xref: /aosp_15_r20/external/eigen/Eigen/src/Core/GenericPacketMath.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008 Gael Guennebaud <[email protected]>
5 // Copyright (C) 2006-2008 Benoit Jacob <[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_GENERIC_PACKET_MATH_H
12 #define EIGEN_GENERIC_PACKET_MATH_H
13 
14 namespace Eigen {
15 
16 namespace internal {
17 
18 /** \internal
19   * \file GenericPacketMath.h
20   *
21   * Default implementation for types not supported by the vectorization.
22   * In practice these functions are provided to make easier the writing
23   * of generic vectorized code.
24   */
25 
26 #ifndef EIGEN_DEBUG_ALIGNED_LOAD
27 #define EIGEN_DEBUG_ALIGNED_LOAD
28 #endif
29 
30 #ifndef EIGEN_DEBUG_UNALIGNED_LOAD
31 #define EIGEN_DEBUG_UNALIGNED_LOAD
32 #endif
33 
34 #ifndef EIGEN_DEBUG_ALIGNED_STORE
35 #define EIGEN_DEBUG_ALIGNED_STORE
36 #endif
37 
38 #ifndef EIGEN_DEBUG_UNALIGNED_STORE
39 #define EIGEN_DEBUG_UNALIGNED_STORE
40 #endif
41 
42 struct default_packet_traits
43 {
44   enum {
45     HasHalfPacket = 0,
46 
47     HasAdd       = 1,
48     HasSub       = 1,
49     HasShift     = 1,
50     HasMul       = 1,
51     HasNegate    = 1,
52     HasAbs       = 1,
53     HasArg       = 0,
54     HasAbs2      = 1,
55     HasAbsDiff   = 0,
56     HasMin       = 1,
57     HasMax       = 1,
58     HasConj      = 1,
59     HasSetLinear = 1,
60     HasBlend     = 0,
61     // This flag is used to indicate whether packet comparison is supported.
62     // pcmp_eq, pcmp_lt and pcmp_le should be defined for it to be true.
63     HasCmp       = 0,
64 
65     HasDiv    = 0,
66     HasSqrt   = 0,
67     HasRsqrt  = 0,
68     HasExp    = 0,
69     HasExpm1  = 0,
70     HasLog    = 0,
71     HasLog1p  = 0,
72     HasLog10  = 0,
73     HasPow    = 0,
74 
75     HasSin    = 0,
76     HasCos    = 0,
77     HasTan    = 0,
78     HasASin   = 0,
79     HasACos   = 0,
80     HasATan   = 0,
81     HasSinh   = 0,
82     HasCosh   = 0,
83     HasTanh   = 0,
84     HasLGamma = 0,
85     HasDiGamma = 0,
86     HasZeta = 0,
87     HasPolygamma = 0,
88     HasErf = 0,
89     HasErfc = 0,
90     HasNdtri = 0,
91     HasBessel = 0,
92     HasIGamma = 0,
93     HasIGammaDerA = 0,
94     HasGammaSampleDerAlpha = 0,
95     HasIGammac = 0,
96     HasBetaInc = 0,
97 
98     HasRound  = 0,
99     HasRint   = 0,
100     HasFloor  = 0,
101     HasCeil   = 0,
102     HasSign   = 0
103   };
104 };
105 
106 template<typename T> struct packet_traits : default_packet_traits
107 {
108   typedef T type;
109   typedef T half;
110   enum {
111     Vectorizable = 0,
112     size = 1,
113     AlignedOnScalar = 0,
114     HasHalfPacket = 0
115   };
116   enum {
117     HasAdd    = 0,
118     HasSub    = 0,
119     HasMul    = 0,
120     HasNegate = 0,
121     HasAbs    = 0,
122     HasAbs2   = 0,
123     HasMin    = 0,
124     HasMax    = 0,
125     HasConj   = 0,
126     HasSetLinear = 0
127   };
128 };
129 
130 template<typename T> struct packet_traits<const T> : packet_traits<T> { };
131 
132 template<typename T> struct unpacket_traits
133 {
134   typedef T type;
135   typedef T half;
136   enum
137   {
138     size = 1,
139     alignment = 1,
140     vectorizable = false,
141     masked_load_available=false,
142     masked_store_available=false
143   };
144 };
145 
146 template<typename T> struct unpacket_traits<const T> : unpacket_traits<T> { };
147 
148 template <typename Src, typename Tgt> struct type_casting_traits {
149   enum {
150     VectorizedCast = 0,
151     SrcCoeffRatio = 1,
152     TgtCoeffRatio = 1
153   };
154 };
155 
156 /** \internal Wrapper to ensure that multiple packet types can map to the same
157     same underlying vector type. */
158 template<typename T, int unique_id = 0>
159 struct eigen_packet_wrapper
160 {
161   EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
162   EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
163   EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
164   EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
165   EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
166     m_val = v;
167     return *this;
168   }
169 
170   T m_val;
171 };
172 
173 
174 /** \internal A convenience utility for determining if the type is a scalar.
