewood Cli s,
New Je sey. P en ice-Hall Inc.,
1982.
Mu oids.
Hol ,
Rineha and Wins on, 1976.
[I21 P.M. Came ini and H.W. Hamache , "In e sec ion
o
Two
Ma oids: (Condensed) Bo de G aph and Ranking",
SIAM
J.
DISC.
Mu h.,
Vol. 2, pp. 16-27, Feb.
1989.
[I31
Q.
Yu
and C. Sechen, "E icien App oxima ion o Symbolic
Ne wo k Func ions Using Ma oid In e sec ion Algo i hms,"
P oc.
IEEE
ISCAS,
pp. 2088-2091,
1995.
1787