forked from Mirrors/openclonk
Use Julien Pommier's SSE vector math functions
parent
a858678248
commit
27b544e778
|
@ -526,6 +526,7 @@ set(OC_CLONK_SOURCES
|
|||
src/tinyxml/tinyxmlparser.cpp
|
||||
src/zlib/gzio.c
|
||||
src/zlib/zutil.h
|
||||
thirdparty/simdmath/sse_mathfun.h
|
||||
)
|
||||
mark_as_advanced(OC_CLONK_SOURCES)
|
||||
mark_as_advanced(OC_SYSTEM_SOURCES)
|
||||
|
@ -663,6 +664,7 @@ include_directories(
|
|||
${CMAKE_CURRENT_SOURCE_DIR}/src/game/object
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/src/lib/texture
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/src/script
|
||||
${CMAKE_CURRENT_SOURCE_DIR}/thirdparty
|
||||
)
|
||||
|
||||
############################################################################
|
||||
|
|
|
@ -55,8 +55,10 @@
|
|||
#define C4REAL_MODE C4REAL_MODE_SSE_FLOAT
|
||||
#endif
|
||||
|
||||
#include <boost/utility/enable_if.hpp>
|
||||
#include <boost/type_traits/is_class.hpp>
|
||||
#include <boost/type_traits/is_pod.hpp>
|
||||
#include <boost/type_traits/is_arithmetic.hpp>
|
||||
|
||||
template<class C4RealImpl>
|
||||
class C4RealBase
|
||||
|
@ -65,14 +67,17 @@ class C4RealBase
|
|||
|
||||
friend C4RealBase Sin(const C4RealBase &);
|
||||
friend C4RealBase Cos(const C4RealBase &);
|
||||
friend typename C4RealImpl;
|
||||
|
||||
public:
|
||||
inline C4RealBase(int32_t val = 0) : value(val) { }
|
||||
inline C4RealBase(float val) : value(val) {}
|
||||
inline C4RealBase(int32_t val, int32_t prec) : value(val) { operator/=(C4RealBase(prec)); }
|
||||
template <class T>
|
||||
inline C4RealBase(typename boost::enable_if<boost::is_arithmetic<T>, T>::type val) : value(val) { }
|
||||
inline C4RealBase(const C4RealImpl &val) : value(val) { }
|
||||
// Conversion between different implementations of C4RealBase
|
||||
template<class T>
|
||||
inline C4RealBase(const C4RealBase<T> &val) : value(static_cast<float>(val)) { }
|
||||
/* template<class T>
|
||||
inline C4RealBase(const C4RealBase<T> &val) : value(static_cast<float>(val)) { }*/
|
||||
|
||||
// Copy ctor and assignment
|
||||
inline C4RealBase(const C4RealBase &rhs) : value(rhs.value) {}
|
||||
|
|
|
@ -31,11 +31,11 @@
|
|||
|
||||
inline C4Real_FPU_Float Sin(const C4Real_FPU_Float &real)
|
||||
{
|
||||
return C4Real_FPU_Float(std::sin(real.value * static_cast<float>(M_PI) / 180.0f));
|
||||
return std::sin(real.value * static_cast<float>(M_PI) / 180.0f);
|
||||
}
|
||||
inline C4Real_FPU_Float Cos(const C4Real_FPU_Float &real)
|
||||
{
|
||||
return C4Real_FPU_Float(std::cos(real.value * static_cast<float>(M_PI) / 180.0f));
|
||||
return std::cos(real.value * static_cast<float>(M_PI) / 180.0f);
|
||||
}
|
||||
|
||||
// Overload to avoid conversion warning
|
||||
|
|
|
@ -18,101 +18,28 @@
|
|||
#include "C4Include.h"
|
||||
#include "C4Real.h"
|
||||
|
||||
// Constants required for sine approximation
|
||||
const __m128 C4RealImpl_SSE::cephes_deg2rad = _mm_set_ps1(0.017453292f); // pi/180
|
||||
const __m128 C4RealImpl_SSE::cephes_FOPI = _mm_set_ps1(1.27323954473516f); // 4/pi
|
||||
const __m128 C4RealImpl_SSE::cephes_scaling_factors[3] = {
|
||||
_mm_set_ps1(0.78515625f), _mm_set_ps1(2.4187564849853515625e-4f), _mm_set_ps1(3.77489497744594108e-8f)
|
||||
};
|
||||
const __m128 C4RealImpl_SSE::cephes_appx_coeffs[3] = {
|
||||
_mm_set_ps(-1.9515295891e-4f, 2.443315711809948e-5f, 2.443315711809948e-5f, -1.9515295891e-4f),
|
||||
_mm_set_ps(8.3321608736e-3f, -1.388731625493765e-3f, -1.388731625493765e-3f, 8.3321608736e-3f),
|
||||
_mm_set_ps(-1.6666654611e-1f, 4.166664568298827e-2f, 4.166664568298827e-2f, -1.6666654611e-1f)
|
||||
};
|
||||
const __m128 C4RealImpl_SSE::iee754_sign_mask = _mm_set_ps1(-0.0f);
|
||||
|
||||
// Branchless approximation of sine and cosine values
|
||||
C4RealImpl_SSE C4RealImpl_SSE::SinCos(bool cosine) const
|
||||
namespace
|
||||
{
|
||||
// NOTE: Consider storing radians directly instead of degrees to avoid
|
||||
// precision loss due to conversion
|
||||
__m128 radian = _mm_mul_ps(value, cephes_deg2rad);
|
||||
#ifndef NDEBUG
|
||||
float float_radians; _mm_store_ss(&float_radians, radian);
|
||||
#ifdef _MSC_VER
|
||||
# pragma warning(push)
|
||||
# pragma warning(disable:4305) // 'identifier' : truncation from 'type1' to 'type2'
|
||||
#endif
|
||||
// put rad into all parts of the xmm register
|
||||
radian = _mm_shuffle_ps(radian, radian, 0);
|
||||
|
||||
// cephes library sine/cosine implementation
|
||||
// This calculates a Taylor approximation of rank 7 with slightly
|
||||
// modified coefficients to achieve better precision on the reduced
|
||||
// input range -pi/4..pi/4.