175  * This is used to enable some generic packet implementations.
176  */
177 template<typename Packet>
178 struct is_scalar {
179   typedef typename unpacket_traits<Packet>::type Scalar;
180   enum {
181     value = internal::is_same<Packet, Scalar>::value
182   };
183 };
184 
185 /** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
186 template <typename SrcPacket, typename TgtPacket>
187 EIGEN_DEVICE_FUNC inline TgtPacket
188 pcast(const SrcPacket& a) {
189   return static_cast<TgtPacket>(a);
190 }
191 template <typename SrcPacket, typename TgtPacket>
192 EIGEN_DEVICE_FUNC inline TgtPacket
193 pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
194   return static_cast<TgtPacket>(a);
195 }
196 template <typename SrcPacket, typename TgtPacket>
197 EIGEN_DEVICE_FUNC inline TgtPacket
198 pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const SrcPacket& /*d*/) {
199   return static_cast<TgtPacket>(a);
200 }
201 template <typename SrcPacket, typename TgtPacket>
202 EIGEN_DEVICE_FUNC inline TgtPacket
203 pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const SrcPacket& /*d*/,
204       const SrcPacket& /*e*/, const SrcPacket& /*f*/, const SrcPacket& /*g*/, const SrcPacket& /*h*/) {
205   return static_cast<TgtPacket>(a);
206 }
207 
208 /** \internal \returns reinterpret_cast<Target>(a) */
209 template <typename Target, typename Packet>
210 EIGEN_DEVICE_FUNC inline Target
211 preinterpret(const Packet& a); /* { return reinterpret_cast<const Target&>(a); } */
212 
213 /** \internal \returns a + b (coeff-wise) */
214 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
215 padd(const Packet& a, const Packet& b) { return a+b; }
216 // Avoid compiler warning for boolean algebra.
217 template<> EIGEN_DEVICE_FUNC inline bool
218 padd(const bool& a, const bool& b) { return a || b; }
219 
220 /** \internal \returns a - b (coeff-wise) */
221 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
222 psub(const Packet& a, const Packet& b) { return a-b; }
223 
224 /** \internal \returns -a (coeff-wise) */
225 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
226 pnegate(const Packet& a) { return -a; }
227 
228 template<> EIGEN_DEVICE_FUNC inline bool
229 pnegate(const bool& a) { return !a; }
230 
231 /** \internal \returns conj(a) (coeff-wise) */
232 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
233 pconj(const Packet& a) { return numext::conj(a); }
234 
235 /** \internal \returns a * b (coeff-wise) */
236 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
237 pmul(const Packet& a, const Packet& b) { return a*b; }
238 // Avoid compiler warning for boolean algebra.
239 template<> EIGEN_DEVICE_FUNC inline bool
240 pmul(const bool& a, const bool& b) { return a && b; }
241 
242 /** \internal \returns a / b (coeff-wise) */
243 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
244 pdiv(const Packet& a, const Packet& b) { return a/b; }
245 
246 // In the generic case, memset to all one bits.
247 template<typename Packet, typename EnableIf = void>
248 struct ptrue_impl {
249   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& /*a*/){
250     Packet b;
251     memset(static_cast<void*>(&b), 0xff, sizeof(Packet));
252     return b;
253   }
254 };
255 
256 // For non-trivial scalars, set to Scalar(1) (i.e. a non-zero value).
257 // Although this is technically not a valid bitmask, the scalar path for pselect
258 // uses a comparison to zero, so this should still work in most cases. We don't
259 // have another option, since the scalar type requires initialization.
260 template<typename T>
261 struct ptrue_impl<T,
262     typename internal::enable_if<is_scalar<T>::value && NumTraits<T>::RequireInitialization>::type > {
263   static EIGEN_DEVICE_FUNC inline T run(const T& /*a*/){
264     return T(1);
265   }
266 };
267 
268 /** \internal \returns one bits. */
269 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
270 ptrue(const Packet& a) {
271   return ptrue_impl<Packet>::run(a);
272 }
273 
274 // In the general case, memset to zero.
275 template<typename Packet, typename EnableIf = void>
276 struct pzero_impl {
277   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& /*a*/) {
278     Packet b;
279     memset(static_cast<void*>(&b), 0x00, sizeof(Packet));
280     return b;
281   }
282 };
283 
284 // For scalars, explicitly set to Scalar(0), since the underlying representation
285 // for zero may not consist of all-zero bits.
