$treeview $search $mathjax
Eigen
3.2.5
$projectbrief
|
$projectbrief
|
$searchbox |
00001 // This file is part of Eigen, a lightweight C++ template library 00002 // for linear algebra. 00003 // 00004 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr> 00005 // 00006 // This Source Code Form is subject to the terms of the Mozilla 00007 // Public License v. 2.0. If a copy of the MPL was not distributed 00008 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 00009 00010 #ifndef EIGEN_COMPLEX_NEON_H 00011 #define EIGEN_COMPLEX_NEON_H 00012 00013 namespace Eigen { 00014 00015 namespace internal { 00016 00017 static uint32x4_t p4ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET4(0x00000000, 0x80000000, 0x00000000, 0x80000000); 00018 static uint32x2_t p2ui_CONJ_XOR = EIGEN_INIT_NEON_PACKET2(0x00000000, 0x80000000); 00019 00020 //---------- float ---------- 00021 struct Packet2cf 00022 { 00023 EIGEN_STRONG_INLINE Packet2cf() {} 00024 EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} 00025 Packet4f v; 00026 }; 00027 00028 template<> struct packet_traits<std::complex<float> > : default_packet_traits 00029 { 00030 typedef Packet2cf type; 00031 enum { 00032 Vectorizable = 1, 00033 AlignedOnScalar = 1, 00034 size = 2, 00035 00036 HasAdd = 1, 00037 HasSub = 1, 00038 HasMul = 1, 00039 HasDiv = 1, 00040 HasNegate = 1, 00041 HasAbs = 0, 00042 HasAbs2 = 0, 00043 HasMin = 0, 00044 HasMax = 0, 00045 HasSetLinear = 0 00046 }; 00047 }; 00048 00049 template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; }; 00050 00051 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) 00052 { 00053 float32x2_t r64; 00054 r64 = vld1_f32((float *)&from); 00055 00056 return Packet2cf(vcombine_f32(r64, r64)); 00057 } 00058 00059 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v,b.v)); } 00060 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v,b.v)); } 00061 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); } 00062 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) 00063 { 00064 Packet4ui b = vreinterpretq_u32_f32(a.v); 00065 return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR))); 00066 } 00067 00068 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00069 { 00070 Packet4f v1, v2; 00071 00072 // Get the real values of a | a1_re | a1_re | a2_re | a2_re | 00073 v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0)); 00074 // Get the real values of a | a1_im | a1_im | a2_im | a2_im | 00075 v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1)); 00076 // Multiply the real a with b 00077 v1 = vmulq_f32(v1, b.v); 00078 // Multiply the imag a with b 00079 v2 = vmulq_f32(v2, b.v); 00080 // Conjugate v2 00081 v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR)); 00082 // Swap real/imag elements in v2. 00083 v2 = vrev64q_f32(v2); 00084 // Add and return the result 00085 return Packet2cf(vaddq_f32(v1, v2)); 00086 } 00087 00088 template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00089 { 00090 return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00091 } 00092 template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00093 { 00094 return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00095 } 00096 template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00097 { 00098 return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00099 } 00100 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00101 { 00102 return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00103 } 00104 00105 template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } 00106 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } 00107 00108 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } 00109 00110 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); } 00111 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); } 00112 00113 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ARM_PREFETCH((float *)addr); } 00114 00115 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) 00116 { 00117 std::complex<float> EIGEN_ALIGN16 x[2]; 00118 vst1q_f32((float *)x, a.v); 00119 return x[0]; 00120 } 00121 00122 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) 00123 { 00124 float32x2_t a_lo, a_hi; 00125 Packet4f a_r128; 00126 00127 a_lo = vget_low_f32(a.v); 00128 a_hi = vget_high_f32(a.v); 00129 a_r128 = vcombine_f32(a_hi, a_lo); 00130 00131 return Packet2cf(a_r128); 00132 } 00133 00134 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a) 00135 { 00136 return Packet2cf(vrev64q_f32(a.v)); 00137 } 00138 00139 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) 00140 { 00141 float32x2_t a1, a2; 00142 std::complex<float> s; 00143 00144 a1 = vget_low_f32(a.v); 00145 a2 = vget_high_f32(a.v); 00146 a2 = vadd_f32(a1, a2); 00147 vst1_f32((float *)&s, a2); 00148 00149 return s; 00150 } 00151 00152 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) 00153 { 00154 Packet4f sum1, sum2, sum; 00155 00156 // Add the first two 64-bit float32x2_t of vecs[0] 00157 sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v)); 00158 sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v)); 00159 sum = vaddq_f32(sum1, sum2); 00160 00161 return Packet2cf(sum); 00162 } 00163 00164 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) 00165 { 00166 float32x2_t a1, a2, v1, v2, prod; 00167 std::complex<float> s; 00168 00169 a1 = vget_low_f32(a.v); 00170 a2 = vget_high_f32(a.v); 00171 // Get the real values of a | a1_re | a1_re | a2_re | a2_re | 00172 v1 = vdup_lane_f32(a1, 0); 00173 // Get the real values of a | a1_im | a1_im | a2_im | a2_im | 00174 v2 = vdup_lane_f32(a1, 1); 00175 // Multiply the real a with b 00176 v1 = vmul_f32(v1, a2); 00177 // Multiply the imag a with b 00178 v2 = vmul_f32(v2, a2); 00179 // Conjugate v2 00180 v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR)); 00181 // Swap real/imag elements in v2. 00182 v2 = vrev64_f32(v2); 00183 // Add v1, v2 00184 prod = vadd_f32(v1, v2); 00185 00186 vst1_f32((float *)&s, prod); 00187 00188 return s; 00189 } 00190 00191 template<int Offset> 00192 struct palign_impl<Offset,Packet2cf> 00193 { 00194 EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second) 00195 { 00196 if (Offset==1) 00197 { 00198 first.v = vextq_f32(first.v, second.v, 2); 00199 } 00200 } 00201 }; 00202 00203 template<> struct conj_helper<Packet2cf, Packet2cf, false,true> 00204 { 00205 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00206 { return padd(pmul(x,y),c); } 00207 00208 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00209 { 00210 return internal::pmul(a, pconj(b)); 00211 } 00212 }; 00213 00214 template<> struct conj_helper<Packet2cf, Packet2cf, true,false> 00215 { 00216 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00217 { return padd(pmul(x,y),c); } 00218 00219 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00220 { 00221 return internal::pmul(pconj(a), b); 00222 } 00223 }; 00224 00225 template<> struct conj_helper<Packet2cf, Packet2cf, true,true> 00226 { 00227 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00228 { return padd(pmul(x,y),c); } 00229 00230 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00231 { 00232 return pconj(internal::pmul(a, b)); 00233 } 00234 }; 00235 00236 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00237 { 00238 // TODO optimize it for AltiVec 00239 Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); 00240 Packet4f s, rev_s; 00241 00242 // this computes the norm 00243 s = vmulq_f32(b.v, b.v); 00244 rev_s = vrev64q_f32(s); 00245 00246 return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); 00247 } 00248 00249 } // end namespace internal 00250 00251 } // end namespace Eigen 00252 00253 #endif // EIGEN_COMPLEX_NEON_H