![]() |
Eigen
3.3.3
|
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 // Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org> 00006 // 00007 // This Source Code Form is subject to the terms of the Mozilla 00008 // Public License v. 2.0. If a copy of the MPL was not distributed 00009 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 00010 00011 #ifndef EIGEN_COMPLEX_NEON_H 00012 #define EIGEN_COMPLEX_NEON_H 00013 00014 namespace Eigen { 00015 00016 namespace internal { 00017 00018 inline uint32x4_t p4ui_CONJ_XOR() { 00019 // See bug 1325, clang fails to call vld1q_u64. 00020 #if EIGEN_COMP_CLANG 00021 uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; 00022 return ret; 00023 #else 00024 static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 }; 00025 return vld1q_u32( conj_XOR_DATA ); 00026 #endif 00027 } 00028 00029 inline uint32x2_t p2ui_CONJ_XOR() { 00030 static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000 }; 00031 return vld1_u32( conj_XOR_DATA ); 00032 } 00033 00034 //---------- float ---------- 00035 struct Packet2cf 00036 { 00037 EIGEN_STRONG_INLINE Packet2cf() {} 00038 EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {} 00039 Packet4f v; 00040 }; 00041 00042 template<> struct packet_traits<std::complex<float> > : default_packet_traits 00043 { 00044 typedef Packet2cf type; 00045 typedef Packet2cf half; 00046 enum { 00047 Vectorizable = 1, 00048 AlignedOnScalar = 1, 00049 size = 2, 00050 HasHalfPacket = 0, 00051 00052 HasAdd = 1, 00053 HasSub = 1, 00054 HasMul = 1, 00055 HasDiv = 1, 00056 HasNegate = 1, 00057 HasAbs = 0, 00058 HasAbs2 = 0, 00059 HasMin = 0, 00060 HasMax = 0, 00061 HasSetLinear = 0 00062 }; 00063 }; 00064 00065 template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; }; 00066 00067 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) 00068 { 00069 float32x2_t r64; 00070 r64 = vld1_f32((float *)&from); 00071 00072 return Packet2cf(vcombine_f32(r64, r64)); 00073 } 00074 00075 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v,b.v)); } 00076 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v,b.v)); } 00077 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); } 00078 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) 00079 { 00080 Packet4ui b = vreinterpretq_u32_f32(a.v); 00081 return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR()))); 00082 } 00083 00084 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00085 { 00086 Packet4f v1, v2; 00087 00088 // Get the real values of a | a1_re | a1_re | a2_re | a2_re | 00089 v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0)); 00090 // Get the imag values of a | a1_im | a1_im | a2_im | a2_im | 00091 v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1)); 00092 // Multiply the real a with b 00093 v1 = vmulq_f32(v1, b.v); 00094 // Multiply the imag a with b 00095 v2 = vmulq_f32(v2, b.v); 00096 // Conjugate v2 00097 v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR())); 00098 // Swap real/imag elements in v2. 00099 v2 = vrev64q_f32(v2); 00100 // Add and return the result 00101 return Packet2cf(vaddq_f32(v1, v2)); 00102 } 00103 00104 template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00105 { 00106 return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00107 } 00108 template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00109 { 00110 return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00111 } 00112 template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00113 { 00114 return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00115 } 00116 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00117 { 00118 return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v)))); 00119 } 00120 00121 template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); } 00122 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); } 00123 00124 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } 00125 00126 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); } 00127 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); } 00128 00129 template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride) 00130 { 00131 Packet4f res = pset1<Packet4f>(0.f); 00132 res = vsetq_lane_f32(std::real(from[0*stride]), res, 0); 00133 res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1); 00134 res = vsetq_lane_f32(std::real(from[1*stride]), res, 2); 00135 res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3); 00136 return Packet2cf(res); 00137 } 00138 00139 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride) 00140 { 00141 to[stride*0] = std::complex<float>(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1)); 00142 to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3)); 00143 } 00144 00145 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { EIGEN_ARM_PREFETCH((float *)addr); } 00146 00147 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) 00148 { 00149 std::complex<float> EIGEN_ALIGN16 x[2]; 00150 vst1q_f32((float *)x, a.v); 00151 return x[0]; 00152 } 00153 00154 