Eigen  3.2.91
SSE/Complex.h
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@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 // Use the packet_traits defined in AVX/PacketMath.h instead if we're going
26 // to leverage AVX instructions.
27 #ifndef EIGEN_VECTORIZE_AVX
28 template<> struct packet_traits<std::complex<float> > : default_packet_traits
29 {
30  typedef Packet2cf type;
31  typedef Packet2cf half;
32  enum {
33  Vectorizable = 1,
34  AlignedOnScalar = 1,
35  size = 2,
36  HasHalfPacket = 0,
37 
38  HasAdd = 1,
39  HasSub = 1,
40  HasMul = 1,
41  HasDiv = 1,
42  HasNegate = 1,
43  HasAbs = 0,
44  HasAbs2 = 0,
45  HasMin = 0,
46  HasMax = 0,
47  HasSetLinear = 0,
48  HasBlend = 1
49  };
50 };
51 #endif
52 
53 template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
54 
55 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
56 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
57 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a)
58 {
59  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
60  return Packet2cf(_mm_xor_ps(a.v,mask));
61 }
62 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
63 {
64  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
65  return Packet2cf(_mm_xor_ps(a.v,mask));
66 }
67 
68 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
69 {
70  // TODO optimize it for SSE3 and 4
71  #ifdef EIGEN_VECTORIZE_SSE3
72  return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
73  _mm_mul_ps(_mm_movehdup_ps(a.v),
74  vec4f_swizzle1(b.v, 1, 0, 3, 2))));
75 // return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
76 // _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
77 // vec4f_swizzle1(b.v, 1, 0, 3, 2))));
78  #else
79  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000));
80  return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
81  _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
82  vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
83  #endif
84 }
85 
86 template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
87 template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
88 template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
89 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); }
90 
91 template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); }
92 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); }
93 
94 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from)
95 {
96  Packet2cf res;
97 #if EIGEN_GNUC_AT_MOST(4,2)
98  // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
99  res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
100 #elif EIGEN_GNUC_AT_LEAST(4,6)
101  // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6
102  #pragma GCC diagnostic push
103  #pragma GCC diagnostic ignored "-Wuninitialized"
104  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
105  #pragma GCC diagnostic pop
106 #else
107  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
108 #endif
109  return Packet2cf(_mm_movelh_ps(res.v,res.v));
110 }
111 
112 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
113 
114 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); }
115 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> * to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); }
116 
117 
118 template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
119 {
120  return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]),
121  std::imag(from[0*stride]), std::real(from[0*stride])));
122 }
123 
124 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
125 {
126  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
127  _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
128  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
129  _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
130 }
131 
132 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
133 
134 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
135 {
136  #if EIGEN_GNUC_AT_MOST(4,3)
137  // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares...
138  // This workaround also fix invalid code generation with gcc 4.3
139  EIGEN_ALIGN16 std::complex<float> res[2];
140  _mm_store_ps((float*)res, a.v);
141  return res[0];
142  #else
143  std::complex<float> res;
144  _mm_storel_pi((__m64*)&res, a.v);
145  return res;
146  #endif
147 }
148 
149 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); }
150 
151 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
152 {
153  return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
154 }
155 
156 template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
157 {
158  return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
159 }
160 
161 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
162 {
163  return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
164 }
165 
166 template<int Offset>
167 struct palign_impl<Offset,Packet2cf>
168 {
169  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
170  {
171  if (Offset==1)
172  {
173  first.v = _mm_movehl_ps(first.v, first.v);
174  first.v = _mm_movelh_ps(first.v, second.v);
175  }
176  }
177 };
178 
179 template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
180 {
181  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
182  { return padd(pmul(x,y),c); }
183 
184  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
185  {
186  #ifdef EIGEN_VECTORIZE_SSE3
187  return internal::pmul(a, pconj(b));
188  #else
189  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
190  return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
191  _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
192  vec4f_swizzle1(b.v, 1, 0, 3, 2))));
193  #endif
194  }
195 };
196 
197 template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
198 {
199  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
200  { return padd(pmul(x,y),c); }
201 
202  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
203  {
204  #ifdef EIGEN_VECTORIZE_SSE3
205  return internal::pmul(pconj(a), b);
206  #else
207  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
208  return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
209  _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
210  vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
211  #endif
212  }
213 };
214 
215 template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
216 {
217  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
218  { return padd(pmul(x,y),c); }
219 
220  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
221  {
222  #ifdef EIGEN_VECTORIZE_SSE3
223  return pconj(internal::pmul(a, b));
224  #else
225  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
226  return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
227  _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
228  vec4f_swizzle1(b.v, 1, 0, 3, 2))));
229  #endif
230  }
231 };
232 
233 template<> struct conj_helper<Packet4f, Packet2cf, false,false>
234 {
235  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
236  { return padd(c, pmul(x,y)); }
237 
238  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
239  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x, y.v)); }
240 };
241 
242 template<> struct conj_helper<Packet2cf, Packet4f, false,false>
243 {
244  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
245  { return padd(c, pmul(x,y)); }
246 
247  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
248  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x.v, y)); }
249 };
250 
251 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
252 {
253  // TODO optimize it for SSE3 and 4
254  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
255  __m128 s = _mm_mul_ps(b.v,b.v);
256  return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
257 }
258 
259 EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
260 {
261  return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
262 }
263 
264 
265 //---------- double ----------
266 struct Packet1cd
267 {
268  EIGEN_STRONG_INLINE Packet1cd() {}
269  EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
270  __m128d v;
271 };
272 
273 // Use the packet_traits defined in AVX/PacketMath.h instead if we're going
274 // to leverage AVX instructions.