286 template<typename T>
287 struct pzero_impl<T,
288     typename internal::enable_if<is_scalar<T>::value>::type> {
289   static EIGEN_DEVICE_FUNC inline T run(const T& /*a*/) {
290     return T(0);
291   }
292 };
293 
294 /** \internal \returns packet of zeros */
295 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
296 pzero(const Packet& a) {
297   return pzero_impl<Packet>::run(a);
298 }
299 
300 /** \internal \returns a <= b as a bit mask */
301 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
302 pcmp_le(const Packet& a, const Packet& b)  { return a<=b ? ptrue(a) : pzero(a); }
303 
304 /** \internal \returns a < b as a bit mask */
305 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
306 pcmp_lt(const Packet& a, const Packet& b)  { return a<b ? ptrue(a) : pzero(a); }
307 
308 /** \internal \returns a == b as a bit mask */
309 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
310 pcmp_eq(const Packet& a, const Packet& b) { return a==b ? ptrue(a) : pzero(a); }
311 
312 /** \internal \returns a < b or a==NaN or b==NaN as a bit mask */
313 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
314 pcmp_lt_or_nan(const Packet& a, const Packet& b) { return a>=b ? pzero(a) : ptrue(a); }
315 
316 template<typename T>
317 struct bit_and {
318   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR EIGEN_ALWAYS_INLINE T operator()(const T& a, const T& b) const {
319     return a & b;
320   }
321 };
322 
323 template<typename T>
324 struct bit_or {
325   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR EIGEN_ALWAYS_INLINE T operator()(const T& a, const T& b) const {
326     return a | b;
327   }
328 };
329 
330 template<typename T>
331 struct bit_xor {
332   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR EIGEN_ALWAYS_INLINE T operator()(const T& a, const T& b) const {
333     return a ^ b;
334   }
335 };
336 
337 template<typename T>
338 struct bit_not {
339   EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR EIGEN_ALWAYS_INLINE T operator()(const T& a) const {
340     return ~a;
341   }
342 };
343 
344 // Use operators &, |, ^, ~.
345 template<typename T>
346 struct operator_bitwise_helper {
347   EIGEN_DEVICE_FUNC static inline T bitwise_and(const T& a, const T& b) { return bit_and<T>()(a, b); }
348   EIGEN_DEVICE_FUNC static inline T bitwise_or(const T& a, const T& b) { return bit_or<T>()(a, b); }
349   EIGEN_DEVICE_FUNC static inline T bitwise_xor(const T& a, const T& b) { return bit_xor<T>()(a, b); }
350   EIGEN_DEVICE_FUNC static inline T bitwise_not(const T& a) { return bit_not<T>()(a); }
351 };
352 
353 // Apply binary operations byte-by-byte
354 template<typename T>
355 struct bytewise_bitwise_helper {
356   EIGEN_DEVICE_FUNC static inline T bitwise_and(const T& a, const T& b) {
357     return binary(a, b, bit_and<unsigned char>());
358   }
359   EIGEN_DEVICE_FUNC static inline T bitwise_or(const T& a, const T& b) {
360     return binary(a, b, bit_or<unsigned char>());
361    }
362   EIGEN_DEVICE_FUNC static inline T bitwise_xor(const T& a, const T& b) {
363     return binary(a, b, bit_xor<unsigned char>());
364   }
365   EIGEN_DEVICE_FUNC static inline T bitwise_not(const T& a) {
366     return unary(a,bit_not<unsigned char>());
367    }
368 
369  private:
370   template<typename Op>
371   EIGEN_DEVICE_FUNC static inline T unary(const T& a, Op op) {
372     const unsigned char* a_ptr = reinterpret_cast<const unsigned char*>(&a);
373     T c;
374     unsigned char* c_ptr = reinterpret_cast<unsigned char*>(&c);
375     for (size_t i = 0; i < sizeof(T); ++i) {
376       *c_ptr++ = op(*a_ptr++);
377     }
378     return c;
379   }
380 
381   template<typename Op>
382   EIGEN_DEVICE_FUNC static inline T binary(const T& a, const T& b, Op op) {
383     const unsigned char* a_ptr = reinterpret_cast<const unsigned char*>(&a);
384     const unsigned char* b_ptr = reinterpret_cast<const unsigned char*>(&b);
385     T c;
386     unsigned char* c_ptr = reinterpret_cast<unsigned char*>(&c);
387     for (size_t i = 0; i < sizeof(T); ++i) {
388       *c_ptr++ = op(*a_ptr++, *b_ptr++);
389     }
390     return c;
391   }
392 };
393 
394 // In the general case, use byte-by-byte manipulation.
395 template<typename T, typename EnableIf = void>
396 struct bitwise_helper : public bytewise_bitwise_helper<T> {};
397 
398 // For integers or non-trivial scalars, use binary operators.
399 template<typename T>
400 struct bitwise_helper<T,
401   typename internal::enable_if<
402     is_scalar<T>::value && (NumTraits<T>::IsInteger || NumTraits<T>::RequireInitialization)>::type
403   > : public operator_bitwise_helper<T> {};
404 
405 /** \internal \returns the bitwise and of \a a and \a b */
406 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
407 pand(const Packet& a, const Packet& b) {
408   return bitwise_helper<Packet>::bitwise_and(a, b);
409 }
410 
411 /** \internal \returns the bitwise or of \a a and \a b */
412 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
413 por(const Packet& a, const Packet& b) {
414   return bitwise_helper<Packet>::bitwise_or(a, b);
415 }
416 
417 /** \internal \returns the bitwise xor of \a a and \a b */
418 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
419 pxor(const Packet& a, const Packet& b) {
420   return bitwise_helper<Packet>::bitwise_xor(a, b);
421 }
422 
423 /** \internal \returns the bitwise not of \a a */
424 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
425 pnot(const Packet& a) {
426   return bitwise_helper<Packet>::bitwise_not(a);
427 }
428 
429 /** \internal \returns the bitwise and of \a a and not \a b */
430 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
431 pandnot(const Packet& a, const Packet& b) { return pand(a, pnot(b)); }
432 
433 // In the general case, use bitwise select.
