xref: /aosp_15_r20/external/eigen/Eigen/src/Core/arch/NEON/Complex.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2010 Gael Guennebaud <[email protected]>
5 // Copyright (C) 2010 Konstantinos Margaritis <[email protected]>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10 
11 #ifndef EIGEN_COMPLEX_NEON_H
12 #define EIGEN_COMPLEX_NEON_H
13 
14 namespace Eigen {
15 
16 namespace internal {
17 
p4ui_CONJ_XOR()18 inline uint32x4_t p4ui_CONJ_XOR()
19 {
20 // See bug 1325, clang fails to call vld1q_u64.
21 #if EIGEN_COMP_CLANG || EIGEN_COMP_CASTXML
22   uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
23   return ret;
24 #else
25   static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
26   return vld1q_u32( conj_XOR_DATA );
27 #endif
28 }
29 
p2ui_CONJ_XOR()30 inline uint32x2_t p2ui_CONJ_XOR()
31 {
32   static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000 };
33   return vld1_u32( conj_XOR_DATA );
34 }
35 
36 //---------- float ----------
37 
38 struct Packet1cf
39 {
Packet1cfPacket1cf40   EIGEN_STRONG_INLINE Packet1cf() {}
Packet1cfPacket1cf41   EIGEN_STRONG_INLINE explicit Packet1cf(const Packet2f& a) : v(a) {}
42   Packet2f v;
43 };
44 struct Packet2cf
45 {
Packet2cfPacket2cf46   EIGEN_STRONG_INLINE Packet2cf() {}
Packet2cfPacket2cf47   EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
48   Packet4f v;
49 };
50 
51 template<> struct packet_traits<std::complex<float> > : default_packet_traits
52 {
53   typedef Packet2cf type;
54   typedef Packet1cf half;
55   enum
56   {
57     Vectorizable = 1,
58     AlignedOnScalar = 1,
59     size = 2,
60     HasHalfPacket = 1,
61 
62     HasAdd       = 1,
63     HasSub       = 1,
64     HasMul       = 1,
65     HasDiv       = 1,
66     HasNegate    = 1,
67     HasAbs       = 0,
68     HasAbs2      = 0,
69     HasMin       = 0,
70     HasMax       = 0,
71     HasSetLinear = 0
72   };
73 };
74 
75 template<> struct unpacket_traits<Packet1cf>
76 {
77   typedef std::complex<float> type;
78   typedef Packet1cf half;
79   typedef Packet2f as_real;
80   enum
81   {
82     size = 1,
83     alignment = Aligned16,
84     vectorizable = true,
85     masked_load_available = false,
86     masked_store_available = false
87   };
88 };
89 template<> struct unpacket_traits<Packet2cf>
90 {
91   typedef std::complex<float> type;
92   typedef Packet1cf half;
93   typedef Packet4f as_real;
94   enum
95   {
96     size = 2,
97     alignment = Aligned16,
98     vectorizable = true,
99     masked_load_available = false,
100     masked_store_available = false
101   };
102 };
103 
104 template<> EIGEN_STRONG_INLINE Packet1cf pcast<float,Packet1cf>(const float& a)
105 { return Packet1cf(vset_lane_f32(a, vdup_n_f32(0.f), 0)); }
106 template<> EIGEN_STRONG_INLINE Packet2cf pcast<Packet2f,Packet2cf>(const Packet2f& a)
107 { return Packet2cf(vreinterpretq_f32_u64(vmovl_u32(vreinterpret_u32_f32(a)))); }
108 
109 template<> EIGEN_STRONG_INLINE Packet1cf pset1<Packet1cf>(const std::complex<float>& from)
110 { return Packet1cf(vld1_f32(reinterpret_cast<const float*>(&from))); }
111 template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from)
112 {
113   const float32x2_t r64 = vld1_f32(reinterpret_cast<const float*>(&from));
114   return Packet2cf(vcombine_f32(r64, r64));
115 }
116 
117 template<> EIGEN_STRONG_INLINE Packet1cf padd<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
118 { return Packet1cf(padd<Packet2f>(a.v, b.v)); }
119 template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
120 { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
121 
122 template<> EIGEN_STRONG_INLINE Packet1cf psub<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
123 { return Packet1cf(psub<Packet2f>(a.v, b.v)); }
124 template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
125 { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
126 
127 template<> EIGEN_STRONG_INLINE Packet1cf pnegate(const Packet1cf& a) { return Packet1cf(pnegate<Packet2f>(a.v)); }
128 template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); }
129 
130 template<> EIGEN_STRONG_INLINE Packet1cf pconj(const Packet1cf& a)
131 {
132   const Packet2ui b = vreinterpret_u32_f32(a.v);
133   return Packet1cf(vreinterpret_f32_u32(veor_u32(b, p2ui_CONJ_XOR())));
134 }
135 template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
136 {
137   const Packet4ui b = vreinterpretq_u32_f32(a.v);
138   return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR())));
139 }
140 
141 template<> EIGEN_STRONG_INLINE Packet1cf pmul<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
142 {
143   Packet2f v1, v2;
144 
145   // Get the real values of a | a1_re | a1_re |
146   v1 = vdup_lane_f32(a.v, 0);
147   // Get the imag values of a | a1_im | a1_im |
148   v2 = vdup_lane_f32(a.v, 1);
149   // Multiply the real a with b
150   v1 = vmul_f32(v1, b.v);
151   // Multiply the imag a with b
152   v2 = vmul_f32(v2, b.v);
153   // Conjugate v2
154   v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR()));
155   // Swap real/imag elements in v2.
