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
13namespace Eigen {
14
15namespace internal {
16
17//---------- float ----------
18struct 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
28template<> 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
53template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
54
55template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
56template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
57template<> 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}
62template<> 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
68template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
69{
70 #ifdef EIGEN_VECTORIZE_SSE3
71 return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
72 _mm_mul_ps(_mm_movehdup_ps(a.v),
73 vec4f_swizzle1(b.v, 1, 0, 3, 2))));
74// return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
75// _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
76// vec4f_swizzle1(b.v, 1, 0, 3, 2))));
77 #else
78 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000));
79 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
80 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
81 vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
82 #endif
83}
84
85template<> EIGEN_STRONG_INLINE Packet2cf pand <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
86template<> EIGEN_STRONG_INLINE Packet2cf por <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
87template<> EIGEN_STRONG_INLINE Packet2cf pxor <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
88template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); }
89
90template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); }
91template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); }
92
93template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from)
94{
95 Packet2cf res;
96#if EIGEN_GNUC_AT_MOST(4,2)
97 // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
98 res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
99#elif EIGEN_GNUC_AT_LEAST(4,6)
100 // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6
101 #pragma GCC diagnostic push
102 #pragma GCC diagnostic ignored "-Wuninitialized"
103 res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
104 #pragma GCC diagnostic pop
105#else
106 res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
107#endif
108 return Packet2cf(_mm_movelh_ps(res.v,res.v));
109}
110
111template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
112
113template<> 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)); }
114template<> 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)); }
115
116
117template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
118{
119 return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]),
120 std::imag(from[0*stride]), std::real(from[0*stride])));
121}
122
123template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
124{
125 to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
126 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
127 to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
128 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
129}
130
131template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
132
133template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
134{
135 #if EIGEN_GNUC_AT_MOST(4,3)
136 // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares...
137 // This workaround also fix invalid code generation with gcc 4.3
138 EIGEN_ALIGN16 std::complex<float> res[2];
139 _mm_store_ps((float*)res, a.v);
140 return res[0];
141 #else
142 std::complex<float> res;
143 _mm_storel_pi((__m64*)&res, a.v);
144 return res;
145 #endif
146}
147
148template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); }
149
150template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
151{
152 return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
153}
154
155template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
156{
157 return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
158}
159
160template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
161{
162 return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
163}
164
165template<int Offset>
166struct palign_impl<Offset,Packet2cf>
167{
168 static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
169 {
170 if (Offset==1)
171 {
172 first.v = _mm_movehl_ps(first.v, first.v);
173 first.v = _mm_movelh_ps(first.v, second.v);
174 }
175 }
176};
177
178template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
179{
180 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
181 { return padd(pmul(x,y),c); }
182
183 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
184 {
185 #ifdef EIGEN_VECTORIZE_SSE3
186 return internal::pmul(a, pconj(b));
187 #else
188 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
189 return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
190 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
191 vec4f_swizzle1(b.v, 1, 0, 3, 2))));
192 #endif
193 }
194};
195
196template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
197{
198 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
199 { return padd(pmul(x,y),c); }
200
201 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
202 {
203 #ifdef EIGEN_VECTORIZE_SSE3
204 return internal::pmul(pconj(a), b);
205 #else
206 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
207 return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
208 _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
209 vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
210 #endif
211 }
212};
213
214template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
215{
216 EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
217 { return padd(pmul(x,y),c); }
218
219 EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
220 {
221 #ifdef EIGEN_VECTORIZE_SSE3
222 return pconj(internal::pmul(a, b));
223 #else
224 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
225 return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
226 _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
227 vec4f_swizzle1(b.v, 1, 0, 3, 2))));
228 #endif
229 }
230};
231
232EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
233
234template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
235{
236 // TODO optimize it for SSE3 and 4
237 Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
238 __m128 s = _mm_mul_ps(b.v,b.v);
239 return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
240}
241
242EIGEN_STRONG_INLINE Packet2cf pcplxflip/* <Packet2cf> */(const Packet2cf& x)
243{
244 return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
245}
246
247
248//---------- double ----------
249struct Packet1cd
250{
251 EIGEN_STRONG_INLINE Packet1cd() {}
252 EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
253 __m128d v;
254};
255
256// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
257// to leverage AVX instructions.
