compute_vecs_l2sq: Replace scalar L2 Squared norm with SIMD-optimized FastL2NormSquared

[arrayka]

Changes

• diskann-disk/src/utils/math_util.rs:
  • Replaced the private compute_vec_l2sq() helper function with FastL2NormSquared.evaluate(chunk)
• [Unrelated improvements] kmeans_bench.rs (criterion):
  • Switched from iter() to iter_batched with BatchSize::SmallInput to avoid measuring setup costs (clone, allocation)
  • Inlined snrm2_benchmark_rust() helper into the benchmark closure
• [Unrelated improvements] kmeans_bench_iai.rs (iai-callgrind):
  • Moved thread pool creation into setup_data() so it's excluded from measurement
  • Removed unnecessary data.clone() calls, since there are no multiple loops (the benchmark is executed only once)
  • Added black_box to prevent dead-code elimination of outputs

Performance

iai-callgrind shows 60% CPU cost reduction (see Estimated Cycles metric) after switching to FastL2NormSquared.evaluate(chunk).
The measurement was taken in a single-threaded run with 896 dimensions.

bench_main_iai::kmeans_bench_iai::snrm2_benchmark_rust_iai
  Instructions:                    29000796|304003517            (-90.4604%) [-10.4826x]
  L1 Hits:                         34794766|388098122            (-91.0345%) [-11.1539x]
  L2 Hits:                                2|15                   (-86.6667%) [-7.50000x]
  RAM Hits:                         5606322|5606321              (+0.00002%) [+1.00000x]
  Total read+write:                40401090|393704458            (-89.7382%) [-9.74490x]
  Estimated Cycles:               231016046|584319432            (-60.4641%) [-2.52935x]

criterion shows 61% latency reduction.
The measurement was taken in a single-threaded run with 896 dimensions.

.\bench_main-a1fa58ecbc5b2232.exe --bench Snrm2 --color=never
Gnuplot not found, using plotters backend
Benchmarking kmeans-computation/Snrm2 Rust Run
Benchmarking kmeans-computation/Snrm2 Rust Run: Warming up for 3.0000 s

Warning: Unable to complete 50 samples in 5.0s. You may wish to increase target time to 8.3s, or reduce sample count to 30.
Benchmarking kmeans-computation/Snrm2 Rust Run: Collecting 50 samples in estimated 8.2595 s (50 iterations)
Benchmarking kmeans-computation/Snrm2 Rust Run: Analyzing
kmeans-computation/Snrm2 Rust Run
                        time:   [45.687 ms 46.349 ms 47.244 ms]
                        change: [-61.988% -61.410% -60.611%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 13 outliers among 50 measurements (26.00%)
  2 (4.00%) low severe
  4 (8.00%) low mild
  1 (2.00%) high mild
  6 (12.00%) high severe

P.S. Neither tool shows noticeable performance improvements for 4-dimensional vectors.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.46%. Comparing base (3dc4a28) to head (fcad292).
⚠️ Report is 5 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1107      +/-   ##
==========================================
- Coverage   89.48%   89.46%   -0.02%     
==========================================
  Files         474      482       +8     
  Lines       89753    91075    +1322     
==========================================
+ Hits        80316    81481    +1165     
- Misses       9437     9594     +157     

Flag	Coverage Δ
miri	89.46% <100.00%> (-0.02%)	⬇️
unittests	89.11% <100.00%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
diskann-disk/src/utils/math_util.rs	98.82% <100.00%> (-0.02%)	⬇️

... and 15 files with indirect coverage changes

🚀 New features to boost your workflow:

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

[Copilot]

Pull request overview

This PR updates diskann-disk’s squared L2 norm computation to use the SIMD-optimized diskann_vector::norm::FastL2NormSquared implementation, and adjusts the k-means benchmark harnesses to better isolate measured work.

Changes:

• Replaced the scalar per-vector L2-squared loop in compute_vecs_l2sq with FastL2NormSquared.evaluate(...).
• Updated the Criterion benchmark to use iter_batched to avoid including setup (clones/allocations) in timing.
• Updated the iai-callgrind benchmark to move thread-pool creation into setup and add black_box to prevent DCE.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
diskann-disk/src/utils/math_util.rs	Switches compute_vecs_l2sq to use FastL2NormSquared for faster L2-squared norms.
diskann-disk/benches/benchmarks/kmeans_bench.rs	Uses Criterion iter_batched and inlines snrm2 benchmark logic to reduce setup overhead in measurements.
diskann-disk/benches/benchmarks_iai/kmeans_bench_iai.rs	Moves pool creation into setup, removes unnecessary clones, and adds black_box for more reliable callgrind measurements.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.