Speed up ARM fixed-width array scans

[mrecachinas]

What

Adds ARM NEON cross-record scanners for 16- and 32-byte array elements used by the first, best, and all APIs. The optimized scanners preserve first-index ordering, lowest-index best ties, threshold behavior, exact-match termination, and ordered results.

The Rayon crossover for these widths moves from 5 MiB to 16 MiB because the faster serial scanner makes parallel scheduling beneficial later. Benchmark and randomized semantic coverage are expanded for fixed-width workloads.

Closes the ARM/cross-record portion of #51.

Performance

Three runs per revision on the same Apple M4 Max, comparing median run medians:

Case	Before	After	Delta
Rust 16B random/no-match first	1.596 us	397 ns	75% faster
Rust 16B random/no-match all	1.382 us	524 ns	62% faster
Rust 16B exact-mid best	808 ns	217 ns	73% faster
Rust 32B random/no-match first	2.003 us	798 ns	60% faster
Rust 32B random/no-match all	1.955 us	1.048 us	46% faster
Python 16B random/no-match all	1.420 us	565 ns	60% faster
Python 32B random/no-match all	2.037 us	852 ns	58% faster

Testing

• cargo test --no-default-features — 76 unit and 2 doc tests
• python3 -m pytest -vls . — 187 tests
• cargo fmt --check
• ruff format --check .
• Required public API validation scenarios

ruff check . continues to report the pre-existing EXE001 warning for test/test_hexhamming.py.

[github-actions]

📊 Benchmark Comparison Results

Status	Benchmark	Base	PR	Delta
🆕	fixed_width_array_scan_matrix[all-exact_early-16]	N/A	0.0063ms	NEW
🆕	fixed_width_array_scan_matrix[all-exact_early-32]	N/A	0.0072ms	NEW
🆕	fixed_width_array_scan_matrix[all-exact_late-16]	N/A	0.0063ms	NEW
🆕	fixed_width_array_scan_matrix[all-exact_late-32]	N/A	0.0073ms	NEW
🆕	fixed_width_array_scan_matrix[all-exact_mid-16]	N/A	0.0063ms	NEW
🆕	fixed_width_array_scan_matrix[all-exact_mid-32]	N/A	0.0073ms	NEW
🆕	fixed_width_array_scan_matrix[all-random_no_match-16]	N/A	0.0062ms	NEW
🆕	fixed_width_array_scan_matrix[all-random_no_match-32]	N/A	0.0072ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d-16]	N/A	0.0063ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d-32]	N/A	0.0073ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d_minus_1-16]	N/A	0.0062ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d_minus_1-32]	N/A	0.0072ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d_plus_1-16]	N/A	0.0063ms	NEW
🆕	fixed_width_array_scan_matrix[all-threshold_d_plus_1-32]	N/A	0.0073ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_early-16]	N/A	0.0001ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_early-32]	N/A	0.0001ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_late-16]	N/A	0.0066ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_late-32]	N/A	0.0075ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_mid-16]	N/A	0.0034ms	NEW
🆕	fixed_width_array_scan_matrix[best-exact_mid-32]	N/A	0.0039ms	NEW
🆕	fixed_width_array_scan_matrix[best-random_no_match-16]	N/A	0.0065ms	NEW
🆕	fixed_width_array_scan_matrix[best-random_no_match-32]	N/A	0.0075ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d-16]	N/A	0.0066ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d-32]	N/A	0.0075ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d_minus_1-16]	N/A	0.0065ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d_minus_1-32]	N/A	0.0075ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d_plus_1-16]	N/A	0.0066ms	NEW
🆕	fixed_width_array_scan_matrix[best-threshold_d_plus_1-32]	N/A	0.0075ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_early-16]	N/A	0.0001ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_early-32]	N/A	0.0001ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_late-16]	N/A	0.0059ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_late-32]	N/A	0.0068ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_mid-16]	N/A	0.0031ms	NEW
🆕	fixed_width_array_scan_matrix[first-exact_mid-32]	N/A	0.0035ms	NEW
🆕	fixed_width_array_scan_matrix[first-random_no_match-16]	N/A	0.0059ms	NEW
🆕	fixed_width_array_scan_matrix[first-random_no_match-32]	N/A	0.0068ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d-16]	N/A	0.0031ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d-32]	N/A	0.0035ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d_minus_1-16]	N/A	0.0059ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d_minus_1-32]	N/A	0.0071ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d_plus_1-16]	N/A	0.0031ms	NEW
🆕	fixed_width_array_scan_matrix[first-threshold_d_plus_1-32]	N/A	0.0035ms	NEW