258#ifndef EIGEN_VECTORIZE_AVX
259template<> struct packet_traits<std::complex<double> > : default_packet_traits
260{
261 typedef Packet1cd type;
262 typedef Packet1cd half;
263 enum {
264 Vectorizable = 1,
265 AlignedOnScalar = 0,
266 size = 1,
267 HasHalfPacket = 0,
268
269 HasAdd = 1,
270 HasSub = 1,
271 HasMul = 1,
272 HasDiv = 1,
273 HasNegate = 1,
274 HasAbs = 0,
275 HasAbs2 = 0,
276 HasMin = 0,
277 HasMax = 0,
278 HasSetLinear = 0
279 };
280};
281#endif
282
283template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
284
285template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
286template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
287template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
288template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
289{
290 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
291 return Packet1cd(_mm_xor_pd(a.v,mask));
292}
293
294template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
295{
296 #ifdef EIGEN_VECTORIZE_SSE3
297 return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v),
298 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
299 vec2d_swizzle1(b.v, 1, 0))));
300 #else
301 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
302 return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
303 _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
304 vec2d_swizzle1(b.v, 1, 0)), mask)));
305 #endif
306}
307
308template<> EIGEN_STRONG_INLINE Packet1cd pand <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
309template<> EIGEN_STRONG_INLINE Packet1cd por <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
310template<> EIGEN_STRONG_INLINE Packet1cd pxor <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
311template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
312
313// FIXME force unaligned load, this is a temporary fix
314template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from)
315{ EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
316template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
317{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
318template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from)
319{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
320
321template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
322
323// FIXME force unaligned store, this is a temporary fix
324template<> 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)); }
325template<> 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)); }
326
327template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
328
329template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
330{
331 EIGEN_ALIGN16 double res[2];
332 _mm_store_pd(res, a.v);
333 return std::complex<double>(res[0],res[1]);
334}
335
336template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
337
338template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
339{
340 return pfirst(a);
341}
342
343template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
344{
345 return vecs[0];
346}
347
348template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
349{
350 return pfirst(a);
351}
352
353template<int Offset>
354struct palign_impl<Offset,Packet1cd>
355{
356 static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
357 {
358 // FIXME is it sure we never have to align a Packet1cd?
359 // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
360 }
361};
362
363template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
364{
365 EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
366 { return padd(pmul(x,y),c); }
367
368 EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
369 {
370 #ifdef EIGEN_VECTORIZE_SSE3
371 return internal::pmul(a, pconj(b));
372 #else
373 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
374 return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
375 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
376 vec2d_swizzle1(b.v, 1, 0))));
377 #endif
378 }
379};
380
381template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
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(pconj(a), b);
390 #else
391 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
392 return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
393 _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
394 vec2d_swizzle1(b.v, 1, 0)), mask)));
395 #endif
396 }
397};
398
399template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
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 pconj(internal::pmul(a, b));
408 #else
409 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
410 return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
411 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
412 vec2d_swizzle1(b.v, 1, 0))));
413 #endif
414 }
415};
416
417EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
418
419template<> 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
427EIGEN_STRONG_INLINE Packet1cd pcplxflip/* <Packet1cd> */(const Packet1cd& x)
428{
429 return Packet1cd(preverse(Packet2d(x.v)));
430}
431
432EIGEN_DEVICE_FUNC inline void
433ptranspose(PacketBlock<Packet2cf,2>& kernel) {
434 __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
435 __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
436
437 __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
438 kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
439 kernel.packet[1].v = tmp;
440}
441
442template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
443 __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
444 return Packet2cf(_mm_castpd_ps(result));
445}
446
447template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b)
448{
449 return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b)));
450}
451
452template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b)
453{
454 return pset1<Packet1cd>(b);
455}
456
457template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b)
458{
459 return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b)));
460}
461
462template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b)
463{
464 return pset1<Packet1cd>(b);
465}
466
467} // end namespace internal
468
469} // end namespace Eigen
470
471#endif // EIGEN_COMPLEX_SSE_H
472