fast abs/neg/sign for 32bit floats

Type : floating point functions
References : Posted by tobybear[AT]web[DOT]de
Notes :
Haven't seen this elsewhere, probably because it is too obvious? Anyway, these functions are intended for 32-bit floating point numbers only and should work a bit faster than the regular ones.

fastabs() gives you the absolute value of a float
fastneg() gives you the negative number (faster than multiplying with -1)
fastsgn() gives back +1 for 0 or positive numbers, -1 for negative numbers

Comments are welcome (tobybear[AT]web[DOT]de)


Toby (
Code :
// C/C++ code:
float fastabs(float f)
{int i=((*(int*)&f)&0x7fffffff);return (*(float*)&i);}

float fastneg(float f)
{int i=((*(int*)&f)^0x80000000);return (*(float*)&i);}

int fastsgn(float f)
{return 1+(((*(int*)&f)>>31)<<1);}

//Delphi/Pascal code:
function fastabs(f:single):single;
begin i:=longint((@f)^) and $7FFFFFFF;result:=single((@i)^) end;

function fastneg(f:single):single;
begin i:=longint((@f)^) xor $80000000;result:=single((@i)^) end;

function fastsgn(f:single):longint;
begin result:=1+((longint((@f)^) shr 31)shl 1) end;

from : tobybear[AT]web[DOT]de
comment : Matthias (bekkah[AT]web[DOT]de) wrote me a mail with the following further improvements for the C++ parts of the code: // C++ code: inline float fastabs(const float f) {int i=((*(int*)&f)&0x7fffffff);return (*(float*)&i);} inline float fastneg(const float f) {int i=((*(int*)&f)^0x80000000);return (*(float*)&i);} inline int fastsgn(const float f) {return 1+(((*(int*)&f)>>31)<<1);} Thanks!

from : picoder[AT]mail[DOT]ru
comment : Too bad these 'tricks' need two additional FWAITs to work in a raw FPU code. Maybe standard fabs and fneg are better? Although, that fastsgn() could be useful since there's no FPU equivalent for it. Cheers, Aleksey.

from : picoder[AT]mail[DOT]ru
comment : I meant 'fchs' in place of 'fneg'.

from : david[AT]brannvall[DOT]net
comment : I don't know if this is any faster, but atleast you can avoid some typecasting. function fastabs(f: Single): Single; var i: Integer absolute f; begin i := i and $7fffffff; Result := f; end;

from : chris[AT]m-audio[DOT]com
comment : Note that a reasonable compiler should be able to perform these optimizations for you. I seem to recall that GCC in particular has the capability to replace calls to [f]abs() with instructions optimized for the platform.

from : kaleja[AT]estarcion[DOT]com
comment : On MS compilers for x86, just do: #pragma intrinsic(fabs) ...and then use fabs() for doubles, fabsf() for floats. The compiler will generate the FABS instruction, which is generally 1 cycle on modern x86 FPUs. (Internally, the FPU just masks the bit.)