phoebewang wrote: > > We don't have bf16->float conversion instructions. > > Intel added them in AVX_NE_CONVERT, memory input only. Ok, they use DAZ as well. https://github.com/llvm/llvm-project/pull/90425