nanoseconds to milliseconds - fast division by 1000000

Question

I'm wanting to convert the output from gethrtime to milliseconds.

The obvious way to do this is to divide by 1000000. However, I'm doing this quite often and wonder if it could become a bottleneck.

Is there an optimized divide operation when dealing with numbers like 1000000?

Note: Any code must be portable. I'm using gcc and this is generally on Sparc hardware

Some quick testing using the code below... hope that is right.

#include <sys/time.h>
#include <iostream>

using namespace std;

const double NANOSECONDS_TO_MILLISECONDS = 1.0 / 1000000.0;

int main()
{
    hrtime_t start;
    hrtime_t tmp;
    hrtime_t fin;

    start = gethrtime();
    tmp = (hrtime_t)(start * NANOSECONDS_TO_MILLISECONDS);
    fin = gethrtime();

    cout << "Method 1"
    cout << "Original val: " << start << endl;
    cout << "Computed: " << tmp << endl;
    cout << "Time:" << fin - start << endl;

    start = gethrtime();
    tmp = (start / 1000000);
    fin = gethrtime();

    cout "Method 2"    
    cout << "Original val: " << start << endl;
    cout << "Computed: " << tmp << endl;
    cout << "Time:" << fin - start << endl;

    return 0;
}  

Example outputs:

Original val: 3048161553965997
Computed: 3048161553
Time:82082
Original val: 3048161556359586
Computed: 3048161556
Time:31230

Original val: 3048239663018915
Computed: 3048239663
Time:79381
Original val: 3048239665393873
Computed: 3048239665
Time:31321

Original val: 3048249874282285
Computed: 3048249874
Time:81812
Original val: 3048249876664084
Computed: 3048249876
Time:34830

If this is correct, then the multiple by reciprocal is actually slower in this case. It's probably due to using floating point math instead of fixed point math. I will just stick to integer division then which still takes hardly any time at all.

Answer 1

Division is not an expensive operation. I doubt very much if a divide-by-1000000 operation will be anywhere near the main bottleneck in your application. Floating-point processors will be way faster than any sort of "tricks" you can come up with than just doing the single operation.