gzdoom/src/xs_Float.h

224 lines
8.8 KiB
C
Raw Normal View History

// ====================================================================================================================
// ====================================================================================================================
// xs_Float.h
//
// Source: "Know Your FPU: Fixing Floating Fast"
// http://www.stereopsis.com/sree/fpu2006.html
//
// xs_CRoundToInt: Round toward nearest, but ties round toward even (just like FISTP)
// xs_ToInt: Round toward zero, just like the C (int) cast
// xs_FloorToInt: Round down
// xs_CeilToInt: Round up
// xs_RoundToInt: Round toward nearest, but ties round up
// ====================================================================================================================
// ====================================================================================================================
#ifndef _xs_FLOAT_H_
#define _xs_FLOAT_H_
// ====================================================================================================================
// Defines
// ====================================================================================================================
#ifndef _xs_DEFAULT_CONVERSION
#define _xs_DEFAULT_CONVERSION 0
#endif //_xs_DEFAULT_CONVERSION
#if __BIG_ENDIAN__
#define _xs_iexp_ 0
#define _xs_iman_ 1
#else
#define _xs_iexp_ 1 //intel is little endian
#define _xs_iman_ 0
#endif //BigEndian_
#ifdef __GNUC__
#define finline inline
#else
#define finline __forceinline
#endif
union _xs_doubleints
{
real64 val;
uint32 ival[2];
};
#if 0
#define _xs_doublecopysgn(a,b) ((int32*)&a)[_xs_iexp_]&=~(((int32*)&b)[_xs_iexp_]&0x80000000)
#define _xs_doubleisnegative(a) ((((int32*)&a)[_xs_iexp_])|0x80000000)
#endif
// ====================================================================================================================
// Constants
// ====================================================================================================================
const real64 _xs_doublemagic = real64 (6755399441055744.0); //2^52 * 1.5, uses limited precisicion to floor
const real64 _xs_doublemagicdelta = (1.5e-8); //almost .5f = .5f + 1e^(number of exp bit)
const real64 _xs_doublemagicroundeps = (.5f-_xs_doublemagicdelta); //almost .5f = .5f - 1e^(number of exp bit)
// ====================================================================================================================
// Prototypes
// ====================================================================================================================
static int32 xs_CRoundToInt (real64 val, real64 dmr = _xs_doublemagic);
static int32 xs_ToInt (real64 val, real64 dme = -_xs_doublemagicroundeps);
static int32 xs_FloorToInt (real64 val, real64 dme = _xs_doublemagicroundeps);
static int32 xs_CeilToInt (real64 val, real64 dme = _xs_doublemagicroundeps);
static int32 xs_RoundToInt (real64 val);
//int32 versions
finline static int32 xs_CRoundToInt (int32 val) {return val;}
finline static int32 xs_ToInt (int32 val) {return val;}
// ====================================================================================================================
// Fix Class
// ====================================================================================================================
template <int32 N> class xs_Fix
{
public:
typedef int32 Fix;
// ====================================================================================================================
// Basic Conversion from Numbers
// ====================================================================================================================
finline static Fix ToFix (int32 val) {return val<<N;}
finline static Fix ToFix (real64 val) {return xs_ConvertToFixed(val);}
// ====================================================================================================================
// Basic Conversion to Numbers
// ====================================================================================================================
finline static real64 ToReal (Fix f) {return real64(f)/real64(1<<N);}
finline static int32 ToInt (Fix f) {return f>>N;}
protected:
// ====================================================================================================================
// Helper function - mainly to preserve _xs_DEFAULT_CONVERSION
// ====================================================================================================================
finline static int32 xs_ConvertToFixed (real64 val)
{
#if _xs_DEFAULT_CONVERSION==0
return xs_CRoundToInt(val, _xs_doublemagic/(1<<N));
#else
return (long)((val)*(1<<N));
#endif
}
};
// ====================================================================================================================
// ====================================================================================================================
// Inline implementation
// ====================================================================================================================
// ====================================================================================================================
finline static int32 xs_CRoundToInt(real64 val, real64 dmr)
{
#if _xs_DEFAULT_CONVERSION==0
_xs_doubleints uval;
uval.val = val + dmr;
return uval.ival[_xs_iman_];
#else
return int32(floor(val+.5));
#endif
}
// ====================================================================================================================
finline static int32 xs_ToInt(real64 val, real64 dme)
{
/* unused - something else I tried...
_xs_doublecopysgn(dme,val);
return xs_CRoundToInt(val+dme);
return 0;
*/
#if _MSC_VER >= 1400
// VC++ 2005's standard cast is a little bit faster than this
// magic number code. (Which is pretty amazing!) SSE has the
// fastest C-style float->int conversion, but unfortunately,
// checking for SSE support every time you need to do a
// conversion completely negates its performance advantage.
return int32(val);
#else
#if _xs_DEFAULT_CONVERSION==0
return (val<0) ? xs_CRoundToInt(val-dme) :
xs_CRoundToInt(val+dme);
#else
return int32(val);
#endif
#endif
}
// ====================================================================================================================
finline static int32 xs_FloorToInt(real64 val, real64 dme)
{
#if _xs_DEFAULT_CONVERSION==0
return xs_CRoundToInt (val - dme);
#else
return floor(val);
#endif
}
// ====================================================================================================================
finline static int32 xs_CeilToInt(real64 val, real64 dme)
{
#if _xs_DEFAULT_CONVERSION==0
return xs_CRoundToInt (val + dme);
#else
return ceil(val);
#endif
}
// ====================================================================================================================
finline static int32 xs_RoundToInt(real64 val)
{
#if _xs_DEFAULT_CONVERSION==0
// Yes, it is important that two fadds be generated, so you cannot override the dmr
// passed to xs_CRoundToInt with _xs_doublemagic + _xs_doublemagicdelta. If you do,
// you'll end up with Banker's Rounding again.
return xs_CRoundToInt (val + _xs_doublemagicdelta);
#else
return floor(val+.5);
#endif
}
// ====================================================================================================================
// ====================================================================================================================
// Unsigned variants
// ====================================================================================================================
// ====================================================================================================================
finline static uint32 xs_CRoundToUInt(real64 val)
{
return (uint32)xs_CRoundToInt(val);
}
finline static uint32 xs_FloorToUInt(real64 val)
{
return (uint32)xs_FloorToInt(val);
}
finline static uint32 xs_CeilToUInt(real64 val)
{
return (uint32)xs_CeilToInt(val);
}
finline static uint32 xs_RoundToUInt(real64 val)
{
return (uint32)xs_RoundToInt(val);
}
// ====================================================================================================================
// ====================================================================================================================
#endif // _xs_FLOAT_H_