|
||||
|
||||
union m128extract {
|
||||
float f[4];
|
||||
uint32_t i[4];
|
||||
__m128 v;
|
||||
};
|
||||
|
||||
// Store sign and take absolute value of input
|
||||
__m128 sign = _mm_and_ps(radian, iee754_sign_mask);
|
||||
radian = _mm_xor_ps(radian, sign);
|
||||
m128extract sign_bits;
|
||||
sign_bits.v = sign;
|
||||
|
||||
// Select octant of the unit circle
|
||||
__m128 scaling = _mm_mul_ps(radian, C4RealImpl_SSE::cephes_FOPI);
|
||||
int octant = _mm_cvttss_si32(scaling);
|
||||
octant = (octant + 1) & ~1;
|
||||
scaling = _mm_cvtsi32_ss(scaling, octant);
|
||||
scaling = _mm_shuffle_ps(scaling, scaling, 0);
|
||||
uint32_t flip_sign_sine = ((octant & 4) << 29);
|
||||
octant &= 3;
|
||||
uint32_t flip_sign_cosine = flip_sign_sine ^ ((octant & 2) << 30);
|
||||
flip_sign_sine ^= sign_bits.i[0];
|
||||
|
||||
// map input to +-pi/4
|
||||
// note that this get more and more imprecise for abs(radian) > 8192
|
||||
radian = _mm_sub_ps(radian, _mm_mul_ps(scaling, cephes_scaling_factors[0]));
|
||||
radian = _mm_sub_ps(radian, _mm_mul_ps(scaling, cephes_scaling_factors[1]));
|
||||
radian = _mm_sub_ps(radian, _mm_mul_ps(scaling, cephes_scaling_factors[2]));
|
||||
|
||||
// run approximation, calculating four octants at once; correct result
|
||||
// will be selected later
|
||||
__m128 radiansq = _mm_mul_ps(radian, radian);
|
||||
__m128 result = cephes_appx_coeffs[0];
|
||||
result = _mm_mul_ps(result, radiansq);
|
||||
result = _mm_add_ps(result, cephes_appx_coeffs[1]);
|
||||
result = _mm_mul_ps(result, radiansq);
|
||||
result = _mm_add_ps(result, cephes_appx_coeffs[2]);
|
||||
result = _mm_mul_ps(result, radiansq);
|
||||
|
||||
radiansq = _mm_shuffle_ps(radiansq, radian, _MM_SHUFFLE(0,0,0,0));
|
||||
radiansq = _mm_shuffle_ps(radiansq, radiansq, _MM_SHUFFLE(2,0,0,2));
|
||||
result = _mm_mul_ps(result, radiansq);
|
||||
|
||||
radiansq = _mm_mul_ps(radiansq, _mm_set_ps(1.0f, -0.5f, -0.5f, 1.0f));
|
||||
result = _mm_add_ps(result, radiansq);
|
||||
result = _mm_add_ps(result, _mm_set_ps(0.0f, 1.0f, 1.0f, 0.0f));
|
||||
|
||||
// Select correct octant
|
||||
m128extract rv;
|
||||
rv.v = result;
|
||||
int sinidx = octant;
|
||||
int cosidx = octant ^ 2;
|
||||
|
||||
// adjust sign
|
||||
rv.i[sinidx] ^= flip_sign_sine;
|
||||
rv.i[cosidx] ^= flip_sign_cosine;
|
||||
|
||||
// relative error less than 1.1e-6 for input values between -360 deg
|
||||
// and 360 deg
|
||||
// relative error less than 1.2e-7 between -260 deg and 260 deg.
|
||||
// absolute error less than 6.0e-8 between -4.2e7 and 4.2e7 deg.
|
||||
assert(float_radians == 0.0f || std::abs((rv.f[sinidx] - std::sin(float_radians)) / std::sin(float_radians)) < 1.1e-6f);
|
||||
assert(float_radians == 0.0f || std::abs((rv.f[cosidx] - std::cos(float_radians)) / std::cos(float_radians)) < 1.1e-6f);
|
||||
|
||||
uint32_t cosine_mask = cosine * ~0;
|
||||
int idx = (cosine_mask & cosidx) | (~cosine_mask & sinidx);
|
||||
return rv.f[idx];
|
||||
#ifdef _M_X64
|
||||
# define USE_SSE2
|
||||
#endif
|
||||
#include "simdmath/sse_mathfun.h"
|
||||
#ifdef _MSC_VER
|
||||
# pragma warning(pop)
|
||||
#endif
|
||||
const __m128 deg2rad = _mm_set_ps1(0.017453292f);
|
||||
}
|
||||
|
||||
C4Real_SSE_Float Sin(const C4Real_SSE_Float &real)
|
||||
{
|
||||
return C4RealImpl_SSE(sin_ps(_mm_mul_ps(real.value.value, deg2rad)));
|
||||
}
|
||||
C4Real_SSE_Float Cos(const C4Real_SSE_Float &real)
|
||||
{
|
||||
return C4RealImpl_SSE(cos_ps(_mm_mul_ps(real.value.value, deg2rad)));
|
||||
}
|
||||
|
||||
|
|
|
@ -20,6 +20,7 @@
|
|||
|
||||
#ifndef INC_C4Real
|
||||
#error C4RealImpl_SSE.h must not be included by itself; include C4Real.h instead
|
||||
#include "C4Real.h"
|
||||
#endif
|
||||
|
||||
#include <cassert>
|
||||
|
@ -31,12 +32,9 @@ class C4RealImpl_SSE
|
|||
friend C4Real_SSE_Float Cos(const C4Real_SSE_Float &);
|
||||
__m128 value;
|
||||
|
||||
static const __m128 iee754_sign_mask; // -0.0
|
||||
static const __m128 cephes_FOPI; // 4/pi
|
||||
static const __m128 cephes_deg2rad; // pi/180
|
||||
static const __m128 cephes_appx_coeffs[3]; // approximation coefficients
|
||||
static const __m128 cephes_scaling_factors[3]; // factors for quick scaling to -pi/4..pi/4
|
||||
C4RealImpl_SSE SinCos(bool cosine) const; // approximation of sine and cosine
|
||||
inline C4RealImpl_SSE(__m128 rhs)
|
||||
: value(rhs)
|
||||
{}
|
||||
|
||||
public:
|
||||
inline C4RealImpl_SSE()
|
||||
|
@ -51,6 +49,9 @@ public:
|
|||
inline C4RealImpl_SSE(float fVal)
|
||||
: value(_mm_set_ss(fVal))
|
||||
{}
|
||||
inline C4RealImpl_SSE(const C4Real_SSE_Float &rhs)
|
||||
: value(rhs.value.value)
|
||||
{}
|
||||
|
||||
operator int () const
|
||||
{
|
||||
|
@ -111,14 +112,4 @@ public:
|
|||
operator bool () const { return _mm_comineq_ss(value, _mm_setzero_ps()) != 0; }
|
||||
bool operator ! () const { return _mm_comieq_ss(value, _mm_setzero_ps()) != 0; }
|
||||
};
|
||||
|
||||
inline C4Real_SSE_Float Sin(const C4Real_SSE_Float &real)
|
||||
{
|
||||
return C4Real_SSE_Float(static_cast<float>(real.value.SinCos(false)));
|
||||
}
|
||||
inline C4Real_SSE_Float Cos(const C4Real_SSE_Float &real)
|
||||
{
|
||||
return C4Real_SSE_Float(static_cast<float>(real.value.SinCos(true)));
|
||||
}
|
||||
|
||||
#endif
|
||||
|
|
|
@ -0,0 +1,20 @@
|
|||
Copyright (C) 2007 Julien Pommier
|
||||
|
||||
This software is provided 'as-is', without any express or implied
|
||||
warranty. In no event will the authors be held liable for any damages
|
||||
arising from the use of this software.
