Aleksei Safronov

@ashvardanian Does the CRC32 implementation use SSE4.2 instructions? If you don’t care about the specific polynomial, it should be much faster compared to crc32fast. For example: github.com/eloj/crc32c

How do you measure the quality of a hash function? Beyond raw speed, you want statistical sanity. For StringZilla’s new hash, I ran the full SMHasher suite — but also built a smaller, minimalistic Rust suite to cover the core signals: avalanche, distribution, and short-input integer collisions. Avalanche (SAC/BIC): Flip one input bit and observe output bits. Ideally ~50% of them flip (Strict Avalanche Criterion), and flips behave independently across positions (Bit-Independence Criterion). Distribution (χ²): Hash N-grams, drop into fixed buckets, and compute a χ² statistic against uniformity. Lower is better; it surfaces skew that raw collision counts miss. Integral collisions (N ≤ 8 bytes): Treat short byte strings as little-endian integers, hash, then map to n slots for n samples (hash % n). A "perfect hash" yields zero collisions. For typical hashes, collisions converge to ≈ n × (1 − e⁻¹) ≈ 36.8% — the birthday bound. github.com/ashvardanian/H…