Given IEEE-754 format for floats and doubles, we can assume that float -> double promotion is lossless. Given this, we can assume that if we do the standard relative comparison of |lhs - rhs| <= epsilon * max(fabs(lhs), fabs(rhs)), then we get the same result if we do this for floats, as if we do this for doubles that were promoted from floats.
source not stored for this graph (policy: none)