quote:Original post by honayboyz
and at that, do i even need to square them? or cout i just get absolute value, which ever is faster?

quote:Original post by Enigma
If you find you do need a square root, here's an optimised sqrt function based on the Carmack InvSqrt. The magic number can probably be tweaked a little more if anybody fancies playing with it. I tested it on the full range of normalised floating point numbers. The maximum percentage error was 2.84658% and the average was 1.73863%.

float mysqrt(float x){	union{		int intPart;		float floatPart;	} convertor;	union{		int intPart;		float floatPart;	} convertor2;	convertor.floatPart = x;	convertor2.floatPart = x;	convertor.intPart = 0x1FBCF800 + (convertor.intPart >> 1);	convertor2.intPart = 0x5f3759df - (convertor2.intPart >> 1);	return 0.5f*(convertor.floatPart + (x * convertor2.floatPart));}     

I think the accuracy can be slightly improved by changing the 0.5f in the return statement into a 0.4999f because the function always overestimates slightly. Again this number can probably be tweaked. Here's the results I got for timing 500 000 000 sqrts:

sqrt (standard): >38s
fsqrt (asm): >38s
mysqrt: <20s

That includes function call and loop overheads.

Enigma

quote:Original post by shadow12345
EDIT: here is another one from tricks of the 3d game programming gurus (in the far back of the book). It's based on the same idea. I think it's equivalent to the carmack inverse sqrt. It's ridiculously fast, but only within 5% accuracy.

/*	-From tricks of the 3d game programming gurus	I guess this makes me a 3d game programming guru then...right...*/float	Faster_Sqrtf(float f){	float	result;	_asm	{		mov eax, f		sub eax, 0x3f800000		sar eax, 1		add eax, 0x3f800000		mov result, eax	}	return result;}