|
||||
|
||||
Permission is granted to anyone to use this software for any purpose,
|
||||
including commercial applications, and to alter it and redistribute it
|
||||
freely, subject to the following restrictions:
|
||||
|
||||
1. The origin of this software must not be misrepresented; you must not
|
||||
claim that you wrote the original software. If you use this software
|
||||
in a product, an acknowledgment in the product documentation would be
|
||||
appreciated but is not required.
|
||||
2. Altered source versions must be plainly marked as such, and must not be
|
||||
misrepresented as being the original software.
|
||||
3. This notice may not be removed or altered from any source distribution.
|
||||
|
||||
(this is the zlib license)
|
||||
|
|
@ -0,0 +1,760 @@
|
|||
/* SIMD (SSE1+MMX or SSE2) implementation of sin, cos, exp and log
|
||||
|
||||
Inspired by Intel Approximate Math library, and based on the
|
||||
corresponding algorithms of the cephes math library
|
||||
|
||||
The default is to use the SSE1 version. If you define USE_SSE2 the
|
||||
the SSE2 intrinsics will be used in place of the MMX intrinsics. Do
|
||||
not expect any significant performance improvement with SSE2.
|
||||
*/
|
||||
|
||||
/* Copyright (C) 2007 Julien Pommier
|
||||
|
||||
This software is provided 'as-is', without any express or implied
|
||||
warranty. In no event will the authors be held liable for any damages
|
||||
arising from the use of this software.
|
||||
|
||||
Permission is granted to anyone to use this software for any purpose,
|
||||
including commercial applications, and to alter it and redistribute it
|
||||
freely, subject to the following restrictions:
|
||||
|
||||
1. The origin of this software must not be misrepresented; you must not
|
||||
claim that you wrote the original software. If you use this software
|
||||
in a product, an acknowledgment in the product documentation would be
|
||||
appreciated but is not required.
|
||||
2. Altered source versions must be plainly marked as such, and must not be
|
||||
misrepresented as being the original software.
|
||||
3. This notice may not be removed or altered from any source distribution.
|
||||
|
||||
(this is the zlib license)
|
||||
*/
|
||||
|
||||
#include <xmmintrin.h>
|
||||
|
||||
/* yes I know, the top of this file is quite ugly */
|
||||
|
||||
#ifdef _MSC_VER /* visual c++ */
|
||||
# define ALIGN16_BEG __declspec(align(16))
|
||||
# define ALIGN16_END
|
||||
#else /* gcc or icc */
|
||||
# define ALIGN16_BEG
|
||||
# define ALIGN16_END __attribute__((aligned(16)))
|
||||
#endif
|
||||
|
||||
/* __m128 is ugly to write */
|
||||
typedef __m128 v4sf; // vector of 4 float (sse1)
|
||||
|
||||
#ifdef USE_SSE2
|
||||
# include <emmintrin.h>
|
||||
typedef __m128i v4si; // vector of 4 int (sse2)
|
||||
#else
|
||||
typedef __m64 v2si; // vector of 2 int (mmx)
|
||||
#endif
|
||||
|
||||
/* declare some SSE constants -- why can't I figure a better way to do that? */
|
||||
#define _PS_CONST(Name, Val) \
|
||||
static const ALIGN16_BEG float _ps_##Name[4] ALIGN16_END = { Val, Val, Val, Val }
|
||||
#define _PI32_CONST(Name, Val) \
|
||||
static const ALIGN16_BEG int _pi32_##Name[4] ALIGN16_END = { Val, Val, Val, Val }
|
||||
#define _PS_CONST_TYPE(Name, Type, Val) \
|
||||
static const ALIGN16_BEG Type _ps_##Name[4] ALIGN16_END = { Val, Val, Val, Val }
|
||||
|
||||
_PS_CONST(1 , 1.0f);
|
||||
_PS_CONST(0p5, 0.5f);
|
||||
/* the smallest non denormalized float number */
|
||||
_PS_CONST_TYPE(min_norm_pos, int, 0x00800000);
|
||||
_PS_CONST_TYPE(mant_mask, int, 0x7f800000);
|
||||
_PS_CONST_TYPE(inv_mant_mask, int, ~0x7f800000);
|
||||
|
||||
_PS_CONST_TYPE(sign_mask, int, 0x80000000);
|
||||
_PS_CONST_TYPE(inv_sign_mask, int, ~0x80000000);
|
||||
|
||||
_PI32_CONST(1, 1);
|
||||
_PI32_CONST(inv1, ~1);
|
||||
_PI32_CONST(2, 2);
|
||||
_PI32_CONST(4, 4);
|
||||
_PI32_CONST(0x7f, 0x7f);
|
||||
|
||||
_PS_CONST(cephes_SQRTHF, 0.707106781186547524);
|
||||
_PS_CONST(cephes_log_p0, 7.0376836292E-2);
|
||||
_PS_CONST(cephes_log_p1, - 1.1514610310E-1);
|
||||
_PS_CONST(cephes_log_p2, 1.1676998740E-1);
|
||||
_PS_CONST(cephes_log_p3, - 1.2420140846E-1);
|
||||
_PS_CONST(cephes_log_p4, + 1.4249322787E-1);
|
||||
_PS_CONST(cephes_log_p5, - 1.6668057665E-1);
|
||||
_PS_CONST(cephes_log_p6, + 2.0000714765E-1);
|
||||
_PS_CONST(cephes_log_p7, - 2.4999993993E-1);
|
||||
_PS_CONST(cephes_log_p8, + 3.3333331174E-1);
|
||||
_PS_CONST(cephes_log_q1, -2.12194440e-4);
|
||||
_PS_CONST(cephes_log_q2, 0.693359375);
|
||||
|
||||
#if defined (__MINGW32__)
|
||||
|
||||
/* the ugly part below: many versions of gcc used to be completely buggy with respect to some intrinsics
|
||||
The movehl_ps is fixed in mingw 3.4.5, but I found out that all the _mm_cmp* intrinsics were completely
|
||||
broken on my mingw gcc 3.4.5 ...
