Eigen  3.2.92
SSE/PacketMath.h
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_PACKET_MATH_SSE_H
11 #define EIGEN_PACKET_MATH_SSE_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
18 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
19 #endif
20 
21 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
22 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
23 #endif
24 
25 #ifdef __FMA__
26 #ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
27 #define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1
28 #endif
29 #endif
30 
31 #if (defined EIGEN_VECTORIZE_AVX) && EIGEN_COMP_GNUC_STRICT && (__GXX_ABI_VERSION < 1004)
32 // With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot
33 // have overloads for both types without linking error.
34 // One solution is to increase ABI version using -fabi-version=4 (or greater).
35 // Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper
36 // structure:
37 template<typename T>
38 struct eigen_packet_wrapper
39 {
40  EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
41  EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
42  EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
43  EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
44  EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
45  m_val = v;
46  return *this;
47  }
48 
49  T m_val;
50 };
51 typedef eigen_packet_wrapper<__m128> Packet4f;
52 typedef eigen_packet_wrapper<__m128i> Packet4i;
53 typedef eigen_packet_wrapper<__m128d> Packet2d;
54 #else
55 typedef __m128 Packet4f;
56 typedef __m128i Packet4i;
57 typedef __m128d Packet2d;
58 #endif
59 
60 template<> struct is_arithmetic<__m128> { enum { value = true }; };
61 template<> struct is_arithmetic<__m128i> { enum { value = true }; };
62 template<> struct is_arithmetic<__m128d> { enum { value = true }; };
63 
64 #define vec4f_swizzle1(v,p,q,r,s) \
65  (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p)))))
66 
67 #define vec4i_swizzle1(v,p,q,r,s) \
68  (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p))))
69 
70 #define vec2d_swizzle1(v,p,q) \
71  (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2)))))
72 
73 #define vec4f_swizzle2(a,b,p,q,r,s) \
74  (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p))))
75 
76 #define vec4i_swizzle2(a,b,p,q,r,s) \
77  (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p))))))
78 
79 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
80  const Packet4f p4f_##NAME = pset1<Packet4f>(X)
81 
82 #define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
83  const Packet2d p2d_##NAME = pset1<Packet2d>(X)
84 
85 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
86  const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X))
87 
88 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
89  const Packet4i p4i_##NAME = pset1<Packet4i>(X)
90 
91 
92 // Use the packet_traits defined in AVX/PacketMath.h instead if we're going
93 // to leverage AVX instructions.
94 #ifndef EIGEN_VECTORIZE_AVX
95 template<> struct packet_traits<float> : default_packet_traits
96 {
97  typedef Packet4f type;
98  typedef Packet4f half;
99  enum {
100  Vectorizable = 1,
101  AlignedOnScalar = 1,
102  size=4,
103  HasHalfPacket = 0,
104 
105  HasDiv = 1,
106  HasSin = EIGEN_FAST_MATH,
107  HasCos = EIGEN_FAST_MATH,
108  HasLog = 1,
109  HasExp = 1,
110  HasSqrt = 1,
111  HasRsqrt = 1,
112  HasBlend = 1
113 
114 #ifdef EIGEN_VECTORIZE_SSE4_1
115  ,
116  HasRound = 1,
117  HasFloor = 1,
118  HasCeil = 1
119 #endif
120  };
121 };
122 template<> struct packet_traits<double> : default_packet_traits
123 {
124  typedef Packet2d type;
125  typedef Packet2d half;
126  enum {
127  Vectorizable = 1,
128  AlignedOnScalar = 1,
129  size=2,
130  HasHalfPacket = 0,
131 
132  HasDiv = 1,
133  HasExp = 1,
134  HasSqrt = 1,
135  HasRsqrt = 1,
136  HasBlend = 1
137 
138 #ifdef EIGEN_VECTORIZE_SSE4_1
139  ,
140  HasRound = 1,
141  HasFloor = 1,
142  HasCeil = 1
143 #endif
144  };
145 };
146 #endif
147 template<> struct packet_traits<int> : default_packet_traits
148 {
149  typedef Packet4i type;
150  typedef Packet4i half;
151  enum {
152  Vectorizable = 1,
153  AlignedOnScalar = 1,
154  size=4,
155 
156  HasBlend = 1
157  };
158 };
159 
160 template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
161 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
162 template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
163 
164 #if EIGEN_COMP_MSVC==1500
165 // Workaround MSVC 9 internal compiler error.
