CryptoMiniSat SAT solver

This system provides CryptoMiniSat, an advanced incremental SAT solver. The system has 3 interfaces: command-line, C++ library and python. The command-line interface takes a cnf as an input in the DIMACS format with the extension of XOR clauses. The C++ and python interface mimics this and also allows for incremental use: assumptions and multiple solve calls. A C compatible wrapper is also provided.

When citing, always reference our SAT 2009 conference paper, bibtex record is here.

License

Everything that is needed to build by default is MIT licensed. If you specifically instruct the system it can build with Bliss, which are both GPL. However, by default CryptoMiniSat will not build with these.

Compiling in Linux

To build and install, issue:

sudo apt-get install build-essential cmake
# not required but very useful
sudo apt-get install zlib1g-dev
tar xzvf cryptominisat-version.tar.gz
cd cryptominisat-version
mkdir build && cd build
cmake ..
make
sudo make install
sudo ldconfig

Compiling in Mac OSX

First, you must get Homebrew from https://brew.sh/ then:

brew install cmake
tar xzvf cryptominisat-version.tar.gz
cd cryptominisat-version
mkdir build && cd build
cmake ..
make
sudo make install

Compiling in Windows

C:\> [ download and unzip cryptominisat-version.zip ]
C:\> cd cryptominisat
C:\cryptominisat> mkdir build
C:\cryptominisat> cd build
C:\cryptominisat\build> cmake -G "Visual Studio 17 2022" -DCMAKE_BUILD_TYPE=Release -DSTATICCOMPILE=ON ..
C:\cryptominisat\build> cmake --build --config Release .

You now have the static binary under C:\cryptominisat\build\Release\cryptominisat5.exe

Command-line usage

Let's take the file:

p cnf 3 3
1 0
-2 0
-1 2 3 0

The file has 3 variables and 3 clauses, this is reflected in the header p cnf 3 3 which gives the number of variables as the first number and the number of clauses as the second. Every clause is ended by '0'. The clauses say: 1 must be True, 2 must be False, and either 1 has to be False, 2 has to be True or 3 has to be True. The only solution to this problem is:

cryptominisat5 --verb 0 file.cnf
s SATISFIABLE
v 1 -2 3 0

Which means, that setting variable 1 True, variable 2 False and variable 3 True satisfies the set of constraints (clauses) in the CNF. If the file had contained:

p cnf 3 4
1 0
-2 0
-3 0
-1 2 3 0

Then there is no solution and the solver returns s UNSATISFIABLE.

Incremental Python Usage

The python module works with both Python 3. Just execute:

pip3 install pycryptosat

You can then use it in incremental mode as:

>>> from pycryptosat import Solver
>>> s = Solver()
>>> s.add_clause([1])
>>> s.add_clause([-2])
>>> s.add_clause([-1, 2, 3])
>>> sat, solution = s.solve()
>>> print sat
True
>>> print solution
(None, True, False, True)
>>> sat, solution = s.solve([-3])
>> print sat
False
>>> sat, solution = s.solve()
>>> print sat
True
>>> s.add_clause([-3])
>>> sat, solution = s.solve()
>>> print sat
False

We can also try to assume any variable values for a single solver run:

>>> sat, solution = s.solve([-3])
>>> print sat
False
>>> print solution
None
>>> sat, solution = s.solve()
>>> print sat
True
>>> print solution
(None, True, False, True)

If you want to build the python module, you can do this:

sudo apt-get install build-essential
sudo apt-get install python3-setuptools python3-dev
git clone https://github.com/msoos/cryptominisat
python -m build
pip install dist/pycryptosat-*.whl

Incremental Library Usage

The library uses a variable numbering scheme that starts from 0. Since 0 cannot be negated, the class Lit is used as: Lit(variable_number, is_negated). As such, the 1st CNF above would become:

#include <cryptominisat5/cryptominisat.h>
#include <assert.h>
#include <vector>
using std::vector;
using namespace CMSat;

int main()
{
    SATSolver solver;
    vector<Lit> clause;

    //Let's use 4 threads
    solver.set_num_threads(4);

    //We need 3 variables. They will be: 0,1,2
    //Variable numbers are always trivially increasing
    solver.new_vars(3);

    //add "1 0"
    clause.push_back(Lit(0, false));
    solver.add_clause(clause);

    //add "-2 0"
    clause.clear();
    clause.push_back(Lit(1, true));
    solver.add_clause(clause);