|
||||
|
||||
Note that the bug on _mm_cmp* does occur only at -O0 optimization level
|
||||
*/
|
||||
|
||||
inline __m128 my_movehl_ps(__m128 a, const __m128 b) {
|
||||
asm (
|
||||
"movhlps %2,%0\n\t"
|
||||
: "=x" (a)
|
||||
: "0" (a), "x"(b)
|
||||
);
|
||||
return a; }
|
||||
#warning "redefined _mm_movehl_ps (see gcc bug 21179)"
|
||||
#define _mm_movehl_ps my_movehl_ps
|
||||
|
||||
inline __m128 my_cmplt_ps(__m128 a, const __m128 b) {
|
||||
asm (
|
||||
"cmpltps %2,%0\n\t"
|
||||
: "=x" (a)
|
||||
: "0" (a), "x"(b)
|
||||
);
|
||||
return a;
|
||||
}
|
||||
inline __m128 my_cmpgt_ps(__m128 a, const __m128 b) {
|
||||
asm (
|
||||
"cmpnleps %2,%0\n\t"
|
||||
: "=x" (a)
|
||||
: "0" (a), "x"(b)
|
||||
);
|
||||
return a;
|
||||
}
|
||||
inline __m128 my_cmpeq_ps(__m128 a, const __m128 b) {
|
||||
asm (
|
||||
"cmpeqps %2,%0\n\t"
|
||||
: "=x" (a)
|
||||
: "0" (a), "x"(b)
|
||||
);
|
||||
return a;
|
||||
}
|
||||
#warning "redefined _mm_cmpxx_ps functions..."
|
||||
#define _mm_cmplt_ps my_cmplt_ps
|
||||
#define _mm_cmpgt_ps my_cmpgt_ps
|
||||
#define _mm_cmpeq_ps my_cmpeq_ps
|
||||
#endif
|
||||
|
||||
#ifndef USE_SSE2
|
||||
typedef union xmm_mm_union {
|
||||
__m128 xmm;
|
||||
__m64 mm[2];
|
||||
} xmm_mm_union;
|
||||
|
||||
#define COPY_XMM_TO_MM(xmm_, mm0_, mm1_) { \
|
||||
xmm_mm_union u; u.xmm = xmm_; \
|
||||
mm0_ = u.mm[0]; \
|
||||
mm1_ = u.mm[1]; \
|
||||
}
|
||||
|
||||
#define COPY_MM_TO_XMM(mm0_, mm1_, xmm_) { \
|
||||
xmm_mm_union u; u.mm[0]=mm0_; u.mm[1]=mm1_; xmm_ = u.xmm; \
|
||||
}
|
||||
|
||||
#endif // USE_SSE2
|
||||
|
||||
/* natural logarithm computed for 4 simultaneous float
|
||||
return NaN for x <= 0
|
||||
*/
|
||||
v4sf log_ps(v4sf x) {
|
||||
#ifdef USE_SSE2
|
||||
v4si emm0;
|
||||
#else
|
||||
v2si mm0, mm1;
|
||||
#endif
|
||||
v4sf one = *(v4sf*)_ps_1;
|
||||
|
||||
v4sf invalid_mask = _mm_cmple_ps(x, _mm_setzero_ps());
|
||||
|
||||
x = _mm_max_ps(x, *(v4sf*)_ps_min_norm_pos); /* cut off denormalized stuff */
|
||||
|
||||
#ifndef USE_SSE2
|
||||
/* part 1: x = frexpf(x, &e); */
|
||||
COPY_XMM_TO_MM(x, mm0, mm1);
|
||||
mm0 = _mm_srli_pi32(mm0, 23);
|
||||
mm1 = _mm_srli_pi32(mm1, 23);
|
||||
#else
|
||||
emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23);
|
||||
#endif
|
||||
/* keep only the fractional part */
|
||||
x = _mm_and_ps(x, *(v4sf*)_ps_inv_mant_mask);
|
||||
x = _mm_or_ps(x, *(v4sf*)_ps_0p5);
|
||||
|
||||
#ifndef USE_SSE2
|
||||
/* now e=mm0:mm1 contain the really base-2 exponent */
|
||||
mm0 = _mm_sub_pi32(mm0, *(v2si*)_pi32_0x7f);
|
||||
mm1 = _mm_sub_pi32(mm1, *(v2si*)_pi32_0x7f);
|
||||
v4sf e = _mm_cvtpi32x2_ps(mm0, mm1);
|
||||
_mm_empty(); /* bye bye mmx */
|
||||
#else
|
||||
emm0 = _mm_sub_epi32(emm0, *(v4si*)_pi32_0x7f);
|
||||
v4sf e = _mm_cvtepi32_ps(emm0);
|
||||
#endif
|
||||
|
||||
e = _mm_add_ps(e, one);
|
||||
|
||||
/* part2:
|
||||
if( x < SQRTHF ) {
|
||||
e -= 1;
|
||||
x = x + x - 1.0;
|
||||
} else { x = x - 1.0; }
|
||||
*/
|
||||
v4sf mask = _mm_cmplt_ps(x, *(v4sf*)_ps_cephes_SQRTHF);
|
||||
v4sf tmp = _mm_and_ps(x, mask);
|
||||
x = _mm_sub_ps(x, one);
|
||||
e = _mm_sub_ps(e, _mm_and_ps(one, mask));
|
||||
x = _mm_add_ps(x, tmp);
|
||||
|
||||
|
||||
v4sf z = _mm_mul_ps(x,x);
|
||||
|
||||
v4sf y = *(v4sf*)_ps_cephes_log_p0;
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p1);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p2);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p3);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p4);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p5);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p6);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p7);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_log_p8);
|
||||
y = _mm_mul_ps(y, x);
|
||||
|
||||
y = _mm_mul_ps(y, z);
|
||||
|
||||
|
||||
tmp = _mm_mul_ps(e, *(v4sf*)_ps_cephes_log_q1);
|
||||
y = _mm_add_ps(y, tmp);
|
||||
|
||||
|
||||
tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
|
||||
y = _mm_sub_ps(y, tmp);
|
||||
|
||||
tmp = _mm_mul_ps(e, *(v4sf*)_ps_cephes_log_q2);
|
||||
x = _mm_add_ps(x, y);
|
||||
x = _mm_add_ps(x, tmp);
|
||||
x = _mm_or_ps(x, invalid_mask); // negative arg will be NAN
|
||||
return x;
|
||||
}
|
||||
|
||||
_PS_CONST(exp_hi, 88.3762626647949f);
|
||||
_PS_CONST(exp_lo, -88.3762626647949f);
|
||||
|
||||
_PS_CONST(cephes_LOG2EF, 1.44269504088896341);
|
||||
_PS_CONST(cephes_exp_C1, 0.693359375);
|
||||
_PS_CONST(cephes_exp_C2, -2.12194440e-4);
|
||||
|
||||
_PS_CONST(cephes_exp_p0, 1.9875691500E-4);
|
||||
_PS_CONST(cephes_exp_p1, 1.3981999507E-3);
|
||||
_PS_CONST(cephes_exp_p2, 8.3334519073E-3);
|
||||
_PS_CONST(cephes_exp_p3, 4.1665795894E-2);
|
||||
_PS_CONST(cephes_exp_p4, 1.6666665459E-1);
|
||||
_PS_CONST(cephes_exp_p5, 5.0000001201E-1);
|
||||
|
||||
v4sf exp_ps(v4sf x) {
|
||||
v4sf tmp = _mm_setzero_ps(), fx;
|
||||
#ifdef USE_SSE2
|
||||
v4si emm0;
|
||||
#else
|
||||
v2si mm0, mm1;
|
||||
#endif
|
||||
v4sf one = *(v4sf*)_ps_1;
|
||||
|
||||
x = _mm_min_ps(x, *(v4sf*)_ps_exp_hi);
|
||||
x = _mm_max_ps(x, *(v4sf*)_ps_exp_lo);
|
||||
|
||||
/* express exp(x) as exp(g + n*log(2)) */
|
||||
fx = _mm_mul_ps(x, *(v4sf*)_ps_cephes_LOG2EF);
|
||||
fx = _mm_add_ps(fx, *(v4sf*)_ps_0p5);
|
||||
|
||||
/* how to perform a floorf with SSE: just below */
|
||||
#ifndef USE_SSE2
|
||||
/* step 1 : cast to int */
|
||||
tmp = _mm_movehl_ps(tmp, fx);
|
||||
mm0 = _mm_cvttps_pi32(fx);
|
||||
mm1 = _mm_cvttps_pi32(tmp);
|
||||
/* step 2 : cast back to float */
|
||||
tmp = _mm_cvtpi32x2_ps(mm0, mm1);
|
||||
#else
|
||||
emm0 = _mm_cvttps_epi32(fx);
|
||||
tmp = _mm_cvtepi32_ps(emm0);
|
||||
#endif
|
||||
/* if greater, substract 1 */
|
||||
v4sf mask = _mm_cmpgt_ps(tmp, fx);
|
||||
mask = _mm_and_ps(mask, one);
|
||||
fx = _mm_sub_ps(tmp, mask);
|
||||
|
||||
tmp = _mm_mul_ps(fx, *(v4sf*)_ps_cephes_exp_C1);
|
||||
v4sf z = _mm_mul_ps(fx, *(v4sf*)_ps_cephes_exp_C2);
|
||||
x = _mm_sub_ps(x, tmp);
|
||||
x = _mm_sub_ps(x, z);
|
||||
|
||||
z = _mm_mul_ps(x,x);
|
||||
|
||||
v4sf y = *(v4sf*)_ps_cephes_exp_p0;
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p1);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p2);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p3);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p4);
|
||||
y = _mm_mul_ps(y, x);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_cephes_exp_p5);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, x);
|
||||
y = _mm_add_ps(y, one);
|
||||
|
||||
/* build 2^n */
|
||||
#ifndef USE_SSE2
|
||||
z = _mm_movehl_ps(z, fx);
|
||||
mm0 = _mm_cvttps_pi32(fx);
|
||||
mm1 = _mm_cvttps_pi32(z);
|
||||
mm0 = _mm_add_pi32(mm0, *(v2si*)_pi32_0x7f);
|
||||
mm1 = _mm_add_pi32(mm1, *(v2si*)_pi32_0x7f);
|
||||
mm0 = _mm_slli_pi32(mm0, 23);
|
||||
mm1 = _mm_slli_pi32(mm1, 23);
|
||||
|
||||
v4sf pow2n;
|
||||
COPY_MM_TO_XMM(mm0, mm1, pow2n);
|
||||
_mm_empty();
|
||||
#else
|
||||
emm0 = _mm_cvttps_epi32(fx);
|
||||
emm0 = _mm_add_epi32(emm0, *(v4si*)_pi32_0x7f);
|
||||
emm0 = _mm_slli_epi32(emm0, 23);
|
||||
v4sf pow2n = _mm_castsi128_ps(emm0);
|
||||
#endif
|
||||
y = _mm_mul_ps(y, pow2n);
|
||||
return y;
|
||||
}
|
||||
|
||||
_PS_CONST(minus_cephes_DP1, -0.78515625);
|
||||
_PS_CONST(minus_cephes_DP2, -2.4187564849853515625e-4);
|
||||
_PS_CONST(minus_cephes_DP3, -3.77489497744594108e-8);
|
||||
_PS_CONST(sincof_p0, -1.9515295891E-4);
|
||||
_PS_CONST(sincof_p1, 8.3321608736E-3);
|
||||
_PS_CONST(sincof_p2, -1.6666654611E-1);
|
||||
_PS_CONST(coscof_p0, 2.443315711809948E-005);
|
||||
_PS_CONST(coscof_p1, -1.388731625493765E-003);
|
||||
_PS_CONST(coscof_p2, 4.166664568298827E-002);
|
||||
_PS_CONST(cephes_FOPI, 1.27323954473516); // 4 / M_PI
|
||||
|
||||
|
||||
/* evaluation of 4 sines at onces, using only SSE1+MMX intrinsics so
|
||||
it runs also on old athlons XPs and the pentium III of your grand
|
||||
mother.
