Perf: SIMD + FFI speedups across hex/bytes APIs

[mrecachinas]

Summary

Round of performance work across the Rust SIMD kernels and the PyO3 FFI layer. Measured on Apple M-series (NEON); x86 kernels updated in lockstep.

Highlights (measured)

API	Before	After	Speedup
bytes_arrays_all_within_dist (10k × 128 B)	0.186 ms	0.050 ms	3.7×
check_bytes_within_dist (4 KB)	0.284 µs	0.155 µs	1.83×
hamming_distance_bytes (1 KB)	0.227 µs	0.164 µs	1.38×
check_hexstrings_within_dist (1024 ch, similar)	0.75 µs	0.19 µs	4×
check_hexstrings_within_dist (1024 ch, random, tight max)	0.095 µs	0.072 µs	1.3×

memoryview / bytearray / NumPy buffers are now accepted zero-copy.

What's in here

SIMD kernels (src/neon_simd.rs, src/x86_simd.rs, src/native.rs)

• New dedicated NEON byte kernel (previously fell through to native).
• AVX2 hex-string path switched to VPSHUFB nibble-LUT popcount (no more pack → GPR popcnt).
• Unrolled early-exit paths in SSE / AVX2 / AVX-512 (256 B / 512 B / 1024 B per SAD + threshold check).
• Consolidated redundant CMP ops; hoisted the popcount table out of hot loops.
• Dispatch contract: bytes kernels now return distance or u64::MAX sentinel when max_dist is exceeded (kills the old 2-pass pattern in batch helpers).
• New _with_max variants for all 4 SIMD hex-string kernels + dispatcher, with periodic in-loop threshold checks so SIMD gets the same early-exit benefit the scalar path had.

FFI layer (src/python.rs)

• check_hexstrings_within_dist now routes ≥64-char inputs through the SIMD dispatcher (previously always scalar).
• Zero-copy PyBuffer for bytes APIs — accepts bytes, bytearray, memoryview, NumPy arrays without copying.
• Direct-typed parameters, with_capacity on result vecs, dedup/cleanup in set_algo and friends.

Batch APIs (src/api.rs)

• Rayon parallelization for first_within_dist / best_within_dist / all_within_dist behind a 64 KB payload gate. Tie-breaking (lowest index) preserved via fold + reduce. Serial fallbacks retained and tested equivalent.

Tests

• 30+ new tests (Rust + Python): SIMD length boundaries, invalid-char validation on SIMD paths, tie-break ordering, serial-vs-parallel equivalence, memoryview / bytearray / NumPy inputs, set_algo behavior, early-exit correctness.
• All 122 Python + 71 Rust tests pass. cargo fmt, ruff clean.

Notes

An initial attempt to gate SIMD vs. scalar on max_dist >= len was a trap: the right choice depends on the actual distance, not a length heuristic. Pushing periodic threshold checks into the SIMD kernels themselves is the clean answer — it wins similar-string similarity search (common case) and random + tight-threshold rejection (previously scalar's strength).

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

📊 Benchmark Comparison Results

Status	Benchmark	Base	PR	Delta
✅	check_bytes_arrays_first_within_dist[16384 elems,s=64,at 0]	0.0358ms	0.0002ms	99.5% faster
✅	check_bytes_arrays_first_within_dist[16384 elems,s=64,mid]	0.0880ms	0.0440ms	50.0% faster
✅	check_bytes_arrays_all_within_dist[16384 elems,s=64,at 0]	0.1602ms	0.0880ms	45.1% faster
✅	check_bytes_arrays_all_within_dist[16384 elems,s=64,mid]	0.1584ms	0.0892ms	43.7% faster
✅	check_bytes_arrays_all_within_dist[16384 elems,s=64,end]	0.1578ms	0.0891ms	43.5% faster
✅	check_bytes_arrays_best_within_dist[ 512 elems,s=16,at 0]	0.0085ms	0.0048ms	43.3% faster
✅	check_bytes_arrays_best_within_dist[16384 elems,s=64,mid]	0.1530ms	0.0913ms	40.4% faster
✅	check_bytes_arrays_best_within_dist[16384 elems,s=64,at 0]	0.1551ms	0.0926ms	40.3% faster
✅	check_bytes_arrays_best_within_dist[16384 elems,s=64,end]	0.1529ms	0.0923ms	39.7% faster
✅	check_bytes_arrays_first_within_dist[16384 elems,s=64,end]	0.1426ms	0.0883ms	38.1% faster
✅	check_bytes_arrays_best_within_dist[ 512 elems,s=16,mid]	0.0062ms	0.0045ms	27.1% faster