    //add "-1 2 3 0"
    clause.clear();
    clause.push_back(Lit(0, true));
    clause.push_back(Lit(1, false));
    clause.push_back(Lit(2, false));
    solver.add_clause(clause);

    lbool ret = solver.solve();
    assert(ret == l_True);
    std::cout
    << "Solution is: "
    << solver.get_model()[0]
    << ", " << solver.get_model()[1]
    << ", " << solver.get_model()[2]
    << std::endl;

    //assumes 3 = FALSE, no solutions left
    vector<Lit> assumptions;
    assumptions.push_back(Lit(2, true));
    ret = solver.solve(&assumptions);
    assert(ret == l_False);

    //without assumptions we still have a solution
    ret = solver.solve();
    assert(ret == l_True);

    //add "-3 0"
    //No solutions left, UNSATISFIABLE returned
    clause.clear();
    clause.push_back(Lit(2, true));
    solver.add_clause(clause);
    ret = solver.solve();
    assert(ret == l_False);

    return 0;
}

The library usage also allows for assumptions. We can add these lines just before the return 0; above:

vector<Lit> assumptions;
assumptions.push_back(Lit(2, true));
lbool ret = solver.solve(&assumptions);
assert(ret == l_False);

lbool ret = solver.solve();
assert(ret == l_True);

Since we assume that variable 2 must be false, there is no solution. However, if we solve again, without the assumption, we get back the original solution. Assumptions allow us to assume certain literal values for a specific run but not all runs -- for all runs, we can simply add these assumptions as 1-long clauses.

Multiple solutions

To find multiple solutions to your problem, just run the solver in a loop and ban the previous solution found:

while(true) {
    lbool ret = solver->solve();
    if (ret != l_True) {
        assert(ret == l_False);
        //All solutions found.
        exit(0);
    }

    //Use solution here. print it, for example.

    //Banning found solution
    vector<Lit> ban_solution;
    for (uint32_t var = 0; var < solver->nVars(); var++) {
        if (solver->get_model()[var] != l_Undef) {
            ban_solution.push_back(
                Lit(var, (solver->get_model()[var] == l_True)? true : false));
        }
    }
    solver->add_clause(ban_solution);
}

The above loop will run as long as there are solutions. It is highly suggested to only add into the new clause(bad_solutions above) the variables that are "important" or "main" to your problem. Variables that were only used to translate the original problem into CNF should not be added. This way, you will not get spurious solutions that don't differ in the main, important variables.

Rust bindings

To build the Rust bindings:

git clone https://github.com/msoos/cryptominisat-rs/
cd cryptominisat-rs
cargo build --release
cargo test

You can use it as per the README in that repository. To include CryptoMiniSat in your Rust project, add the dependency to your Cargo.toml file:

cryptominisat = { git = "https://github.com/msoos/cryptominisat-rs", branch= "master" }

You can see an example project using CryptoMiniSat in Rust here.

Preprocessing

If you wish to use CryptoMiniSat as a preprocessor, we encourage you to try out our model counting preprocessing framework (Arjun)[https://www.github.com/meelgroup/arjun]. Arjun was conceived to minimize input formulas for model counting, but it can also be used for non-model-counting purposes. However, it cannot produce a solution to the original CNF given a solution to the simplified CNF. This current limitation will eventually be lifted for Arjun.

Please see the README of Arjun for more details. The basic usage of Arjun is: ./arjun --renumber 0 original.cnf simplified.cnf

Gauss-Jordan elimination

Since CryptoMiniSat 5.8, Gauss-Jordan elimination is compiled into the solver by default. However, it will turn off automatically in case the solver observes GJ not to perform too well. To use Gaussian elimination, provide a CNF with xors in it (either in CNF or XOR+CNF form) and either run with default setup, or, tune it to your heart's desire:

Gauss options:
  --iterreduce arg (=1)       Reduce iteratively the matrix that is updated.We
                              effectively are moving the start to the last
                              column updated
  --maxmatrixrows arg (=3000) Set maximum no. of rows for gaussian matrix. Too
                              large matrixes should be discarded for reasons of
                              efficiency
  --autodisablegauss arg (=1) Automatically disable gauss when performing badly
  --minmatrixrows arg (=5)    Set minimum no. of rows for gaussian matrix.
                              Normally, too small matrixes are discarded for
                              reasons of efficiency
  --savematrix arg (=2)       Save matrix every Nth decision level
  --maxnummatrixes arg (=3)   Maximum number of matrixes to treat.

In particular, you may want to set --autodisablegauss 0 in case you are sure it'll help.

