Add fast functions for exp2, pow, log2, root, sin/cos/tan, asin/acos/atan along with abs, neg and sgn.

This commit is contained in:
ehofman
2004-05-08 12:58:29 +00:00
parent ffada9257d
commit 634a2035ee
2 changed files with 282 additions and 8 deletions

View File

@@ -9,9 +9,17 @@
* IDSIA, Lugano, Switzerland
* http://www.inf.ethz.ch/~schraudo/pubs/exp.pdf
*
* Base-2 exp, Laurent de Soras
* http://www.musicdsp.org/archive.php?classid=5#106
*
* Fast log() Function, by Laurent de Soras:
* http://www.flipcode.com/cgi-bin/msg.cgi?showThread=Tip-Fastlogfunction&forum=totd&id=-1
*
* Sin, Cos, Tan approximation
* http://www.musicdsp.org/showArchiveComment.php?ArchiveID=115
*
* fast floating point power computation:
* http://playstation2-linux.com/download/adam/power.c
*/
/*
@@ -21,6 +29,7 @@
#include "fastmath.hxx"
#define SGD_PI_2 1.57079632679489661923
/**
* This function is on avarage 9 times faster than the system exp() function
@@ -46,6 +55,71 @@ double fast_exp(double val) {
return _eco.d;
}
/*
* Linear approx. between 2 integer values of val. Uses 32-bit integers.
* Not very efficient but faster than exp()
*/
double fast_exp2( const double val )
{
int e;
double ret;
if (val >= 0) {
e = int (val);
ret = val - (e - 1);
#if BYTE_ORDER == BIG_ENDIAN
((*((int *) &ret)) &= ~(2047 << 20)) += (e + 1023) << 20;
#else
((*(1 + (int *) &ret)) &= ~(2047 << 20)) += (e + 1023) << 20;
#endif
} else {
e = int (val + 1023);
ret = val - (e - 1024);
#if BYTE_ORDER == BIG_ENDIAN
((*((int *) &ret)) &= ~(2047 << 20)) += e << 20;
#else
((*(1 + (int *) &ret)) &= ~(2047 << 20)) += e << 20;
#endif
}
return ret;
}
/*
*
*/
float _fast_log2(const float val)
{
float result, tmp;
float mp = 0.346607f;
result = *(int*)&val;
result *= 1.0/(1<<23);
result = result - 127;
tmp = result - floor(result);
tmp = (tmp - tmp*tmp) * mp;
return tmp + result;
}
float _fast_pow2(const float val)
{
float result;
float mp = 0.33971f;
float tmp = val - floor(val);
tmp = (tmp - tmp*tmp) * mp;
result = val + 127 - tmp;
result *= (1<<23);
*(int*)&result = (int)result;
return result;
}
/**
* While we're on the subject, someone might have use for these as well?
@@ -63,3 +137,155 @@ void fast_BSR(float &x, register unsigned long shiftAmount) {
}
/*
* fastpow(f,n) gives a rather *rough* estimate of a float number f to the
* power of an integer number n (y=f^n). It is fast but result can be quite a
* bit off, since we directly mess with the floating point exponent.
*
* Use it only for getting rough estimates of the values and where precision
* is not that important.
*/
float fast_pow(const float f, const int n)
{
long *lp,l;
lp=(long*)(&f);
l=*lp;l-=0x3F800000l;l<<=(n-1);l+=0x3F800000l;
*lp=l;
return f;
}
float fast_root(const float f, const int n)
{
long *lp,l;
lp=(long*)(&f);
l=*lp;l-=0x3F800000l;l>>=(n-1);l+=0x3F800000l;
*lp=l;
return f;
}
/*
* Code for approximation of cos, sin, tan and inv sin, etc.
* Surprisingly accurate and very usable.
*
* Domain:
* Sin/Cos [0, pi/2]
* Tan [0,pi/4]
* InvSin/Cos [0, 1]
* InvTan [-1, 1]
*/
float fast_sin(const float val)
{
float fASqr = val*val;
float fResult = -2.39e-08f;
fResult *= fASqr;
fResult += 2.7526e-06f;
fResult *= fASqr;
fResult -= 1.98409e-04f;
fResult *= fASqr;
fResult += 8.3333315e-03f;
fResult *= fASqr;
fResult -= 1.666666664e-01f;
fResult *= fASqr;
fResult += 1.0f;
fResult *= val;
return fResult;
}
float fast_cos(const float val)
{
float fASqr = val*val;
float fResult = -2.605e-07f;
fResult *= fASqr;
fResult += 2.47609e-05f;
fResult *= fASqr;
fResult -= 1.3888397e-03f;
fResult *= fASqr;
fResult += 4.16666418e-02f;
fResult *= fASqr;
fResult -= 4.999999963e-01f;
fResult *= fASqr;
fResult += 1.0f;
return fResult;
}
float fast_tan(const float val)
{
float fASqr = val*val;
float fResult = 9.5168091e-03f;
fResult *= fASqr;
fResult += 2.900525e-03f;
fResult *= fASqr;
fResult += 2.45650893e-02f;
fResult *= fASqr;
fResult += 5.33740603e-02f;
fResult *= fASqr;
fResult += 1.333923995e-01f;
fResult *= fASqr;
fResult += 3.333314036e-01f;
fResult *= fASqr;
fResult += 1.0f;
fResult *= val;
return fResult;
}
float fast_asin(float val)
{
float fRoot = sqrt(1.0f-val);
float fResult = -0.0187293f;
fResult *= val;
fResult += 0.0742610f;
fResult *= val;
fResult -= 0.2121144f;
fResult *= val;
fResult += 1.5707288f;
fResult = SGD_PI_2 - fRoot*fResult;
return fResult;
}
float fast_acos(float val)
{
float fRoot = sqrt(1.0f-val);
float fResult = -0.0187293f;
fResult *= val;
fResult += 0.0742610f;
fResult *= val;
fResult -= 0.2121144f;
fResult *= val;
fResult += 1.5707288f;
fResult *= fRoot;
return fResult;
}
float fast_atan(float val)
{
float fVSqr = val*val;
float fResult = 0.0028662257f;
fResult *= fVSqr;
fResult -= 0.0161657367f;
fResult *= fVSqr;
fResult += 0.0429096138f;
fResult *= fVSqr;
fResult -= 0.0752896400f;
fResult *= fVSqr;
fResult += 0.1065626393f;
fResult *= fVSqr;
fResult -= 0.1420889944f;
fResult *= fVSqr;
fResult += 0.1999355085f;
fResult *= fVSqr;
fResult -= 0.3333314528f;
fResult *= fVSqr;
fResult += 1.0f;
fResult *= val;
return fResult;
}

