Ruthless

A UCI chess engine written in Rust. It uses an NNUE evaluation and a fairly standard alpha-beta search.

Features

• Bitboards, magic or PEXT slider attacks (picked at startup)
• NNUE: 10 king-bucketed inputs, 512 hidden units, 8 output buckets, SCReLU. AVX2, AVX-512 and NEON paths
• PVS with null move pruning, LMR, futility pruning, razoring, ProbCut, singular extensions
• Quiescence search, SEE, check extensions
• Quiet, noisy, continuation and correction histories
• Shared transposition table, multithreaded search
• Syzygy WDL probing (optional, through Fathom)

Building

You need a recent stable Rust toolchain.

cargo build --release

The binary ends up in target/release/. By default it is built with target-cpu=native, so it may not run on older CPUs. For something you want to share, pick an x86-64 level explicitly:

RUSTFLAGS="-C target-cpu=x86-64-v3" cargo build --release

Use x86-64-v2 for the baseline or x86-64-v4 for AVX-512. make release-v2, release-v3 and release-v4 do the same thing. The NNUE kernels pick AVX2 at runtime, so a v2 build still uses the vector paths on AVX2 hardware.

For the last few percent, build with profile-guided optimization:

cargo pgo instrument
cargo pgo run -- bench
cargo pgo optimize

(make pgo on Unix, pgo.bat on Windows.) This needs llvm-tools (rustup component add llvm-tools) and cargo-pgo (cargo install cargo-pgo). The optimized binary is copied to target/release/ and kept as Ruthless-pgo next to it, so a later plain cargo build --release cannot overwrite the only copy.

To enable Syzygy support, put Fathom in deps/Fathom before building. Without it the engine builds fine, just without tablebases.

Running

Ruthless only speaks UCI, so use a GUI such as Cute Chess, Nibbler or Banksia. You can also talk to it directly:

./target/release/Ruthless
uci
position startpos moves e2e4 e7e5
go movetime 1000

Ruthless bench runs a fixed-depth benchmark.

Options: Hash, Threads, SyzygyPath, NumaBind, plus a handful of search parameters (RfpBase, NmpBase, LmrDivisor, ...) that I use for tuning.

License

MIT

_{net trained on master-binpacks_relabel (maybe switch to my own datagen later).}