434 template<typename Packet, typename EnableIf = void>
435 struct pselect_impl {
436   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& mask, const Packet& a, const Packet& b) {
437     return por(pand(a,mask),pandnot(b,mask));
438   }
439 };
440 
441 // For scalars, use ternary select.
442 template<typename Packet>
443 struct pselect_impl<Packet,
444     typename internal::enable_if<is_scalar<Packet>::value>::type > {
445   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& mask, const Packet& a, const Packet& b) {
446     return numext::equal_strict(mask, Packet(0)) ? b : a;
447   }
448 };
449 
450 /** \internal \returns \a or \b for each field in packet according to \mask */
451 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
452 pselect(const Packet& mask, const Packet& a, const Packet& b) {
453   return pselect_impl<Packet>::run(mask, a, b);
454 }
455 
456 template<> EIGEN_DEVICE_FUNC inline bool pselect<bool>(
457     const bool& cond, const bool& a, const bool& b) {
458   return cond ? a : b;
459 }
460 
461 /** \internal \returns the min or of \a a and \a b (coeff-wise)
462     If either \a a or \a b are NaN, the result is implementation defined. */
463 template<int NaNPropagation>
464 struct pminmax_impl {
465   template <typename Packet, typename Op>
466   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& a, const Packet& b, Op op) {
467     return op(a,b);
468   }
469 };
470 
471 /** \internal \returns the min or max of \a a and \a b (coeff-wise)
472     If either \a a or \a b are NaN, NaN is returned. */
473 template<>
474 struct pminmax_impl<PropagateNaN> {
475   template <typename Packet, typename Op>
476   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& a, const Packet& b, Op op) {
477   Packet not_nan_mask_a = pcmp_eq(a, a);
478   Packet not_nan_mask_b = pcmp_eq(b, b);
479   return pselect(not_nan_mask_a,
480                  pselect(not_nan_mask_b, op(a, b), b),
481                  a);
482   }
483 };
484 
485 /** \internal \returns the min or max of \a a and \a b (coeff-wise)
486     If both \a a and \a b are NaN, NaN is returned.
487     Equivalent to std::fmin(a, b).  */
488 template<>
489 struct pminmax_impl<PropagateNumbers> {
490   template <typename Packet, typename Op>
491   static EIGEN_DEVICE_FUNC inline Packet run(const Packet& a, const Packet& b, Op op) {
492   Packet not_nan_mask_a = pcmp_eq(a, a);
493   Packet not_nan_mask_b = pcmp_eq(b, b);
494   return pselect(not_nan_mask_a,
495                  pselect(not_nan_mask_b, op(a, b), a),
496                  b);
497   }
498 };
499 
500 
501 #ifndef SYCL_DEVICE_ONLY
502 #define EIGEN_BINARY_OP_NAN_PROPAGATION(Type, Func) Func
503 #else
504 #define EIGEN_BINARY_OP_NAN_PROPAGATION(Type, Func) \
505 [](const Type& a, const Type& b) { \
506         return Func(a, b);}
507 #endif
508 
509 /** \internal \returns the min of \a a and \a b  (coeff-wise).
510     If \a a or \b b is NaN, the return value is implementation defined. */
511 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
512 pmin(const Packet& a, const Packet& b) { return numext::mini(a,b); }
513 
514 /** \internal \returns the min of \a a and \a b  (coeff-wise).
515     NaNPropagation determines the NaN propagation semantics. */
516 template <int NaNPropagation, typename Packet>
517 EIGEN_DEVICE_FUNC inline Packet pmin(const Packet& a, const Packet& b) {
518   return pminmax_impl<NaNPropagation>::run(a, b, EIGEN_BINARY_OP_NAN_PROPAGATION(Packet, (pmin<Packet>)));
519 }
520 
521 /** \internal \returns the max of \a a and \a b  (coeff-wise)
522     If \a a or \b b is NaN, the return value is implementation defined. */
523 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
524 pmax(const Packet& a, const Packet& b) { return numext::maxi(a, b); }
525 
526 /** \internal \returns the max of \a a and \a b  (coeff-wise).
