Fix the precision loss in float -> 32bit conversion case, introduced by my earlier commit. Instead use method proposed by Reimar. Also, avoid using ldexp, it is slower than multiply with constant.