Skip to content

refactor(rm): Consolidate dtype utilities

Ian Tayler Lessa requested to merge refactor/rm/consolidate-dtype-utils into main
  • Remove AccPrecision, DtypeLimits, QMax and QMin
  • Add dtypeInfo function, which centralises the information provided by all of the above.
  • Adjust code where necessary to account for the fact that dtypeInfo is not constexpr unlike the original utilities
  • Use fp8_e4m3 and fp8_e5m2 consistently instead of fp8e4m3 and fp8e5m2.
  • NOTE: replacing bf16 with ct::bfloat16 requires more subtlety because it could collide with Eigen::bfloat16 in many of the files. Therefore, bf16 is left untouched in this patch.

Signed-off-by: Ian Tayler Lessa ian.taylerlessa@arm.com

Merge request reports