|
||||
|
||||
The code is the exact rewriting of the cephes sinf function.
|
||||
Precision is excellent as long as x < 8192 (I did not bother to
|
||||
take into account the special handling they have for greater values
|
||||
-- it does not return garbage for arguments over 8192, though, but
|
||||
the extra precision is missing).
|
||||
|
||||
Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
|
||||
surprising but correct result.
|
||||
|
||||
Performance is also surprisingly good, 1.33 times faster than the
|
||||
macos vsinf SSE2 function, and 1.5 times faster than the
|
||||
__vrs4_sinf of amd's ACML (which is only available in 64 bits). Not
|
||||
too bad for an SSE1 function (with no special tuning) !
|
||||
However the latter libraries probably have a much better handling of NaN,
|
||||
Inf, denormalized and other special arguments..
|
||||
|
||||
On my core 1 duo, the execution of this function takes approximately 95 cycles.
|
||||
|
||||
From what I have observed on the experiments with Intel AMath lib, switching to an
|
||||
SSE2 version would improve the perf by only 10%.
|
||||
|
||||
Since it is based on SSE intrinsics, it has to be compiled at -O2 to
|
||||
deliver full speed.
|
||||
*/
|
||||
v4sf sin_ps(v4sf x) { // any x
|
||||
v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y;
|
||||
|
||||
#ifdef USE_SSE2
|
||||
v4si emm0, emm2;
|
||||
#else
|
||||
v2si mm0, mm1, mm2, mm3;
|
||||
#endif
|
||||
sign_bit = x;
|
||||
/* take the absolute value */
|
||||
x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
|
||||
/* extract the sign bit (upper one) */
|
||||
sign_bit = _mm_and_ps(sign_bit, *(v4sf*)_ps_sign_mask);
|
||||
|
||||
/* scale by 4/Pi */
|
||||
y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
|
||||
|
||||
//printf("plop:"); print4(y);
|
||||
#ifdef USE_SSE2
|
||||
/* store the integer part of y in mm0 */
|
||||
emm2 = _mm_cvttps_epi32(y);
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
|
||||
y = _mm_cvtepi32_ps(emm2);
|
||||
/* get the swap sign flag */
|
||||
emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
|
||||
emm0 = _mm_slli_epi32(emm0, 29);
|
||||
/* get the polynom selection mask
|
||||
there is one polynom for 0 <= x <= Pi/4
|
||||
and another one for Pi/4<x<=Pi/2
|
||||
|
||||
Both branches will be computed.
|
||||
*/
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
|
||||
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
|
||||
|
||||
v4sf swap_sign_bit = _mm_castsi128_ps(emm0);
|
||||
v4sf poly_mask = _mm_castsi128_ps(emm2);
|
||||
sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
|
||||
#else
|
||||
/* store the integer part of y in mm0:mm1 */
|
||||
xmm2 = _mm_movehl_ps(xmm2, y);
|
||||
mm2 = _mm_cvttps_pi32(y);
|
||||
mm3 = _mm_cvttps_pi32(xmm2);
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
|
||||
mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
|
||||
y = _mm_cvtpi32x2_ps(mm2, mm3);
|
||||
/* get the swap sign flag */
|
||||
mm0 = _mm_and_si64(mm2, *(v2si*)_pi32_4);
|
||||
mm1 = _mm_and_si64(mm3, *(v2si*)_pi32_4);
|
||||
mm0 = _mm_slli_pi32(mm0, 29);
|
||||
mm1 = _mm_slli_pi32(mm1, 29);
|
||||
/* get the polynom selection mask */
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
|
||||
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
|
||||
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
|
||||
v4sf swap_sign_bit, poly_mask;
|
||||
COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit);
|
||||
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
|
||||
sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
|
||||
_mm_empty(); /* good-bye mmx */
|
||||
#endif
|
||||
|
||||
/* The magic pass: "Extended precision modular arithmetic"
|
||||
x = ((x - y * DP1) - y * DP2) - y * DP3; */
|
||||
xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
|
||||
xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
|
||||
xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
|
||||
xmm1 = _mm_mul_ps(y, xmm1);
|
||||
xmm2 = _mm_mul_ps(y, xmm2);
|
||||
xmm3 = _mm_mul_ps(y, xmm3);
|
||||
x = _mm_add_ps(x, xmm1);
|
||||
x = _mm_add_ps(x, xmm2);
|
||||
x = _mm_add_ps(x, xmm3);
|
||||
|
||||
/* Evaluate the first polynom (0 <= x <= Pi/4) */
|
||||
y = *(v4sf*)_ps_coscof_p0;
|
||||
v4sf z = _mm_mul_ps(x,x);
|
||||
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p1);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p2);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_mul_ps(y, z);
|
||||