156   v2 = vrev64_f32(v2);
157   // Add and return the result
158   return Packet1cf(vadd_f32(v1, v2));
159 }
160 template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
161 {
162   Packet4f v1, v2;
163 
164   // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
165   v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0));
166   // Get the imag values of a | a1_im | a1_im | a2_im | a2_im |
167   v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1));
168   // Multiply the real a with b
169   v1 = vmulq_f32(v1, b.v);
170   // Multiply the imag a with b
171   v2 = vmulq_f32(v2, b.v);
172   // Conjugate v2
173   v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR()));
174   // Swap real/imag elements in v2.
175   v2 = vrev64q_f32(v2);
176   // Add and return the result
177   return Packet2cf(vaddq_f32(v1, v2));
178 }
179 
180 template<> EIGEN_STRONG_INLINE Packet1cf pcmp_eq(const Packet1cf& a, const Packet1cf& b)
181 {
182   // Compare real and imaginary parts of a and b to get the mask vector:
183   // [re(a[0])==re(b[0]), im(a[0])==im(b[0])]
184   Packet2f eq = pcmp_eq<Packet2f>(a.v, b.v);
185   // Swap real/imag elements in the mask in to get:
186   // [im(a[0])==im(b[0]), re(a[0])==re(b[0])]
187   Packet2f eq_swapped = vrev64_f32(eq);
188   // Return re(a)==re(b) && im(a)==im(b) by computing bitwise AND of eq and eq_swapped
189   return Packet1cf(pand<Packet2f>(eq, eq_swapped));
190 }
191 template<> EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b)
192 {
193   // Compare real and imaginary parts of a and b to get the mask vector:
194   // [re(a[0])==re(b[0]), im(a[0])==im(b[0]), re(a[1])==re(b[1]), im(a[1])==im(b[1])]
195   Packet4f eq = pcmp_eq<Packet4f>(a.v, b.v);
196   // Swap real/imag elements in the mask in to get:
197   // [im(a[0])==im(b[0]), re(a[0])==re(b[0]), im(a[1])==im(b[1]), re(a[1])==re(b[1])]
198   Packet4f eq_swapped = vrev64q_f32(eq);
199   // Return re(a)==re(b) && im(a)==im(b) by computing bitwise AND of eq and eq_swapped
200   return Packet2cf(pand<Packet4f>(eq, eq_swapped));
201 }
202 
203 template<> EIGEN_STRONG_INLINE Packet1cf pand<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
204 { return Packet1cf(vreinterpret_f32_u32(vand_u32(vreinterpret_u32_f32(a.v), vreinterpret_u32_f32(b.v)))); }
205 template<> EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
206 { return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v), vreinterpretq_u32_f32(b.v)))); }
207 
208 template<> EIGEN_STRONG_INLINE Packet1cf por<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
209 { return Packet1cf(vreinterpret_f32_u32(vorr_u32(vreinterpret_u32_f32(a.v), vreinterpret_u32_f32(b.v)))); }
210 template<> EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
211 { return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v), vreinterpretq_u32_f32(b.v)))); }
212 
213 template<> EIGEN_STRONG_INLINE Packet1cf pxor<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
214 { return Packet1cf(vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(a.v), vreinterpret_u32_f32(b.v)))); }
215 template<> EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
216 { return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v), vreinterpretq_u32_f32(b.v)))); }
217 
218 template<> EIGEN_STRONG_INLINE Packet1cf pandnot<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
219 { return Packet1cf(vreinterpret_f32_u32(vbic_u32(vreinterpret_u32_f32(a.v), vreinterpret_u32_f32(b.v)))); }
220 template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
221 { return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v), vreinterpretq_u32_f32(b.v)))); }
222 
223 template<> EIGEN_STRONG_INLINE Packet1cf pload<Packet1cf>(const std::complex<float>* from)
224 { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cf(pload<Packet2f>((const float*)from)); }
225 template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from)
226 { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(reinterpret_cast<const float*>(from))); }
227 
228 template<> EIGEN_STRONG_INLINE Packet1cf ploadu<Packet1cf>(const std::complex<float>* from)
229 { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cf(ploadu<Packet2f>((const float*)from)); }
230 template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)
231 { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(reinterpret_cast<const float*>(from))); }
232 