527     NaNPropagation determines the NaN propagation semantics. */
528 template <int NaNPropagation, typename Packet>
529 EIGEN_DEVICE_FUNC inline Packet pmax(const Packet& a, const Packet& b) {
530   return pminmax_impl<NaNPropagation>::run(a, b, EIGEN_BINARY_OP_NAN_PROPAGATION(Packet,(pmax<Packet>)));
531 }
532 
533 /** \internal \returns the absolute value of \a a */
534 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
535 pabs(const Packet& a) { return numext::abs(a); }
536 template<> EIGEN_DEVICE_FUNC inline unsigned int
537 pabs(const unsigned int& a) { return a; }
538 template<> EIGEN_DEVICE_FUNC inline unsigned long
539 pabs(const unsigned long& a) { return a; }
540 template<> EIGEN_DEVICE_FUNC inline unsigned long long
541 pabs(const unsigned long long& a) { return a; }
542 
543 /** \internal \returns the addsub value of \a a,b */
544 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
545 paddsub(const Packet& a, const Packet& b) {
546   return pselect(peven_mask(a), padd(a, b), psub(a, b));
547  }
548 
549 /** \internal \returns the phase angle of \a a */
550 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
551 parg(const Packet& a) { using numext::arg; return arg(a); }
552 
553 
554 /** \internal \returns \a a logically shifted by N bits to the right */
555 template<int N> EIGEN_DEVICE_FUNC inline int
556 parithmetic_shift_right(const int& a) { return a >> N; }
557 template<int N> EIGEN_DEVICE_FUNC inline long int
558 parithmetic_shift_right(const long int& a) { return a >> N; }
559 
560 /** \internal \returns \a a arithmetically shifted by N bits to the right */
561 template<int N> EIGEN_DEVICE_FUNC inline int
562 plogical_shift_right(const int& a) { return static_cast<int>(static_cast<unsigned int>(a) >> N); }
563 template<int N> EIGEN_DEVICE_FUNC inline long int
564 plogical_shift_right(const long int& a) { return static_cast<long>(static_cast<unsigned long>(a) >> N); }
565 
566 /** \internal \returns \a a shifted by N bits to the left */
567 template<int N> EIGEN_DEVICE_FUNC inline int
568 plogical_shift_left(const int& a) { return a << N; }
569 template<int N> EIGEN_DEVICE_FUNC inline long int
570 plogical_shift_left(const long int& a) { return a << N; }
571 
572 /** \internal \returns the significant and exponent of the underlying floating point numbers
573   * See https://en.cppreference.com/w/cpp/numeric/math/frexp
574   */
575 template <typename Packet>
576 EIGEN_DEVICE_FUNC inline Packet pfrexp(const Packet& a, Packet& exponent) {
577   int exp;
578   EIGEN_USING_STD(frexp);
579   Packet result = static_cast<Packet>(frexp(a, &exp));
580   exponent = static_cast<Packet>(exp);
581   return result;
582 }
583 
584 /** \internal \returns a * 2^((int)exponent)
585   * See https://en.cppreference.com/w/cpp/numeric/math/ldexp
586   */
587 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
588 pldexp(const Packet &a, const Packet &exponent) {
589   EIGEN_USING_STD(ldexp)
590   return static_cast<Packet>(ldexp(a, static_cast<int>(exponent)));
591 }
592 
593 /** \internal \returns the min of \a a and \a b  (coeff-wise) */
594 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
595 pabsdiff(const Packet& a, const Packet& b) { return pselect(pcmp_lt(a, b), psub(b, a), psub(a, b)); }
596 
597 /** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */
598 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
599 pload(const typename unpacket_traits<Packet>::type* from) { return *from; }
600 
601 /** \internal \returns a packet version of \a *from, (un-aligned load) */
602 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
603 ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; }
604 
605 /** \internal \returns a packet version of \a *from, (un-aligned masked load)
606  * There is no generic implementation. We only have implementations for specialized
607  * cases. Generic case should not be called.
608  */
609 template<typename Packet> EIGEN_DEVICE_FUNC inline
610 typename enable_if<unpacket_traits<Packet>::masked_load_available, Packet>::type
611 ploadu(const typename unpacket_traits<Packet>::type* from, typename unpacket_traits<Packet>::mask_t umask);
612 
613 /** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */
614 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
615 pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
616 
617 /** \internal \returns a packet with constant coefficients set from bits */
618 template<typename Packet,typename BitsType> EIGEN_DEVICE_FUNC inline Packet
619 pset1frombits(BitsType a);
620 
621 /** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */
622 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
623 pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(*a); }
624 
625 /** \internal \returns a packet with elements of \a *from duplicated.
626   * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and
627   * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]}
628   * Currently, this function is only used for scalar * complex products.
629   */
630 template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
631 ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
632 
633 /** \internal \returns a packet with elements of \a *from quadrupled.
634   * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and
635   * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]}
636   * Currently, this function is only used in matrix products.