v4sf tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
|
||||
y = _mm_sub_ps(y, tmp);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_1);
|
||||
|
||||
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
|
||||
|
||||
v4sf y2 = *(v4sf*)_ps_sincof_p0;
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_mul_ps(y2, x);
|
||||
y2 = _mm_add_ps(y2, x);
|
||||
|
||||
/* select the correct result from the two polynoms */
|
||||
xmm3 = poly_mask;
|
||||
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
|
||||
y = _mm_andnot_ps(xmm3, y);
|
||||
y = _mm_add_ps(y,y2);
|
||||
/* update the sign */
|
||||
y = _mm_xor_ps(y, sign_bit);
|
||||
|
||||
return y;
|
||||
}
|
||||
|
||||
/* almost the same as sin_ps */
|
||||
v4sf cos_ps(v4sf x) { // any x
|
||||
v4sf xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
|
||||
#ifdef USE_SSE2
|
||||
v4si emm0, emm2;
|
||||
#else
|
||||
v2si mm0, mm1, mm2, mm3;
|
||||
#endif
|
||||
/* take the absolute value */
|
||||
x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
|
||||
|
||||
/* scale by 4/Pi */
|
||||
y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
|
||||
|
||||
#ifdef USE_SSE2
|
||||
/* store the integer part of y in mm0 */
|
||||
emm2 = _mm_cvttps_epi32(y);
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
|
||||
y = _mm_cvtepi32_ps(emm2);
|
||||
|
||||
emm2 = _mm_sub_epi32(emm2, *(v4si*)_pi32_2);
|
||||
|
||||
/* get the swap sign flag */
|
||||
emm0 = _mm_andnot_si128(emm2, *(v4si*)_pi32_4);
|
||||
emm0 = _mm_slli_epi32(emm0, 29);
|
||||
/* get the polynom selection mask */
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
|
||||
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
|
||||
|
||||
v4sf sign_bit = _mm_castsi128_ps(emm0);
|
||||
v4sf poly_mask = _mm_castsi128_ps(emm2);
|
||||
#else
|
||||
/* store the integer part of y in mm0:mm1 */
|
||||
xmm2 = _mm_movehl_ps(xmm2, y);
|
||||
mm2 = _mm_cvttps_pi32(y);
|
||||
mm3 = _mm_cvttps_pi32(xmm2);
|
||||
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
|
||||
mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
|
||||
|
||||
y = _mm_cvtpi32x2_ps(mm2, mm3);
|
||||
|
||||
|
||||
mm2 = _mm_sub_pi32(mm2, *(v2si*)_pi32_2);
|
||||
mm3 = _mm_sub_pi32(mm3, *(v2si*)_pi32_2);
|
||||
|
||||
/* get the swap sign flag in mm0:mm1 and the
|
||||
polynom selection mask in mm2:mm3 */
|
||||
|
||||
mm0 = _mm_andnot_si64(mm2, *(v2si*)_pi32_4);
|
||||
mm1 = _mm_andnot_si64(mm3, *(v2si*)_pi32_4);
|
||||
mm0 = _mm_slli_pi32(mm0, 29);
|
||||
mm1 = _mm_slli_pi32(mm1, 29);
|
||||
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
|
||||
|
||||
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
|
||||
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
|
||||
|
||||
v4sf sign_bit, poly_mask;
|
||||
COPY_MM_TO_XMM(mm0, mm1, sign_bit);
|
||||
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
|
||||
_mm_empty(); /* good-bye mmx */
|
||||
#endif
|
||||
/* The magic pass: "Extended precision modular arithmetic"
|
||||
x = ((x - y * DP1) - y * DP2) - y * DP3; */
|
||||
xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
|
||||
xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
|
||||
xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
|
||||
xmm1 = _mm_mul_ps(y, xmm1);
|
||||
xmm2 = _mm_mul_ps(y, xmm2);
|
||||
xmm3 = _mm_mul_ps(y, xmm3);
|
||||
x = _mm_add_ps(x, xmm1);
|
||||
x = _mm_add_ps(x, xmm2);
|
||||
x = _mm_add_ps(x, xmm3);
|
||||
|
||||
/* Evaluate the first polynom (0 <= x <= Pi/4) */
|
||||
y = *(v4sf*)_ps_coscof_p0;
|
||||
v4sf z = _mm_mul_ps(x,x);
|
||||
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p1);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p2);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_mul_ps(y, z);
|
||||
v4sf tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
|
||||
y = _mm_sub_ps(y, tmp);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_1);
|
||||
|
||||
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
|
||||
|
||||
v4sf y2 = *(v4sf*)_ps_sincof_p0;
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_mul_ps(y2, x);
|
||||
y2 = _mm_add_ps(y2, x);
|
||||
|
||||
/* select the correct result from the two polynoms */
|
||||
xmm3 = poly_mask;
|
||||
y2 = _mm_and_ps(xmm3, y2); //, xmm3);
|
||||
y = _mm_andnot_ps(xmm3, y);
|
||||
y = _mm_add_ps(y,y2);
|
||||
/* update the sign */
|
||||
y = _mm_xor_ps(y, sign_bit);
|
||||
|
||||
return y;
|
||||
}
|
||||
|
||||
/* since sin_ps and cos_ps are almost identical, sincos_ps could replace both of them..
