sudoku-norvig-rs

This repository contains a Rust version of Peter Norvig's "Solve Every Sudoku Puzzle".

The rabbit hole

Sudoku was said to be a "denial of service attack on the brain" (apparently by Ben Laurie, of Apache httpd fame). Writing a Sudoku solver is an even bigger denial of service attack on the brain. However, as a side effect of writing this library, I was able to learn some software development tricks in Rust, Go, and Python.

I improved Peter Norvig's version slightly by using Python lists instead of dictionaries. This turns out not to complicate the code too much while improving performance.

Then I turned to Go and it resulted in significantly better performance. Another really neat repo is go-sudoku by Eli Bendersky. His code is much faster than the Norvig Python version, and was faster than this Rust version. Then I switched to using the same u16 representation as him to represent squares of the puzzle, and now my version is faster. (Rust is also more performant than Go, in general.)

Making a fast Sudoku solver is more complicated than perhaps meets the eye. Much has been written, and there's a devoted gang of solver nerds that have cut their teeth on this problem, including the current prime minister of Singapore. The fastest solver I could find was tdoku by t-dillon. His writeup is a very deep dive.

The fastest Rust version I found was https://github.com/Emerentius/sudoku. My version is not nearly as fast because it sticks to Norvig's basic approach.

The coolest Sudoku thing ever is the augmented reality Sudoku solver that makes uses of the fast Rust solver by Emerentius.

Run

Install Rust with rustup, which installs cargo. Then run:

cargo run --release --example bench

This will compile the library and run it against a set of puzzles.

Library

Use this code as a Rust library. It currently doesn't reside in crates.io, so you'll need to include it in your Cargo.toml file like:

[dependencies]
sudoku = { git = "https://github.com/mlbright/sudoku-norvig-rs" }

Performance

Run:

python3 sudoku.py # For the Python benchmarks
cargo run --release --example bench # For the Rust benchmarks

MacOS

On my MacBook Air (Retina, 13-inch, 2020), I get:

Solved 50 of 50 easy puzzles (avg. 0.00007972 secs (12543.9351 Hz), max 0.00015145 secs).
Solved 95 of 95 hard puzzles (avg. 0.00035310 secs (2832.0641 Hz), max 0.00800655 secs).
Solved 11 of 11 hardest puzzles (avg. 0.00006083 secs (16438.6428 Hz), max 0.00013210 secs).
Solved 20 of 20 hardest20 puzzles (avg. 0.00074833 secs (1336.3037 Hz), max 0.00288490 secs).
Solved 1000 of 1000 hardest20x50 puzzles (avg. 0.00074095 secs (1349.6268 Hz), max 0.00347086 secs).
Solved 87 of 87 topn87 puzzles (avg. 0.00048929 secs (2043.7616 Hz), max 0.00257739 secs).
Solved 1 of 1 most-difficult puzzles (avg. 0.55276984 secs (1.8091 Hz), max 0.55276984 secs).
Solved 1264 of 1264 combined puzzles (avg. 0.00109279 secs (915.0886 Hz), max 0.55694445 secs).