233 template<> EIGEN_STRONG_INLINE Packet1cf ploaddup<Packet1cf>(const std::complex<float>* from)
234 { return pset1<Packet1cf>(*from); }
235 template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from)
236 { return pset1<Packet2cf>(*from); }
237 
238 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *to, const Packet1cf& from)
239 { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
240 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *to, const Packet2cf& from)
241 { EIGEN_DEBUG_ALIGNED_STORE pstore(reinterpret_cast<float*>(to), from.v); }
242 
243 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *to, const Packet1cf& from)
244 { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
245 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *to, const Packet2cf& from)
246 { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<float*>(to), from.v); }
247 
248 template<> EIGEN_DEVICE_FUNC inline Packet1cf pgather<std::complex<float>, Packet1cf>(
249     const std::complex<float>* from, Index stride)
250 {
251   const Packet2f tmp = vdup_n_f32(std::real(from[0*stride]));
252   return Packet1cf(vset_lane_f32(std::imag(from[0*stride]), tmp, 1));
253 }
254 template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(
255     const std::complex<float>* from, Index stride)
256 {
257   Packet4f res = vdupq_n_f32(std::real(from[0*stride]));
258   res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1);
259   res = vsetq_lane_f32(std::real(from[1*stride]), res, 2);
260   res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3);
261   return Packet2cf(res);
262 }
263 
264 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet1cf>(
265     std::complex<float>* to, const Packet1cf& from, Index stride)
266 { to[stride*0] = std::complex<float>(vget_lane_f32(from.v, 0), vget_lane_f32(from.v, 1)); }
267 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(
268     std::complex<float>* to, const Packet2cf& from, Index stride)
269 {
270   to[stride*0] = std::complex<float>(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1));
271   to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3));
272 }
273 
274 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *addr)
275 { EIGEN_ARM_PREFETCH(reinterpret_cast<const float*>(addr)); }
276 
277 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet1cf>(const Packet1cf& a)
278 {
279   EIGEN_ALIGN16 std::complex<float> x;
280   vst1_f32(reinterpret_cast<float*>(&x), a.v);
281   return x;
282 }
283 template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
284 {
285   EIGEN_ALIGN16 std::complex<float> x[2];
286   vst1q_f32(reinterpret_cast<float*>(x), a.v);
287   return x[0];
288 }
289 
290 template<> EIGEN_STRONG_INLINE Packet1cf preverse(const Packet1cf& a) { return a; }
291 template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
292 { return Packet2cf(vcombine_f32(vget_high_f32(a.v), vget_low_f32(a.v))); }
293 
294 template<> EIGEN_STRONG_INLINE Packet1cf pcplxflip<Packet1cf>(const Packet1cf& a)
295 { return Packet1cf(vrev64_f32(a.v)); }
296 template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a)
297 { return Packet2cf(vrev64q_f32(a.v)); }
298 
299 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet1cf>(const Packet1cf& a)
300 {
301   std::complex<float> s;
302   vst1_f32((float *)&s, a.v);
303   return s;
304 }
305 template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
306 {
307   std::complex<float> s;
308   vst1_f32(reinterpret_cast<float*>(&s), vadd_f32(vget_low_f32(a.v), vget_high_f32(a.v)));
309   return s;
310 }
311 
312 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet1cf>(const Packet1cf& a)
313 {
314   std::complex<float> s;
315   vst1_f32((float *)&s, a.v);
316   return s;
317 }
318 template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
319 {
320   float32x2_t a1, a2, v1, v2, prod;
321   std::complex<float> s;
322 
323   a1 = vget_low_f32(a.v);
324   a2 = vget_high_f32(a.v);
325    // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
326   v1 = vdup_lane_f32(a1, 0);
327   // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
328   v2 = vdup_lane_f32(a1, 1);
329   // Multiply the real a with b
330   v1 = vmul_f32(v1, a2);
331   // Multiply the imag a with b
332   v2 = vmul_f32(v2, a2);
333   // Conjugate v2
334   v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR()));
335   // Swap real/imag elements in v2.
