* remove raw NEON/SSE2 implementation as much as possible
  * replace them to universal intrinsic in InRange/Compare/AddWeighted