➖ 34 benchmarks within noise (click to expand)

Benchmark	Base	PR	Delta
check_bytes_arrays_all_within_dist[ 512 elems,s=16,at 0]	0.0040ms	0.0039ms	2.1% faster
check_bytes_arrays_all_within_dist[ 512 elems,s=16,end]	0.0040ms	0.0039ms	2.1% faster
check_bytes_arrays_all_within_dist[ 512 elems,s=16,mid]	0.0040ms	0.0039ms	2.5% faster
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,at 0]	0.0087ms	0.0084ms	3.1% faster
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,end]	0.0088ms	0.0085ms	3.4% faster
check_bytes_arrays_all_within_dist[ 1024 elems,s=32,mid]	0.0088ms	0.0085ms	3.4% faster
check_bytes_arrays_best_within_dist[ 512 elems,s=16,end]	0.0038ms	0.0041ms	6.7% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,at 0]	0.0077ms	0.0087ms	12.5% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,end]	0.0083ms	0.0087ms	4.6% slower
check_bytes_arrays_best_within_dist[ 1024 elems,s=32,mid]	0.0080ms	0.0087ms	8.7% slower
check_bytes_arrays_first_within_dist[ 512 elems,s=16,at 0]	0.0004ms	0.0002ms	51.0% faster
check_bytes_arrays_first_within_dist[ 512 elems,s=16,end]	0.0034ms	0.0038ms	9.9% slower
check_bytes_arrays_first_within_dist[ 512 elems,s=16,mid]	0.0019ms	0.0020ms	8.5% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,at 0]	0.0010ms	0.0002ms	81.1% faster
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,end]	0.0077ms	0.0083ms	7.9% slower
check_bytes_arrays_first_within_dist[ 1024 elems,s=32,mid]	0.0043ms	0.0043ms	1.0% slower
check_bytes_within_dist[127 bytes,d=500]	0.0002ms	0.0003ms	58.8% slower
check_bytes_within_dist[16 bytes,d=0]	0.0003ms	0.0003ms	7.3% faster
check_bytes_within_dist[64 bytes,d=100]	0.0002ms	0.0003ms	64.7% slower
hex_within_dist	0.0001ms	0.0001ms	19.3% faster
hd_bytes[1000-diff]	0.0002ms	0.0002ms	3.7% slower
hd_bytes[1000-same]	0.0002ms	0.0002ms	9.6% slower
hd_bytes[1024-diff]	0.0002ms	0.0002ms	5.1% slower
hd_bytes[1024-same]	0.0002ms	0.0002ms	6.2% slower
hd_bytes[3-diff]	0.0002ms	0.0002ms	18.6% slower
hd_bytes[3-same]	0.0001ms	0.0002ms	18.4% slower
hd_bytes[64-diff]	0.0002ms	0.0002ms	16.1% slower
hd_string[1000-diff]	0.0003ms	0.0002ms	25.4% faster
hd_string[1000-same]	0.0003ms	0.0002ms	25.4% faster
hd_string[1024-diff]	0.0003ms	0.0002ms	25.7% faster
hd_string[1024-same]	0.0003ms	0.0002ms	20.9% faster
hd_string[3-diff]	0.0003ms	0.0003ms	3.6% slower
hd_string[3-same]	0.0002ms	0.0001ms	3.6% faster
hd_string[64-diff]	0.0002ms	0.0001ms	8.8% faster

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