|
||||
it is almost as fast, and gives you a free cosine with your sine */
|
||||
void sincos_ps(v4sf x, v4sf *s, v4sf *c) {
|
||||
v4sf xmm1, xmm2, xmm3 = _mm_setzero_ps(), sign_bit_sin, y;
|
||||
#ifdef USE_SSE2
|
||||
v4si emm0, emm2, emm4;
|
||||
#else
|
||||
v2si mm0, mm1, mm2, mm3, mm4, mm5;
|
||||
#endif
|
||||
sign_bit_sin = x;
|
||||
/* take the absolute value */
|
||||
x = _mm_and_ps(x, *(v4sf*)_ps_inv_sign_mask);
|
||||
/* extract the sign bit (upper one) */
|
||||
sign_bit_sin = _mm_and_ps(sign_bit_sin, *(v4sf*)_ps_sign_mask);
|
||||
|
||||
/* scale by 4/Pi */
|
||||
y = _mm_mul_ps(x, *(v4sf*)_ps_cephes_FOPI);
|
||||
|
||||
#ifdef USE_SSE2
|
||||
/* store the integer part of y in emm2 */
|
||||
emm2 = _mm_cvttps_epi32(y);
|
||||
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
emm2 = _mm_add_epi32(emm2, *(v4si*)_pi32_1);
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_inv1);
|
||||
y = _mm_cvtepi32_ps(emm2);
|
||||
|
||||
emm4 = emm2;
|
||||
|
||||
/* get the swap sign flag for the sine */
|
||||
emm0 = _mm_and_si128(emm2, *(v4si*)_pi32_4);
|
||||
emm0 = _mm_slli_epi32(emm0, 29);
|
||||
v4sf swap_sign_bit_sin = _mm_castsi128_ps(emm0);
|
||||
|
||||
/* get the polynom selection mask for the sine*/
|
||||
emm2 = _mm_and_si128(emm2, *(v4si*)_pi32_2);
|
||||
emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
|
||||
v4sf poly_mask = _mm_castsi128_ps(emm2);
|
||||
#else
|
||||
/* store the integer part of y in mm2:mm3 */
|
||||
xmm3 = _mm_movehl_ps(xmm3, y);
|
||||
mm2 = _mm_cvttps_pi32(y);
|
||||
mm3 = _mm_cvttps_pi32(xmm3);
|
||||
|
||||
/* j=(j+1) & (~1) (see the cephes sources) */
|
||||
mm2 = _mm_add_pi32(mm2, *(v2si*)_pi32_1);
|
||||
mm3 = _mm_add_pi32(mm3, *(v2si*)_pi32_1);
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_inv1);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_inv1);
|
||||
|
||||
y = _mm_cvtpi32x2_ps(mm2, mm3);
|
||||
|
||||
mm4 = mm2;
|
||||
mm5 = mm3;
|
||||
|
||||
/* get the swap sign flag for the sine */
|
||||
mm0 = _mm_and_si64(mm2, *(v2si*)_pi32_4);
|
||||
mm1 = _mm_and_si64(mm3, *(v2si*)_pi32_4);
|
||||
mm0 = _mm_slli_pi32(mm0, 29);
|
||||
mm1 = _mm_slli_pi32(mm1, 29);
|
||||
v4sf swap_sign_bit_sin;
|
||||
COPY_MM_TO_XMM(mm0, mm1, swap_sign_bit_sin);
|
||||
|
||||
/* get the polynom selection mask for the sine */
|
||||
|
||||
mm2 = _mm_and_si64(mm2, *(v2si*)_pi32_2);
|
||||
mm3 = _mm_and_si64(mm3, *(v2si*)_pi32_2);
|
||||
mm2 = _mm_cmpeq_pi32(mm2, _mm_setzero_si64());
|
||||
mm3 = _mm_cmpeq_pi32(mm3, _mm_setzero_si64());
|
||||
v4sf poly_mask;
|
||||
COPY_MM_TO_XMM(mm2, mm3, poly_mask);
|
||||
#endif
|
||||
|
||||
/* The magic pass: "Extended precision modular arithmetic"
|
||||
x = ((x - y * DP1) - y * DP2) - y * DP3; */
|
||||
xmm1 = *(v4sf*)_ps_minus_cephes_DP1;
|
||||
xmm2 = *(v4sf*)_ps_minus_cephes_DP2;
|
||||
xmm3 = *(v4sf*)_ps_minus_cephes_DP3;
|
||||
xmm1 = _mm_mul_ps(y, xmm1);
|
||||
xmm2 = _mm_mul_ps(y, xmm2);
|
||||
xmm3 = _mm_mul_ps(y, xmm3);
|
||||
x = _mm_add_ps(x, xmm1);
|
||||
x = _mm_add_ps(x, xmm2);
|
||||
x = _mm_add_ps(x, xmm3);
|
||||
|
||||
#ifdef USE_SSE2
|
||||
emm4 = _mm_sub_epi32(emm4, *(v4si*)_pi32_2);
|
||||
emm4 = _mm_andnot_si128(emm4, *(v4si*)_pi32_4);
|
||||
emm4 = _mm_slli_epi32(emm4, 29);
|
||||
v4sf sign_bit_cos = _mm_castsi128_ps(emm4);
|
||||
#else
|
||||
/* get the sign flag for the cosine */
|
||||
mm4 = _mm_sub_pi32(mm4, *(v2si*)_pi32_2);
|
||||
mm5 = _mm_sub_pi32(mm5, *(v2si*)_pi32_2);
|
||||
mm4 = _mm_andnot_si64(mm4, *(v2si*)_pi32_4);
|
||||
mm5 = _mm_andnot_si64(mm5, *(v2si*)_pi32_4);
|
||||
mm4 = _mm_slli_pi32(mm4, 29);
|
||||
mm5 = _mm_slli_pi32(mm5, 29);
|
||||
v4sf sign_bit_cos;
|
||||
COPY_MM_TO_XMM(mm4, mm5, sign_bit_cos);
|
||||
_mm_empty(); /* good-bye mmx */
|
||||
#endif
|
||||
|
||||
sign_bit_sin = _mm_xor_ps(sign_bit_sin, swap_sign_bit_sin);
|
||||
|
||||
|
||||
/* Evaluate the first polynom (0 <= x <= Pi/4) */
|
||||
v4sf z = _mm_mul_ps(x,x);
|
||||
y = *(v4sf*)_ps_coscof_p0;
|
||||
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p1);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_coscof_p2);
|
||||
y = _mm_mul_ps(y, z);
|
||||
y = _mm_mul_ps(y, z);
|
||||
v4sf tmp = _mm_mul_ps(z, *(v4sf*)_ps_0p5);
|
||||
y = _mm_sub_ps(y, tmp);
|
||||
y = _mm_add_ps(y, *(v4sf*)_ps_1);
|
||||
|
||||
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
|
||||
|
||||
v4sf y2 = *(v4sf*)_ps_sincof_p0;
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p1);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_add_ps(y2, *(v4sf*)_ps_sincof_p2);
|
||||
y2 = _mm_mul_ps(y2, z);
|
||||
y2 = _mm_mul_ps(y2, x);
|
||||
y2 = _mm_add_ps(y2, x);
|
||||
|
||||
/* select the correct result from the two polynoms */
|
||||
xmm3 = poly_mask;
|
||||
v4sf ysin2 = _mm_and_ps(xmm3, y2);
|
||||
v4sf ysin1 = _mm_andnot_ps(xmm3, y);
|
||||
y2 = _mm_sub_ps(y2,ysin2);
|
||||
y = _mm_sub_ps(y, ysin1);
|
||||
|
||||
xmm1 = _mm_add_ps(ysin1,ysin2);
|
||||
xmm2 = _mm_add_ps(y,y2);
|
||||
|
||||
/* update the sign */
|
||||
*s = _mm_xor_ps(xmm1, sign_bit_sin);
|
||||
*c = _mm_xor_ps(xmm2, sign_bit_cos);
|
||||
}
|
||||
|
Loading…
Reference in New Issue