Store NNUE accumulators in the stack

[bdmendes]

Summary by CodeRabbit

• Refactor
  • Optimized neural network evaluation efficiency through improved memory management and simplified indexing logic.

[coderabbitai]

Walkthrough

Modified the neural network evaluator to replace heap-allocated accumulator vectors with a fixed-size stack array and simplified indexing by removing the perspective_index field, replaced with direct perspective-to-index casting. Updated imports accordingly.

Changes

Cohort / File(s)	Summary
Neural Network Accumulator Optimization src/evaluation/nnue.rs	Replaced [Vec<f64>; 2] accumulator with [[f64; HIDDEN_LAYER_SIZE]; 2] stack array; removed perspective_index field and replaced all usages with perspective as usize indexing in set, clear, forward, forward_and_cache, and evaluate_unscaled methods; adjusted imports by removing std::array.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~15 minutes

Possibly related PRs

• Migrate to storch and use new exported model #315 — Overlapping modifications to src/evaluation/nnue.rs affecting accumulator indexing and the set/clear/forward code paths.
• Remove NNUE score caching and Option boxing #324 — Changes to accumulator representation and initialization logic in the same neural network evaluator file.

Poem

🐰 A rabbit hops through stack allocation land,
No more heaps bouncing far and wide!
Perspective casting, hand in paw so grand,
Fixed-size arrays now our guide,
NNUE optimization with pride! ✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'Store NNUE accumulators in the stack' directly and accurately summarizes the main change: replacing heap-allocated accumulator vectors with a fixed-size stack array.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

src/evaluation/nnue.rs (1)

131-166: 🧹 Nitpick | 🔵 Trivial

Encapsulate the perspective-to-index conversion to prevent silent reordering bugs.

The code uses perspective as usize directly to index two-element arrays (acc, last_seen). This works only because Color::White = 0 and Color::Black = 1 (enforced by primitive_enum!), but an explicit conversion helper makes the invariant local and exhaustively checked. This pattern appears throughout the codebase; consider applying the same pattern here as a starting point.

Proposed fix

 impl NeuralNetwork {
+    fn perspective_index(perspective: Perspective) -> usize {
+        match perspective {
+            Color::White => 0,
+            Color::Black => 1,
+        }
+    }
+
     fn set(&mut self, piece: Piece, color: Color, square: Square, perspective: Perspective) {
         let input_idx = Self::input_index(
             piece,
             Self::transformed_color(color, perspective),
             Self::transformed_square(square, perspective),
         );
+        let perspective_idx = Self::perspective_index(perspective);
         for i in 0..HIDDEN_LAYER_SIZE {
-            self.acc[perspective as usize][i] += self.params.acc_weights[i * INPUT_SIZE + input_idx];
+            self.acc[perspective_idx][i] += self.params.acc_weights[i * INPUT_SIZE + input_idx];
         }
     }
 
     fn clear(&mut self, piece: Piece, color: Color, square: Square, perspective: Perspective) {
         let input_idx = Self::input_index(
             piece,
             Self::transformed_color(color, perspective),
             Self::transformed_square(square, perspective),
         );
+        let perspective_idx = Self::perspective_index(perspective);
         for i in 0..HIDDEN_LAYER_SIZE {
-            self.acc[perspective as usize][i] -= self.params.acc_weights[i * INPUT_SIZE + input_idx];
+            self.acc[perspective_idx][i] -= self.params.acc_weights[i * INPUT_SIZE + input_idx];
         }
     }
 
     fn forward(&self, perspective: Perspective) -> f64 {
         let mut eval: f64 = 0.0;
+        let perspective_idx = Self::perspective_index(perspective);
 
         for i in 0..HIDDEN_LAYER_SIZE {
-            let hidden_out = Self::relu(self.acc[perspective as usize][i] + self.params.acc_biases[i]);
+            let hidden_out = Self::relu(self.acc[perspective_idx][i] + self.params.acc_biases[i]);
             eval += hidden_out * self.params.out_weights[i];
         }
 
         eval + self.params.out_bias
@@
     fn forward_and_cache(&mut self, position: &Position) -> f64 {
         let perspective = position.side_to_move();
         let res = self.forward(perspective);
-        self.last_seen[perspective as usize] = *position;
+        self.last_seen[Self::perspective_index(perspective)] = *position;
         res
     }
 
     fn evaluate_unscaled(&mut self, position: &Position) -> f64 {
         let perspective = position.side_to_move();
-        let diff = position.diff(&self.last_seen[perspective as usize]);
+        let diff = position.diff(&self.last_seen[Self::perspective_index(perspective)]);

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: e34dfefd-8cd8-4e39-86d3-7807411a1f80

📥 Commits

Reviewing files that changed from the base of the PR and between 4d4486c and e4f37a3.

📒 Files selected for processing (1)

• src/evaluation/nnue.rs

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 91%. Comparing base (4d4486c) to head (e4f37a3).
⚠️ Report is 2 commits behind head on master.

Additional details and impacted files

@@          Coverage Diff          @@
##           master   #443   +/-   ##
=====================================
- Coverage      92%    91%   -0%     
=====================================
  Files          33     33           
  Lines        3760   3754    -6     
  Branches     3760   3754    -6     
=====================================
- Hits         3422   3416    -6     
  Misses        319    319           
  Partials       19     19           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Against master [hash=64]: 🆗 Elo: 2.78 +/- 15.98, nElo: 3.75 +/- 21.53

[github-actions]

Against v1.6.0 [hash=64]: ✅ Elo: 77.34 +/- 19.23, nElo: 89.60 +/- 21.53