Re: [Algorithms] How to get 3dvector largest coordinate index?

[Sylvain G. V. <vi...@ii...>]

What's about something like:
yx = SignBit(x-y)>>31 // 1 if y > x
yz = SignBit(z-y)>>31
zx = SignBit(x-z)>>31
zy = SignBit(y-z)>>31
return (yx&yz) | ((zx&zy)<<1);

4 subs (int or float, doesn't matter), a few simple bit instructions, no
jump (untested, just got back from a diner)?

----- Original Message -----
From: Matt J <mjo...@gm...>
Date: Saturday, February 28, 2009 6:17 pm
Subject: Re: [Algorithms] How to get 3dvector largest coordinate index?
To: Game Development Algorithms <gda...@li...>

> I did some profiling on MSVC 2009, and integer methods are ~10% 
> faster:
>  unsigned int x = 0x7FFFFFFF & *(unsigned int *)&n.x;
>  unsigned int y = 0x7FFFFFFF & *(unsigned int *)&n.y;
>  unsigned int z = 0x7FFFFFFF & *(unsigned int *)&n.z;
> 
>  return (x > y) ? (x > z) ? 0 : 2 : (y > z) ? 1 : 2;
> 
> For some odd reason only when using integers can it simplify the 
> two inner
> ternary methods using bit 'magic'. If we use floats or use SSE the 
> integermethods are about ~5% faster because this compiler generates 
> a bunch of
> jumps. The method above (as well as mine) generates 1 jump.
> 
> I actually eliminated all jumps but since I use one too many 
> registers it
> slows it down. I think the CPU is smart enough to look ahead if you 
> get down
> to one conditional ..so yeah I think if you use the stack it can 
> slow things
> down slightly but you do see a gain if you use integers. I assume
> conditional jumps would be the best way, assuming the compiler can 
> do it
> depth first because you need 3 conditionals - the inner two, and 
> then a
> conditional move on the results of those two
> 
> Asm:
> 
> 
> ; _n$ = edx
> 
> ; 61   :   unsigned int x = 0x7FFFFFFF & *(unsigned int *)&n.x;
> 
> mov eax, DWORD PTR [edx]
> 
> ; 62   :   unsigned int y = 0x7FFFFFFF & *(unsigned int *)&n.y;
> 
> mov ecx, DWORD PTR [edx+4]
> 
> ; 63   :   unsigned int z = 0x7FFFFFFF & *(unsigned int *)&n.z;
> 
> mov edx, DWORD PTR [edx+8]
> and eax, 2147483647    ; 7fffffffH
> and ecx, 2147483647    ; 7fffffffH
> and edx, 2147483647    ; 7fffffffH
> 
> ; 69   :   return (x > y) ? (x > z) ? 0 : 2 : (y > z) ? 1 : 2;
> 
> cmp eax, ecx
> jbe SHORT $LN3@coord_inde
> cmp edx, eax
> sbb eax, eax
> and eax, -2     ; fffffffeH
> add eax, 2
> 
> ; 70   : }
> 
> ret 0
> $LN3@coord_inde:
> 
> ; 69   :   return (x > y) ? (x > z) ? 0 : 2 : (y > z) ? 1 : 2;
> 
> cmp edx, ecx
> sbb eax, eax
> add eax, 2
> 
> ; 70   : }
> 
> ret
>