336   v2 = vrev64_f32(v2);
337   // Add v1, v2
338   prod = vadd_f32(v1, v2);
339 
340   vst1_f32(reinterpret_cast<float*>(&s), prod);
341 
342   return s;
343 }
344 
345 EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cf,Packet2f)
346 EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
347 
348 template<> EIGEN_STRONG_INLINE Packet1cf pdiv<Packet1cf>(const Packet1cf& a, const Packet1cf& b)
349 {
350   // TODO optimize it for NEON
351   Packet1cf res = pmul(a, pconj(b));
352   Packet2f s, rev_s;
353 
354   // this computes the norm
355   s = vmul_f32(b.v, b.v);
356   rev_s = vrev64_f32(s);
357 
358   return Packet1cf(pdiv<Packet2f>(res.v, vadd_f32(s, rev_s)));
359 }
360 template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
361 {
362   // TODO optimize it for NEON
363   Packet2cf res = pmul(a,pconj(b));
364   Packet4f s, rev_s;
365 
366   // this computes the norm
367   s = vmulq_f32(b.v, b.v);
368   rev_s = vrev64q_f32(s);
369 
370   return Packet2cf(pdiv<Packet4f>(res.v, vaddq_f32(s, rev_s)));
371 }
372 
373 EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet1cf, 1>& /*kernel*/) {}
374 EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel)
375 {
376   Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v));
377   kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v));
378   kernel.packet[1].v = tmp;
379 }
380 
381 template<> EIGEN_STRONG_INLINE Packet1cf psqrt<Packet1cf>(const Packet1cf& a) {
382   return psqrt_complex<Packet1cf>(a);
383 }
384 
385 template<> EIGEN_STRONG_INLINE Packet2cf psqrt<Packet2cf>(const Packet2cf& a) {
386   return psqrt_complex<Packet2cf>(a);
387 }
388 
389 //---------- double ----------
390 #if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
391 
392 // See bug 1325, clang fails to call vld1q_u64.
393 #if EIGEN_COMP_CLANG || EIGEN_COMP_CASTXML
394   static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000};
395 #else
396   const uint64_t  p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
397   static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA );
398 #endif
399 
400 struct Packet1cd
401 {
402   EIGEN_STRONG_INLINE Packet1cd() {}
403   EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
404   Packet2d v;
405 };
406 
407 template<> struct packet_traits<std::complex<double> >  : default_packet_traits
408 {
409   typedef Packet1cd type;
410   typedef Packet1cd half;
411   enum
412   {
413     Vectorizable = 1,
414     AlignedOnScalar = 0,
415     size = 1,
416     HasHalfPacket = 0,
417 
418     HasAdd    = 1,
419     HasSub    = 1,
420     HasMul    = 1,
421     HasDiv    = 1,
422     HasNegate = 1,
423     HasAbs    = 0,
424     HasAbs2   = 0,
425     HasMin    = 0,
426     HasMax    = 0,
427     HasSetLinear = 0
428   };
429 };
430 
431 template<> struct unpacket_traits<Packet1cd>
432 {
433   typedef std::complex<double> type;
434   typedef Packet1cd half;
435   typedef Packet2d as_real;
436   enum
437   {
438     size=1,
439     alignment=Aligned16,
440     vectorizable=true,
441     masked_load_available=false,
442     masked_store_available=false
443   };
444 };
445 
446 template<> EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from)
447 { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>(reinterpret_cast<const double*>(from))); }
448 
449 template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
450 { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>(reinterpret_cast<const double*>(from))); }
451 
452 template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>& from)
453 {
454   /* here we really have to use unaligned loads :( */
455   return ploadu<Packet1cd>(&from);
456 }
457 
458 template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
459 { return Packet1cd(padd<Packet2d>(a.v, b.v)); }
460 
461 template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
462 { return Packet1cd(psub<Packet2d>(a.v, b.v)); }
463 
464 template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a)
465 { return Packet1cd(pnegate<Packet2d>(a.v)); }
466 
467 template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
468 { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); }
469 
470 template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
471 {
472   Packet2d v1, v2;
473 
474   // Get the real values of a
475   v1 = vdupq_lane_f64(vget_low_f64(a.v), 0);
476   // Get the imag values of a
477   v2 = vdupq_lane_f64(vget_high_f64(a.v), 0);
478   // Multiply the real a with b
479   v1 = vmulq_f64(v1, b.v);
480   // Multiply the imag a with b
481   v2 = vmulq_f64(v2, b.v);
482   // Conjugate v2
483   v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR));
484   // Swap real/imag elements in v2.
