Skip to content

AVX2/AVX512 wrapper fixes + new AVX256 backend

Arthur Marius Hennequin requested to merge ahennequ_avx512_fixes into master
  • Few fixes to SIMD wrapper.
  • Introduction of AVX256 backend: uses avx512 new instructions but with a vector size of 256. This allow to benefit from the increased register count, the new mask registers and new instructions (compressstore) but without the problem of frequency scaling.
  • Added masked gathers

@sponce @cattanem @clemenci @bcouturi

Edited by Marco Cattaneo

Merge request reports