Java 1.6 on an x86 (Intel Core 2 Duo in this case), how much faster are the 4 basic math operations +-*/ when it comes to integer vs floating point operations? We’ve all been taught that integer operations are vastly faster than floating point operations, but that was probably only true when you were taught that back in school.

Now a days 32-bit floating point operations are on par with their integer counter parts. Most modern processors can do most floating point operations in an single instruction, just like integers.

I wrote a micro benchmark doing the same operation 4,000,000 times on the same set of vectors and here are the results in milliseconds:

client	+	-	*	/	sqrt	sin	cos
integer	16	16	15	26
long	30	30	30	120
float	15	15	16	30
double	29	26	27	52	242	4205	4205
fixed	13	12	40	191	427	60	59

Fixed is fixed-point library I am testing. The fixed point sin/cos functions are so fast, since they are pre-calculated tables at start-up. On the other hand, Java’s sin/cos functions are abysmal. Java has strict requirements on sin/cos functions and this issue has been brought up as bug 5005861. This is probably one of the last major Java performance issue. If you need a fast sin, then approximate it.

The table above shows there is no performance difference between integers/floats and longs/doubles, except that double’s divide is faster than long’s!

The above was running without the -server flag. Below we have the flag turned on.

server	+	-	*	/	sqrt	sin	cos
integer	21	115	14	21
long	24	25	25	106
float	12	12	12	29
double	24	24	25	54	87	3317	3260
fixed	14	13	32	189	314	57	64

With the server flag turned on, there is some performance issues at the start of the benchmark, which erroneously shows integer addition and subtracting being worst than with the server flag turned off. I believe this is just a loading issue, and in the real world you’ll see performance at the same level as multiplication. One very interesting data point is Math.sqrt is much faster with the server flag on. This benchmark shows an improvement of almost 3x!

I would say it is probably safe for everybody to use doubles for math intensive applications. I not saying just use doubles everywhere, floats and doubles have their pitfalls as well, but that is for another article (the one sentence summary is “Time and money are valuable, so don’t use a float or double to represent it!”) Performance wise, don’t fret it, go double.