Skip to content

v0.9.15 — SIMD ReLU/Exp/Sigmoid/SiLU + wgpu v0.30.9

Latest

Choose a tag to compare

@kolkov kolkov released this 05 Jul 14:43
da9caea

Added

  • SIMD ReLU — AVX + AVX-512 branchless max(0, x). Up to 22.9× speedup (#119 by @bennibbelink)
  • SIMD Exp — Cephes/SLEEF minimax polynomial with FMA Horner evaluation. AVX2+FMA. Edge-domain scalar fallback. Up to 813× vs scalar math.Exp (#119 by @bennibbelink)
  • SIMD Sigmoid1/(1+exp(-x)) composing per-vector SIMD Exp. AVX2+FMA. Up to 726× (#119 by @bennibbelink)
  • SIMD SiLUx*sigmoid(x) composing per-vector SIMD Exp. AVX2+FMA. Up to 600× (#119 by @bennibbelink)
  • Exp fuzz tests with domain-specific seed corpus

Changed

  • gogpu/wgpu v0.30.7 → v0.30.9
  • goffi v0.5.5 → v0.5.6 (callback stack-move corruption fix)

Contributors


Full Changelog: v0.9.14...v0.9.15

Sponsor Born ML: https://opencollective.com/born-ml