00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011 #ifndef EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
00012 #define EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
00013
00014 namespace Eigen {
00015
00016 namespace internal {
00017
00019 template<typename Packet> inline static Packet pasin(Packet a) { return std::asin(a); }
00020
00021 #ifdef EIGEN_VECTORIZE_SSE
00022
00023 template<> EIGEN_DONT_INLINE Packet4f pasin(Packet4f x)
00024 {
00025 _EIGEN_DECLARE_CONST_Packet4f(half, 0.5);
00026 _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5);
00027 _EIGEN_DECLARE_CONST_Packet4f(3half, 1.5);
00028
00029 _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000);
00030
00031 _EIGEN_DECLARE_CONST_Packet4f(pi, 3.141592654);
00032 _EIGEN_DECLARE_CONST_Packet4f(pi_over_2, 3.141592654*0.5);
00033
00034 _EIGEN_DECLARE_CONST_Packet4f(asin1, 4.2163199048E-2);
00035 _EIGEN_DECLARE_CONST_Packet4f(asin2, 2.4181311049E-2);
00036 _EIGEN_DECLARE_CONST_Packet4f(asin3, 4.5470025998E-2);
00037 _EIGEN_DECLARE_CONST_Packet4f(asin4, 7.4953002686E-2);
00038 _EIGEN_DECLARE_CONST_Packet4f(asin5, 1.6666752422E-1);
00039
00040 Packet4f a = pabs(x);
00041
00042 Packet4f sign_bit= _mm_and_ps(x, p4f_sign_mask);
00043
00044 Packet4f z1,z2;
00045
00046
00047
00048
00049
00050 Packet4f branch_mask= _mm_cmpgt_ps(a, p4f_half);
00051
00052
00053
00054 {
00055
00056 z1=pmadd(p4f_minus_half,a,p4f_half);
00057 Packet4f x1=psqrt(z1);
00058 Packet4f s1=pmadd(p4f_asin1, z1, p4f_asin2);
00059 Packet4f s2=pmadd(s1, z1, p4f_asin3);
00060 Packet4f s3=pmadd(s2,z1, p4f_asin4);
00061 Packet4f s4=pmadd(s3,z1, p4f_asin5);
00062 Packet4f temp=pmul(s4,z1);
00063 z1=pmadd(temp,x1,x1);
00064 z1=padd(z1,z1);
00065 z1=psub(p4f_pi_over_2,z1);
00066 }
00067
00068 {
00069
00070 Packet4f x2=a;
00071 z2=pmul(x2,x2);
00072 Packet4f s1=pmadd(p4f_asin1, z2, p4f_asin2);
00073 Packet4f s2=pmadd(s1, z2, p4f_asin3);
00074 Packet4f s3=pmadd(s2,z2, p4f_asin4);
00075 Packet4f s4=pmadd(s3,z2, p4f_asin5);
00076 Packet4f temp=pmul(s4,z2);
00077 z2=pmadd(temp,x2,x2);
00078 }
00079
00080
00081 z1 = _mm_and_ps(branch_mask, z1);
00082 z2 = _mm_andnot_ps(branch_mask, z2);
00083 Packet4f z = _mm_or_ps(z1,z2);
00084
00085
00086 return _mm_xor_ps(z, sign_bit);
00087 }
00088
00089 #endif // EIGEN_VECTORIZE_SSE
00090
00091 }
00092
00093 }
00094
00095 #endif // EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H