Drop unused and unportable "fast math" code

This commit is contained in:
sfan5 2021-07-16 23:52:26 +02:00
parent 22c3219ff0
commit 1d66c921c2
3 changed files with 2 additions and 252 deletions

@ -779,17 +779,6 @@ ones. */
#undef __IRR_COMPILE_WITH_WAD_ARCHIVE_LOADER_
#endif
//! Set FPU settings
/** Irrlicht should use approximate float and integer fpu techniques
precision will be lower but speed higher. currently X86 only
*/
#if !defined(_IRR_OSX_PLATFORM_) && !defined(_IRR_SOLARIS_PLATFORM_)
//#define IRRLICHT_FAST_MATH
#ifdef NO_IRRLICHT_FAST_MATH
#undef IRRLICHT_FAST_MATH
#endif
#endif
// Some cleanup and standard stuff
#ifdef _IRR_WINDOWS_API_

@ -12,29 +12,6 @@
#include <stdlib.h> // for abs() etc.
#include <limits.h> // For INT_MAX / UINT_MAX
#if defined(_IRR_SOLARIS_PLATFORM_) || defined(__BORLANDC__) || defined (__BCPLUSPLUS__) || defined (_WIN32_WCE)
#define sqrtf(X) (irr::f32)sqrt((irr::f64)(X))
#define sinf(X) (irr::f32)sin((irr::f64)(X))
#define cosf(X) (irr::f32)cos((irr::f64)(X))
#define asinf(X) (irr::f32)asin((irr::f64)(X))
#define acosf(X) (irr::f32)acos((irr::f64)(X))
#define atan2f(X,Y) (irr::f32)atan2((irr::f64)(X),(irr::f64)(Y))
#define ceilf(X) (irr::f32)ceil((irr::f64)(X))
#define floorf(X) (irr::f32)floor((irr::f64)(X))
#define powf(X,Y) (irr::f32)pow((irr::f64)(X),(irr::f64)(Y))
#define fmodf(X,Y) (irr::f32)fmod((irr::f64)(X),(irr::f64)(Y))
#define fabsf(X) (irr::f32)fabs((irr::f64)(X))
#define logf(X) (irr::f32)log((irr::f64)(X))
#endif
#ifndef FLT_MAX
#define FLT_MAX 3.402823466E+38F
#endif
#ifndef FLT_MIN
#define FLT_MIN 1.17549435e-38F
#endif
namespace irr
{
namespace core
@ -379,45 +356,14 @@ namespace core
#define F32_VALUE_0 0x00000000
#define F32_VALUE_1 0x3f800000
#define F32_SIGN_BIT 0x80000000U
#define F32_EXPON_MANTISSA 0x7FFFFFFFU
//! code is taken from IceFPU
//! Integer representation of a floating-point value.
#ifdef IRRLICHT_FAST_MATH
#define IR(x) ((u32&)(x))
#else
inline u32 IR(f32 x) {inttofloat tmp; tmp.f=x; return tmp.u;}
#endif
//! Absolute integer representation of a floating-point value
#define AIR(x) (IR(x)&0x7fffffff)
//! Floating-point representation of an integer value.
