Optimize NEON packet math with AArch64 FMA, fix benchmark naming

Michael Teschrequested to merge
feature/neon-fma-optimization into master

Summary

  • Optimize NEON dual/dual packet math (pmul, pmadd, pdiv) using AArch64 FMA intrinsics
  • Add pmadd specializations for Eigen's GEMM inner loop — fixes dualf MatMat regression (0.82x → 1.13x)
  • Optimize complex<dual> NEON packet math (pmul, pmadd)
  • Add double-precision benchmarks to bench_eigen
  • Add NEON benchmark comparison tables to README (single + double precision)
  • Fix inverted _novec benchmark binary naming in CMakeLists (base name was getting EIGEN_DONT_VECTORIZE, _novec was vectorized)

All AArch64-specific code is guarded with #if EIGEN_ARCH_ARM64 with original ARMv7 code preserved in #else.

Merge request reports