637   * For packet-size smaller or equal to 4, this function is equivalent to pload1
638   */
639 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
640 ploadquad(const typename unpacket_traits<Packet>::type* from)
641 { return pload1<Packet>(from); }
642 
643 /** \internal equivalent to
644   * \code
645   * a0 = pload1(a+0);
646   * a1 = pload1(a+1);
647   * a2 = pload1(a+2);
648   * a3 = pload1(a+3);
649   * \endcode
650   * \sa pset1, pload1, ploaddup, pbroadcast2
651   */
652 template<typename Packet> EIGEN_DEVICE_FUNC
653 inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a,
654                         Packet& a0, Packet& a1, Packet& a2, Packet& a3)
655 {
656   a0 = pload1<Packet>(a+0);
657   a1 = pload1<Packet>(a+1);
658   a2 = pload1<Packet>(a+2);
659   a3 = pload1<Packet>(a+3);
660 }
661 
662 /** \internal equivalent to
663   * \code
664   * a0 = pload1(a+0);
665   * a1 = pload1(a+1);
666   * \endcode
667   * \sa pset1, pload1, ploaddup, pbroadcast4
668   */
669 template<typename Packet> EIGEN_DEVICE_FUNC
670 inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
671                         Packet& a0, Packet& a1)
672 {
673   a0 = pload1<Packet>(a+0);
674   a1 = pload1<Packet>(a+1);
675 }
676 
677 /** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
678 template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
679 plset(const typename unpacket_traits<Packet>::type& a) { return a; }
680 
681 /** \internal \returns a packet with constant coefficients \a a, e.g.: (x, 0, x, 0),
682      where x is the value of all 1-bits. */
683 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
684 peven_mask(const Packet& /*a*/) {
685   typedef typename unpacket_traits<Packet>::type Scalar;
686   const size_t n = unpacket_traits<Packet>::size;
687   EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) Scalar elements[n];
688   for(size_t i = 0; i < n; ++i) {
689     memset(elements+i, ((i & 1) == 0 ? 0xff : 0), sizeof(Scalar));
690   }
691   return ploadu<Packet>(elements);
692 }
693 
694 
695 /** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
696 template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
697 { (*to) = from; }
698 
699 /** \internal copy the packet \a from to \a *to, (un-aligned store) */
700 template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from)
701 {  (*to) = from; }
702 
703 /** \internal copy the packet \a from to \a *to, (un-aligned store with a mask)
704  * There is no generic implementation. We only have implementations for specialized
705  * cases. Generic case should not be called.
706  */
707 template<typename Scalar, typename Packet>
708 EIGEN_DEVICE_FUNC inline
709 typename enable_if<unpacket_traits<Packet>::masked_store_available, void>::type
710 pstoreu(Scalar* to, const Packet& from, typename unpacket_traits<Packet>::mask_t umask);
711 
712  template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/)
713  { return ploadu<Packet>(from); }
714 
715  template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/)
716  { pstore(to, from); }
717 
718 /** \internal tries to do cache prefetching of \a addr */
719 template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
720 {
721 #if defined(EIGEN_HIP_DEVICE_COMPILE)
722   // do nothing
723 #elif defined(EIGEN_CUDA_ARCH)
724 #if defined(__LP64__) || EIGEN_OS_WIN64
725   // 64-bit pointer operand constraint for inlined asm
726   asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
727 #else
728   // 32-bit pointer operand constraint for inlined asm
729   asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr));
730 #endif
731 #elif (!EIGEN_COMP_MSVC) && (EIGEN_COMP_GNUC || EIGEN_COMP_CLANG || EIGEN_COMP_ICC)
732   __builtin_prefetch(addr);
733 #endif
734 }
735 
736 /** \internal \returns the reversed elements of \a a*/
737 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
738 { return a; }
739 
740 /** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
741 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)
742 {
743   return Packet(numext::imag(a),numext::real(a));
744 }
745 
746 /**************************
747 * Special math functions
748 ***************************/
749 
750 /** \internal \returns the sine of \a a (coeff-wise) */
751 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
752 Packet psin(const Packet& a) { EIGEN_USING_STD(sin); return sin(a); }
753 
754 /** \internal \returns the cosine of \a a (coeff-wise) */
755 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
756 Packet pcos(const Packet& a) { EIGEN_USING_STD(cos); return cos(a); }
757 
758 /** \internal \returns the tan of \a a (coeff-wise) */
759 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
760 Packet ptan(const Packet& a) { EIGEN_USING_STD(tan); return tan(a); }
761 
762 /** \internal \returns the arc sine of \a a (coeff-wise) */
763 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
764 Packet pasin(const Packet& a) { EIGEN_USING_STD(asin); return asin(a); }
765 
766 /** \internal \returns the arc cosine of \a a (coeff-wise) */
767 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
768 Packet pacos(const Packet& a) { EIGEN_USING_STD(acos); return acos(a); }
769 
770 /** \internal \returns the arc tangent of \a a (coeff-wise) */
771 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
772 Packet patan(const Packet& a) { EIGEN_USING_STD(atan); return atan(a); }
773 
774 /** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
775 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
776 Packet psinh(const Packet& a) { EIGEN_USING_STD(sinh); return sinh(a); }
777 
778 /** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
779 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
780 Packet pcosh(const Packet& a) { EIGEN_USING_STD(cosh); return cosh(a); }
781 
782 /** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
783 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
784 Packet ptanh(const Packet& a) { EIGEN_USING_STD(tanh); return tanh(a); }
