Prelude Chess Engine

A superhuman UCI chess engine with NNUE evaluation

Prelude is a UCI-compatible chess engine utilizing advanced search* algorithms and a neural network (NNUE) for evaluation. Designed for high performance** and configurability, Prelude is not ideal for chess enthusiasts and developers.

*depends on a person's definition of advanced
**performance is unlikely

Feature List:

CLI Commands:

1. Functional UCI implementation with custom commands:
   • d: Display the current board position.
   • move <move>: Applies a move in long algebraic notation.
   • bulk <depth>: Starts a perft test from the current position using semi-bulk counting.
   • perft <depth>: Performs a perft test from the current position.
   • perftsuite <suite>: Executes a suite of perft tests with multithreading.
   • bench <depth>: Benchmarks engine performance on test positions.
   • datagen <threads>: Starts datagen with the given number of threads.
   • position kiwipete: Loads the "Kiwipete" position, commonly used for debugging.
   • Some other commands are supported, but are mostly for debugging. See Prelude.cpp for the full list.

Evaluation:

1. NNUE evaluation trained with bullet
2. Uses QA of 255, QB of 64, and eval scale of 400
3. SCReLU activation function
4. (768->1024)x2->1x8

More information can be found in config.h

Build From Source

Prerequisites

• Make: Uses make to build the executable.
• Clang++ Compiler: Requires support for C++20 or later.
• CPU Architecture: AVX2 or later recommended.
• Neural Network File: Provide an NNUE and update the code to load it. If no network is given, it will download the latest Prelude network.

Compilation

1. Clone the repository:

   git clone https://github.com/Quinniboi10/Prelude.git
   cd Prelude
   

2. Compile using make:

   make
   

3. Run the engine:

   ./Prelude
   

Usage

Prelude uses the UCI protocol but supports custom debugging and testing commands. Compatible with GUIs like Cute Chess, En Croissant, or any UCI-compatible interface. See above commands.

Configuration

Prelude supports customizable options via the setoption command:

• Threads: Number of threads to use (1 to 1024). Default: 1.
• Hash: Configurable hash table size (1 to 4096 MB). Default: 16 MB.
• Move Overhead: Adjusts time overhead per move (0 to 1000 ms). Default: 20 ms.
• EvalFile: Path to the NNUE file. Default: internal.
• SyzygyPath: Path to the Syzygy tablebases. Default:
• SyzygyProbeDepth: Minimal internal search depth to probe the TBs. Default: 1
• SyzygyProbeLimit: Max number of pieces to probe the TBs for. Default: 32*
• UCI_Chess960: Represents moves as king takes rook. Default: false.
• Softnodes: Only check if the node count has been reached after completing a depth. Default: false

*Currently the largest TB is only 7 pieces

Neural Network

Ensure a NNUE file is placed relative to this file, as a .nnue will be if you build with make.

When building run:

make EVALFILE=<path to your file>

In the case of a different architecture, modifications may be made to the NNUE section of the config:

In config.h:

// ************ NNUE ************
constexpr i16    QA             = 255;
constexpr i16    QB             = 64;
constexpr i16    EVAL_SCALE     = 400;
...

Special Thanks

• Swedishchef: Teaching me about CPUs, high performance programming, and much more
• Dan: Allowing me to use his OpenBench instance while MattBench was down
• Vast: Help hunting for bugs and explaining concepts when I was new
• Fury: Integral is a big source of inspiration, and the fantastic comments are a great source
• Ciekce: Lots of guidance and test NNUEs
• Shawn_xu: Explaining NNUEs and other common concepts
• A_randomnoob: Hunting for bugs and explaining concepts
• Matt: Providing help and allowing me to use his git instance, as well as sharing OpenBench
• Kan: Making the Prelude TCEC logo
• jw: Helping me with NNUE training and developer of bullet
• Cosmo: Creator of Viri binpacks, and the person who added them to bullet
• All other members of the Stockfish discord: So many people who I didn't name have helped with this project, and even those who I did mention by name certainly did more than I listed here. Thanks to everyone who contributed!