275 #ifndef EIGEN_VECTORIZE_AVX
276 template<> struct packet_traits<std::complex<double> > : default_packet_traits
277 {
278  typedef Packet1cd type;
279  typedef Packet1cd half;
280  enum {
281  Vectorizable = 1,
282  AlignedOnScalar = 0,
283  size = 1,
284  HasHalfPacket = 0,
285 
286  HasAdd = 1,
287  HasSub = 1,
288  HasMul = 1,
289  HasDiv = 1,
290  HasNegate = 1,
291  HasAbs = 0,
292  HasAbs2 = 0,
293  HasMin = 0,
294  HasMax = 0,
295  HasSetLinear = 0
296  };
297 };
298 #endif
299 
300 template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
301 
302 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
303 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
304 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
305 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
306 {
307  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
308  return Packet1cd(_mm_xor_pd(a.v,mask));
309 }
310 
311 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
312 {
313  // TODO optimize it for SSE3 and 4
314  #ifdef EIGEN_VECTORIZE_SSE3
315  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
316  _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
317  vec2d_swizzle1(b.v, 1, 0))));
318  #else
319  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
320  return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
321  _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
322  vec2d_swizzle1(b.v, 1, 0)), mask)));
323  #endif
324 }
325 
326 template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
327 template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
328 template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
329 template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
330 
331 // FIXME force unaligned load, this is a temporary fix
332 template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from)
333 { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
334 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
335 { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
336 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from)
337 { /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
338 
339 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
340 
341 // FIXME force unaligned store, this is a temporary fix
342 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
343 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> * to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
344 
345 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
346 
347 template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
348 {
349  EIGEN_ALIGN16 double res[2];
350  _mm_store_pd(res, a.v);
351  return std::complex<double>(res[0],res[1]);
352 }
353 
354 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
355 
356 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
357 {
358  return pfirst(a);
359 }
360 
361 template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
362 {
363  return vecs[0];
364 }
365 
366 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
367 {
368  return pfirst(a);
369 }
370 
371 template<int Offset>
372 struct palign_impl<Offset,Packet1cd>
373 {
374  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
375  {
376  // FIXME is it sure we never have to align a Packet1cd?
377  // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
378  }
379 };
380 
381 template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
382 {
383  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
384  { return padd(pmul(x,y),c); }
385 
386  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
387  {
388  #ifdef EIGEN_VECTORIZE_SSE3
389  return internal::pmul(a, pconj(b));
390  #else
391  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
392  return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
393  _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
394  vec2d_swizzle1(b.v, 1, 0))));
395  #endif
396  }
397 };
398 
399 template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
400 {
401  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
402  { return padd(pmul(x,y),c); }
403 
404  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
405  {
406  #ifdef EIGEN_VECTORIZE_SSE3
407  return internal::pmul(pconj(a), b);
408  #else
409  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
410  return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
411  _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
412  vec2d_swizzle1(b.v, 1, 0)), mask)));
413  #endif
414  }
415 };
416 
417 template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
418 {
419  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
420  { return padd(pmul(x,y),c); }
421 
422  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
423  {
424  #ifdef EIGEN_VECTORIZE_SSE3
425  return pconj(internal::pmul(a, b));
426  #else
427  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
428  return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
429  _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
430  vec2d_swizzle1(b.v, 1, 0))));
431  #endif
432  }
433 };
434 
435 template<> struct conj_helper<Packet2d, Packet1cd, false,false>
436 {
437  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
438  { return padd(c, pmul(x,y)); }
439 
440  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
441  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x, y.v)); }
442 };
443 
444 template<> struct conj_helper<Packet1cd, Packet2d, false,false>
445 {
446  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
447  { return padd(c, pmul(x,y)); }
448 
449  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
450  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x.v, y)); }
451 };
452 
453 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
454 {
455  // TODO optimize it for SSE3 and 4
456  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
457  __m128d s = _mm_mul_pd(b.v,b.v);
458  return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1))));
459 }
460 
461 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
462 {
463  return Packet1cd(preverse(Packet2d(x.v)));
464 }
465 
466 EIGEN_DEVICE_FUNC inline void
467 ptranspose(PacketBlock<Packet2cf,2>& kernel) {
468  __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
469  __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
470 
471  __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
472  kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
473  kernel.packet[1].v = tmp;
474 }
475 
476 template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
477  __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
478  return Packet2cf(_mm_castpd_ps(result));
479 }
480 
481 } // end namespace internal
482 
483 } // end namespace Eigen
484 
485 #endif // EIGEN_COMPLEX_SSE_H
Definition: LDLT.h:16
Definition: StdDeque.h:58
Definition: Constants.h:222
Definition: Eigen_Colamd.h:54