Eigen  5.0.1
 
Loading...
Searching...
No Matches
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// SPDX-License-Identifier: MPL-2.0
10
11#ifndef EIGEN_COMPLEX_SSE_H
12#define EIGEN_COMPLEX_SSE_H
13
14// IWYU pragma: private
15#include "../../InternalHeaderCheck.h"
16
17namespace Eigen {
18
19namespace internal {
20
21EIGEN_GCC_FAST_MATH_COMPLEX_VECTORIZE_WORKAROUND_PUSH
22
23//---------- float ----------
24struct Packet2cf {
25 EIGEN_STRONG_INLINE Packet2cf() {}
26 EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {}
27 Packet4f v;
28};
29
30// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
31// to leverage AVX instructions.
32#ifndef EIGEN_VECTORIZE_AVX
33template <>
34struct packet_traits<std::complex<float> > : default_packet_traits {
35 using type = Packet2cf;
36 using half = Packet2cf;
37 enum {
38 Vectorizable = 1,
39 AlignedOnScalar = 1,
40 size = 2,
41
42 HasAdd = 1,
43 HasSub = 1,
44 HasMul = 1,
45 HasDiv = 1,
46 HasNegate = 1,
47 HasSqrt = 1,
48 HasLog = 1,
49 HasExp = 1,
50 HasAbs = 0,
51 HasAbs2 = 0,
52 HasMin = 0,
53 HasMax = 0,
54 HasSetLinear = 0,
55 };
56};
57#endif
58
59template <>
60struct unpacket_traits<Packet2cf> {
61 using type = std::complex<float>;
62 using half = Packet2cf;
63 using as_real = Packet4f;
64 enum {
65 size = 2,
66 alignment = Aligned16,
67 vectorizable = true,
68 masked_load_available = false,
69 masked_store_available = false
70 };
71};
72
73template <>
74EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
75 return Packet2cf(_mm_add_ps(a.v, b.v));
76}
77template <>
78EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
79 return Packet2cf(_mm_sub_ps(a.v, b.v));
80}
81
82template <>
83EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) {
84 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000));
85 return Packet2cf(_mm_xor_ps(a.v, mask));
86}
87template <>
88EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) {
89 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000, 0x80000000, 0x00000000, 0x80000000));
90 return Packet2cf(_mm_xor_ps(a.v, mask));
91}
92
93template <>
94EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) {
95#ifdef EIGEN_VECTORIZE_SSE3
96 __m128 tmp1 = _mm_mul_ps(_mm_movehdup_ps(a.v), vec4f_swizzle1(b.v, 1, 0, 3, 2));
97 __m128 tmp2 = _mm_moveldup_ps(a.v);
98#else
99 __m128 tmp1 = _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), vec4f_swizzle1(b.v, 1, 0, 3, 2));
100 __m128 tmp2 = vec4f_swizzle1(a.v, 0, 0, 2, 2);
101#endif
102#ifdef EIGEN_VECTORIZE_FMA
103 __m128 result = _mm_fmaddsub_ps(tmp2, b.v, tmp1);
104#else
105#ifdef EIGEN_VECTORIZE_SSE3
106 __m128 result = _mm_addsub_ps(_mm_mul_ps(tmp2, b.v), tmp1);
107#else
108 const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000, 0x00000000, 0x80000000, 0x00000000));
109 __m128 result = _mm_add_ps(_mm_mul_ps(tmp2, b.v), _mm_xor_ps(tmp1, mask));
110#endif
111#endif
112 return Packet2cf(result);
113}
114
115template <>
116EIGEN_STRONG_INLINE Packet2cf ptrue<Packet2cf>(const Packet2cf& a) {
117 return Packet2cf(ptrue(Packet4f(a.v)));
118}
119template <>
120EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
121 return Packet2cf(_mm_and_ps(a.v, b.v));
122}
123template <>
124EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
125 return Packet2cf(_mm_or_ps(a.v, b.v));
126}
127template <>
128EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
129 return Packet2cf(_mm_xor_ps(a.v, b.v));
130}
131template <>
132EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
133 return Packet2cf(_mm_andnot_ps(b.v, a.v));
134}
135
136template <>
137EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) {
138 EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(_mm_load_ps(&numext::real_ref(*from)));
139}
140template <>
141EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) {
142 EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(_mm_loadu_ps(&numext::real_ref(*from)));
143}
144
145template <>
146EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) {
147 const float re = std::real(from);
148 const float im = std::imag(from);
149 return Packet2cf(_mm_set_ps(im, re, im, re));
150}
151
152template <>
153EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) {
154 return pset1<Packet2cf>(*from);
155}
156
157template <>
158EIGEN_STRONG_INLINE void pstore<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
159 EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(&numext::real_ref(*to), from.v);
160}
161template <>
162EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
163 EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(&numext::real_ref(*to), from.v);
164}
165
166template <>
167EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from,
168 Index stride) {
