This is a library that's useful for doing tree decompositions. I have been using these tools for my tool Ganak, but I want to use it in more projects. Now it's a library.

Authors

• Ben Strasser -- original FlowCutter library
• Tuukka Korhonen and Matti Jarvisalo -- graph libraries
• Kenji Hashimoto -- graph libraries

Standalone treedecomp binary

Alongside the library, build/treedecomp is a CLI that reads a CNF on stdin (or a file), builds the primal graph (one edge between every pair of variables that share a clause), runs FlowCutter, and prints a per-variable tree decomposition score — the same score ganak uses internally to bias branching.

Input

Standard DIMACS CNF:

p cnf <nvars> <nclauses>
2 30 0
-1 4 5 0
...

Output

Comment lines prefixed with c (parse stats and the TD width), followed by one line per variable on stdout in the form:

<var> <tdscore>

The first column is the variable (1-indexed, matching the CNF numbering). The second column is its tree-decomposition score, in the range 0..100 — higher means the variable is closer to the centroid of the tree decomposition (this is the compute_td_score_using_raw path from ganak's counter.cpp).

Example

A small CNF with a mix of ternary clauses:

$ cat example.cnf
p cnf 7 8
1 2 3 0
2 3 4 0
3 4 5 0
-4 5 6 0
5 6 7 0
-1 3 5 0
2 4 6 0
-5 -6 7 0

$ ./build/treedecomp --tdvis example.dot example.cnf
c parsed nvars=7 clauses=8
c primal nodes=7 edges=13 density=0.265306 edge/var=1.85714
c TD width: 3
c centroid bag: 2
c wrote DOT file: example.dot (render: dot -Tpdf example.dot -o td.pdf)
1 0
2 100
3 50
4 100
5 100
6 100
7 50

Each output line is <variable> <score>: column 1 is the CNF variable id (1-indexed), column 2 is its TD score in 0..100. Higher means the variable sits closer to the centroid of the tree decomposition — in this run, vars 2, 4, 5, 6 are central (score 100), vars 3 and 7 are one step out, and var 1 is farthest (score 0).

Passing --tdvis <file> writes the tree decomposition as a Graphviz DOT file, with one node per bag (labeled with its size and the CNF variables it contains). Render it to an image with:

$ dot -Tpng example.dot -o example.png

The resulting tree for the CNF above has 4 bags of width 3 (i.e. 4 variables each); bag 2 {2, 4, 5, 6} is reported as the centroid and is highlighted in the rendered graph:

Usage

./treedecomp input.cnf              # read from file
./treedecomp < input.cnf            # read from stdin
./treedecomp --tditers 300 input.cnf

Options

Cutoffs and tuning knobs mirror ganak's conf.td_* settings so results line up with what ganak computes on the same instance:

flag	meaning
--tdsteps N	FlowCutter max steps (default 100000)
--tditers N	FlowCutter iterations / restarts (default 900)
--tdmaxedges N	skip TD if primal has more than N edges
--tdmaxdensity F	skip TD if primal density > F
--tdmaxedgeratio N	skip TD if edge/var ratio > N
--tdvarlim N	skip TD if #vars > N
--tdcontract 0/1	contract high-numbered vars before running TD
--tdvis FILE	write the tree decomposition as a Graphviz DOT file
-v N	verbosity

The binary links against the treedecomp library and uses argparse.hpp for option parsing.

Notes

Basically, if you find a bug, it's likely been introduced by me. Just open an issue and I'll fix.