[1/6] Improve PSQT evaluation performance

[luccabb]

Summary

• Use Zobrist hash instead of FEN string for evaluation cache keys (faster integer hashing vs string comparison)
• Add cache size limit (1M entries) to prevent unbounded memory growth during long searches
• Optimize board_evaluation() to iterate over actual pieces (~16-32) instead of all 64 squares

Details

The board evaluation function is called frequently during search. These optimizations reduce overhead:

1. Zobrist hashing: chess.polyglot.zobrist_hash() returns an integer, which is faster to hash and compare than FEN strings
2. Cache eviction: Simple clear-on-full policy prevents memory exhaustion in long games
3. Piece iteration: Using board.pieces(piece_type, color) returns only occupied squares, reducing iterations by 50-75%

Test plan

• All 64 unit tests pass
• No changes to search behavior, only evaluation speed

---

🤖 Generated with Claude Code

[github-actions]

🔬 Stockfish Benchmark Results

vs Stockfish Skill Level 3

Metric	Wins	Losses	Draws	Total	Win %
Overall	32	60	8	100	32.0%
As White	15	31	4	50	30.0%
As Black	17	29	4	50	34.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 6

vs Stockfish Skill Level 4

Metric	Wins	Losses	Draws	Total	Win %
Overall	25	70	5	100	25.0%
As White	11	37	2	50	22.0%
As Black	14	33	3	50	28.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 5

vs Stockfish Skill Level 5

Metric	Wins	Losses	Draws	Total	Win %
Overall	9	83	8	100	9.0%
As White	6	39	5	50	12.0%
As Black	3	44	3	50	6.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 5

Configuration

• 5 chunks × 20 rounds × 3 skill levels = 300 total games
• Each opening played with colors reversed (-repeat) for fairness
• Moonfish: 60s per move
• Stockfish: 60+5 time control

[greptile-apps]

Greptile Overview

Greptile Summary

This PR optimizes the board evaluation performance through three key improvements:

• Replaced FEN string-based cache keys with Zobrist hash integers for faster lookups
• Added 1M entry cache limit with clear-on-full eviction to prevent memory exhaustion
• Refactored evaluation loop to iterate over actual pieces (16-32) instead of all 64 squares

The changes are well-implemented and maintain evaluation correctness. The optimizations reduce computational overhead without altering search behavior. The cache eviction strategy is simple but effective for preventing unbounded memory growth during extended searches.

Confidence Score: 5/5

• Safe to merge - performance optimizations with no behavioral changes
• All optimizations are sound: Zobrist hashing is standard in chess engines, piece iteration reduces unnecessary work, and cache limiting prevents memory issues. No logical errors or breaking changes detected.
• No files require special attention

Important Files Changed

Filename	Overview
moonfish/psqt.py	Optimized evaluation caching with Zobrist hashing and piece iteration; no logical issues found

Sequence Diagram

sequenceDiagram
    participant Engine as Alpha-Beta Engine
    participant Cache as board_evaluation_cache
    participant Eval as board_evaluation
    participant Board as chess.Board
    
    Engine->>Cache: call board_evaluation(board)
    Cache->>Board: zobrist_hash(board)
    Board-->>Cache: int hash key
    
    alt Cache hit
        Cache-->>Engine: return cached value
    else Cache miss
        alt Cache full (≥1M entries)
            Cache->>Cache: clear entire cache
        end
        Cache->>Eval: call board_evaluation(board)
        Eval->>Eval: get_phase(board)
        loop For each piece type
            Eval->>Board: pieces(piece_type, WHITE)
            Board-->>Eval: SquareSet of white pieces
            loop For each white piece
                Eval->>Eval: mg_white += table[square^56] + value
                Eval->>Eval: eg_white += table[square^56] + value
            end
            Eval->>Board: pieces(piece_type, BLACK)
            Board-->>Eval: SquareSet of black pieces
            loop For each black piece
                Eval->>Eval: mg_black += table[square] + value
                Eval->>Eval: eg_black += table[square] + value
            end
        end
        Eval->>Eval: calculate tapered eval based on phase
        Eval-->>Cache: evaluation score
        Cache->>Cache: store in cache[hash]
        Cache-->>Engine: return evaluation score
    end

Loading

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}

[github-actions]

🔬 Stockfish Benchmark Results

vs Stockfish Skill Level 3

Metric	Wins	Losses	Draws	Total	Win %
Overall	27	60	13	100	27.0%
As White	19	23	8	50	38.0%
As Black	8	37	5	50	16.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 11

vs Stockfish Skill Level 4

Metric	Wins	Losses	Draws	Total	Win %
Overall	17	77	6	100	17.0%
As White	11	35	4	50	22.0%
As Black	6	42	2	50	12.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 6

vs Stockfish Skill Level 5

Metric	Wins	Losses	Draws	Total	Win %
Overall	8	89	3	100	8.0%
As White	3	45	2	50	6.0%
As Black	5	44	1	50	10.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 2
• Draw by insufficient mating material: 1