166 // TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
167 // TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
168 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); }
169 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
170 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); }
171 #else
172 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps1(from); }
173 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
174 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); }
175 #endif
176 
177 // GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
178 // However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
179 // Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
180 // Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply.
181 // Also note that with AVX, we want it to generate a vbroadcastss.
182 #if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
183 template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) {
184  return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
185 }
186 #endif
187 
188 template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
189 template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
190 template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
191 
192 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
193 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); }
194 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); }
195 
196 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); }
197 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); }
198 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); }
199 
200 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
201 {
202  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
203  return _mm_xor_ps(a,mask);
204 }
205 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a)
206 {
207  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
208  return _mm_xor_pd(a,mask);
209 }
210 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a)
211 {
212  return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
213 }
214 
215 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
216 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
217 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
218 
219 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); }
220 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); }
221 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
222 {
223 #ifdef EIGEN_VECTORIZE_SSE4_1
224  return _mm_mullo_epi32(a,b);
225 #else
226  // this version is slightly faster than 4 scalar products
227  return vec4i_swizzle1(
228  vec4i_swizzle2(
229  _mm_mul_epu32(a,b),
230  _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
231  vec4i_swizzle1(b,1,0,3,2)),
232  0,2,0,2),
233  0,2,1,3);
234 #endif
235 }
236 
237 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
238 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
239 
240 // for some weird raisons, it has to be overloaded for packet of integers
241 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
242 #ifdef __FMA__
243 template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
244 template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
245 #endif
246 
247 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); }
248 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); }
249 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b)
250 {
251 #ifdef EIGEN_VECTORIZE_SSE4_1
252  return _mm_min_epi32(a,b);
253 #else
254  // after some bench, this version *is* faster than a scalar implementation
255  Packet4i mask = _mm_cmplt_epi32(a,b);
256  return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
257 #endif
258 }
259 
260 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); }
261 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); }
262 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b)
263 {
264 #ifdef EIGEN_VECTORIZE_SSE4_1
265  return _mm_max_epi32(a,b);
266 #else
267  // after some bench, this version *is* faster than a scalar implementation
268  Packet4i mask = _mm_cmpgt_epi32(a,b);
269  return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
270 #endif
271 }
272 
273 #ifdef EIGEN_VECTORIZE_SSE4_1
274 template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
275 template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
276 
277 template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
278 template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
279 
280 template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
281 template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
282 #endif
283 
284 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
285 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
286 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
287 
288 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); }
289 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); }
290 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); }
291 
292 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); }
293 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); }
294 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); }
295 
296 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); }
297 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); }
298 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); }
299 
300 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); }
301 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); }
302 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
303 
304 #if EIGEN_COMP_MSVC
305  template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) {
306  EIGEN_DEBUG_UNALIGNED_LOAD
307  #if (EIGEN_COMP_MSVC==1600)
308  // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
309  // (i.e., it does not generate an unaligned load!!
310  __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
311  res = _mm_loadh_pi(res, (const __m64*)(from+2));
312  return res;
313  #else
314  return _mm_loadu_ps(from);
315  #endif
316  }
317  template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); }
318  template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from)); }
319 #else
320 // NOTE: with the code below, MSVC's compiler crashes!
