Hello Luc.
> >> [...]
> >>+ public static double pow(double d, int e) {
> >>+ if (e == 0) {
> >>+ return 1.0;
> >>+ } else if (e < 0) {
> >>+ e = -e;
> >>+ d = 1.0 / d;
> >>+ }
> >>+
> >>+ double result = 1;
> >>+ double d2p = d;
> >>+ while (e != 0) {
> >>+ if ((e & 0x1) != 0) {
> >>+ result *= d2p;
> >>+ }
> >>+ d2p *= d2p;
> >>+ e = e >> 1;
> >>+ }
> >>+
> >>+ return result;
> >>+ }
> >
> >I've added a unit test for this function. It shows that the result
> >is not
> >the same as with "pow(double, double)" (cf. tolerance in
> >"assertEquals").
> >I didn't check which one is more accurate, but I thought that we
> >should be
> >aware of the discrepancy.
>
> I have checked this further. The pow(double, double) seems to be
> always slightly
> better than pow(double, int). I have used a comparison with Dfp set
> up at 60 digits
> and 100 digits too. I have also checked depending on the initial
> number being representable
> or not.
>
> I expected the method to be slightly more accurate for small powers
> (say up to 20) and most
> importantly be much faster. The first part of my assumption is not
> true, so it is a major
> drawback. Even for very small powers I quickly have one ulp.
>
> I'll take a further look at the current speed of pow(double,double)
> and if it is reasonable,
> I'll remove the new method. Adding a method that decreases accuracy
> is never good.
I had checked the performance too. ;-)
The "int" method is indeed about 8 times faster:
pow (calls per timed block: 200000, timed blocks: 100, time unit: ms)
name time/call std error total time ratio
difference
double 2.92909569e-04 6.40278315e-05 5.8582e+03 1.0000e+00
0.00000000e+00
int 3.72931144e-05 1.10545444e-05 7.4586e+02 1.2732e-01
-5.11232909e+03
double (Math) 3.52308890e-04 6.92417994e-05 7.0462e+03 1.2028e+00
1.18798641e+03
So, it's probably good to keep it, as long as people are aware of the
trade-off.
Best,
Gilles
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
For additional commands, e-mail: dev-h...@commons.apache.org
