Support BFloat16 in Eigen

Teng Lurequested to merge
ShengYang1/eigen:bf16_in_eigen into master

This is a preview of supporting BFloat16 in Eigen from Intel. What we have done:

  • define BFloat16 data type(similar with Half)
  • support basic operation for BFloat16
  • support vectorization for BFloat16
  • add new compiler flag to enable new BFloat16 instruction on CPU

We submit this MR to discuss how we can better support this feature, suggestions and feedback are welcome, thanks.

Edited by Antonio Sánchez

Merge request reports