321 
322 #if EIGEN_COMP_GNUC && (EIGEN_ARCH_i386 || (EIGEN_ARCH_x86_64 && EIGEN_GNUC_AT_LEAST(4, 8)))
323  // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd
324  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
325 #elif EIGEN_COMP_CLANG
326  // bug 201: Segfaults in __mm_loadh_pd with clang 2.8
327  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
328 #else
329  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0
330 #endif
331 
332 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
333 {
334  EIGEN_DEBUG_UNALIGNED_LOAD
335 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
336  return _mm_loadu_ps(from);
337 #else
338  __m128d res;
339  res = _mm_load_sd((const double*)(from)) ;
340  res = _mm_loadh_pd(res, (const double*)(from+2)) ;
341  return _mm_castpd_ps(res);
342 #endif
343 }
344 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
345 {
346  EIGEN_DEBUG_UNALIGNED_LOAD
347 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
348  return _mm_loadu_pd(from);
349 #else
350  __m128d res;
351  res = _mm_load_sd(from) ;
352  res = _mm_loadh_pd(res,from+1);
353  return res;
354 #endif
355 }
356 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
357 {
358  EIGEN_DEBUG_UNALIGNED_LOAD
359 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
360  return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
361 #else
362  __m128d res;
363  res = _mm_load_sd((const double*)(from)) ;
364  res = _mm_loadh_pd(res, (const double*)(from+2)) ;
365  return _mm_castpd_si128(res);
366 #endif
367 }
368 #endif
369 
370 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
371 {
372  return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
373 }
374 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from)
375 { return pset1<Packet2d>(from[0]); }
376 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
377 {
378  Packet4i tmp;
379  tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
380  return vec4i_swizzle1(tmp, 0, 0, 1, 1);
381 }
382 
383 template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
384 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
385 template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
386 
387 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
388 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
389 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
390 
391 template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
392 {
393  return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
394 }
395 template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
396 {
397  return _mm_set_pd(from[1*stride], from[0*stride]);
398 }
399 template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
400 {
401  return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
402  }
403 
404 template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
405 {
406  to[stride*0] = _mm_cvtss_f32(from);
407  to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
408  to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
409  to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
410 }
411 template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
412 {
413  to[stride*0] = _mm_cvtsd_f64(from);
414  to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
415 }
416 template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
417 {
418  to[stride*0] = _mm_cvtsi128_si32(from);
419  to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
420  to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
421  to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
422 }
423 
424 // some compilers might be tempted to perform multiple moves instead of using a vector path.
425 template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
426 {
427  Packet4f pa = _mm_set_ss(a);
428  pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
429 }
430 // some compilers might be tempted to perform multiple moves instead of using a vector path.
431 template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a)
432 {
433  Packet2d pa = _mm_set_sd(a);
434  pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
435 }
436 
437 #ifndef EIGEN_VECTORIZE_AVX
438 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
439 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
440 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
441 #endif
442 
443 #if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
444 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
445 // Direct of the struct members fixed bug #62.
446 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; }
447 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; }
448 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
449 #elif EIGEN_COMP_MSVC_STRICT
450 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
451 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; }
452 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; }
453 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
454 #else
455 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); }
456 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); }
457 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); }
458 #endif
459 
460 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
461 { return _mm_shuffle_ps(a,a,0x1B); }
462 template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
463 { return _mm_shuffle_pd(a,a,0x1); }
464 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
465 { return _mm_shuffle_epi32(a,0x1B); }
466 
467 template<size_t offset>
468 struct protate_impl<offset, Packet4f>
469 {
470  static Packet4f run(const Packet4f& a) {
471  return vec4f_swizzle1(a, offset, (offset + 1) % 4, (offset + 2) % 4, (offset + 3) % 4);
472  }
473 };
474 
475 template<size_t offset>
476 struct protate_impl<offset, Packet4i>
477 {
478  static Packet4i run(const Packet4i& a) {
479  return vec4i_swizzle1(a, offset, (offset + 1) % 4, (offset + 2) % 4, (offset + 3) % 4);
480  }
481 };
482 
483 template<size_t offset>
484 struct protate_impl<offset, Packet2d>
485 {
486  static Packet2d run(const Packet2d& a) {
487  return vec2d_swizzle1(a, offset, (offset + 1) % 2);
488  }
489 };
490 
491 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a)
492 {
493  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
494  return _mm_and_ps(a,mask);
495 }
496 template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a)
497 {
498  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
499  return _mm_and_pd(a,mask);
500 }
501 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
502 {
503  #ifdef EIGEN_VECTORIZE_SSSE3
504  return _mm_abs_epi32(a);
505  #else
