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
13namespace Eigen {
14
15namespace 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 || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)) || EIGEN_OS_QNX
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:
37template<typename T>
38struct 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};
51typedef eigen_packet_wrapper<__m128> Packet4f;
52typedef eigen_packet_wrapper<__m128i> Packet4i;
53typedef eigen_packet_wrapper<__m128d> Packet2d;
54#else
55typedef __m128 Packet4f;
56typedef __m128i Packet4i;
57typedef __m128d Packet2d;
58#endif
59
60template<> struct is_arithmetic<__m128> { enum { value = true }; };
61template<> struct is_arithmetic<__m128i> { enum { value = true }; };
62template<> 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
95template<> 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 HasTanh = EIGEN_FAST_MATH,
113 HasBlend = 1
114
115#ifdef EIGEN_VECTORIZE_SSE4_1
116 ,
117 HasRound = 1,
118 HasFloor = 1,
119 HasCeil = 1
120#endif
121 };
122};
123template<> struct packet_traits<double> : default_packet_traits
124{
125 typedef Packet2d type;
126 typedef Packet2d half;
127 enum {
128 Vectorizable = 1,
129 AlignedOnScalar = 1,
130 size=2,
131 HasHalfPacket = 0,
132
133 HasDiv = 1,
134 HasExp = 1,
135 HasSqrt = 1,
136 HasRsqrt = 1,
137 HasBlend = 1
138
139#ifdef EIGEN_VECTORIZE_SSE4_1
140 ,
141 HasRound = 1,
142 HasFloor = 1,
143 HasCeil = 1
144#endif
145 };
146};
147#endif
148template<> struct packet_traits<int> : default_packet_traits
149{
150 typedef Packet4i type;
151 typedef Packet4i half;
152 enum {
153 Vectorizable = 1,
154 AlignedOnScalar = 1,
155 size=4,
156
157 HasBlend = 1
158 };
159};
160
161template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
162template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
163template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
164
165#ifndef EIGEN_VECTORIZE_AVX
166template<> struct scalar_div_cost<float,true> { enum { value = 7 }; };
167template<> struct scalar_div_cost<double,true> { enum { value = 8 }; };
168#endif
169
170#if EIGEN_COMP_MSVC==1500
171// Workaround MSVC 9 internal compiler error.
172// TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
173// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
174template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); }
175template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
176template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); }
177#else
178template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps1(from); }
179template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
180template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); }
181#endif
182
183// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
184// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
185// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
186// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply.
187// Also note that with AVX, we want it to generate a vbroadcastss.
188#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
189template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) {
190 return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
191}
192#endif
193
194template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
195template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
196template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
197
198template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
199template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); }
200template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); }
201
202template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); }
203template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); }
204template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); }
205
206template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
207{
208 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
209 return _mm_xor_ps(a,mask);
210}
211template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a)
212{
213 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
214 return _mm_xor_pd(a,mask);
215}
216template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a)
217{
218 return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
219}
220
221template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
222template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
223template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
224
225template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); }
226template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); }
227template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
228{
229#ifdef EIGEN_VECTORIZE_SSE4_1
230 return _mm_mullo_epi32(a,b);
231#else
232 // this version is slightly faster than 4 scalar products
233 return vec4i_swizzle1(
234 vec4i_swizzle2(
235 _mm_mul_epu32(a,b),
236 _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
237 vec4i_swizzle1(b,1,0,3,2)),
238 0,2,0,2),
239 0,2,1,3);
240#endif
241}
242
243template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
244template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
245
246// for some weird raisons, it has to be overloaded for packet of integers
247template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
248#ifdef __FMA__
249template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
250template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
251#endif
252
253template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); }
254template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); }
255template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b)
256{
257#ifdef EIGEN_VECTORIZE_SSE4_1
258 return _mm_min_epi32(a,b);
259#else
260 // after some bench, this version *is* faster than a scalar implementation
261 Packet4i mask = _mm_cmplt_epi32(a,b);
262 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
263#endif
264}
265
266template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); }
267template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); }
268template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b)
269{
270#ifdef EIGEN_VECTORIZE_SSE4_1
271 return _mm_max_epi32(a,b);
272#else
273 // after some bench, this version *is* faster than a scalar implementation
274 Packet4i mask = _mm_cmpgt_epi32(a,b);
275 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
276#endif
277}
278
279#ifdef EIGEN_VECTORIZE_SSE4_1
280template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
281template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
282
283template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
284template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
285
286template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
287template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
288#endif
289
290template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
291template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
292template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
293
294template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); }
295template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); }
296template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); }
297
298template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); }
299template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); }
300template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); }
301
302template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); }
303template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); }
304template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); }
305
306template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); }
307template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); }
308template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
309
310#if EIGEN_COMP_MSVC
311 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) {
312 EIGEN_DEBUG_UNALIGNED_LOAD
313 #if (EIGEN_COMP_MSVC==1600)
314 // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
315 // (i.e., it does not generate an unaligned load!!
