ELVIS A Scalable, Loadable Custom Programmable Logic Device for Solving Boolean Satisfiability Probl - PowerPoint PPT Presentation

1 / 20
About This Presentation
Title:

ELVIS A Scalable, Loadable Custom Programmable Logic Device for Solving Boolean Satisfiability Probl

Description:

ELVIS - A Scalable, Loadable Custom Programmable Logic Device for Solving ... ELVIS. Easily Loaded Variable Implication Solver ... – PowerPoint PPT presentation

Number of Views:81
Avg rating:3.0/5.0
Slides: 21
Provided by: usersS
Category:

less

Transcript and Presenter's Notes

Title: ELVIS A Scalable, Loadable Custom Programmable Logic Device for Solving Boolean Satisfiability Probl


1
ELVIS - A Scalable, Loadable Custom Programmable
Logic Device for Solving Boolean Satisfiability
Problems
  • Mark Boyd and Tracy Larrabee

Jack Baskin School of Engineering
University of California Santa Cruz, CA
95064 mjboyd, larrabee_at_cse.ucsc.edu
2
Satisfiability
  • Satisfiability (SAT) is an important problem
  • test pattern generation
  • computer aided design
  • logic minimization
  • formal verification
  • Given n bombers (with a limited range) and m
    targets to bomb, can we service all the targets?

3
Satisfiability
  • k-SAT formulas are of the form
  • (ABC)(AB)(BC)(ABC)
  • n3, of variables
  • m4, of clauses
  • k3, max of variables per clause
  • For k 3, SAT is NP-Hard
  • there is no known algorithm which determines the
    satisfiability of an arbitrary SAT problem in
    polynomial time

4
Satisfiability
  • The partial truth assignment A1, C1
  • (TBT)(FB)(BF)(FBF)
  • causes transitive implications
  • B0 and B1, a contradiction

5
Serial vs. Parallel Design
  • Calculating transitive implications is critical
    for speedup
  • GRASP (1996, Silva and Sakallah)
  • Previous serial processor approach
  • Calculates transitive implications serially
  • Other excellent heuristics provide speedup
  • Zhong (1998, Martonosi, Ashar, Malik)
  • Calculates and broadcasts all direct transitive
    implications simultaneously in parallel

6
Zhongs Design
  • Shows significant speedup over GRASP by using
    massive, fine-grained parallelism
  • Routes a unique implication circuit (IMP) for
    each SAT problem instance

7
Zhongs Satisfiability
  • (ABC)(AB)(BC)(ABC)

Bit Encoding B B Free 0
0 False 0 1 True 1 0 Contra 1 1
8
ELVIS
  • Easily Loaded Variable Implication Solver
  • Two significant improvements over Zhongs
    original approach
  • ONEACT reduces logic use by factor of k
  • Bus Mask avoids NP-hard routing

9
ELVIS ONEACT
  • One Active Encoder, k-to-1 encoder
  • Binary fan-in, ?(k) space complexity
  • internal logic
  • 16-to-1 ONEACT encoder

10
Zhong Hole10
  • For the clause (12345678910)
  • Space complexity is (k-1)(k), ?(k )

2
11
ELVIS Hole10
  • ONEACT for (12345678910)
  • Space complexity is k2k, ?(k)
  • Key Result logic reduced by a factor of k

12
ELVIS Design
  • Clause Evaluation Circuit (CEC)

13
ELVIS Design
  • Variable Evaluation Circuit (VEC)

14
ELVIS Components Wiring
15
ELVIS Layout
  • Programmable Logic Device
  • 2-level AND-OR, like PLD
  • bus mask for routing
  • Majority CAM
  • bus mask ternary CAM
  • fanout to each cache entry

16
Hardware Requirements


2
2
n, of variables m, of clauses k, max of
variables per clause
where k is uniform, otherwise
17
Latency Characteristics
n, of variables m, of clauses k, max of
variables per clause
18
Future Work
  • Implementation
  • fast prototyping on Xilinx Virtex FPGAs
  • heuristic version with priority variables
  • SAT-server, accessible to other researchers
  • Complete pipelining, ?(1) delay

19
Competing Approach vs. ELVIS
  • Design a place and route compiler which
  • accepts k 3, n m/4
  • always terminates in O(nm) time
  • improves ELVIS space complexity
  • loads new problems in O(m) time
  • loads new clauses serially in constant time
  • Is this impossible?

20
Conclusion
  • Purely polynomial design.
  • Known timing and area characteristics.
  • Loadable in ?(km) cycles, same as the size of the
    formula.
  • IMP design compatible with any FSM.
  • Reduces logic by factor of k vs. Zhong
Write a Comment
User Comments (0)
About PowerShow.com