506  Packet4i aux = _mm_srai_epi32(a,31);
507  return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
508  #endif
509 }
510 
511 // with AVX, the default implementations based on pload1 are faster
512 #ifndef __AVX__
513 template<> EIGEN_STRONG_INLINE void
514 pbroadcast4<Packet4f>(const float *a,
515  Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
516 {
517  a3 = pload<Packet4f>(a);
518  a0 = vec4f_swizzle1(a3, 0,0,0,0);
519  a1 = vec4f_swizzle1(a3, 1,1,1,1);
520  a2 = vec4f_swizzle1(a3, 2,2,2,2);
521  a3 = vec4f_swizzle1(a3, 3,3,3,3);
522 }
523 template<> EIGEN_STRONG_INLINE void
524 pbroadcast4<Packet2d>(const double *a,
525  Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
526 {
527 #ifdef EIGEN_VECTORIZE_SSE3
528  a0 = _mm_loaddup_pd(a+0);
529  a1 = _mm_loaddup_pd(a+1);
530  a2 = _mm_loaddup_pd(a+2);
531  a3 = _mm_loaddup_pd(a+3);
532 #else
533  a1 = pload<Packet2d>(a);
534  a0 = vec2d_swizzle1(a1, 0,0);
535  a1 = vec2d_swizzle1(a1, 1,1);
536  a3 = pload<Packet2d>(a+2);
537  a2 = vec2d_swizzle1(a3, 0,0);
538  a3 = vec2d_swizzle1(a3, 1,1);
539 #endif
540 }
541 #endif
542 
543 EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
544 {
545  vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
546  vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
547  vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
548  vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
549 }
550 
551 #ifdef EIGEN_VECTORIZE_SSE3
552 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
553 {
554  return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
555 }
556 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
557 {
558  return _mm_hadd_pd(vecs[0], vecs[1]);
559 }
560 
561 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
562 {
563  Packet4f tmp0 = _mm_hadd_ps(a,a);
564  return pfirst<Packet4f>(_mm_hadd_ps(tmp0, tmp0));
565 }
566 
567 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst<Packet2d>(_mm_hadd_pd(a, a)); }
568 #else
569 // SSE2 versions
570 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
571 {
572  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
573  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
574 }
575 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
576 {
577  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
578 }
579 
580 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
581 {
582  Packet4f tmp0, tmp1, tmp2;
583  tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
584  tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
585  tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
586  tmp0 = _mm_add_ps(tmp0, tmp1);
587  tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
588  tmp1 = _mm_add_ps(tmp1, tmp2);
589  tmp2 = _mm_movehl_ps(tmp1, tmp0);
590  tmp0 = _mm_movelh_ps(tmp0, tmp1);
591  return _mm_add_ps(tmp0, tmp2);
592 }
593 
594 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
595 {
596  return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
597 }
598 #endif // SSE3
599 
600 
601 #ifdef EIGEN_VECTORIZE_SSSE3
602 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
603 {
604  return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
605 }
606 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
607 {
608  Packet4i tmp0 = _mm_hadd_epi32(a,a);
609  return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
610 }
611 #else
612 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
613 {
614  Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
615  return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
616 }
617 
618 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
619 {
620  Packet4i tmp0, tmp1, tmp2;
621  tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
622  tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
623  tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
624  tmp0 = _mm_add_epi32(tmp0, tmp1);
625  tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
626  tmp1 = _mm_add_epi32(tmp1, tmp2);
627  tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
628  tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
629  return _mm_add_epi32(tmp0, tmp2);
630 }
631 #endif
632 // Other reduction functions:
633 
634 // mul
635 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
636 {
637  Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
638  return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
639 }
640 template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
641 {
642  return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
643 }
644 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
645 {
646  // after some experiments, it is seems this is the fastest way to implement it
647  // for GCC (eg., reusing pmul is very slow !)
648  // TODO try to call _mm_mul_epu32 directly
649  EIGEN_ALIGN16 int aux[4];
650  pstore(aux, a);
651  return (aux[0] * aux[1]) * (aux[2] * aux[3]);;
652 }
653 
654 // min
655 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
656 {
657  Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
658  return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
659 }
660 template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
661 {
662  return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
663 }
664 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
665 {
666 #ifdef EIGEN_VECTORIZE_SSE4_1
667  Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
668  return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
669 #else
670  // after some experiments, it is seems this is the fastest way to implement it
671  // for GCC (eg., it does not like using std::min after the pstore !!)
672  EIGEN_ALIGN16 int aux[4];
673  pstore(aux, a);
674  int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
675  int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
676  return aux0<aux2 ? aux0 : aux2;
677 #endif // EIGEN_VECTORIZE_SSE4_1
678 }
679 
680 // max
681 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
682 {
683  Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
684  return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
685 }
686 template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
687 {
688  return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
689 }
690 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
691 {
692 #ifdef EIGEN_VECTORIZE_SSE4_1
693  Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
694  return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
695 #else
696  // after some experiments, it is seems this is the fastest way to implement it
697  // for GCC (eg., it does not like using std::min after the pstore !!)