Configuration

• 5 chunks × 20 rounds × 3 skill levels = 300 total games
• Each opening played with colors reversed (-repeat) for fairness
• Moonfish: 60s per move
• Stockfish: 60+5 time control

[github-actions]

🔬 Stockfish Benchmark Results

vs Stockfish Skill Level 3

Metric	Wins	Losses	Draws	Total	Win %
Overall	27	62	11	100	27.0%
As White	15	29	6	50	30.0%
As Black	12	33	5	50	24.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 7
• Draw by insufficient mating material: 1
• Draw by fifty moves rule: 1

vs Stockfish Skill Level 4

Metric	Wins	Losses	Draws	Total	Win %
Overall	26	65	9	100	26.0%
As White	18	26	6	50	36.0%
As Black	8	39	3	50	16.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 9

vs Stockfish Skill Level 5

Metric	Wins	Losses	Draws	Total	Win %
Overall	8	87	5	100	8.0%
As White	4	44	2	50	8.0%
As Black	4	43	3	50	8.0%

Non-checkmate endings:

• Draw by 3-fold repetition: 4

Configuration

• 5 chunks × 20 rounds × 3 skill levels = 300 total games
• Each opening played with colors reversed (-repeat) for fairness
• Moonfish: 60s per move
• Stockfish: 60+5 time control

[luccabb]

/run-nps-benchmark

[github-actions]

⚡ NPS Benchmark Results

Metric	Value
Depth	5
Positions	48
Total nodes	20949384
Total time	2321.35s
Nodes/second	9024

Node count is the primary signal — it's deterministic and catches search behavior changes. If the node count changes, the PR changed search behavior. NPS is informational only (CI runner performance varies).

Per-position breakdown

Position  1/48: nodes=155074     time=12.73s  nps=12177
Position  2/48: nodes=606593     time=61.64s  nps=9840
Position  3/48: nodes=13225      time=0.89s  nps=14846
Position  4/48: nodes=750165     time=81.23s  nps=9235
Position  5/48: nodes=109167     time=13.25s  nps=8241
Position  6/48: nodes=500245     time=56.34s  nps=8879
Position  7/48: nodes=283405     time=31.49s  nps=8999
Position  8/48: nodes=344459     time=28.16s  nps=12230
Position  9/48: nodes=2113446    time=211.09s  nps=10011
Position 10/48: nodes=408907     time=33.82s  nps=12089
Position 11/48: nodes=516720     time=60.90s  nps=8485
Position 12/48: nodes=1134648    time=140.22s  nps=8091
Position 13/48: nodes=515044     time=56.92s  nps=9049
Position 14/48: nodes=838712     time=84.27s  nps=9952
Position 15/48: nodes=982014     time=101.99s  nps=9628
Position 16/48: nodes=315039     time=30.18s  nps=10439
Position 17/48: nodes=10112      time=0.67s  nps=15024
Position 18/48: nodes=15323      time=0.78s  nps=19666
Position 19/48: nodes=48718      time=3.87s  nps=12572
Position 20/48: nodes=89369      time=6.59s  nps=13561
Position 21/48: nodes=18259      time=1.15s  nps=15914
Position 22/48: nodes=662        time=0.03s  nps=19393
Position 23/48: nodes=11402      time=0.68s  nps=16715
Position 24/48: nodes=25010      time=1.74s  nps=14343
Position 25/48: nodes=8781       time=0.52s  nps=16972
Position 26/48: nodes=60646      time=5.13s  nps=11812
Position 27/48: nodes=74122      time=5.13s  nps=14458
Position 28/48: nodes=258609     time=23.33s  nps=11083
Position 29/48: nodes=301542     time=32.24s  nps=9353
Position 30/48: nodes=2451       time=0.19s  nps=13212
Position 31/48: nodes=1431889    time=148.15s  nps=9665
Position 32/48: nodes=854162     time=78.23s  nps=10918
Position 33/48: nodes=2479630    time=417.18s  nps=5943
Position 34/48: nodes=1469013    time=208.57s  nps=7043
Position 35/48: nodes=307495     time=25.95s  nps=11849
Position 36/48: nodes=1770804    time=171.99s  nps=10295
Position 37/48: nodes=1289243    time=114.63s  nps=11247
Position 38/48: nodes=14229      time=0.67s  nps=21387
Position 39/48: nodes=7885       time=0.29s  nps=27030
Position 40/48: nodes=23313      time=0.39s  nps=59040
Position 41/48: nodes=87269      time=6.34s  nps=13756
Position 42/48: nodes=81449      time=5.85s  nps=13911
Position 43/48: nodes=29486      time=2.05s  nps=14415
Position 44/48: nodes=125063     time=11.13s  nps=11240
Position 45/48: nodes=46664      time=3.71s  nps=12571
Position 46/48: nodes=419921     time=39.04s  nps=10755
Position 47/48: nodes=0          time=0.00s  nps=0  (terminal)
Position 48/48: nodes=0          time=0.00s  nps=0  (terminal)