dpll-sat is a SAT solver implementing the classic DPLL algorithm. I wrote this program for learning purposes and also for comparing the performance of a naive DPLL solver and modern solvers.

• Build
• Usage
• Implementation notes
• Benchmarks
• References

$ git clone https://github.com/snsinfu/dpll-sat
$ cd dpll-sat
$ cargo build --release
$ cp target/release/dpll-sat ~/bin/

dpll-sat command reads a simplified DIMACS CNF from stdin. It prints "sat" followed by an assignment and exits with exit code 0 if the formula is satisfiable. Otherwise, it prints "unsat" and exits with exit code 1.

$ dpll-sat < examples/qg3-08.cnf
sat
1 2 3 4 5 6 7 8 -9 -10 -11 -12 -13 -14 -15 -16 -17 ...

The second line shows an assignment. A positive number i means that the i-th variable is true. A negative number -i means that the i-th variable is false. The output format is essentially the same as that of z3.

The algorithm is implemented in the following way in pseudocode:

function DPLL(formula, vars) {
    // Simplify formula by determining variables in unit clauses.
    formula = copy(formula);
    unit_propagate(formula, vars);

    // Stopping conditions.
    if (formula is empty) {
        return "SAT";
    }
    if (empty clause in formula) {
        return "UNSAT";
    }

    // Choose a branching literal and recurse.
    v = choose_literal(formula);

    return DPLL(formula + [v], vars) || DPLL(formula + [not(v)], vars);
}

• Pure literal elimination is not implemented due to the complexity of keeping track of the polarity of every literal in a formula.
• The branching literal is chosen to be the most-used variable in a formula. The literal is eliminated in the next recursion, so this strategy eagerly reduces the size of the formula.
• Every recursion creates a new copy of a formula. This is inefficient. The copy is necessary because the algorithm eliminates some clauses and literals in a formula and later revert it for backtracking. But, most clauses are untouched. There would be a clever data structure that can reduce the number of copies. Maybe deque?
• A formula is represented as a vector-of-vectors. It's horribly inefficient since a formula tends to consist of many small clauses, making memory access extremely scattered. It would be much better to use a flat vector with sentinel values.

Run time for some benchmark instances found on the satlib site and a benchmark page. The run time is the total "USER" time spent from a program startup to termination. Benchmarks ran on Fedora 33 Linux with AMD Ryzen 3600 CPU @ 4.2GHz (boost).

Z3 does pretty good job considering CaDiCaL is the top-1 winner in the SAT track of the SAT Race 2019.

This lecture note is useful to understand the DPLL algorithm:

• http://www.cs.cornell.edu/courses/cs4860/2009sp/lec-04.pdf

This book has a ton of examples:

• https://sat-smt.codes/