I can't tell from the paper, but are they using maximally precise double-precision libraries? Implementations may have errors much larger than 1ULP for transce

I can't tell from the paper, but are they using maximally precise double-precision libraries? Implementations may have errors much larger than 1ULP for transcendental functions and roots.