For the bf16bf16of32bf16 lpgemm api, inside the micro-kernels in order
to convert the accumulated float values to bfloat16 before storing,
the _mm512_cvtneps_pbh intrinsic (vcvtneps2bf16) is used. This
intrinsic rounds the value based on a rounding bias logic. Replicating
the same rounding logic inside the bf16 bench accuracy check function
to get proper one to one comparison of output values.
AMD Internal: [SWLCSG-2948]
Change-Id: I135ac39ac8484769b6c0fe5b3e351dd22d7ca1d8
baa1ff98da