485   v2 = preverse<Packet2d>(v2);
486   // Add and return the result
487   return Packet1cd(vaddq_f64(v1, v2));
488 }
489 
490 template<> EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b)
491 {
492   // Compare real and imaginary parts of a and b to get the mask vector:
493   // [re(a)==re(b), im(a)==im(b)]
494   Packet2d eq = pcmp_eq<Packet2d>(a.v, b.v);
495   // Swap real/imag elements in the mask in to get:
496   // [im(a)==im(b), re(a)==re(b)]
497   Packet2d eq_swapped = vreinterpretq_f64_u32(vrev64q_u32(vreinterpretq_u32_f64(eq)));
498   // Return re(a)==re(b) & im(a)==im(b) by computing bitwise AND of eq and eq_swapped
499   return Packet1cd(pand<Packet2d>(eq, eq_swapped));
500 }
501 
502 template<> EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
503 { return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); }
504 
505 template<> EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
506 { return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); }
507 
508 template<> EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
509 { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); }
510 
511 template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
512 { return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v)))); }
513 
514 template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from)
515 { return pset1<Packet1cd>(*from); }
516 
517 template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *to, const Packet1cd& from)
518 { EIGEN_DEBUG_ALIGNED_STORE pstore(reinterpret_cast<double*>(to), from.v); }
519 
520 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *to, const Packet1cd& from)
521 { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), from.v); }
522 
523 template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *addr)
524 { EIGEN_ARM_PREFETCH(reinterpret_cast<const double*>(addr)); }
525 
526 template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(
527     const std::complex<double>* from, Index stride)
528 {
529   Packet2d res = pset1<Packet2d>(0.0);
530   res = vsetq_lane_f64(std::real(from[0*stride]), res, 0);
531   res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1);
532   return Packet1cd(res);
533 }
534 
535 template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(
536     std::complex<double>* to, const Packet1cd& from, Index stride)
537 { to[stride*0] = std::complex<double>(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1)); }
538 
539 template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a)
540 {
541   EIGEN_ALIGN16 std::complex<double> res;
542   pstore<std::complex<double> >(&res, a);
543   return res;
544 }
545 
546 template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
547 
548 template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
549 
550 template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
551 
552 EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
553 
554 template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
555 {
556   // TODO optimize it for NEON
557   Packet1cd res = pmul(a,pconj(b));
558   Packet2d s = pmul<Packet2d>(b.v, b.v);
559   Packet2d rev_s = preverse<Packet2d>(s);
560 
561   return Packet1cd(pdiv(res.v, padd<Packet2d>(s,rev_s)));
562 }
563 
564 EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
565 { return Packet1cd(preverse(Packet2d(x.v))); }
566 
567 EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
568 {
569   Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v));
570   kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v));
571   kernel.packet[1].v = tmp;
572 }
573 
574 template<> EIGEN_STRONG_INLINE Packet1cd psqrt<Packet1cd>(const Packet1cd& a) {
575   return psqrt_complex<Packet1cd>(a);
576 }
577 
578 #endif // EIGEN_ARCH_ARM64
579 
580 } // end namespace internal
581 
582 } // end namespace Eigen
583 
584 #endif // EIGEN_COMPLEX_NEON_H
585