View File

@@ -32,24 +32,41 @@
double fast_exp(double val);
double fast_exp2(const double val);
float fast_pow(const float val1, const float val2);
float fast_log2(const float cal);
float fast_root(const float f, const int n);
float _fast_pow2(const float cal);
float _fast_log2(const float val);
float fast_sin(const float val);
float fast_cos(const float val);
float fast_tan(const float val);
float fast_asin(const float val);
float fast_acos(const float val);
float fast_atan(const float val);
void fast_BSL(float &x, register unsigned long shiftAmount);
void fast_BSR(float &x, register unsigned long shiftAmount);
inline float fast_log2 (float val)
{
int * const exp_ptr = reinterpret_cast <int *> (&val);
int x = *exp_ptr;
const int log_2 = ((x >> 23) & 255) - 128;
x &= ~(255 << 23);
x += 127 << 23;
*exp_ptr = x;
int * const exp_ptr = reinterpret_cast <int *> (&val);
int x = *exp_ptr;
const int log_2 = ((x >> 23) & 255) - 128;
x &= ~(255 << 23);
x += 127 << 23;
*exp_ptr = x;
val = ((-1.0f/3) * val + 2) * val - 2.0f/3; // (1)
val = ((-1.0f/3) * val + 2) * val - 2.0f/3; // (1)
return (val + log_2);
return (val + log_2);
}
/**
* This function is about 3 times faster than the system log() function
* and has an error of about 0.01%
@@ -65,5 +82,36 @@ inline float fast_log10 (const float &val)
}
/**
* This function is about twice as fast as the system pow(x,y) function
*/
inline float fast_pow(const float val1, const float val2)
{
return _fast_pow2(val2 * _fast_log2(val1));
}
/*
* 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.
*/
inline float fast_abs(float f)
{
int i=((*(int*)&f)&0x7fffffff);
return (*(float*)&i);
}
inline float fast_neg(float f)
{
int i=((*(int*)&f)^0x80000000);
return (*(float*)&i);
}
inline int fast_sgn(float f)
{
return 1+(((*(int*)&f)>>31)<<1);
}
#endif // !_SG_FMATH_HXX