316 __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
317 res = _mm_loadh_pi(res, (const __m64*)(from+2));
318 return res;
319 #else
320 return _mm_loadu_ps(from);
321 #endif
322 }
323#else
324// NOTE: with the code below, MSVC's compiler crashes!
325
326template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
327{
328 EIGEN_DEBUG_UNALIGNED_LOAD
329 return _mm_loadu_ps(from);
330}
331#endif
332
333template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
334{
335 EIGEN_DEBUG_UNALIGNED_LOAD
336 return _mm_loadu_pd(from);
337}
338template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
339{
340 EIGEN_DEBUG_UNALIGNED_LOAD
341 return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
342}
343
344
345template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
346{
347 return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
348}
349template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from)
350{ return pset1<Packet2d>(from[0]); }
351template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
352{
353 Packet4i tmp;
354 tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
355 return vec4i_swizzle1(tmp, 0, 0, 1, 1);
356}
357
358template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
359template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
360template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
361
362template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
363template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
364template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
365
366template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
367{
368 return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
369}
370template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
371{
372 return _mm_set_pd(from[1*stride], from[0*stride]);
373}
374template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
375{
376 return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
377 }
378
379template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
380{
381 to[stride*0] = _mm_cvtss_f32(from);
382 to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
383 to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
384 to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
385}
386template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
387{
388 to[stride*0] = _mm_cvtsd_f64(from);
389 to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
390}
391template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
392{
393 to[stride*0] = _mm_cvtsi128_si32(from);
394 to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
395 to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
396 to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
397}
398
399// some compilers might be tempted to perform multiple moves instead of using a vector path.
400template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
401{
402 Packet4f pa = _mm_set_ss(a);
403 pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
404}
405// some compilers might be tempted to perform multiple moves instead of using a vector path.
406template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a)
407{
408 Packet2d pa = _mm_set_sd(a);
409 pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
410}
411
412#if EIGEN_COMP_PGI
413typedef const void * SsePrefetchPtrType;
414#else
415typedef const char * SsePrefetchPtrType;
416#endif
417
418#ifndef EIGEN_VECTORIZE_AVX
419template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
420template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
421template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
422#endif
423
424#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
425// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
426// Direct of the struct members fixed bug #62.
427template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; }
428template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; }
429template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
430#elif EIGEN_COMP_MSVC_STRICT
431// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
432template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; }
433template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; }
434template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
435#else
436template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); }
437template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); }
438template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); }
439#endif
440
441template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
442{ return _mm_shuffle_ps(a,a,0x1B); }
443template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
444{ return _mm_shuffle_pd(a,a,0x1); }
445template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
446{ return _mm_shuffle_epi32(a,0x1B); }
447
448template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a)
449{
450 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
451 return _mm_and_ps(a,mask);
452}
453template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a)
454{
455 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
456 return _mm_and_pd(a,mask);
457}
458template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
459{
460 #ifdef EIGEN_VECTORIZE_SSSE3
461 return _mm_abs_epi32(a);
462 #else
463 Packet4i aux = _mm_srai_epi32(a,31);
464 return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
465 #endif
466}
467
468// with AVX, the default implementations based on pload1 are faster
469#ifndef __AVX__
470template<> EIGEN_STRONG_INLINE void
471pbroadcast4<Packet4f>(const float *a,
472 Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
473{
474 a3 = pload<Packet4f>(a);
475 a0 = vec4f_swizzle1(a3, 0,0,0,0);
476 a1 = vec4f_swizzle1(a3, 1,1,1,1);
477 a2 = vec4f_swizzle1(a3, 2,2,2,2);
478 a3 = vec4f_swizzle1(a3, 3,3,3,3);
479}
480template<> EIGEN_STRONG_INLINE void
481pbroadcast4<Packet2d>(const double *a,
482 Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
483{
484#ifdef EIGEN_VECTORIZE_SSE3
485 a0 = _mm_loaddup_pd(a+0);
486 a1 = _mm_loaddup_pd(a+1);
487 a2 = _mm_loaddup_pd(a+2);
488 a3 = _mm_loaddup_pd(a+3);
489#else
490 a1 = pload<Packet2d>(a);
491 a0 = vec2d_swizzle1(a1, 0,0);
492 a1 = vec2d_swizzle1(a1, 1,1);
493 a3 = pload<Packet2d>(a+2);
494 a2 = vec2d_swizzle1(a3, 0,0);
495 a3 = vec2d_swizzle1(a3, 1,1);
496#endif
497}
498#endif
499
500EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
501{
502 vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
503 vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
504 vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
505 vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
506}
507
508#ifdef EIGEN_VECTORIZE_SSE3
509template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
510{
511 return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
512}
513
514template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
515{
516 return _mm_hadd_pd(vecs[0], vecs[1]);
517}
518
519#else
520template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
521{
522 Packet4f tmp0, tmp1, tmp2;
523 tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