#ifdef IRRLICHT_FAST_MATH
#define FR(x) ((f32&)(x))
#else
inline f32 FR(u32 x) {inttofloat tmp; tmp.u=x; return tmp.f;}
inline f32 FR(s32 x) {inttofloat tmp; tmp.s=x; return tmp.f;}
#endif
//! integer representation of 1.0
#define IEEE_1_0 0x3f800000
//! integer representation of 255.0
#define IEEE_255_0 0x437f0000
#ifdef IRRLICHT_FAST_MATH
#define F32_LOWER_0(f) (F32_AS_U32(f) > F32_SIGN_BIT)
#define F32_LOWER_EQUAL_0(f) (F32_AS_S32(f) <= F32_VALUE_0)
#define F32_GREATER_0(f) (F32_AS_S32(f) > F32_VALUE_0)
#define F32_GREATER_EQUAL_0(f) (F32_AS_U32(f) <= F32_SIGN_BIT)
#define F32_EQUAL_1(f) (F32_AS_U32(f) == F32_VALUE_1)
#define F32_EQUAL_0(f) ( (F32_AS_U32(f) & F32_EXPON_MANTISSA ) == F32_VALUE_0)
// only same sign
#define F32_A_GREATER_B(a,b) (F32_AS_S32((a)) > F32_AS_S32((b)))
#else
#define F32_LOWER_0(n) ((n) < 0.0f)
#define F32_LOWER_EQUAL_0(n) ((n) <= 0.0f)
@ -426,8 +372,6 @@ namespace core
#define F32_EQUAL_1(n) ((n) == 1.0f)
#define F32_EQUAL_0(n) ((n) == 0.0f)
#define F32_A_GREATER_B(a,b) ((a) > (b))
#endif
#ifndef REALINLINE
#ifdef _MSC_VER
@ -437,23 +381,6 @@ namespace core
#endif
#endif
#if defined(__BORLANDC__) || defined (__BCPLUSPLUS__)
// 8-bit bools in Borland builder
//! conditional set based on mask and arithmetic shift
REALINLINE u32 if_c_a_else_b ( const c8 condition, const u32 a, const u32 b )
{
return ( ( -condition >> 7 ) & ( a ^ b ) ) ^ b;
}
//! conditional set based on mask and arithmetic shift
REALINLINE u32 if_c_a_else_0 ( const c8 condition, const u32 a )
{
return ( -condition >> 31 ) & a;
}
#else
//! conditional set based on mask and arithmetic shift
REALINLINE u32 if_c_a_else_b ( const s32 condition, const u32 a, const u32 b )
{
@ -471,7 +398,6 @@ namespace core
{
return ( -condition >> 31 ) & a;
}
#endif
/*
if (condition) state |= m; else state &= ~m;
@ -526,30 +452,7 @@ namespace core
// calculate: 1 / sqrtf ( x )
REALINLINE f32 reciprocal_squareroot(const f32 f)
{
#if defined ( IRRLICHT_FAST_MATH )
// NOTE: Unlike comment below says I found inaccuracies already at 4'th significant bit.
// p.E: Input 1, expected 1, got 0.999755859
#if defined(_MSC_VER) && !defined(_WIN64)
// SSE reciprocal square root estimate, accurate to 12 significant
// bits of the mantissa
f32 recsqrt;
__asm rsqrtss xmm0, f // xmm0 = rsqrtss(f)
__asm movss recsqrt, xmm0 // return xmm0
return recsqrt;
/*
// comes from Nvidia
u32 tmp = (u32(IEEE_1_0 << 1) + IEEE_1_0 - *(u32*)&x) >> 1;
f32 y = *(f32*)&tmp;
return y * (1.47f - 0.47f * x * y * y);
*/
#else
return 1.f / sqrtf(f);
#endif
#else // no fast math
return 1.f / sqrtf(f);
#endif
}
// calculate: 1 / sqrtf( x )
@ -561,37 +464,7 @@ namespace core
// calculate: 1 / x
REALINLINE f32 reciprocal( const f32 f )
{
#if defined (IRRLICHT_FAST_MATH)
// NOTE: Unlike with 1.f / f the values very close to 0 return -nan instead of inf
// SSE Newton-Raphson reciprocal estimate, accurate to 23 significant
// bi ts of the mantissa
// One Newton-Raphson Iteration:
// f(i+1) = 2 * rcpss(f) - f * rcpss(f) * rcpss(f)
#if defined(_MSC_VER) && !defined(_WIN64)
f32 rec;
__asm rcpss xmm0, f // xmm0 = rcpss(f)
__asm movss xmm1, f // xmm1 = f
__asm mulss xmm1, xmm0 // xmm1 = f * rcpss(f)
__asm mulss xmm1, xmm0 // xmm2 = f * rcpss(f) * rcpss(f)
__asm addss xmm0, xmm0 // xmm0 = 2 * rcpss(f)
__asm subss xmm0, xmm1 // xmm0 = 2 * rcpss(f)
// - f * rcpss(f) * rcpss(f)
__asm movss rec, xmm0 // return xmm0
return rec;
#else // no support yet for other compilers
return 1.f / f;
#endif
//! i do not divide through 0.. (fpu expection)
// instead set f to a high value to get a return value near zero..
// -1000000000000.f.. is use minus to stay negative..
