Compiler runtimes for float16 basic operations might not be correctly rounded for all rounding modes.