➖ 62 benchmarks within noise (click to expand)

Benchmark	Base	PR	Delta
check_bytes_arrays_all_within_dist[ 512 elems,s=16,at 0]	0.0033ms	0.0033ms	no change
check_bytes_arrays_all_within_dist[ 512 elems,s=16,end]	0.0033ms	0.0033ms	0.0% faster
check_bytes_arrays_all_within_dist[ 512 elems,s=16,mid]	0.0033ms	0.0033ms	0.0% faster
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,at 0]	0.0072ms	0.0072ms	0.1% slower
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,end]	0.0073ms	0.0073ms	0.0% faster
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,mid]	0.0073ms	0.0073ms	0.1% faster
check_bytes_arrays_all_within_dist[16384 elems,s=64,at 0]	0.1035ms	0.1034ms	0.0% faster
check_bytes_arrays_all_within_dist[16384 elems,s=64,end]	0.1084ms	0.1086ms	0.2% slower
check_bytes_arrays_all_within_dist[16384 elems,s=64,mid]	0.1038ms	0.1082ms	4.2% slower
check_bytes_arrays_best_within_dist[ 512 elems,s=16,at 0]	0.0001ms	0.0001ms	3.2% slower
check_bytes_arrays_best_within_dist[ 512 elems,s=16,end]	0.0036ms	0.0034ms	4.5% faster
check_bytes_arrays_best_within_dist[ 512 elems,s=16,mid]	0.0017ms	0.0017ms	0.2% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,at 0]	0.0075ms	0.0075ms	0.1% faster
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,end]	0.0075ms	0.0075ms	0.0% faster
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,mid]	0.0075ms	0.0075ms	no change
check_bytes_arrays_best_within_dist[16384 elems,s=64,at 0]	0.1132ms	0.1083ms	4.4% faster
check_bytes_arrays_best_within_dist[16384 elems,s=64,end]	0.1134ms	0.1087ms	4.1% faster
check_bytes_arrays_best_within_dist[16384 elems,s=64,mid]	0.1132ms	0.1084ms	4.2% faster
check_bytes_arrays_first_within_dist[ 512 elems,s=16,at 0]	0.0001ms	0.0001ms	0.4% slower
check_bytes_arrays_first_within_dist[ 512 elems,s=16,end]	0.0030ms	0.0030ms	1.0% faster
check_bytes_arrays_first_within_dist[ 512 elems,s=16,mid]	0.0016ms	0.0016ms	0.4% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,at 0]	0.0001ms	0.0001ms	2.3% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,end]	0.0068ms	0.0072ms	4.5% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,mid]	0.0035ms	0.0035ms	0.6% slower
check_bytes_arrays_first_within_dist[16384 elems,s=64,at 0]	0.0002ms	0.0002ms	12.1% slower
check_bytes_arrays_first_within_dist[16384 elems,s=64,end]	0.1081ms	0.1134ms	4.9% slower
check_bytes_arrays_first_within_dist[16384 elems,s=64,mid]	0.0542ms	0.0540ms	0.3% faster
check_bytes_within_dist[127 bytes,d=500]	0.0001ms	0.0001ms	9.4% faster
check_bytes_within_dist[16 bytes,d=0]	0.0001ms	0.0001ms	9.7% faster
check_bytes_within_dist[64 bytes,d=100]	0.0001ms	0.0001ms	10.8% faster
