You can use casting to do the integer calculation in 32 bit and safely put the result into 16 bit.
This avoids the occasional .999999 is 1 floating point funnies and run a magnitude or two faster than float on AVR.
int RPM_A;
..........
RPM_A = (int) ( 60L * ((long) newpositionA - (long ) oldpositionA ) * 1000L ) / (((long) newtimeA - (long) oldtimeA ) * 16L * 2L );