Influence of Professor T C Hus Works on Fundamental Approaches in Layout Andrew B Kahng CSE and ECE Departments UC San Diego httpvlsicaducsdedu Professor T C Hu Introduced combinatorial optimization and mathematical programming formulations and methods to VLSI Layout ID: 935411
Download Presentation The PPT/PDF document "Professor T. C. Hu ISPD-2018 Lifetime A..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Professor T. C. Hu ISPD-2018 Lifetime Achievement Commemoration
Slide2Influence of Professor T. C. Hu’s Works on Fundamental Approaches in Layout
Andrew B. KahngCSE and ECE DepartmentsUC San Diegohttp://vlsicad.ucsd.edu
Slide3Professor T. C. HuIntroduced combinatorial optimization, and mathematical programming formulations and methods, to VLSI LayoutMany works reflect unique ability to combine geometric, graph-theoretic, and combinatorial-algorithmic ideas
1961: Gomory-Hu cut tree1973: Adolphson-Hu cut-based linear placement1985: Hu-Moerder hyperedge net model1985: Hu-Shing - routingApplications of duality: flows and cuts, shadow price Professor C.-K. Cheng in next talk
Slide4Professor T. C. Hu
Slide5A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide6A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide7TACP and Shadow Price (1)
TACP: tentative assignment and competitive pricingApplication: Fixed-outline floorplanningFixed die, fixed block aspect ratio classical “packing” that minimizes whitespace, etc. !!!Seeks “perfect” rectilinear floorplanning: zero whitespaceIrregular block shapeOverlapping blocks
Slide8TACP and Shadow Price (2)
Shadow price in linear programming dualityPrimal-dual iterations in global routingLocal density in global placementGlobal densityMore recent: constraint-oriented local density
Better cell spreading, better wirelength!
Slide9A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide10Linear Placement
Optimal linear ordering (O.L.O.) problem pins in holes, pin per holeHoles in a line, unit distance apart
Minimize the wirelength
Gomory
and Hu /
Adolphson
and Hu
max flow values between any source / sink nodes can be obtained with
max flow problems, giving
fundamental cuts
fundamental cuts are lower bound for O.L.O.
The min-cut defines an ordered partition that is consistent with an optimal vertex order in the linear placement problem.
Slide11Minimum Cuts in Placement
Recursive min-cut[Cheng87]: universal application to VLSI placementCapo: top-down, min-cut bisectionFeng Shui: general purpose mixed-size placerDuality between max flows and min cuts[Yang96]: flow-based balanced netlist bipartitionMLPart: multilevel KL-FM/ flat KL-FM / flow-based partitioning
Slide12Linear Placements Today
Single-row placementVariable cell widthFixed row length with free sitesFixed cell orderingMulti-row placementLocal layout effect-awareReorderable cellsSupport of multi-height cells
Slide13A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide14Net Modeling
“Multiterminal Flows in a Hypergraph”, Hu and
Moerder
, 1985
Challenging question:
How should a hyperedge of a hypergraph be modeled by graph edges in a graph model of the hypergraph?
Applications for analytic placement, for exploiting sparse-matrix codes for layouts
New hyperedge net model -
p pin nodes and one star node to represent a p-pin hyperedge
Slide15Transform netlist hypergraphAdd one star node for each signal netConnect star node to each pin node (via a graph edge)
Sparse, symmetric + exactly captures true cut costStar model: [Brenner01], BonnPlace [Brenner08]Example Transformation
Example circuit with 5 modules
and 3 nets
Equivalent hypergraph model
Slide16A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide17The Prim-Dijkstra Tradeoff
Prim’s Minimum Spanning Tree (MST)Iteratively add edge eij to T, such that vi ϵ T, v
i
∉ T and
d
ij
is minimum
Minimizes tree wirelength (WL)
Dijkstra’s Shortest Path Tree (SPT)
Iteratively add edge
e
ij
to T, such that v
i
ϵ
T, v
i
∉ T and l
i + dij is minimum (where
is source-to-sink pathlength of
)Minimizes source-to-sink pathlengths (PLs) Prim-Dijkstra Tradeoff (Alpert, Hu, Huang, Kahng, 1993)“PD1” tradeoff: iteratively add eij to T that minimizes c li + dijc = 0 Prim’s MSTc = 1 Dijkstra’s SPT // c enables balancing of tree WL, source-sink PLs“PD2” tradeoff: iteratively add eij to T that minimizes ( lip + dij )1/p
p = ∞ Prim’s MST;p = 1 Dijkstra’s
SPT
Prim-Dijkstra Construction
Prim’s Minimum Spanning Tree (MST)
Minimizes wirelength
Dijkstra’s Shortest Path Tree (SPT)
Minimizes source-sink pathlengths
Prim-Dijkstra (PD) tradeoff
Directly trades off the
Prim, Dijkstra constructions
0
But large pathlengths to nodes 3,4,5
2
1
3
4
5
2
1
0
3
4
5
But large tree
wirelength!
