Skip to content

v2.1.0 — Clean-room NEON kernels (erf/exp/log/cos) & SIMD accuracy fixes

Latest
Latest

Choose a tag to compare

View all tags
@OldCrow OldCrow released this 21 Jul 04:04
v2.1.0
This tag was signed with the committer’s verified signature.
OldCrow Gary Wolfman
GPG key ID: BCE891EA63357A90
Verified
Learn about vigilant mode.
48f83e8
This commit was signed with the committer’s verified signature.
OldCrow Gary Wolfman
GPG key ID: BCE891EA63357A90
Verified
Learn about vigilant mode.

What's changed

Security

vector_erf_neon (shipped since v1.5.0) was found to be a near-verbatim port of glibc's LGPL-2.1+ erf_advsimd.c. It was replaced with an independently authored clean-room implementation (local Taylor expansion with a compensated hi/lo table), built with zero access to the prior code or the glibc source. The replacement carries no third-party license obligations and is strictly more accurate: max 1 ULP (was ~2.29 ULP), 7.1–7.4x speedup (on par with the ~8.0x it replaced). See docs/NEON_ERF_DERIVATION.md / NEON_ERF_DIVERGENCE_AUDIT.md.

Added

  • Clean-room vector_exp_neon: tail-corrected N=128 table + order-5 polynomial kernel. <1 ULP accuracy, ~20–27% faster than the prior SLEEF polynomial.
  • Clean-room vector_log_neon: compensated (L_hi, L_lo) table kernel. Max 0.52 ULP (was 2.0), 1.74x scalar (was 1.28x).
  • Clean-room vector_cos_neon: quadrant-reduction kernel fixing sign errors near k·π/2 in the prior 7-term Taylor kernel (was ~6e8 ULP in [-2π, 2π]) — a correctness fix for the Von Mises PDF/CDF path. New kernel: max 0.78 ULP.

Fixed

  • x86 vector_exp_* subnormal underflow (SSE2/AVX/AVX2/AVX-512): deep-negative inputs now flush gracefully through subnormals to +0 (previously clamped at -708.0, up to ~6.7e15 ULP error); also fixes incorrect +inf for x > ~709.44.
  • x86 vector_exp_* non-finite edge cases: NaN/overflow now correctly propagate instead of clamping to a finite value, matching std::exp.
  • vector_log_sse2 subnormal inputs (#74): ported the subnormal-input 2^54 scaling step already present in AVX/AVX2/AVX-512.
  • AVX-512 vector_exp/vector_cos rounding: added the missing _MM_FROUND_NO_EXC bit for parity with the other three x86 tiers.
  • Windows/MSVC LIBSTATS_MAX_SIMD_TIER link failures: the global /arch: flag now cascades AVX-512 → AVX2 → AVX → SSE2, fixing LNK2019 errors when capping the compiled tier below AVX-512.

Removed

  • Stale vector_erfc SIMD stub (unused; no batch path calls erfc() directly). Scalar erfc() is retained.

Validation

46/46 correctness tests pass on Kaby Lake AVX2+FMA, Mac Mini M1 NEON, and Asus TUF A16 AVX-512. All four x86 SIMD tiers (AVX-512, AVX2, AVX, SSE2) are now natively validated on both x86 machines.

Full changelog: https://github.com/OldCrow/libstats/blob/main/CHANGELOG.md

Assets 2
Loading