Skip to main content

Crate nnue_rs

Crate nnue_rs 

Source
Expand description

Load and evaluate NNUE (Efficiently Updatable Neural Network) chess networks.

The crate is small, dependency-free and cross-platform, with an optional AVX2 fast path (used automatically at runtime when available, with a scalar fallback everywhere else).

§Architectures

The architecture is detected automatically from the network file’s header:

  • SFNNv10 — Stockfish 18 big nets with threat inputs.
  • HalfKAv2_hm — Stockfish SFNNv5-v9 (Stockfish 16/17 and SF 18 small nets).
  • HalfKAv2 — the non-mirrored predecessor (Stockfish SFNNv2-v4, SF 14).
  • HalfKP — the classic Stockfish NNUE feature set (Stockfish 12-14 nets).

§Integration

Feed positions either by FEN or by implementing the Board trait on your own position type.

By FEN:

use nnue_rs::Network;

let net = Network::from_file("net.nnue")?;
let cp = net.evaluate_fen("rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1")?;

By implementing Board:

use nnue_rs::{Board, Color, Piece, Network};

impl Board for MyPosition {
    fn side_to_move(&self) -> Color { /* ... */ todo!() }
    fn king_square(&self, color: Color) -> u8 { /* ... */ todo!() }
    fn for_each_piece(&self, f: &mut dyn FnMut(u8, Piece)) { /* ... */ todo!() }
}

let score = net.evaluate(&position);

§Incremental evaluation

For a search, advance an Accumulator as moves are made instead of recomputing from scratch. See Network::accumulator, Network::update (or Network::update_changes for known piece changes) and Network::evaluate_accumulator. Common updates use fused arithmetic; HalfKAv2_hm and SFNNv10 inference uses a packed first layer on AVX2 CPUs. Public APIs and network-file formats are unchanged from 0.4.1.

Structs§

Accumulator
A pair of feature-transformer accumulators (one per side-to-move perspective) plus the piece count for the position they describe.
FenBoard
A minimal board built from a FEN string, implementing Board.
Network
A loaded NNUE network.
Piece
A colored chess piece.

Enums§

Arch
The feature set (input layer) an NNUE network is built on.
Color
Side of a piece or the side to move.
Error
Errors returned when loading a network or parsing a position.
PieceKind
Type of a chess piece, color-independent.

Traits§

Board
A chess position, as seen by the network.

Type Aliases§

Result
Convenience alias for results returned by this crate.