Skip to content

Field alignment on ARM

Michael Kuron requested to merge arm64 into master
  • conditionally remove -march=native, which Apple's clang doesn't support or need (see !415 (merged))
  • don't use __BIGGEST_ALIGNMENT__ on ARM because the architecture does not require alignment, yet has a performance penalty for unaligned access
  • check alignment in generated code, should be aligned to vector width or, in the nontemporal case, to the cacheline size (see pycodegen/pystencils!230 (merged))
  • SVE vectorization currently requires specifying the width (see pycodegen/pystencils!232 (merged)), so autodetect it. Even with the width-agnostic version (pycodegen/pystencils!234 (merged)), explicitly specifying a width can be considered a WALBERLA_OPTIMIZE_FOR_LOCALHOST-style optimization

Merge request reports