hex_within_dist	0.0001ms	0.0001ms	7.7% faster
hd_bytes[1000-diff]	0.0001ms	0.0001ms	10.3% faster
hd_bytes[1000-same]	0.0001ms	0.0001ms	12.3% faster
hd_bytes[1024-diff]	0.0001ms	0.0001ms	1.2% faster
hd_bytes[1024-same]	0.0002ms	0.0001ms	55.1% faster
hd_bytes[3-diff]	0.0001ms	0.0001ms	2.8% faster
hd_bytes[3-same]	0.0001ms	0.0001ms	3.6% faster
hd_bytes[64-diff]	0.0001ms	0.0001ms	0.7% faster
hd_bytes_buffer[bytearray]	0.0001ms	0.0002ms	84.3% slower
hd_bytes_buffer[memoryview]	0.0001ms	0.0001ms	7.4% slower
hd_bytes_gil_boundary[16376]	0.0006ms	0.0006ms	3.2% slower
hd_bytes_gil_boundary[16384]	0.0006ms	0.0006ms	7.5% slower
hd_bytes_gil_boundary[16392]	0.0006ms	0.0006ms	7.5% slower
hd_bytes_random[1024]	0.0001ms	0.0001ms	1.6% faster
hd_bytes_random[16]	0.0001ms	0.0001ms	1.2% faster
hd_bytes_random[64]	0.0001ms	0.0001ms	11.5% faster
hd_string[1000-diff]	0.0002ms	0.0002ms	0.5% slower
hd_string[1000-same]	0.0002ms	0.0002ms	0.6% slower
hd_string[1024-diff]	0.0002ms	0.0002ms	2.8% faster
hd_string[1024-same]	0.0002ms	0.0002ms	2.9% faster
hd_string[3-diff]	0.0002ms	0.0002ms	5.0% faster
hd_string[3-same]	0.0001ms	0.0001ms	9.0% faster
hd_string[64-diff]	0.0001ms	0.0001ms	3.9% faster
hd_string_random[1024]	0.0002ms	0.0002ms	4.4% faster
hd_string_random[16]	0.0001ms	0.0001ms	1.0% slower
hd_string_random[64]	0.0001ms	0.0001ms	4.3% faster
stdlib_bytes_distance_random[1024]	0.0028ms	0.0028ms	0.4% slower
stdlib_bytes_distance_random[16]	0.0004ms	0.0004ms	1.4% slower
stdlib_bytes_distance_random[64]	0.0005ms	0.0005ms	0.9% slower
stdlib_hex_distance_random[1024]	0.0036ms	0.0036ms	0.3% slower
stdlib_hex_distance_random[16]	0.0004ms	0.0004ms	0.5% faster
stdlib_hex_distance_random[64]	0.0005ms	0.0005ms	1.3% slower

Legend: ✅ Faster (>5%) · ⚠️ Slower (5-30%) · ❌ Regression (>30%) · 🆕 New · 🗑️ Removed

[mrecachinas]

Python/PyO3 overhead context

The M4 Max measurements indicate that tiny operations have largely reached the Python/PyO3 call-overhead floor, approximately 30–40 ns per call. This includes Python dispatch, buffer extraction, and result construction.

Operation	Raw Rust	Python
3-byte distance	~1 ns	33–40 ns
64-byte distance	~2 ns	34 ns
512×16 first, match at start	1.9 ns	35.6 ns
512×16 first, match at end	402 ns	441 ns
1024×16 random scan	398–524 ns	443–565 ns
1024×32 random scan	798–1,048 ns	837–852 ns

The fixed wrapper cost dominates tiny and immediate-match calls, but becomes negligible for complete scans and larger inputs. Further meaningful Python-side improvements would likely require batching more work per call or changing the API shape rather than continuing to optimize the per-record Rust kernels.