⚡ Bolt: Optimize RANSAC RNG and rounding in tight loops

[suraj-ranganath]

💡 What: Optimized Fisher-Yates shuffle implementations in rand_sample and rand_permutation by vectorizing random number generation and using faster scalar rounding.

🎯 Why: Frequent scalar calls to stream.rand() and round_mat() in tight loops were a performance bottleneck.

📊 Impact: Approximately 25-40% speedup for permutation and sampling operations (e.g., rand_permutation for n=100k reduced from ~1.4s to ~1.0s).

🔬 Measurement: Verified with tests/test_parity_rng.py for correctness and custom benchmarks for performance.

---

PR created automatically by Jules for task 13896802785975683596 started by @suraj-ranganath

[google-labs-jules]

👋 Jules, reporting for duty! I'm here to lend a hand with this pull request.

When you start a review, I'll add a 👀 emoji to each comment to let you know I've read it. I'll focus on feedback directed at me and will do my best to stay out of conversations between you and other bots or reviewers to keep the noise down.

I'll push a commit with your requested changes shortly after. Please note there might be a delay between these steps, but rest assured I'm on the job!

For more direct control, you can switch me to Reactive Mode. When this mode is on, I will only act on comments where you specifically mention me with @jules. You can find this option in the Pull Request section of your global Jules UI settings. You can always switch back!

New to Jules? Learn more at jules.google/docs.

---

For security, I will only act on instructions from the user who triggered this task.

[claude]

Claude finished @suraj-ranganath's task in 1m 50s —— View job

---

PR Review: Optimize RANSAC RNG and rounding

• Read the diff against origin/develop
• Verify RNG-vectorization parity
• Verify rounding parity (round_mat → math.floor)
• Check edge cases and the added .jules/bolt.md

Verdict

The core optimization is correct and parity-preserving. Two things worth addressing — one likely-unintended file, one minor edge case.

Parity analysis (looks good ✅)

1. Vectorized RNG — stream.rand(m) draws m values sequentially from the same MT19937 stream, producing the identical sequence to m successive scalar stream.rand() calls. Because the loop consumes rands[k] in order, the RNG stream consumption is unchanged, so MATLAB parity (the whole point of these functions) is preserved.
2. Rounding — for decimals == 0, round_mat(x) = copysign(floor(abs(x) + 0.5), x). The rounded inputs here are always non-negative ((remaining - 1) * rands[k] with remaining ≥ 1, and k * rands[i] with k ≥ 1), so copysign(...) is a no-op and round_mat(x) == math.floor(x + 0.5) exactly. The # Faster scalar rounding for non-negative values comment correctly scopes this — good that the non-negativity assumption is documented.

So both micro-optimizations are behavior-preserving on the actual input domain. 👍

Issues

1. .jules/bolt.md appears to be an unintended scratch/notes file (+3/-0)

This is an agent learning-log, not part of the functional change, and it isn't covered by .gitignore. It doesn't belong in the package source tree. Recommend either removing it from the PR or adding .jules/ to .gitignore.

Fix this →

2. Minor edge-case regression for n = 0 in rand_permutation (ransac.py:93)

rands = stream.rand(n - 1) is evaluated before the loop. For n = 0 this becomes stream.rand(-1), which raises ValueError: negative dimensions are not allowed. The previous code never called rand() for n = 0 (the range(-1, 0, -1) loop is empty), so it returned an empty array without error. n = 1 is fine (rand(0) → empty array). rand_sample is unaffected since m ≥ 0 always.

This is almost certainly unreachable in practice (these are ICA/RANSAC channel counts), so it's a low-priority nit — but if you want to fully preserve the old contract, guard with n > 1:

rands = stream.rand(n - 1) if n > 1 else np.empty(0)

Nits

• rand_permutation docstring still says "by using rand() + round_mat()" — now uses math.floor. Stale but harmless.
• The import math / removed round_mat import are clean; no remaining round_mat uses in the file. ✅

Testing

The parity tests in tests/test_parity_rng.py exercise rand_sample/round_mat and require MATLAB, so they'll be skipped in CI without an engine. Since the RNG-consumption order and rounding domain are unchanged, parity is preserved by construction — but it's worth confirming the existing RANSAC/clean_rawdata tests still pass locally (uv run pytest -m "not slow" -k ransac or the relevant clean_rawdata tests).
· branch bolt/optimize-ransac-rng-rounding-13896802785975683596