00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026 #ifndef EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
00027 #define EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H
00028
00029 namespace internal {
00030
00032 template<typename Packet> inline static Packet pasin(Packet a) { return std::asin(a); }
00033
00034 #ifdef EIGEN_VECTORIZE_SSE
00035
00036 template<> EIGEN_DONT_INLINE Packet4f pasin(Packet4f x)
00037 {
00038 _EIGEN_DECLARE_CONST_Packet4f(half, 0.5);
00039 _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5);
00040 _EIGEN_DECLARE_CONST_Packet4f(3half, 1.5);
00041
00042 _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000);
00043
00044 _EIGEN_DECLARE_CONST_Packet4f(pi, 3.141592654);
00045 _EIGEN_DECLARE_CONST_Packet4f(pi_over_2, 3.141592654*0.5);
00046
00047 _EIGEN_DECLARE_CONST_Packet4f(asin1, 4.2163199048E-2);
00048 _EIGEN_DECLARE_CONST_Packet4f(asin2, 2.4181311049E-2);
00049 _EIGEN_DECLARE_CONST_Packet4f(asin3, 4.5470025998E-2);
00050 _EIGEN_DECLARE_CONST_Packet4f(asin4, 7.4953002686E-2);
00051 _EIGEN_DECLARE_CONST_Packet4f(asin5, 1.6666752422E-1);
00052
00053 Packet4f a = pabs(x);
00054
00055 Packet4f sign_bit= _mm_and_ps(x, p4f_sign_mask);
00056
00057 Packet4f z1,z2;
00058
00059
00060
00061
00062
00063 Packet4f branch_mask= _mm_cmpgt_ps(a, p4f_half);
00064
00065
00066
00067 {
00068
00069 z1=pmadd(p4f_minus_half,a,p4f_half);
00070 Packet4f x1=psqrt(z1);
00071 Packet4f s1=pmadd(p4f_asin1, z1, p4f_asin2);
00072 Packet4f s2=pmadd(s1, z1, p4f_asin3);
00073 Packet4f s3=pmadd(s2,z1, p4f_asin4);
00074 Packet4f s4=pmadd(s3,z1, p4f_asin5);
00075 Packet4f temp=pmul(s4,z1);
00076 z1=pmadd(temp,x1,x1);
00077 z1=padd(z1,z1);
00078 z1=psub(p4f_pi_over_2,z1);
00079 }
00080
00081 {
00082
00083 Packet4f x2=a;
00084 z2=pmul(x2,x2);
00085 Packet4f s1=pmadd(p4f_asin1, z2, p4f_asin2);
00086 Packet4f s2=pmadd(s1, z2, p4f_asin3);
00087 Packet4f s3=pmadd(s2,z2, p4f_asin4);
00088 Packet4f s4=pmadd(s3,z2, p4f_asin5);
00089 Packet4f temp=pmul(s4,z2);
00090 z2=pmadd(temp,x2,x2);
00091 }
00092
00093
00094 z1 = _mm_and_ps(branch_mask, z1);
00095 z2 = _mm_andnot_ps(branch_mask, z2);
00096 Packet4f z = _mm_or_ps(z1,z2);
00097
00098
00099 return _mm_xor_ps(z, sign_bit);
00100 }
00101
00102 }
00103
00104 #endif
00105
00106 #endif // EIGEN_MOREVECTORIZATION_MATHFUNCTIONS_H