[Feature request] Support translation of CUDA instructions similar to `__bfloat1622float2`

[yukuai26]

      __3.x = (float)(((nv_bfloat162*)(&(v_.x)))->x);
      __3.y = (float)(((nv_bfloat162*)(&(v_.x)))->y);
      __3.z = (float)(((nv_bfloat162*)(&(v_.y)))->x);
      __3.w = (float)(((nv_bfloat162*)(&(v_.y)))->y);

We can use __bfloat1622float2 to improve performance. Similar, __float22bfloat162_rn , __nv_cvt_float2_to_fp8x2 and more CUDA instructions like this can also improve performance.

[Rachmanino]

Hi, @yukuai26! Thanks for your valuable suggestions. Actually we've already supported __bfloat1622float2 and __float22bfloat162_rn, but you have to manually enable them currently by adding "-DENABLE_BF16" to compile_flags. We'll consider supporting more such kinda conversions when availble.

[Rachmanino]

Closed as resolved via pr #1095

[Rachmanino]

Now vectorized cast between bf16 and fp32 is automatically applied, without the need of the compile flag "-DENABLE_BF16"