Testing

For testing you will need the GIT checkout and build as per:

sudo apt-get install build-essential cmake git
sudo apt-get install zlib1g-dev libboost-program-options-dev libsqlite3-dev
sudo apt-get install git python3-pip python3-setuptools python3-dev
sudo pip3 install --upgrade pip
sudo pip3 install lit
git clone https://github.com/msoos/cryptominisat.git
cd cryptominisat
git submodule update --init
mkdir build && cd build
cmake -DENABLE_TESTING=ON ..
make -j4
make test
sudo make install
sudo ldconfig

Fuzzing

Build for test as per above, then:

cd ../cryptominisat/scripts/fuzz/
./fuzz_test.py

CrystalBall

Build and use instructions below. Please see the associated blog post for more information.

# prerequisites on a modern Debian/Ubuntu installation
sudo apt-get install build-essential cmake git
sudo apt-get install zlib1g-dev libsqlite3-dev
sudo apt-get install libboost-program-options-dev libboost-serialization-dev
sudo apt-get install python3-pip
sudo pip3 install sklearn pandas numpy lit matplotlib

# build and install Louvain Communities
git clone https://github.com/meelgroup/louvain-community
cd louvain-community
mkdir build && cd build
cmake ..
make -j10
sudo make install
cd ../..

# build and install XGBoost
git clone https://github.com/dmlc/xgboost
cd xgboost
mkdir build && cd build
cmake ..
make -j10
sudo make install
cd ../..

# build and install LightGBM
git clone https://github.com/microsoft/LightGBM
cd LightGBM
mkdir build && cd build
cmake ..
make -j10
sudo make install
cd ../..

# getting the code
git clone https://github.com/msoos/cryptominisat
cd cryptominisat
git checkout crystalball
git submodule update --init
mkdir build && cd build
ln -s ../scripts/crystal/* .
ln -s ../scripts/build_scripts/* .

# Let's get an unsatisfiable CNF
wget https://www.msoos.org/largefiles/goldb-heqc-i10mul.cnf.gz
gunzip goldb-heqc-i10mul.cnf.gz

# Gather the data, denormalize, label,
# create the classifier, generate C++,
# and build the final SAT solver
./ballofcrystal.sh goldb-heqc-i10mul.cnf
[...compilations and the full data pipeline...]

# let's use our newly built tool
./cryptominisat5 goldb-heqc-i10mul.cnf
[ ... ]
s UNSATISFIABLE

# Let's look at the data
cd goldb-heqc-i10mul.cnf-dir
sqlite3 mydata.db
sqlite> select count() from sum_cl_use;
94507

Configuring a build for a minimal binary&library

The following configures the system to build a bare minimal binary&library. It needs a compiler, but nothing much else:

cmake -DONLY_SIMPLE=ON -DNOZLIB=ON -DSTATS=OFF -DENABLE_TESTING=OFF .

CMake Arguments

The following arguments to cmake configure the generated build artifacts. To use, specify options prior to running make in a clean subdirectory: cmake <options> ..

-DSTATICCOMPILE=<ON/OFF> -- statically linked library and binary. You must build&link BreakID with the same DSTATICCOMPILE=<ON/OFF> setting as well. You can get the BreakID library from our GitHub repository
-DSTATS=<ON/OFF> -- advanced statistics (slower). Needs louvain communities installed.
-DENABLE_TESTING=<ON/OFF> -- test suite support
-DMIT=<ON/OFF> -- MIT licensed components only
-DNOMPI=<ON/OFF> -- without MPI support
-DNOZLIB=<ON/OFF> -- no gzip DIMACS input support
-DONLY_SIMPLE=<ON/OFF> -- only the simple binary is built
-DLARGEMEM=<ON/OFF> -- more memory available for clauses (but slower on most problems)
-DIPASIR=<ON/OFF> -- Build libipasircryptominisat.so for IPASIR interface support

Getting learnt clauses

As an experimental feature, you can get the learnt clauses from the system with the following code, where lits is filled with learnt clauses every time get_next_small_clause is called. The example below will eventually return all clauses of size 4 or less. You can call end_getting_small_clauses at any time.

SATSolver s;
//fill the solver, run solve, etc.

//Get all clauses of size 4 or less

s->start_getting_small_clauses(4);

vector<Lit> lits;
bool ret = true;
while (ret) {
    bool ret = s->get_next_small_clause(lits);
    if (ret) {
        //deal with clause in "lits"
        add_to_my_db(lits);
    }
}
s->end_getting_small_clauses();