Eigen  5.0.1-dev+master
 
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#if EIGEN_GCC15_SLP_BUG && defined(EIGEN_VECTORIZE_SSE3)
148 // With FMA (which implies SSE3), GCC 15.1/15.2 miscompile _mm_set_ps(im, re, im, re) when from is a complex product
149 // computed in the same function, e.g. Matrix2cf v * v.inverse(); see EIGEN_GCC15_SLP_BUG. Broadcasting the complex as
150 // one 64-bit element (a single movddup) keeps re and im out of the SLP group. _mm_loadl_epi64 loads through __m128i,
151 // which is __may_alias__.
152 const __m128i lo64 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&from));
153 return Packet2cf(_mm_castpd_ps(_mm_movedup_pd(_mm_castsi128_pd(lo64))));
154#else
155 const float re = std::real(from);
156 const float im = std::imag(from);
157 return Packet2cf(_mm_set_ps(im, re, im, re));
158#endif
159}
160
161template <>
162EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) {
163 return pset1<Packet2cf>(*from);
164}
165
166template <>
167EIGEN_STRONG_INLINE void pstore<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
168 EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(&numext::real_ref(*to), from.v);
169}
170template <>
171EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
172 EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(&numext::real_ref(*to), from.v);
173}
174
175template <>
176EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from,
177 Index stride) {
178#if EIGEN_GCC15_SLP_BUG
179 // Two 64-bit element loads; the scalar lanes miscompile when both elements hold one complex product, see pset1.
180 const __m128i lo = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
181 const __m128i hi = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from + stride));
182 return Packet2cf(_mm_castsi128_ps(_mm_unpacklo_epi64(lo, hi)));
183#else
184 return Packet2cf(_mm_set_ps(std::imag(from[1 * stride]), std::real(from[1 * stride]), std::imag(from[0 * stride]),
185 std::real(from[0 * stride])));
186#endif
187}
188
189template <>
190EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from,
191 Index stride) {
192 to[stride * 0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
193 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
194 to[stride * 1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
195 _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
196}
197
198template <>
199EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float>* addr) {
200 _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
201}
202
203template <>
204EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) {
205 return std::complex<float>(_mm_cvtss_f32(a.v), _mm_cvtss_f32(_mm_shuffle_ps(a.v, a.v, _MM_SHUFFLE(1, 1, 1, 1))));
206}
207
208template <>
209EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) {
210 return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v)))));
211}
212
213template <>
214EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) {
215 return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v, a.v))));
216}
217
218template <>
219EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) {
220 return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v, a.v))));
221}
222
223EIGEN_STRONG_INLINE Packet2cf pcplxflip /* <Packet2cf> */ (const Packet2cf& x) {
224 return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
225}
226
227EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
228
229template <>
230EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
231 return pdiv_complex(a, b);
232}
233
234//---------- double ----------
235struct Packet1cd {
236 EIGEN_STRONG_INLINE Packet1cd() {}
237 EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
238 Packet2d v;
239};
240
241// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
242// to leverage AVX instructions.
243#ifndef EIGEN_VECTORIZE_AVX
244template <>
245struct packet_traits<std::complex<double> > : default_packet_traits {
246 using type = Packet1cd;
247 using half = Packet1cd;
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 HasSqrt = 1,
259 HasLog = 1,
260 HasAbs = 0,
261 HasAbs2 = 0,
262 HasMin = 0,
263 HasMax = 0,
264 HasSetLinear = 0,
265 HasExp = 1
266 };
267};
268#endif
269
270template <>
271struct unpacket_traits<Packet1cd> {
272 using type = std::complex<double>;
273 using half = Packet1cd;
274 using as_real = Packet2d;
275 enum {
276 size = 1,
277 alignment = Aligned16,
278 vectorizable = true,
279 masked_load_available = false,
280 masked_store_available = false
281 };
282};
283
284template <>
285EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
286 return Packet1cd(_mm_add_pd(a.v, b.v));
287}
288template <>
289EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
290 return Packet1cd(_mm_sub_pd(a.v, b.v));
291}
292template <>
293EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) {
294 return Packet1cd(pnegate(Packet2d(a.v)));
295}
296template <>
297EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) {
298 const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(static_cast<int32_t>(0x80000000), 0x0, 0x0, 0x0));
299 return Packet1cd(_mm_xor_pd(a.v, mask));
300}
301
302template <>
303EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) {
304 __m128d tmp1 = _mm_mul_pd(_mm_unpackhi_pd(a.v, a.v), vec2d_swizzle1(b.v, 1, 0));
305#ifdef EIGEN_VECTORIZE_SSE3
306 __m128d tmp2 = _mm_movedup_pd(a.v);
307#else
308 __m128d tmp2 = _mm_unpacklo_pd(a.v, a.v);
309#endif
310#ifdef EIGEN_VECTORIZE_FMA
