Arc cosine calculation performance problem

[pm44xl22]

In my previous post I figured out (thanks to D_Drmmr) how to calculate arc cosine with real numbers (including numbers not in the range -1;1).

_acos = -c*log(dotprods[j]+c*pow(one-pow(dotprods[j],2),0.5));

Anyway my application response times have increased by four times with this line

Is there any other way to calculate arc cosine (I am using VS2010 in a standard C++ project) faster?

Or is there any other possible engineering to make the program faster?


Thanks in advance

[SkyNetTo]

isnt this equal to:

_acos = -c*log(dotprods[j]+c*sqrt(one-(dotprods[j]*dotprods[j])));

[pm44xl22]

Just a little faster, but they are both slow

[D_Drmmr]

Is there any other way to calculate arc cosine (I am using VS2010 in a standard C++ project) faster?

Or is there any other possible engineering to make the program faster?

Are you testing with a release build? I'm not sure about VS2010, but in previous versions the default project settings did not perform full optimization. Check the project settings, most importantly those under C/C++ --> Optimization.

[pm44xl22]

Yes release mode with optimization at "Favors fast code"

[Lindley]

sqrt(one-(dotprods[j]*dotprods[j])));

If these were dot products of unit vectors, then dotprods[i] == cos(theta), where theta is the angle between the unit vectors.

In that specific case, this entire expression simplifies to abs(sin(theta)).

[pm44xl22]

I'll try this as solution but the logarithm still takes part. Don't know whether I can get better response times

[OReubens]

calculating a log() is slow. not much you can do about that.

If you're doing lots of acos calculations, then you can speed things up by making a table of acos over a certain range with a certain step and use interpolation to calculate the value of an acos() call.

Depending on needed accuracy and range, you may end up needing a very big table though.


Since you mention 'slow' I'm asuming you're calling acos() thousands/millions of times? You probably need to look at ways to remove the acos() call out of your inner loops rather than trying to optimise the acos() call itself.

[pm44xl22]

Hundreds of thousands, I think you're right and a table is the only way to speed up the process