🔒 Locked

Fast Cosine Approximation

Started by arriu Sep 27, 2010 at 2:11 PM 4 replies 3.6k views

Original Post

Sep 27, 2010 at 2:11 PM

Hey guys,

I'm trying to remove what seems to be a bottle neck in my noise function. I have a call to Math.Cos() which calculates the cosine in double precision but also adds another function call and cast to float. I have seen simple methods of estimating the cosine function in the past (using a parabola I think) but I'm having trouble tracking them down.

Anyways, here is the function I wish to speed up, any help is appreciated:

// cosine interpolation between a and b by a factor of x
public float lerp(float a, float b, float x)
{
float ft = x * 3.1415927f;
float f = (1.0f - (float)Math.Cos(ft)) * 0.5f;
return a * (1.0f - f) + b * f;
}

Kambiz

100

Kambiz

Sep 27, 2010 at 3:11 PM

In the range -pi/2 to pi/2 you can use the approximatoion
cos(x) ~ -(x-pi/2)(x+pi/2)/(pi*pi/4)

(This approximation is different from the Taylor expansion 1-x^2/2+... but has the nice property of having roots at exactly -pi/2 and pi/2)

arriu

Author

arriu

Sep 27, 2010 at 3:44 PM

Thanks! I may end up using that.

Also I managed to dig up the article I was looking for: http://www.devmaster.net/forums/showthread.php?t=5784

Is there any way to remove the abs()? I am working with the compact .net framework which does not inline any function calls. Calling Math.Abs() may make it pointless.

float sine(float x){    const float B = 4/pi;    const float C = -4/(pi*pi);    float y = B * x + C * x * abs(x);    #ifdef EXTRA_PRECISION    //  const float Q = 0.775;        const float P = 0.225;        y = P * (y * abs(y) - y) + y;   // Q * y + P * y * abs(y)    #endif}

Dave Eberly

127

Dave Eberly

Sep 27, 2010 at 3:49 PM

This file has several approximations for trigonometric functions. The
header has the restrictions on the domain of these functions.

Matias Goldberg

366

Matias Goldberg

Sep 27, 2010 at 4:12 PM

If you are working with 32 bit floats all the time, you may want to try changing the precision and rounding mode at the start of the application:

_controlfp(_PC_24, _MCW_PC)_controlfp(_RC_NEAR, _MCW_RC)

Every now and then you will have to check a DLL (i.e. Direct3D) didn't change it back:

assert( (_controlfp(0, 0) & _MCW_PC) == _PC_24 );assert( (_controlfp(0, 0) & _MCW_RC) == _RC_NEAR );

By the way, what happens when you just use cosf()?

If precision isn't your concern, try changing your compiler optimizations from Precise to Fast (in MSVC, Alt+F7 -> C++ -> Code Generation -> Floating Point Model)

Cheers
Dark Sylinc

Edit: Oh I see, this seems C#, then you're probably f**ked.

arriu

Author

arriu

Sep 27, 2010 at 4:30 PM

Yeah, I'm stuck with c#. Which is why I do not want to use the Math.Cos class. It's overkill for what I need.

I guess I am going to use this:

        public static float Cos(float x)        {            const float B = 4 / PI;            const float C = -4 / (TwoPi);            x -= PIby2;            return -(B * x + C * x * ((x < 0) ? -x : x));        }        public static float CosLerp(float a, float b, float x)        {            const float B = 4 / PI;            const float C = -4 / (TwoPi);            float ft = x * PI;            //float f = (1.0f - (float)Math.Cos(ft)) * 0.5f;            float c = ft - PIby2;            float f = (1.0f + (B * c + C * c * ((c < 0) ? -c : c))) * 0.5f;            return a * (1.0f - f) + b * f;        }

edit: even without the extra precision this will do nicely

[Edited by - arriu on September 27, 2010 5:30:08 PM]

Topic Locked

This topic has been locked by a moderator. New replies are not allowed.