SirioC: Architecture Analysis of a Modern Chess Engine Based on C++20

SirioC is an open-source UCI chess engine developed by Jorge Ruiz Centelles, implemented in C++20, and characterized by its clear modular architecture. This article will analyze its core mechanisms (bitboard, Alpha-Beta search, NNUE neural network evaluation, etc.) and discuss its value as a reference for learning modern engines.

SirioC was inspired by the Rustic Chess educational path, aiming to build a clear and modular engine foundation. Unlike commercial engines that pursue extreme Elo ratings, it focuses more on code readability and auditability. The project is open-sourced under the MIT license and follows the principle of "learning architecture is allowed, but copying external source code is not". Its name pays tribute to the Dogon people of Burkina Faso and Mali, reflecting respect for African culture, while representing the concept of "documented technical inheritance": public technologies can be learned, but code independence and traceability must be maintained.

Bitboard Representation

Uses 64-bit integer bitboards, leverages CPU bitwise operations to improve the efficiency of move generation and attack detection, supports full FEN parsing and serialization, distinguishes between pseudo-legal/legal moves, and verifies the legality of castling and en passant captures.

Search Algorithms

Based on the iterative deepening Negamax framework + Alpha-Beta pruning, including strategies such as quiescence search (to avoid the horizon effect), null move pruning, late move reduction (LMR), futility pruning, and aspiration windows; move ordering optimization uses transposition table best move first, killer moves, and quiet move history heuristics.

Classic Evaluation Function

Retains handcrafted evaluation as a fallback, covering material position scoring, mobility, king safety, pawn structure analysis, bishop pair bonus, etc., and uses tapered scoring to dynamically adjust midgame/endgame weights.

Implements Lazy SMP parallel search, with multiple threads sharing a transposition table, and thread count configurable via UCI Threads; integrates Syzygy endgame tablebases (detected via the Fathom library), providing perfect endgame information for up to 7 pieces, and giving precise checkmate/stalemate move counts.

The proprietary sparse neural network evaluation path SirioNNUE2 is under development, emphasizing design from scratch: sparse feature representation, deterministic feature index contract, perspective-aware feature state, accumulator-oriented inference, etc. It focuses on engineering ethics—validation of encoders, loaders, etc., must be completed before it becomes the default backend. Each network is required to provide a MODEL_CARD.md, and datasets must provide a DATASET.md (including checksums, training configurations, etc.).

Follows professional software engineering best practices: behavior-preserving refactoring (refactor before optimizing), test-first (establish tests to ensure correct modifications), small-step merging (avoid large-scale risks), and transparency (no silent changes). Includes a complete time manager (adaptive allocation, move cost compensation), opening book support, UCI protocol implementation, unit tests, and reproducible benchmark tests.

Compared to complex codebases like Stockfish, SirioC's modular design makes component responsibilities clearer, making it suitable for understanding modern engine principles (bitwise operation techniques, Alpha-Beta pruning, transposition table design, etc.). It also demonstrates the boundary between "learning public technologies" and "copying code" in the open-source ecosystem, which has exemplary significance for the healthy development of the community.

SirioC adheres to "clear architecture first", not pursuing the strongest engine but becoming an easy-to-understand, easy-to-audit, and easy-to-learn reference. With the advancement of SirioNNUE2, it is expected to improve chess strength while maintaining code quality. It is recommended for AI algorithm researchers, C++ performance optimization enthusiasts, and those interested in game tree search to pay attention and learn.