Rounding
When working with integers, fixedpoint math or when converting floats to integers we need to round somehow, but which way is best …
In school they probably told you to round to nearest and round halfway cases away from zero +-x.0000 … +-x.4999 to +-x and +-x.5 … +-x.9999 to +-x+-1 but thats actually not optimal
First decission, round to nearest or not, rounding to nearest is pretty clear already, it means to well round to the nearest if there is a single nearest value and thats obviously more accurate then not doing so, not rounding to nearest has one big advantage, its often but not always (PAVGB MMX instruction for example) faster (x>>2 vs. (x+2)>>2)
Second decission, what to do with x.5, it seems to not matter much but actually it often does matter, the round toward +inf and round to -inf cases are easy to implement (x+2)>>2 and (x+1)>>2 for example but they add some ugly systematic bias which can be problematic if values are used in long calculations and get rounded many times
the solution, round to even (or odd) doesnt add such a bias and is thus better, alternatively you could also repeatedly flip the direction in which rounding is done, mpeg4 for example does that in the motion compensation code …
to round to even you could do something like (x + ((x>>1)&1))>>1, (x+1+((x>>2)&1))>>2
to round to odd you could do something like (x>>1)|(x&1), (x+2-((x>>2)&1))>>2
you might also be stuck with instructions which do something like (a*constant)>>16 to reduce the rounding error of these you could simple add (1<<15) / constant to a