Catalyst

A UCI-compliant chess engine written in C++20, featuring NNUE evaluation and advanced search.

Catalyst uses a (768 → 256)×2 → 8 neural network with incremental accumulator updates and SIMD-accelerated inference (SSE4.1 / AVX2 / AVX-512), combined with a robust search implementation including PVS, Lazy SMP, and an extensive suite of pruning, extension, and reduction techniques.

---

Strength

Official CCRL ratings for each release are listed below.

Version	CCRL 40/15	CCRL 2+1
v1.0.0	-	3080
v2.0.0	-	-
v2.1.0	3161	-
v2.2.0	3222	-
v3.0.0	-	3332

---

Features

Search

• Principal Variation Search (PVS) with iterative deepening
• Quiescence search
• Aspiration windows
• Transposition table with aging, rule50 counter, and huge page support
• Pruning
  • Reverse futility pruning
  • Null move pruning with verification
  • Razoring
  • Futility pruning
  • Capture futility pruning
  • Late move pruning (LMP)
  • SEE pruning (quiets and noisy moves)
  • History pruning
  • ProbCut
  • Small ProbCut (early TT-based return when fail-high margin is large)
• Extensions
  • Singular extensions
  • Double and triple extensions
  • Negative extensions
  • Check extensions
• Reductions
  • Late move reductions (LMR) with history-based adjustments
  • Internal iterative reduction (IIR)
  • Alpha-raise depth reduction
• Move Ordering
  • TT move
  • Staged move picker (good captures → killers → countermove → quiets → bad captures)
  • MVV-LVA for captures
  • Threat-based quiet move scoring with precomputed opponent attack maps
  • Threat escape bonus/malus for quiet moves (queen, rook, minor pieces)
  • Continuation history–weighted quiet move ordering (1-ply × 2 > 2-ply > 3-ply > 4-ply)
  • Dynamic SEE thresholds for capture classification based on move score
  • Bad captures ordered by SEE loss (least-losing first)
  • Killer move heuristic (2 per ply)
  • Countermove heuristic
  • Butterfly history
  • Capture history
  • Pawn history
  • 1-ply, 2-ply, 3-ply, and 4-ply continuation history
• History
  • Butterfly history (threat-indexed)
  • Capture history (threat-indexed)
  • Pawn history
  • Continuation history (1-ply, 2-ply, 3-ply, 4-ply)
  • Correction history (main, pawn, non-pawn white, non-pawn black, continuation)
  • Eval history (applies malus to opponent's quiet move when our eval improves)
• Miscellaneous
  • Mate distance pruning
  • Hindsight depth adjustment
  • Shuffling detection in singular extension
  • ttPv propagation on fail-low
  • TT move bonus/malus on cutoff and fail-low
  • Fifty-move rule eval scaling
  • Draw score randomization (anti-repetition)
  • Stalemate avoidance at root
  • Lazy SMP (multi-threaded search)

Evaluation

• NNUE
  • Architecture: (768 → 256)×2 → 8 (8 material-count output buckets)
  • Incremental accumulator updates
  • SIMD-accelerated inference (SSE4.1 / AVX2 / AVX-512)
  • Embedded network (catalyst-v2.nnue)
  • SCReLU activation
  • Correction history applied on top of raw NNUE score

Time Management

• Soft and hard time limits
• Best-move stability scaling (less time when best move is stable)
• Score instability scaling (more time when eval is volatile)
• Node fraction scaling (more time when best-move node fraction is low)
• Complexity estimate scaling
• Pondering support (go ponder / ponderhit)

---

UCI Options

Name	Type	Default	Valid values	Description
Hash	integer	64	[1, 65536]	Transposition table size in MiB.
Clear Hash	button	—	—	Clears the transposition table.
Threads	integer	1	[1, hardware max]	Number of search threads (Lazy SMP).
Move Overhead	integer	50	[0, 5000]	Time overhead per move in ms.
Ponder	check	false	true, false	Enable pondering.
EvalFile	string	catalyst.nnue	any path	External NNUE file to load (overrides embedded).

---

Non-standard Commands

Command	Description
d	Display the current board position.
eval	Print NNUE evaluation for the current position.
perft <depth>	Run a perft test from the current position.
bench [depth <n>] [threads <n>]	Run a benchmark. Default depth: 13.
datagen [output <file>] [threads <n>] [nodes <n>] [games <n>] [book <file>]	Generate training data.

---

Builds

Choose the binary that matches your CPU's highest supported instruction set:

Binary	Requirements	Notes
avx512vnni	AVX-512 + VNNI (Cascade Lake, Zen 4+)	Fastest — use if supported
avx512	AVX-512 + BMI2 (Ice Lake, Rocket Lake+)
bmi2	AVX2 + BMI2 (Intel Haswell+, AMD Zen 3+)	Recommended for Intel Haswell+ and AMD Zen 3+
avx2	AVX2 (Broadwell+, AMD Excavator+)	Use for AMD Zen 1/2 or older Intel
sse41	SSE4.1 (Core 2 Penryn+, Phenom II+)	Fallback for pre-AVX2 CPUs
x86-64	x86-64 + POPCNT	Widest compatibility, slowest

AMD Zen 1 / Zen 2 users: use the avx2 build even if your CPU supports BMI2. These CPUs implement pext/pdep in microcode, making them very slow for Catalyst's purposes.

---

Building from Source

Requires make, a C++20 compiler (GCC ≥ 13 or Clang ≥ 16), and objcopy. The Makefile will automatically download the NNUE file and embed it into the binary.

# Clone the repository
git clone https://github.com/AnanyTanwar/Catalyst
cd Catalyst

# Build optimised for the host CPU (recommended for local use)
make native

# Build a specific architecture
make linux-x86-64
make linux-sse41
make linux-avx2
make linux-bmi2
make linux-avx512
make linux-avx512vnni

# Build all Linux release binaries
make release-linux

# Build all Windows release binaries (requires MinGW cross-compiler)
make release-win

# Build with PGO (profile-guided optimisation)
make pgo                  # native CPU
make pgo ARCH=avx2
make pgo ARCH=bmi2
make pgo ARCH=sse41
make pgo ARCH=avx512

# Install to system
make install              # installs to /usr/local/bin
make install PREFIX=/usr  # custom prefix

# Debug build
make debug
make sanitize             # debug + ASan + UBSan
make debug SANITIZE=-fsanitize=address,undefined

# Clean build artifacts
make clean

# Clean build artifacts and downloaded NNUE file
make distclean

All binaries are placed in bin/.

---

NNUE

Catalyst's network is stored at CatalystNet. The Makefile fetches it automatically at build time and embeds it into the binary via objcopy, so no external .nnue file is needed at runtime.

You can override the embedded network at runtime using the EvalFile UCI option.

---

License

Catalyst is free software distributed under the GNU General Public License v3.0.

---

Credits

Special thanks to the Stockfish Discord community for their invaluable help with debugging, NNUE guidance, and overall support during Catalyst's development.

Catalyst would not exist without the broader chess programming community. In no particular order, these engines and projects were notable sources of ideas and inspiration:

• Stockfish
• Stormphrax
• Alexandria
• Tarnished
• Integral
• Obsidian
• bullet — NNUE trainer