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) 00155 { 00156 float32x2_t a_lo, a_hi; 00157 Packet4f a_r128; 00158 00159 a_lo = vget_low_f32(a.v); 00160 a_hi = vget_high_f32(a.v); 00161 a_r128 = vcombine_f32(a_hi, a_lo); 00162 00163 return Packet2cf(a_r128); 00164 } 00165 00166 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a) 00167 { 00168 return Packet2cf(vrev64q_f32(a.v)); 00169 } 00170 00171 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) 00172 { 00173 float32x2_t a1, a2; 00174 std::complex<float> s; 00175 00176 a1 = vget_low_f32(a.v); 00177 a2 = vget_high_f32(a.v); 00178 a2 = vadd_f32(a1, a2); 00179 vst1_f32((float *)&s, a2); 00180 00181 return s; 00182 } 00183 00184 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) 00185 { 00186 Packet4f sum1, sum2, sum; 00187 00188 // Add the first two 64-bit float32x2_t of vecs[0] 00189 sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v)); 00190 sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v)); 00191 sum = vaddq_f32(sum1, sum2); 00192 00193 return Packet2cf(sum); 00194 } 00195 00196 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) 00197 { 00198 float32x2_t a1, a2, v1, v2, prod; 00199 std::complex<float> s; 00200 00201 a1 = vget_low_f32(a.v); 00202 a2 = vget_high_f32(a.v); 00203 // Get the real values of a | a1_re | a1_re | a2_re | a2_re | 00204 v1 = vdup_lane_f32(a1, 0); 00205 // Get the real values of a | a1_im | a1_im | a2_im | a2_im | 00206 v2 = vdup_lane_f32(a1, 1); 00207 // Multiply the real a with b 00208 v1 = vmul_f32(v1, a2); 00209 // Multiply the imag a with b 00210 v2 = vmul_f32(v2, a2); 00211 // Conjugate v2 00212 v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR())); 00213 // Swap real/imag elements in v2. 00214 v2 = vrev64_f32(v2); 00215 // Add v1, v2 00216 prod = vadd_f32(v1, v2); 00217 00218 vst1_f32((float *)&s, prod); 00219 00220 return s; 00221 } 00222 00223 template<int Offset> 00224 struct palign_impl<Offset,Packet2cf> 00225 { 00226 EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second) 00227 { 00228 if (Offset==1) 00229 { 00230 first.v = vextq_f32(first.v, second.v, 2); 00231 } 00232 } 00233 }; 00234 00235 template<> struct conj_helper<Packet2cf, Packet2cf, false,true> 00236 { 00237 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00238 { return padd(pmul(x,y),c); } 00239 00240 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00241 { 00242 return internal::pmul(a, pconj(b)); 00243 } 00244 }; 00245 00246 template<> struct conj_helper<Packet2cf, Packet2cf, true,false> 00247 { 00248 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00249 { return padd(pmul(x,y),c); } 00250 00251 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00252 { 00253 return internal::pmul(pconj(a), b); 00254 } 00255 }; 00256 00257 template<> struct conj_helper<Packet2cf, Packet2cf, true,true> 00258 { 00259 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 00260 { return padd(pmul(x,y),c); } 00261 00262 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 00263 { 00264 return pconj(internal::pmul(a, b)); 00265 } 00266 }; 00267 00268 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 00269 { 00270 // TODO optimize it for NEON 00271 Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); 00272 Packet4f s, rev_s; 00273 00274 // this computes the norm 00275 s = vmulq_f32(b.v, b.v); 00276 rev_s = vrev64q_f32(s); 00277 00278 return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); 00279 } 00280 00281 EIGEN_DEVICE_FUNC inline void 00282 ptranspose(PacketBlock<Packet2cf,2>& kernel) { 00283 Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v)); 00284 kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v)); 00285 kernel.packet[1].v = tmp; 00286 } 00287 00288 //---------- double ---------- 00289 #if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG 00290 00291 // See bug 1325, clang fails to call vld1q_u64. 00292 #if EIGEN_COMP_CLANG 00293 static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000}; 00294 #else 00295 const uint64_t p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 }; 00296 static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA ); 00297 #endif 00298 00299 struct Packet1cd 00300 { 00301 EIGEN_STRONG_INLINE Packet1cd() {} 00302 EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {} 00303 Packet2d v; 00304 }; 00305 00306 template<> struct packet_traits<std::complex<double> > : default_packet_traits 00307 { 00308 typedef Packet1cd type; 00309 typedef Packet1cd half; 00310 enum { 00311 Vectorizable = 1, 00312 AlignedOnScalar = 0, 00313 size = 1, 00314 HasHalfPacket = 0, 00315 00316 HasAdd = 1, 00317 HasSub = 1, 00318 HasMul = 1, 00319 HasDiv = 1, 00320 HasNegate = 1, 00321 HasAbs = 0, 00322 HasAbs2 = 0, 00323 HasMin = 0, 00324 HasMax = 0, 00325 HasSetLinear = 0 00326 }; 00327 }; 00328 00329 template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; }; 00330 00331 template<> EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); } 00332 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); } 00333 00334 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) 