311 __m128d result = _mm_fmaddsub_pd(tmp2, b.v, tmp1);
312#else
313#ifdef EIGEN_VECTORIZE_SSE3
314 __m128d result = _mm_addsub_pd(_mm_mul_pd(tmp2, b.v), tmp1);
315#else
316 const __m128d mask = _mm_castsi128_pd(_mm_set_epi64x(0x0, 0x8000000000000000ull));
317 __m128d result = _mm_add_pd(_mm_mul_pd(tmp2, b.v), _mm_xor_pd(tmp1, mask));
318#endif
319#endif
320 return Packet1cd(result);
321}
322
323template <>
324EIGEN_STRONG_INLINE Packet1cd ptrue<Packet1cd>(const Packet1cd& a) {
325 return Packet1cd(ptrue(Packet2d(a.v)));
326}
327template <>
328EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
329 return Packet1cd(_mm_and_pd(a.v, b.v));
330}
331template <>
332EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
333 return Packet1cd(_mm_or_pd(a.v, b.v));
334}
335template <>
336EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
337 return Packet1cd(_mm_xor_pd(a.v, b.v));
338}
339template <>
340EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
341 return Packet1cd(_mm_andnot_pd(b.v, a.v));
342}
343
344template <>
345EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) {
346 EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(_mm_load_pd((const double*)from));
347}
348template <>
349EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) {
350 EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(_mm_loadu_pd((const double*)from));
351}
352template <>
353EIGEN_STRONG_INLINE Packet1cd
354pset1<Packet1cd>(const std::complex<double>& from) { /* here we really have to use unaligned loads :( */
355 return ploadu<Packet1cd>(&from);
356}
357
358template <>
359EIGEN_STRONG_INLINE void pstore<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
360 EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd((double*)to, from.v);
361}
362template <>
363EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
364 EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd((double*)to, from.v);
365}
366
367template <>
368EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double>* addr) {
369 _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
370}
371
372template <>
373EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) {
374 EIGEN_ALIGN16 double res[2];
375 _mm_store_pd(res, a.v);
376 return std::complex<double>(res[0], res[1]);
377}
378
379EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
380
381template <>
382EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
383 return pdiv_complex(a, b);
384}
385
386EIGEN_STRONG_INLINE Packet1cd pcplxflip /* <Packet1cd> */ (const Packet1cd& x) {
387 return Packet1cd(preverse(Packet2d(x.v)));
388}
389
390EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
391 __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
392 __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
393
394 __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
395 kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
396 kernel.packet[1].v = tmp;
397}
398
399template <>
400EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) {
401 __m128 eq = _mm_cmpeq_ps(a.v, b.v);
402 return Packet2cf(pand<Packet4f>(eq, vec4f_swizzle1(eq, 1, 0, 3, 2)));
403}
404
405template <>
406EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) {
407 __m128d eq = _mm_cmpeq_pd(a.v, b.v);
408 return Packet1cd(pand<Packet2d>(eq, vec2d_swizzle1(eq, 1, 0)));
409}
410
411EIGEN_INSTANTIATE_COMPLEX_MATH_FUNCS(Packet1cd)
412EIGEN_INSTANTIATE_COMPLEX_MATH_FUNCS(Packet2cf)
413
414#ifdef EIGEN_VECTORIZE_FMA
415// std::complex<float>
416template <>
417EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
418 __m128 a_odd = _mm_movehdup_ps(a.v);
419 __m128 a_even = _mm_moveldup_ps(a.v);
420 __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
421 __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmaddsub_ps(a_odd, b_swap, c.v));
422 return Packet2cf(result);
423}
424template <>
425EIGEN_STRONG_INLINE Packet2cf pmsub(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
426 __m128 a_odd = _mm_movehdup_ps(a.v);
427 __m128 a_even = _mm_moveldup_ps(a.v);
428 __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
429 __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmsubadd_ps(a_odd, b_swap, c.v));
430 return Packet2cf(result);
431}
432template <>
433EIGEN_STRONG_INLINE Packet2cf pnmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
434 return pnegate(pmsub(a, b, c));
435}
436// std::complex<double>
437template <>
438EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
439 __m128d a_odd = _mm_permute_pd(a.v, 0x3);
440 __m128d a_even = _mm_movedup_pd(a.v);
441 __m128d b_swap = _mm_permute_pd(b.v, 0x1);
442 __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmaddsub_pd(a_odd, b_swap, c.v));
443 return Packet1cd(result);
444}
445template <>
446EIGEN_STRONG_INLINE Packet1cd pmsub(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
447 __m128d a_odd = _mm_permute_pd(a.v, 0x3);
448 __m128d a_even = _mm_movedup_pd(a.v);
449 __m128d b_swap = _mm_permute_pd(b.v, 0x1);
450 __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmsubadd_pd(a_odd, b_swap, c.v));
451 return Packet1cd(result);
452}
453template <>
454EIGEN_STRONG_INLINE Packet1cd pnmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
455 return pnegate(pmsub(a, b, c));
456}
457#endif
458
459EIGEN_GCC_FAST_MATH_COMPLEX_VECTORIZE_WORKAROUND_POP
460} // end namespace internal
461} // end namespace Eigen
462
463#endif // EIGEN_COMPLEX_SSE_H
@ Aligned16
Definition Constants.h:238