785 
786 /** \internal \returns the exp of \a a (coeff-wise) */
787 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
788 Packet pexp(const Packet& a) { EIGEN_USING_STD(exp); return exp(a); }
789 
790 /** \internal \returns the expm1 of \a a (coeff-wise) */
791 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
792 Packet pexpm1(const Packet& a) { return numext::expm1(a); }
793 
794 /** \internal \returns the log of \a a (coeff-wise) */
795 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
796 Packet plog(const Packet& a) { EIGEN_USING_STD(log); return log(a); }
797 
798 /** \internal \returns the log1p of \a a (coeff-wise) */
799 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
800 Packet plog1p(const Packet& a) { return numext::log1p(a); }
801 
802 /** \internal \returns the log10 of \a a (coeff-wise) */
803 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
804 Packet plog10(const Packet& a) { EIGEN_USING_STD(log10); return log10(a); }
805 
806 /** \internal \returns the log10 of \a a (coeff-wise) */
807 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
808 Packet plog2(const Packet& a) {
809   typedef typename internal::unpacket_traits<Packet>::type Scalar;
810   return pmul(pset1<Packet>(Scalar(EIGEN_LOG2E)), plog(a));
811 }
812 
813 /** \internal \returns the square-root of \a a (coeff-wise) */
814 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
815 Packet psqrt(const Packet& a) { return numext::sqrt(a); }
816 
817 /** \internal \returns the reciprocal square-root of \a a (coeff-wise) */
818 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
819 Packet prsqrt(const Packet& a) {
820   typedef typename internal::unpacket_traits<Packet>::type Scalar;
821   return pdiv(pset1<Packet>(Scalar(1)), psqrt(a));
822 }
823 
824 /** \internal \returns the rounded value of \a a (coeff-wise) */
825 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
826 Packet pround(const Packet& a) { using numext::round; return round(a); }
827 
828 /** \internal \returns the floor of \a a (coeff-wise) */
829 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
830 Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
831 
832 /** \internal \returns the rounded value of \a a (coeff-wise) with current
833  * rounding mode */
834 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
835 Packet print(const Packet& a) { using numext::rint; return rint(a); }
836 
837 /** \internal \returns the ceil of \a a (coeff-wise) */
838 template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
839 Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
840 
841 /** \internal \returns the first element of a packet */
842 template<typename Packet>
843 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type
844 pfirst(const Packet& a)
845 { return a; }
846 
847 /** \internal \returns the sum of the elements of upper and lower half of \a a if \a a is larger than 4.
848   * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7}
849   * For packet-size smaller or equal to 4, this boils down to a noop.
850   */
851 template<typename Packet>
852 EIGEN_DEVICE_FUNC inline typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
853 predux_half_dowto4(const Packet& a)
854 { return a; }
855 
856 // Slow generic implementation of Packet reduction.
857 template <typename Packet, typename Op>
858 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type
859 predux_helper(const Packet& a, Op op) {
860   typedef typename unpacket_traits<Packet>::type Scalar;
861   const size_t n = unpacket_traits<Packet>::size;
862   EIGEN_ALIGN_TO_BOUNDARY(sizeof(Packet)) Scalar elements[n];
863   pstoreu<Scalar>(elements, a);
864   for(size_t k = n / 2; k > 0; k /= 2)  {
865     for(size_t i = 0; i < k; ++i) {
866       elements[i] = op(elements[i], elements[i + k]);
867     }
868   }
869   return elements[0];
870 }
871 
872 /** \internal \returns the sum of the elements of \a a*/
873 template<typename Packet>
874 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type
875 predux(const Packet& a)
876 {
877   return a;
878 }
879 
880 /** \internal \returns the product of the elements of \a a */
881 template <typename Packet>
882 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_mul(
883     const Packet& a) {
884   typedef typename unpacket_traits<Packet>::type Scalar;
885   return predux_helper(a, EIGEN_BINARY_OP_NAN_PROPAGATION(Scalar, (pmul<Scalar>)));
886 }
887 
888 /** \internal \returns the min of the elements of \a a */
889 template <typename Packet>
890 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(
891     const Packet &a) {
892   typedef typename unpacket_traits<Packet>::type Scalar;
893   return predux_helper(a, EIGEN_BINARY_OP_NAN_PROPAGATION(Scalar, (pmin<PropagateFast, Scalar>)));
894 }
895 
896 template <int NaNPropagation, typename Packet>
897 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(
898     const Packet& a) {
899   typedef typename unpacket_traits<Packet>::type Scalar;
900   return predux_helper(a, EIGEN_BINARY_OP_NAN_PROPAGATION(Scalar, (pmin<NaNPropagation, Scalar>)));
901 }
902 
903 /** \internal \returns the min of the elements of \a a */
904 template <typename Packet>
905 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(
906     const Packet &a) {
907   typedef typename unpacket_traits<Packet>::type Scalar;
908   return predux_helper(a, EIGEN_BINARY_OP_NAN_PROPAGATION(Scalar, (pmax<PropagateFast, Scalar>)));
909 }
910 
911 template <int NaNPropagation, typename Packet>
912 EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(
913     const Packet& a) {
914   typedef typename unpacket_traits<Packet>::type Scalar;
915   return predux_helper(a, EIGEN_BINARY_OP_NAN_PROPAGATION(Scalar, (pmax<NaNPropagation, Scalar>)));
916 }
917 
918 #undef EIGEN_BINARY_OP_NAN_PROPAGATION
919 
920 /** \internal \returns true if all coeffs of \a a means "true"
921   * It is supposed to be called on values returned by pcmp_*.
