⚡ Bolt: [성능 개선] 오디오 섹션 분할에서의 행렬 대각선 패치 추출 최적화

.jules/bolt.md

@@ -61,3 +61,6 @@
 ## 2026-07-13 - Array.from mapping optimization
 **Learning:** Using `Array.from({ length: N }).map(...)` creates an intermediate array of `undefined` values which requires memory allocation and garbage collection, adding O(N) unnecessary overhead in frequently re-rendered UI components.
 **Action:** Use `Array.from({ length: N }, (_, index) => ...)` to map elements directly during array creation, avoiding intermediate allocations.
+## 2024-05-20 - [Performance Optimization] Vectorizing diagonal patch extraction in Checkerboard Novelty Kernel
+**Learning:** When performing windowed operations strictly along the diagonal of large matrices in Python/NumPy, using Python array slicing loops incurs linear time constant overhead due to inner loop array allocation and summation.
+**Action:** Use sub-matrix diagonal vectorization (`numpy.lib.stride_tricks.sliding_window_view` coupled with `np.diagonal` and `np.einsum`) instead of nested Python loops to shift heavy computations to fast C-level operations.

services/analysis-engine/src/bandscope_analysis/sections/segmenter.py

@@ -135,15 +135,11 @@ def _checkerboard_novelty_reference(
     kernel[half:, half:] = 1.0
 
     # Sum each checkerboard offset across all valid diagonal windows at once.
-    valid = novelty[half : n - half]
-    for di in range(-half, half):
-        for dj in range(-half, half):
-            value = kernel[di + half, dj + half]
-            diagonal = np.diagonal(ssm[half + di : n - half + di, half + dj : n - half + dj])
-            if value > 0:
-                valid += diagonal
-            else:
-                valid -= diagonal
+    # Uses sliding_window_view and einsum to avoid nested python loops and
+    # minimize array allocation overhead.
+    windows = np.lib.stride_tricks.sliding_window_view(ssm, (kernel_size, kernel_size))
+    diag_windows = np.diagonal(windows, axis1=0, axis2=1)[:, :, : n - 2 * half]
+    novelty[half : n - half] = np.einsum("ijk,ij->k", diag_windows, kernel)
 
     # Normalize by peak absolute magnitude, preserving sign.
     max_val = np.max(np.abs(novelty))