Skip to content

v0.7.1

Latest

Choose a tag to compare

@jvdd jvdd released this 20 Oct 07:14
· 1 commit to main since this release

What's Changed

  • refactor: rename the f16 source files to half, before bf16 shares them by @jvdd in #121
  • feat: support bf16 by making the f16 implementation generic over a Half trait by @jvdd in #119
  • test: cover the chunk offset of the overflow-safe loops, merged in one place by @jvdd in #115
  • perf: unsigned index lanes for the 8-bit cores by @jvdd in #117
  • perf: native min and max on the AVX2 16 and 32-bit integer lanes by @jvdd in #120
  • refactor: implement SIMDArgMinMax for all return-NaN floats with one macro by @jvdd in #124
  • docs: say why the return-NaN argmin and argmax use argminmax, and test it by @jvdd in #125
  • docs: fix the grammar of the README sentence on the branchless inner loop by @jvdd in #126
  • refactor: call the i8 and i16 horizontal reductions from the x86 u8 and u16 code by @jvdd in #127
  • refactor: call the i16 horizontal reductions from the f16 and bf16 code by @jvdd in #128
  • refactor: call the i32 and i64 horizontal reductions from the u32, u64 and return-NaN f32 and f64 code by @jvdd in #129
  • perf: branch-free horizontal min and max for the 32 and 64-bit types by @jvdd in #130

Full Changelog: v0.7.0...v0.7.1