Skip to content

SIMD_MATH.md

Latest commit

 

History

History
37 lines (26 loc) · 1.49 KB

SIMD_MATH.md

File metadata and controls

37 lines (26 loc) · 1.49 KB

SIMD Math Status

SIMDeez includes a native, pure-Rust SIMD math surface exposed through simdeez::math and simdeez::prelude.

The public surface covers:

  • core log/exp: log2_u35, exp2_u35, ln_u35, exp_u35
  • trig and inverse trig: sin_u35, cos_u35, tan_u35, asin_u35, acos_u35, atan_u35, atan2_u35
  • hyperbolic and inverse hyperbolic: sinh_u35, cosh_u35, tanh_u35, asinh_u35, acosh_u35, atanh_u35
  • binary misc: log10_u35, hypot_u35, fmod

Current Shape

  • Most f32 families use portable SIMD by default.
  • f32 log2_u35 keeps an AVX2 override where local benchmarks justify it.
  • Revived f64 log/exp, inverse trig, and binary-misc families keep SIMD defaults.
  • Some f64 families intentionally remain mixed or scalar-reference where local rescue passes did not justify a SIMD default.

Implementation Pattern

The maintained pattern is:

  1. start with portable SIMD kernels
  2. add dispatch glue without changing the public API
  3. add backend-specific overrides only where profiling justifies them
  4. keep scalar-lane patching centralized for exceptional semantics

The restored f32 path in src/math/f32/ is the clearest reference implementation of that layering.

Benchmarks

Criterion targets for the math surface:

cargo bench --bench simd_math
cargo bench --bench simd_math_remaining_baseline

These benches report native scalar, runtime-selected, and forced backend variants such as scalar, sse2, sse41, avx2, and avx512 when available.