922   */
923 // not needed yet
924 // template<typename Packet> EIGEN_DEVICE_FUNC inline bool predux_all(const Packet& a)
925 // { return bool(a); }
926 
927 /** \internal \returns true if any coeffs of \a a means "true"
928   * It is supposed to be called on values returned by pcmp_*.
929   */
930 template<typename Packet> EIGEN_DEVICE_FUNC inline bool predux_any(const Packet& a)
931 {
932   // Dirty but generic implementation where "true" is assumed to be non 0 and all the sames.
933   // It is expected that "true" is either:
934   //  - Scalar(1)
935   //  - bits full of ones (NaN for floats),
936   //  - or first bit equals to 1 (1 for ints, smallest denormal for floats).
937   // For all these cases, taking the sum is just fine, and this boils down to a no-op for scalars.
938   typedef typename unpacket_traits<Packet>::type Scalar;
939   return numext::not_equal_strict(predux(a), Scalar(0));
940 }
941 
942 /***************************************************************************
943 * The following functions might not have to be overwritten for vectorized types
944 ***************************************************************************/
945 
946 /** \internal copy a packet with constant coefficient \a a (e.g., [a,a,a,a]) to \a *to. \a to must be 16 bytes aligned */
947 // NOTE: this function must really be templated on the packet type (think about different packet types for the same scalar type)
948 template<typename Packet>
949 inline void pstore1(typename unpacket_traits<Packet>::type* to, const typename unpacket_traits<Packet>::type& a)
950 {
951   pstore(to, pset1<Packet>(a));
952 }
953 
954 /** \internal \returns a * b + c (coeff-wise) */
955 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
956 pmadd(const Packet&  a,
957          const Packet&  b,
958          const Packet&  c)
959 { return padd(pmul(a, b),c); }
960 
961 /** \internal \returns a packet version of \a *from.
962   * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
963 template<typename Packet, int Alignment>
964 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from)
965 {
966   if(Alignment >= unpacket_traits<Packet>::alignment)
967     return pload<Packet>(from);
968   else
969     return ploadu<Packet>(from);
970 }
971 
972 /** \internal copy the packet \a from to \a *to.
973   * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
974 template<typename Scalar, typename Packet, int Alignment>
975 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from)
976 {
977   if(Alignment >= unpacket_traits<Packet>::alignment)
978     pstore(to, from);
979   else
980     pstoreu(to, from);
981 }
982 
983 /** \internal \returns a packet version of \a *from.
984   * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the
985   * hardware if available to speedup the loading of data that won't be modified
986   * by the current computation.
987   */
988 template<typename Packet, int LoadMode>
989 EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
990 {
991   return ploadt<Packet, LoadMode>(from);
992 }
993 
994 /***************************************************************************
995 * Fast complex products (GCC generates a function call which is very slow)
996 ***************************************************************************/
997 
998 // Eigen+CUDA does not support complexes.
999 #if !defined(EIGEN_GPUCC)
1000 
1001 template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
1002 { return std::complex<float>(a.real()*b.real() - a.imag()*b.imag(), a.imag()*b.real() + a.real()*b.imag()); }
1003 
1004 template<> inline std::complex<double> pmul(const std::complex<double>& a, const std::complex<double>& b)
1005 { return std::complex<double>(a.real()*b.real() - a.imag()*b.imag(), a.imag()*b.real() + a.real()*b.imag()); }
1006 
1007 #endif
1008 
1009 
1010 /***************************************************************************
1011  * PacketBlock, that is a collection of N packets where the number of words
1012  * in the packet is a multiple of N.
1013 ***************************************************************************/
1014 template <typename Packet,int N=unpacket_traits<Packet>::size> struct PacketBlock {
1015   Packet packet[N];
1016 };
1017 
1018 template<typename Packet> EIGEN_DEVICE_FUNC inline void
1019 ptranspose(PacketBlock<Packet,1>& /*kernel*/) {
1020   // Nothing to do in the scalar case, i.e. a 1x1 matrix.
1021 }
1022 
1023 /***************************************************************************
1024  * Selector, i.e. vector of N boolean values used to select (i.e. blend)
1025  * words from 2 packets.
1026 ***************************************************************************/
1027 template <size_t N> struct Selector {
1028   bool select[N];
1029 };
1030 
1031 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
1032 pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) {
1033   return ifPacket.select[0] ? thenPacket : elsePacket;
1034 }
1035 
1036 } // end namespace internal
1037 
1038 } // end namespace Eigen
1039 
1040 #endif // EIGEN_GENERIC_PACKET_MATH_H
1041