698  EIGEN_ALIGN16 int aux[4];
699  pstore(aux, a);
700  int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
701  int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
702  return aux0>aux2 ? aux0 : aux2;
703 #endif // EIGEN_VECTORIZE_SSE4_1
704 }
705 
706 #if EIGEN_COMP_GNUC
707 // template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
708 // {
709 // Packet4f res = b;
710 // asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c));
711 // return res;
712 // }
713 // EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i)
714 // {
715 // Packet4i res = a;
716 // asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i));
717 // return res;
718 // }
719 #endif
720 
721 #ifdef EIGEN_VECTORIZE_SSSE3
722 // SSSE3 versions
723 template<int Offset>
724 struct palign_impl<Offset,Packet4f>
725 {
726  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
727  {
728  if (Offset!=0)
729  first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
730  }
731 };
732 
733 template<int Offset>
734 struct palign_impl<Offset,Packet4i>
735 {
736  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
737  {
738  if (Offset!=0)
739  first = _mm_alignr_epi8(second,first, Offset*4);
740  }
741 };
742 
743 template<int Offset>
744 struct palign_impl<Offset,Packet2d>
745 {
746  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
747  {
748  if (Offset==1)
749  first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
750  }
751 };
752 #else
753 // SSE2 versions
754 template<int Offset>
755 struct palign_impl<Offset,Packet4f>
756 {
757  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
758  {
759  if (Offset==1)
760  {
761  first = _mm_move_ss(first,second);
762  first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39));
763  }
764  else if (Offset==2)
765  {
766  first = _mm_movehl_ps(first,first);
767  first = _mm_movelh_ps(first,second);
768  }
769  else if (Offset==3)
770  {
771  first = _mm_move_ss(first,second);
772  first = _mm_shuffle_ps(first,second,0x93);
773  }
774  }
775 };
776 
777 template<int Offset>
778 struct palign_impl<Offset,Packet4i>
779 {
780  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
781  {
782  if (Offset==1)
783  {
784  first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
785  first = _mm_shuffle_epi32(first,0x39);
786  }
787  else if (Offset==2)
788  {
789  first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first)));
790  first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
791  }
792  else if (Offset==3)
793  {
794  first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
795  first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93));
796  }
797  }
798 };
799 
800 template<int Offset>
801 struct palign_impl<Offset,Packet2d>
802 {
803  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
804  {
805  if (Offset==1)
806  {
807  first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first)));
808  first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second)));
809  }
810  }
811 };
812 #endif
813 
814 EIGEN_DEVICE_FUNC inline void
815 ptranspose(PacketBlock<Packet4f,4>& kernel) {
816  _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
817 }
818 
819 EIGEN_DEVICE_FUNC inline void
820 ptranspose(PacketBlock<Packet2d,2>& kernel) {
821  __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
822  kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
823  kernel.packet[1] = tmp;
824 }
825 
826 EIGEN_DEVICE_FUNC inline void
827 ptranspose(PacketBlock<Packet4i,4>& kernel) {
828  __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
829  __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
830  __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
831  __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
832 
833  kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
834  kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
835  kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
836  kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
837 }
838 
839 template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
840  const __m128i zero = _mm_setzero_si128();
841  const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
842  __m128i false_mask = _mm_cmpeq_epi32(select, zero);
843 #ifdef EIGEN_VECTORIZE_SSE4_1
844  return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
845 #else
846  return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
847 #endif
848 }
849 template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
850  const __m128 zero = _mm_setzero_ps();
851  const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
852  __m128 false_mask = _mm_cmpeq_ps(select, zero);
853 #ifdef EIGEN_VECTORIZE_SSE4_1
854  return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
855 #else
856  return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
857 #endif
858 }
859 template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
860  const __m128d zero = _mm_setzero_pd();
861  const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
862  __m128d false_mask = _mm_cmpeq_pd(select, zero);
863 #ifdef EIGEN_VECTORIZE_SSE4_1
864  return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
865 #else
866  return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
867 #endif
868 }
869 
870 } // end namespace internal
871 
872 } // end namespace Eigen
873 
874 #endif // EIGEN_PACKET_MATH_SSE_H
Definition: LDLT.h:16
Definition: Constants.h:230
Definition: Eigen_Colamd.h:54