0
2
1
3
4
5
Slide19PD Tradeoff: 25 Years of Impact
Widely used In EDA for timing estimation, buffer tree construction and global routingIn flood control, biomedical, military, wireless sensor networks, etc.Simple and fast – O(n log n)Alpert et al., DAC06: PD is practically ‘free’
Yesterday: “PD Revisited”
Iterative repair of spanning tree
Detour-aware
Steinerization
Better WL, PL tradeoff
Slide20A Few ExamplesTentative Assignment / Competitive PricingOptimal Linear Ordering
Hyperedge Net ModelThe Prim-Dijkstra TradeoffThe Discrete Plateau Problem and Finding a Wide Path
Slide21Connection Finding
Basic element of any routing approach- routing (Hu and Shing, 1985)Find connections given edge and vertex costsComprehend existence of “turn” at vertexProvide unified elements for
Dijkstra’s algorithm
Best-first (A*) search
Proc. Nat. Acad. Sci., October 1992
Discrete version of Plateau’s minimum-surface problem
Solved using duality of cuts and flow
Slide23Towards Robust (Wide) Path Finding
Robust path finding problemSource-destination routing with prescribed widthSeek minimum-cost path that has robustness (width) = dE.g., a mobile agent with finite width
Slide24Network Flow Approach
Discretize routing environmentA minimum cut in flow networkContain all vertices and edges on a robust pathCorrespond to a maximum flow by dualityReturn a robust path
Slide25Applications Today
Relevant to many difficult problemsBus routing, bus feedthrough determination, etc.IC package routingPer-net PI/SI requirementNeed traces of various widthWide path finding (with multiple commodities) can be useful
Slide26Conclusion
Slide27T. C. Hu
W. T. Torres
K-C. Tan
Y. S.
Kuo
P. A. Tucker
A. B. Kahng
K. E. Moerder
M.-T.
Shing
F. Ruskey
D. Adolphson
B. N. Tien
D.R. van
Baronaigien
Y.
Koda
P. Evans
A. Smith
K. Wong
J. Sawada
S. Chow
M.R. Kindl
M.M. Cordeiro
J. Chen
G. Robins
C.J. Alpert
K.D. Boese
Y. Chen
K. Masuko
P. Gupta
L. Hagen
D.J. Huang
B. Liu
S. Mantik
I. Markov
S. Muddu
S. Reda
C-W.A. Tsao
Q. Wang
X. Xu
M. Alexander
T. Zhang
A Ramani
S.
Adya
G.
Pruesse
G. Thomas
A. Zaki
S-J. Su
Y-H. Hsu
C-C. Jung
Professor Hu’s 96 Ph.D. Descendants
S. Kang
C. H. Park
W. Chan
S. Muddu
K. Samadi
K. Jeong
R. O. Topaloglu
T. Chan
P. Sharma
S. Nath
J. Li
M. Weston
B.
Bultena
A. Erickson
A.
Mamakani
V. Irvine
A. Williams
L. Bolotnyy
C. Taylor
K. Chawla
R. Layer
N. Brunelle
A. M. Eren
G. Xu
V. Maffei
G. Viamontes
K-H. Chang
S. Krishnaswamy
S. Plaza
J. Roy
D. Papa
D. Lee
M. Kim
J. Hu
H. J. Garcia
R. Cochran
N. Abdullah
K. Nepal
K. Dev
X. Zhan
S. Hashemi
R. Azimi
J. Lee
L. Cheng
R. Ghaida
A. A. Kagalwalla
L. Lai
M. Gottscho
S. Wang
Y. Badr
Slide28Thank you, Professor Hu.