169 return Packet2cf(_mm_set_ps(std::imag(from[1 * stride]), std::real(from[1 * stride]), std::imag(from[0 * stride]),
170 std::real(from[0 * stride])));
171}
172
173template <>
174EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from,
175 Index stride) {
176 to[stride * 0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
177 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
178 to[stride * 1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
179 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
180}
181
182template <>
183EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float>* addr) {
184 _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
185}
186
187template <>
188EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) {
189 alignas(alignof(__m64)) std::complex<float> res;
190 _mm_storel_pi((__m64*)&res, a.v);
191 return res;
192}
193
194template <>
195EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) {
196 return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v)))));
197}
198
199template <>
200EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) {
201 return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v, a.v))));
202}
203
204template <>
205EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) {
206 return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v, a.v))));
207}
208
209EIGEN_STRONG_INLINE Packet2cf pcplxflip /* <Packet2cf> */ (const Packet2cf& x) {
210 return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
211}
212
213EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
214
215template <>
216EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
217 return pdiv_complex(a, b);
218}
219
220//---------- double ----------
221struct Packet1cd {
222 EIGEN_STRONG_INLINE Packet1cd() {}
223 EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
224 Packet2d v;
225};
226
227// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
228// to leverage AVX instructions.
229#ifndef EIGEN_VECTORIZE_AVX
230template <>
231struct packet_traits<std::complex<double> > : default_packet_traits {
232 using type = Packet1cd;
233 using half = Packet1cd;
234 enum {
235 Vectorizable = 1,
236 AlignedOnScalar = 0,
237 size = 1,
238
239 HasAdd = 1,
240 HasSub = 1,
241 HasMul = 1,
242 HasDiv = 1,
243 HasNegate = 1,
244 HasSqrt = 1,
245 HasLog = 1,
246 HasAbs = 0,
247 HasAbs2 = 0,
248 HasMin = 0,
249 HasMax = 0,
250 HasSetLinear = 0,
251 HasExp = 1
252 };
253};
254#endif
255
256template <>
257struct unpacket_traits<Packet1cd> {
258 using type = std::complex<double>;
259 using half = Packet1cd;
260 using as_real = Packet2d;
261 enum {
262 size = 1,
263 alignment = Aligned16,
264 vectorizable = true,
265 masked_load_available = false,
266 masked_store_available = false
267 };
268};
269
270template <>
271EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
272 return Packet1cd(_mm_add_pd(a.v, b.v));
273}
274template <>
275EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
276 return Packet1cd(_mm_sub_pd(a.v, b.v));
277}
278template <>
279EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) {
280 return Packet1cd(pnegate(Packet2d(a.v)));
281}
282template <>
283EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) {
284 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(static_cast<int32_t>(0x80000000), 0x0, 0x0, 0x0));
285 return Packet1cd(_mm_xor_pd(a.v, mask));
286}
287
288template <>
289EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) {
290 __m128d tmp1 = _mm_mul_pd(_mm_unpackhi_pd(a.v, a.v), vec2d_swizzle1(b.v, 1, 0));
291#ifdef EIGEN_VECTORIZE_SSE3
292 __m128d tmp2 = _mm_movedup_pd(a.v);
293#else
294 __m128d tmp2 = _mm_unpacklo_pd(a.v, a.v);
295#endif
296#ifdef EIGEN_VECTORIZE_FMA
297 __m128d result = _mm_fmaddsub_pd(tmp2, b.v, tmp1);
298#else
299#ifdef EIGEN_VECTORIZE_SSE3
300 __m128d result = _mm_addsub_pd(_mm_mul_pd(tmp2, b.v), tmp1);
301#else
302 const __m128d mask = _mm_castsi128_pd(_mm_set_epi64x(0x0, 0x8000000000000000ull));
303 __m128d result = _mm_add_pd(_mm_mul_pd(tmp2, b.v), _mm_xor_pd(tmp1, mask));
304#endif
305#endif
306 return Packet1cd(result);
307}
308
309template <>
310EIGEN_STRONG_INLINE Packet1cd ptrue<Packet1cd>(const Packet1cd& a) {
311 return Packet1cd(ptrue(Packet2d(a.v)));
312}
313template <>
314EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
315 return Packet1cd(_mm_and_pd(a.v, b.v));
316}
317template <>
318EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
319 return Packet1cd(_mm_or_pd(a.v, b.v));
320}
321template <>
322EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
323 return Packet1cd(_mm_xor_pd(a.v, b.v));
324}
325template <>
326EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
327 return Packet1cd(_mm_andnot_pd(b.v, a.v));
328}
329
330template <>
331EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) {
332 EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(_mm_load_pd((const double*)from));
333}
334template <>
335EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) {
336 EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(_mm_loadu_pd((const double*)from));
337}
338template <>
339EIGEN_STRONG_INLINE Packet1cd
340pset1<Packet1cd>(const std::complex<double>& from) { /* here we really have to use unaligned loads :( */
341 return ploadu<Packet1cd>(&from);
342}
343
344template <>
345EIGEN_STRONG_INLINE void pstore<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
346 EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd((double*)to, from.v);
347}
348template <>
349EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
350 EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd((double*)to, from.v);
351}
352
353template <>
354EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double>* addr) {
355 _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
356}
357
358template <>
359EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) {
360 EIGEN_ALIGN16 double res[2];
361 _mm_store_pd(res, a.v);
362 return std::complex<double>(res[0], res[1]);
363}
364
365EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
366
367template <>
368EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
369 return pdiv_complex(a, b);
370}
371
372EIGEN_STRONG_INLINE Packet1cd pcplxflip /* <Packet1cd> */ (const Packet1cd& x) {
373 return Packet1cd(preverse(Packet2d(x.v)));
374}
375
376EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
377 __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
378 __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
379
380 __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
381 kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
382 kernel.packet[1].v = tmp;
383}
384
385template <>
386EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) {
387 __m128 eq = _mm_cmpeq_ps(a.v, b.v);
388 return Packet2cf(pand<Packet4f>(eq, vec4f_swizzle1(eq, 1, 0, 3, 2)));
389}
390
391template <>
392EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) {
393 __m128d eq = _mm_cmpeq_pd(a.v, b.v);
394 return Packet1cd(pand<Packet2d>(eq, vec2d_swizzle1(eq, 1, 0)));
395}
396
397EIGEN_INSTANTIATE_COMPLEX_MATH_FUNCS(Packet1cd)
398EIGEN_INSTANTIATE_COMPLEX_MATH_FUNCS(Packet2cf)
399
400#ifdef EIGEN_VECTORIZE_FMA
401// std::complex<float>
402template <>
403EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
404 __m128 a_odd = _mm_movehdup_ps(a.v);
405 __m128 a_even = _mm_moveldup_ps(a.v);
406 __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
407 __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmaddsub_ps(a_odd, b_swap, c.v));
408 return Packet2cf(result);
409}
410template <>
411EIGEN_STRONG_INLINE Packet2cf pmsub(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
412 __m128 a_odd = _mm_movehdup_ps(a.v);
413 __m128 a_even = _mm_moveldup_ps(a.v);
414 __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
415 __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmsubadd_ps(a_odd, b_swap, c.v));
416 return Packet2cf(result);
417}
418template <>
419EIGEN_STRONG_INLINE Packet2cf pnmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
420 return pnegate(pmsub(a, b, c));
421}
422template <>
423EIGEN_STRONG_INLINE Packet2cf pnmsub(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
424 return pnegate(pmadd(a, b, c));
425}
426// std::complex<double>
427template <>
428EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
429 __m128d a_odd = _mm_permute_pd(a.v, 0x3);
430 __m128d a_even = _mm_movedup_pd(a.v);
431 __m128d b_swap = _mm_permute_pd(b.v, 0x1);
432 __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmaddsub_pd(a_odd, b_swap, c.v));
433 return Packet1cd(result);
434}
435template <>
436EIGEN_STRONG_INLINE Packet1cd pmsub(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
437 __m128d a_odd = _mm_permute_pd(a.v, 0x3);
438 __m128d a_even = _mm_movedup_pd(a.v);
439 __m128d b_swap = _mm_permute_pd(b.v, 0x1);
440 __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmsubadd_pd(a_odd, b_swap, c.v));
441 return Packet1cd(result);
442}
443template <>
444EIGEN_STRONG_INLINE Packet1cd pnmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
445 return pnegate(pmsub(a, b, c));
446}
447template <>
448EIGEN_STRONG_INLINE Packet1cd pnmsub(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
449 return pnegate(pmadd(a, b, c));
450}
451#endif
452
453EIGEN_GCC_FAST_MATH_COMPLEX_VECTORIZE_WORKAROUND_POP
454} // end namespace internal
455} // end namespace Eigen
456
457#endif // EIGEN_COMPLEX_SSE_H
Eigen::Aligned16
@ Aligned16
Definition Constants.h:238