// must test's here (plane.normal dot anything ) checks on <= 0.f
//u32 x = (-(AIR(f) != 0 ) >> 31 ) & ( IR(f) ^ 0xd368d4a5 ) ^ 0xd368d4a5;
//return 1.f / FR ( x );
#else // no fast math
return 1.f / f;
#endif
}
// calculate: 1 / x
@ -604,45 +477,9 @@ namespace core
// calculate: 1 / x, low precision allowed
REALINLINE f32 reciprocal_approxim ( const f32 f )
{
#if defined( IRRLICHT_FAST_MATH)
// SSE Newton-Raphson reciprocal estimate, accurate to 23 significant
// bi ts of the mantissa
// One Newton-Raphson Iteration:
// f(i+1) = 2 * rcpss(f) - f * rcpss(f) * rcpss(f)
#if defined(_MSC_VER) && !defined(_WIN64)
f32 rec;
__asm rcpss xmm0, f // xmm0 = rcpss(f)
__asm movss xmm1, f // xmm1 = f
__asm mulss xmm1, xmm0 // xmm1 = f * rcpss(f)
__asm mulss xmm1, xmm0 // xmm2 = f * rcpss(f) * rcpss(f)
__asm addss xmm0, xmm0 // xmm0 = 2 * rcpss(f)
__asm subss xmm0, xmm1 // xmm0 = 2 * rcpss(f)
// - f * rcpss(f) * rcpss(f)
__asm movss rec, xmm0 // return xmm0
return rec;
#else // no support yet for other compilers
return 1.f / f;
#endif
/*
// SSE reciprocal estimate, accurate to 12 significant bits of
f32 rec;
__asm rcpss xmm0, f // xmm0 = rcpss(f)
__asm movss rec , xmm0 // return xmm0
return rec;
*/
/*
u32 x = 0x7F000000 - IR ( p );
const f32 r = FR ( x );
return r * (2.0f - p * r);
*/
#else // no fast math
return 1.f / f;
#endif
}
REALINLINE s32 floor32(f32 x)
{
return (s32) floorf ( x );
@ -677,9 +514,7 @@ namespace core
} // end namespace core
} // end namespace irr
#ifndef IRRLICHT_FAST_MATH
using irr::core::IR;
using irr::core::FR;
#endif
using irr::core::IR;
using irr::core::FR;
#endif

@ -24,33 +24,7 @@ namespace core
// Input -1.40129846e-45, expected -1, got 0
REALINLINE s32 floor32_fast(f32 x)
{
#ifdef IRRLICHT_FAST_MATH
const f32 h = 0.5f;
s32 t;
#if defined(_MSC_VER) && !defined(_WIN64)
__asm
{
fld x
fsub h
fistp t
}
#elif defined(__GNUC__)
__asm__ __volatile__ (
"fsub %2 \n\t"
"fistpl %0"
: "=m" (t)
: "t" (x), "f" (h)
: "st"
);
#else
return (s32) floorf ( x );
#endif
return t;
#else // no fast math
return (s32) floorf ( x );
#endif
}
// Some examples for unexpected results when using this with IRRLICHT_FAST_MATH:
@ -59,33 +33,7 @@ namespace core
// Input -3, expected -3, got -2
REALINLINE s32 ceil32_fast ( f32 x )
{
#ifdef IRRLICHT_FAST_MATH
const f32 h = 0.5f;
s32 t;
#if defined(_MSC_VER) && !defined(_WIN64)
__asm
{
fld x
fadd h
fistp t
}
#elif defined(__GNUC__)
__asm__ __volatile__ (
"fadd %2 \n\t"
"fistpl %0 \n\t"
: "=m"(t)
: "t"(x), "f"(h)
: "st"
);
#else
return (s32) ceilf ( x );
#endif
return t;
#else // not fast math
return (s32) ceilf ( x );
#endif
}
// Some examples for unexpected results when using this with IRRLICHT_FAST_MATH:
@ -95,29 +43,7 @@ namespace core
// Input -2.80259693e-45, expected -nan(ind), got -inf
REALINLINE s32 round32_fast(f32 x)
{
#if defined(IRRLICHT_FAST_MATH)
s32 t;
#if defined(_MSC_VER) && !defined(_WIN64)
__asm
{
fld x
fistp t
}
#elif defined(__GNUC__)
__asm__ __volatile__ (
"fistpl %0 \n\t"
: "=m"(t)
: "t"(x)
: "st"
);
#else
return (s32) round_(x);
#endif
return t;
#else // no fast math
return (s32) round_(x);
#endif
}
} // end namespace core