User talk:Rednaxela/FastTrig

Sound good! But, will it skipped turns at first tick of the first round? » Nat | Talk » 10:41, 3 March 2009 (UTC)

It shouldn't. It will always be as fast or faster than standard trig functions at least so long as you do the following:

• 'inline' the index calculations like was said and used in the example testfor
• Run init() BEFORE the battle begins (i.e. in static initialization code of your bot class)

An example of that static initialization code is like this:

public class SomeBot extends AdvancedRobot {
    static {
        FastTrig.init();
    }

    public void run() {
        // Not in here
    }
{

The initialization code doesn't take long (0.0017 seconds on my computer at 2880 divisions) but just to be sure it doesn't hurt to place the initialization outside of where the bot is timed --Rednaxela 16:27, 3 March 2009 (UTC)

My take

Nice work. From what I remember Simonton did something similar. I'm not sure if he had any luck though. One thing I noticed, I'm not sure if it works properly for negative angles, you will be rounding the wrong way. Also, the most useful trig function would be atan2. --Skilgannon 20:43, 3 March 2009 (UTC)

Ahh yes, sine doesn't work right at all yet for negative numbers. Cosine does work correctly except for rounding. I'll fix this now. :) --Rednaxela 21:51, 3 March 2009 (UTC)

Bugs

I was going to change the User:Rednaxela/FastTrig to fix the bug that Skilgannon said, and another one that happens with large DIVISION numbers and saw your edit, I haven't tested the fixed version, but I think it is still buggy. I think it should be:

  public static final double sin(double value) {
    return sineTable[(int)(((value/K + 0.5) % DIVISIONS + DIVISIONS)%DIVISIONS)];
  }
  public static final double cos(double value) {
    return cosineTable[(int)(((value/K + 0.5) % DIVISIONS + DIVISIONS)%DIVISIONS)];
  }

I'm testing it with 62832 DIVISIONS, and much bigger loops than yours and is much faster. Here is the test I used:

package ags.util;
public class FastTrigTest {
  public static void main(String[] args) {
    FastTrig.init();
    double v=0;
    long ms;
    int A = 10000;
    int B = 1000;
    double absolute = 0;
    for ( int i = 0; i < A; ++i ) for ( int j = 0; j < B; ++j ) {
      double angle = Math.random() * Math.PI * 200 - Math.PI * 100;
      absolute = Math.max(absolute, Math.abs(Math.sin(angle) - FastTrig.sin(angle)));
      absolute = Math.max(absolute, Math.abs(Math.cos(angle) - FastTrig.cos(angle)));
    }
    System.out.printf("Wrose error: %.12f\n", absolute);
    v=0;
    ms = -System.nanoTime();
    for ( int i = 0; i < A; ++i ) for ( int j = 0; j < B; ++j )
      v += FastTrig.sin(i * j * Math.PI - Math.PI / 3);
    ms += System.nanoTime();
    System.out.printf("FastTrig time: %.2f seconds\n", ms / 1E9);
    v = 0;
    ms = -System.nanoTime();
    for ( int i = 0; i < A; ++i ) for ( int j = 0; j < B; ++j )
      v += Math.sin(i * j * Math.PI - Math.PI / 3);
    ms += System.nanoTime();
    System.out.printf("Math time: %.2f seconds\n", ms/1E9);
  }
}

It also checks the error, so it was easy to the bugs. But anyway, including this will be a priority for me now :). Good work. --zyx 22:15, 3 March 2009 (UTC)

Thanks! Well, my +40*DIVISIONS one DOES work unless the angle is extremely negative, to a magnitude you wouldn't generally expect to see, but your double-modulus solution is more robust. Now making an updated version that includes that increased robustness, and atan(), and maybe atan2(). Will update the page when that's ready. --Rednaxela 00:05, 4 March 2009 (UTC)

Also discovered speed was further increased by using value*K instead of value/K and adjusting the constant to compensate, and also that the cosine table is redundant. Using the sine table for cosine with a shifted index doesn't affect speed but decreases init time and memory use. --Rednaxela 00:12, 4 March 2009 (UTC)