First, Anatoly's tip helped me (see "WA#4 - I don't know why" thread).

Another one tip for those receiving WA with "swear correct algo": I've multiplyed whole input data to some koef (tried 500, 800) and divided result by this koef before printing — got AC while without it got WA on 8-th test. It's all about error of calculations.

And, of course, I've used double, not float.