00335 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } 00336 00337 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd<Packet2d>(a.v,b.v)); } 00338 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub<Packet2d>(a.v,b.v)); } 00339 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate<Packet2d>(a.v)); } 00340 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); } 00341 00342 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00343 { 00344 Packet2d v1, v2; 00345 00346 // Get the real values of a 00347 v1 = vdupq_lane_f64(vget_low_f64(a.v), 0); 00348 // Get the imag values of a 00349 v2 = vdupq_lane_f64(vget_high_f64(a.v), 0); 00350 // Multiply the real a with b 00351 v1 = vmulq_f64(v1, b.v); 00352 // Multiply the imag a with b 00353 v2 = vmulq_f64(v2, b.v); 00354 // Conjugate v2 00355 v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR)); 00356 // Swap real/imag elements in v2. 00357 v2 = preverse<Packet2d>(v2); 00358 // Add and return the result 00359 return Packet1cd(vaddq_f64(v1, v2)); 00360 } 00361 00362 template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00363 { 00364 return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); 00365 } 00366 template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00367 { 00368 return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); 00369 } 00370 template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00371 { 00372 return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); 00373 } 00374 template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00375 { 00376 return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); 00377 } 00378 00379 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } 00380 00381 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); } 00382 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); } 00383 00384 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { EIGEN_ARM_PREFETCH((double *)addr); } 00385 00386 template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride) 00387 { 00388 Packet2d res = pset1<Packet2d>(0.0); 00389 res = vsetq_lane_f64(std::real(from[0*stride]), res, 0); 00390 res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1); 00391 return Packet1cd(res); 00392 } 00393 00394 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride) 00395 { 00396 to[stride*0] = std::complex<double>(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1)); 00397 } 00398 00399 00400 template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) 00401 { 00402 std::complex<double> EIGEN_ALIGN16 res; 00403 pstore<std::complex<double> >(&res, a); 00404 00405 return res; 00406 } 00407 00408 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } 00409 00410 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); } 00411 00412 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; } 00413 00414 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); } 00415 00416 template<int Offset> 00417 struct palign_impl<Offset,Packet1cd> 00418 { 00419 static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) 00420 { 00421 // FIXME is it sure we never have to align a Packet1cd? 00422 // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... 00423 } 00424 }; 00425 00426 template<> struct conj_helper<Packet1cd, Packet1cd, false,true> 00427 { 00428 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 00429 { return padd(pmul(x,y),c); } 00430 00431 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 00432 { 00433 return internal::pmul(a, pconj(b)); 00434 } 00435 }; 00436 00437 template<> struct conj_helper<Packet1cd, Packet1cd, true,false> 00438 { 00439 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 00440 { return padd(pmul(x,y),c); } 00441 00442 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 00443 { 00444 return internal::pmul(pconj(a), b); 00445 } 00446 }; 00447 00448 template<> struct conj_helper<Packet1cd, Packet1cd, true,true> 00449 { 00450 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 00451 { return padd(pmul(x,y),c); } 00452 00453 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 00454 { 00455 return pconj(internal::pmul(a, b)); 00456 } 00457 }; 00458 00459 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) 00460 { 00461 // TODO optimize it for NEON 00462 Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); 00463 Packet2d s = pmul<Packet2d>(b.v, b.v); 00464 Packet2d rev_s = preverse<Packet2d>(s); 00465 00466 return Packet1cd(pdiv(res.v, padd<Packet2d>(s,rev_s))); 00467 } 00468 00469 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) 00470 { 00471 return Packet1cd(preverse(Packet2d(x.v))); 00472 } 00473 00474 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel) 00475 { 00476 Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v)); 00477 kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v)); 00478 kernel.packet[1].v = tmp; 00479 } 00480 #endif // EIGEN_ARCH_ARM64 00481 00482 } // end namespace internal 00483 00484 } // end namespace Eigen 00485 00486 #endif // EIGEN_COMPLEX_NEON_H