sidekick/src/sat
Simon Cruanes 233df98229 perf(sat): proper GC for clauses
GC is now eager in the cleanup of watchlists and more elaborate.

It starts by marking all clauses justifying literals on the trail,
because we cannot remove them yet (they might be used in resolution
steps during clause analysis).

Then, we sort set of learnt clauses by decreasing activity, and pop them
one by one until the target size is reached.  Clauses that are marked,
because they are on the trail, are kept on the side so we can push them
back at the end.

For each clause we want to remove, we start by marking it "dead".
Then we mark its two watch literals are "dirty", which means we will
need to purge their watchlists from dead clauses.

Then we purge watchlists (remove deadclauses entirely). At this point
the dead clauses are unreachable from both trail and watchlists.
We can remove all data for each such clause, and put their index for
recycling so another new clause can reuse their slot.

Finally we unmark dirty literals and saved-by-trail-explanation
explanation clauses.
2021-07-21 20:29:27 -04:00
..
.gitignore prepare for vendoring 2021-07-18 01:24:04 -04:00
CHANGELOG.md prepare for vendoring 2021-07-18 01:24:04 -04:00
dune wip: refactor SAT solver 2021-07-19 09:57:02 -04:00
Heap.ml perf(solver): use VecI32 in the heap 2021-07-20 10:07:43 -04:00
Heap.mli wip: refactor(sat): use struct-of-array for atom/var 2021-07-19 09:57:02 -04:00
Heap_intf.ml perf(solver): use VecI32 in the heap 2021-07-20 10:07:43 -04:00
README.md prepare for vendoring 2021-07-18 01:24:04 -04:00
Sidekick_sat.ml refactor(sat): use first-class modules instead of records 2021-07-18 19:18:42 -04:00
Solver.ml perf(sat): proper GC for clauses 2021-07-21 20:29:27 -04:00
Solver.mli feat: mli for the SAT solver 2021-07-19 09:17:20 -04:00
Solver_intf.ml more stats for main 2021-07-21 20:24:27 -04:00

MSAT Build Status

MSAT is an OCaml library that features a modular SAT-solver and some extensions (including SMT), derived from Alt-Ergo Zero.

It was presented at ICFP 2017, using a poster

This program is distributed under the Apache Software License version 2.0. See the enclosed file LICENSE.

Documentation

See https://gbury.github.io/mSAT/

INSTALLATION

Via opam

Once the package is on opam, just opam install msat. For the development version, use:

opam pin add msat https://github.com/Gbury/mSAT.git

Manual installation

You will need dune and iter. The command is:

$ make install

USAGE

Generic SAT/SMT Solver

A modular implementation of the SMT algorithm can be found in the Msat.Solver module, as a functor which takes two modules :

  • A representation of formulas (which implements the Formula_intf.S signature)

  • A theory (which implements the Theory_intf.S signature) to check consistence of assertions.

  • A dummy empty module to ensure generativity of the solver (solver modules heavily relies on side effects to their internal state)

Sat Solver

A ready-to-use SAT solver is available in the Msat_sat module using the msat.sat library. It can be loaded as shown in the following code :

# #require "msat";;
# #require "msat.sat";;
# #print_depth 0;; (* do not print details *)

Then we can create a solver and create some boolean variables:

module Sat = Msat_sat
module E = Sat.Int_lit (* expressions *)

let solver = Sat.create()

(* We create here two distinct atoms *)
let a = E.fresh ()    (* A 'new_atom' is always distinct from any other atom *)
let b = E.make 1      (* Atoms can be created from integers *)

We can try and check the satisfiability of some clauses — here, the clause a or b. Sat.assume adds a list of clauses to the solver. Calling Sat.solve will check the satisfiability of the current set of clauses, here "Sat".

# a <> b;;
- : bool = true
# Sat.assume solver [[a; b]] ();;
- : unit = ()
# let res = Sat.solve solver;;
val res : Sat.res = Sat.Sat ...

The Sat solver has an incremental mutable state, so we still have the clause a or b in our assumptions. We add not a and not b to the state, and get "Unsat".

# Sat.assume solver [[E.neg a]; [E.neg b]] () ;;
- : unit = ()
# let res = Sat.solve solver ;;
val res : Sat.res = Sat.Unsat ...

Formulas API

Writing clauses by hand can be tedious and error-prone. The functor Msat_tseitin.Make in the library msat.tseitin proposes a formula AST (parametrized by atoms) and a function to convert these formulas into clauses:

# #require "msat.tseitin";;
(* Module initialization *)
module F = Msat_tseitin.Make(E)

let solver = Sat.create ()

(* We create here two distinct atoms *)
let a = E.fresh ()    (* A fresh atom is always distinct from any other atom *)
let b = E.make 1      (* Atoms can be created from integers *)

(* Let's create some formulas *)
let p = F.make_atom a
let q = F.make_atom b
let r = F.make_and [p; q]
let s = F.make_or [F.make_not p; F.make_not q]

We can try and check the satisfiability of the given formulas, by turning it into clauses using make_cnf:

# Sat.assume solver (F.make_cnf r) ();;
- : unit = ()
# Sat.solve solver;;
- : Sat.res = Sat.Sat ...
# Sat.assume solver (F.make_cnf s) ();;
- : unit = ()
# Sat.solve solver ;;
- : Sat.res = Sat.Unsat ...

CDCL(T): a Sudoku solver as an example

The directory src/sudoku/ contains a simple Sudoku solver that uses the interface Msat.Make_cdcl_t. In essence, it implements the logical theory CDCL(Sudoku). The script sudoku_solve.sh compiles and runs the solver, as does dune exec src/sudoku/sudoku_solve.exe.

It's able to parse sudoku grids denoted as 81 integers (see tests/sudoku/sudoku.txt for example).

Here is a sample grid and the output from the solver (in roughly .5s):

$ echo '..............3.85..1.2.......5.7.....4...1...9.......5......73..2.1........4...9' > sudoku.txt
$ dune exec src/sudoku/sudoku_solve.exe -- sudoku.txt
...
#########################
solve grid:
  .........
  .....3.85
  ..1.2....
  ...5.7...
  ..4...1..
  .9.......
  5......73
  ..2.1....
  ....4...9
  
...
  987654321
  246173985
  351928746
  128537694
  634892157
  795461832
  519286473
  472319568
  863745219
  
###################
...