1 // This file is part of Eigen, a lightweight C++ template library 2 // for linear algebra. 3 // 4 // Copyright (C) 2010 Gael Guennebaud <gael.guennebaud (at) inria.fr> 5 // 6 // This Source Code Form is subject to the terms of the Mozilla 7 // Public License v. 2.0. If a copy of the MPL was not distributed 8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 9 10 #ifndef EIGEN_COMPLEX_SSE_H 11 #define EIGEN_COMPLEX_SSE_H 12 13 namespace Eigen { 14 15 namespace internal { 16 17 //---------- float ---------- 18 struct Packet2cf 19 { 20 EIGEN_STRONG_INLINE Packet2cf() {} 21 EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {} 22 __m128 v; 23 }; 24 25 template<> struct packet_traits<std::complex<float> > : default_packet_traits 26 { 27 typedef Packet2cf type; 28 enum { 29 Vectorizable = 1, 30 AlignedOnScalar = 1, 31 size = 2, 32 33 HasAdd = 1, 34 HasSub = 1, 35 HasMul = 1, 36 HasDiv = 1, 37 HasNegate = 1, 38 HasAbs = 0, 39 HasAbs2 = 0, 40 HasMin = 0, 41 HasMax = 0, 42 HasSetLinear = 0 43 }; 44 }; 45 46 template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2}; }; 47 48 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); } 49 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); } 50 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) 51 { 52 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); 53 return Packet2cf(_mm_xor_ps(a.v,mask)); 54 } 55 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) 56 { 57 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); 58 return Packet2cf(_mm_xor_ps(a.v,mask)); 59 } 60 61 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 62 { 63 // TODO optimize it for SSE3 and 4 64 #ifdef EIGEN_VECTORIZE_SSE3 65 return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v), 66 _mm_mul_ps(_mm_movehdup_ps(a.v), 67 vec4f_swizzle1(b.v, 1, 0, 3, 2)))); 68 // return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), 69 // _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), 70 // vec4f_swizzle1(b.v, 1, 0, 3, 2)))); 71 #else 72 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000)); 73 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), 74 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), 75 vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); 76 #endif 77 } 78 79 template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); } 80 template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); } 81 template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); } 82 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); } 83 84 template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&real_ref(*from))); } 85 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&real_ref(*from))); } 86 87 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) 88 { 89 Packet2cf res; 90 #if EIGEN_GNUC_AT_MOST(4,2) 91 // workaround annoying "may be used uninitialized in this function" warning with gcc 4.2 92 res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from)); 93 #else 94 res.v = _mm_loadl_pi(res.v, (const __m64*)&from); 95 #endif 96 return Packet2cf(_mm_movelh_ps(res.v,res.v)); 97 } 98 99 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); } 100 101 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&real_ref(*to), from.v); } 102 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&real_ref(*to), from.v); } 103 104 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 105 106 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) 107 { 108 #if EIGEN_GNUC_AT_MOST(4,3) 109 // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares... 110 // This workaround also fix invalid code generation with gcc 4.3 111 EIGEN_ALIGN16 std::complex<float> res[2]; 112 _mm_store_ps((float*)res, a.v); 113 return res[0]; 114 #else 115 std::complex<float> res; 116 _mm_storel_pi((__m64*)&res, a.v); 117 return res; 118 #endif 119 } 120 121 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(_mm_castps_pd(a.v)))); } 122 123 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) 124 { 125 return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v)))); 126 } 127 128 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs) 129 { 130 return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v))); 131 } 132 133 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) 134 { 135 return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v)))); 136 } 137 138 template<int Offset> 139 struct palign_impl<Offset,Packet2cf> 140 { 141 static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second) 142 { 143 if (Offset==1) 144 { 145 first.v = _mm_movehl_ps(first.v, first.v); 146 first.v = _mm_movelh_ps(first.v, second.v); 147 } 148 } 149 }; 150 151 template<> struct conj_helper<Packet2cf, Packet2cf, false,true> 152 { 153 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 154 { return padd(pmul(x,y),c); } 155 156 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 157 { 158 #ifdef EIGEN_VECTORIZE_SSE3 159 return internal::pmul(a, pconj(b)); 160 #else 161 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); 162 return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), 163 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), 164 vec4f_swizzle1(b.v, 1, 0, 3, 2)))); 165 #endif 166 } 167 }; 168 169 template<> struct conj_helper<Packet2cf, Packet2cf, true,false> 170 { 171 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 172 { return padd(pmul(x,y),c); } 173 174 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 175 { 176 #ifdef EIGEN_VECTORIZE_SSE3 177 return internal::pmul(pconj(a), b); 178 #else 179 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); 180 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), 181 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), 182 vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask))); 183 #endif 184 } 185 }; 186 187 template<> struct conj_helper<Packet2cf, Packet2cf, true,true> 188 { 189 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const 190 { return padd(pmul(x,y),c); } 191 192 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const 193 { 194 #ifdef EIGEN_VECTORIZE_SSE3 195 return pconj(internal::pmul(a, b)); 196 #else 197 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000)); 198 return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask), 199 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), 200 vec4f_swizzle1(b.v, 1, 0, 3, 2)))); 201 #endif 202 } 203 }; 204 205 template<> struct conj_helper<Packet4f, Packet2cf, false,false> 206 { 207 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const 208 { return padd(c, pmul(x,y)); } 209 210 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const 211 { return Packet2cf(Eigen::internal::pmul(x, y.v)); } 212 }; 213 214 template<> struct conj_helper<Packet2cf, Packet4f, false,false> 215 { 216 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const 217 { return padd(c, pmul(x,y)); } 218 219 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const 220 { return Packet2cf(Eigen::internal::pmul(x.v, y)); } 221 }; 222 223 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) 224 { 225 // TODO optimize it for SSE3 and 4 226 Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b); 227 __m128 s = _mm_mul_ps(b.v,b.v); 228 return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1))))); 229 } 230 231 EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x) 232 { 233 return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2)); 234 } 235 236 237 //---------- double ---------- 238 struct Packet1cd 239 { 240 EIGEN_STRONG_INLINE Packet1cd() {} 241 EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {} 242 __m128d v; 243 }; 244 245 template<> struct packet_traits<std::complex<double> > : default_packet_traits 246 { 247 typedef Packet1cd type; 248 enum { 249 Vectorizable = 1, 250 AlignedOnScalar = 0, 251 size = 1, 252 253 HasAdd = 1, 254 HasSub = 1, 255 HasMul = 1, 256 HasDiv = 1, 257 HasNegate = 1, 258 HasAbs = 0, 259 HasAbs2 = 0, 260 HasMin = 0, 261 HasMax = 0, 262 HasSetLinear = 0 263 }; 264 }; 265 266 template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1}; }; 267 268 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); } 269 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); } 270 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(a.v)); } 271 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) 272 { 273 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); 274 return Packet1cd(_mm_xor_pd(a.v,mask)); 275 } 276 277 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b) 278 { 279 // TODO optimize it for SSE3 and 4 280 #ifdef EIGEN_VECTORIZE_SSE3 281 return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), 282 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), 283 vec2d_swizzle1(b.v, 1, 0)))); 284 #else 285 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0)); 286 return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), 287 _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), 288 vec2d_swizzle1(b.v, 1, 0)), mask))); 289 #endif 290 } 291 292 template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); } 293 template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); } 294 template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); } 295 template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); } 296 297 // FIXME force unaligned load, this is a temporary fix 298 template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) 299 { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); } 300 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) 301 { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); } 302 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from) 303 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); } 304 305 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); } 306 307 // FIXME force unaligned store, this is a temporary fix 308 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); } 309 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); } 310 311 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 312 313 template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) 314 { 315 EIGEN_ALIGN16 double res[2]; 316 _mm_store_pd(res, a.v); 317 return std::complex<double>(res[0],res[1]); 318 } 319 320 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; } 321 322 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) 323 { 324 return pfirst(a); 325 } 326 327 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) 328 { 329 return vecs[0]; 330 } 331 332 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) 333 { 334 return pfirst(a); 335 } 336 337 template<int Offset> 338 struct palign_impl<Offset,Packet1cd> 339 { 340 static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/) 341 { 342 // FIXME is it sure we never have to align a Packet1cd? 343 // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary... 344 } 345 }; 346 347 template<> struct conj_helper<Packet1cd, Packet1cd, false,true> 348 { 349 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 350 { return padd(pmul(x,y),c); } 351 352 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 353 { 354 #ifdef EIGEN_VECTORIZE_SSE3 355 return internal::pmul(a, pconj(b)); 356 #else 357 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); 358 return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), 359 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), 360 vec2d_swizzle1(b.v, 1, 0)))); 361 #endif 362 } 363 }; 364 365 template<> struct conj_helper<Packet1cd, Packet1cd, true,false> 366 { 367 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 368 { return padd(pmul(x,y),c); } 369 370 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 371 { 372 #ifdef EIGEN_VECTORIZE_SSE3 373 return internal::pmul(pconj(a), b); 374 #else 375 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); 376 return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), 377 _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), 378 vec2d_swizzle1(b.v, 1, 0)), mask))); 379 #endif 380 } 381 }; 382 383 template<> struct conj_helper<Packet1cd, Packet1cd, true,true> 384 { 385 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const 386 { return padd(pmul(x,y),c); } 387 388 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const 389 { 390 #ifdef EIGEN_VECTORIZE_SSE3 391 return pconj(internal::pmul(a, b)); 392 #else 393 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0)); 394 return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask), 395 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1), 396 vec2d_swizzle1(b.v, 1, 0)))); 397 #endif 398 } 399 }; 400 401 template<> struct conj_helper<Packet2d, Packet1cd, false,false> 402 { 403 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const 404 { return padd(c, pmul(x,y)); } 405 406 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const 407 { return Packet1cd(Eigen::internal::pmul(x, y.v)); } 408 }; 409 410 template<> struct conj_helper<Packet1cd, Packet2d, false,false> 411 { 412 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const 413 { return padd(c, pmul(x,y)); } 414 415 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const 416 { return Packet1cd(Eigen::internal::pmul(x.v, y)); } 417 }; 418 419 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) 420 { 421 // TODO optimize it for SSE3 and 4 422 Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b); 423 __m128d s = _mm_mul_pd(b.v,b.v); 424 return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1)))); 425 } 426 427 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x) 428 { 429 return Packet1cd(preverse(x.v)); 430 } 431 432 } // end namespace internal 433 434 } // end namespace Eigen 435 436 #endif // EIGEN_COMPLEX_SSE_H 437