Floating-point arithmetics for low-precision computations