524 tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
525 tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
526 tmp0 = _mm_add_ps(tmp0, tmp1);
527 tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
528 tmp1 = _mm_add_ps(tmp1, tmp2);
529 tmp2 = _mm_movehl_ps(tmp1, tmp0);
530 tmp0 = _mm_movelh_ps(tmp0, tmp1);
531 return _mm_add_ps(tmp0, tmp2);
532}
533
534template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
535{
536 return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
537}
538#endif // SSE3
539
540template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
541{
542 // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
543 // (from Nehalem to Haswell)
544// #ifdef EIGEN_VECTORIZE_SSE3
545// Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3));
546// return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp));
547// #else
548 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
549 return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
550// #endif
551}
552
553template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
554{
555 // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
556 // (from Nehalem to Haswell)
557// #ifdef EIGEN_VECTORIZE_SSE3
558// return pfirst<Packet2d>(_mm_hadd_pd(a, a));
559// #else
560 return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
561// #endif
562}
563
564#ifdef EIGEN_VECTORIZE_SSSE3
565template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
566{
567 return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
568}
569template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
570{
571 Packet4i tmp0 = _mm_hadd_epi32(a,a);
572 return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
573}
574#else
575template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
576{
577 Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
578 return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
579}
580
581template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
582{
583 Packet4i tmp0, tmp1, tmp2;
584 tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
585 tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
586 tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
587 tmp0 = _mm_add_epi32(tmp0, tmp1);
588 tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
589 tmp1 = _mm_add_epi32(tmp1, tmp2);
590 tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
591 tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
592 return _mm_add_epi32(tmp0, tmp2);
593}
594#endif
595// Other reduction functions:
596
597// mul
598template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
599{
600 Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
601 return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
602}
603template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
604{
605 return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
606}
607template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
608{
609 // after some experiments, it is seems this is the fastest way to implement it
610 // for GCC (eg., reusing pmul is very slow !)
611 // TODO try to call _mm_mul_epu32 directly
612 EIGEN_ALIGN16 int aux[4];
613 pstore(aux, a);
614 return (aux[0] * aux[1]) * (aux[2] * aux[3]);;
615}
616
617// min
618template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
619{
620 Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
621 return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
622}
623template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
624{
625 return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
626}
627template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
628{
629#ifdef EIGEN_VECTORIZE_SSE4_1
630 Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
631 return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
632#else
633 // after some experiments, it is seems this is the fastest way to implement it
634 // for GCC (eg., it does not like using std::min after the pstore !!)
635 EIGEN_ALIGN16 int aux[4];
636 pstore(aux, a);
637 int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
638 int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
639 return aux0<aux2 ? aux0 : aux2;
640#endif // EIGEN_VECTORIZE_SSE4_1
641}
642
643// max
644template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
645{
646 Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
647 return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
648}
649template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
650{
651 return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
652}
653template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
654{
655#ifdef EIGEN_VECTORIZE_SSE4_1
656 Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
657 return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
658#else
659 // after some experiments, it is seems this is the fastest way to implement it
660 // for GCC (eg., it does not like using std::min after the pstore !!)
661 EIGEN_ALIGN16 int aux[4];
662 pstore(aux, a);
663 int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
664 int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
665 return aux0>aux2 ? aux0 : aux2;
666#endif // EIGEN_VECTORIZE_SSE4_1
667}
668
669#if EIGEN_COMP_GNUC
670// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
671// {
672// Packet4f res = b;
673// asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c));
674// return res;
675// }
676// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i)
677// {
678// Packet4i res = a;
679// asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i));
680// return res;
681// }
682#endif
683
684#ifdef EIGEN_VECTORIZE_SSSE3
685// SSSE3 versions
686template<int Offset>
687struct palign_impl<Offset,Packet4f>
688{
689 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
690 {
691 if (Offset!=0)
692 first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
693 }
694};
695
696template<int Offset>
697struct palign_impl<Offset,Packet4i>
698{
699 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
700 {
701 if (Offset!=0)
702 first = _mm_alignr_epi8(second,first, Offset*4);
703 }
704};
705
706template<int Offset>
707struct palign_impl<Offset,Packet2d>
708{
709 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
710 {
711 if (Offset==1)
712 first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
713 }
714};
715#else
716// SSE2 versions
717template<int Offset>
718struct palign_impl<Offset,Packet4f>
719{
720 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
721 {
722 if (Offset==1)
723 {
724 first = _mm_move_ss(first,second);
725 first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39));
726 }
727 else if (Offset==2)
728 {
729 first = _mm_movehl_ps(first,first);
730 first = _mm_movelh_ps(first,second);
731 }
732 else if (Offset==3)
733 {
734 first = _mm_move_ss(first,second);
735 first = _mm_shuffle_ps(first,second,0x93);
736 }
737 }
738};
739
740template<int Offset>
741struct palign_impl<Offset,Packet4i>
742{
743 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
744 {
745 if (Offset==1)
746 {
747 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
748 first = _mm_shuffle_epi32(first,0x39);
749 }
750 else if (Offset==2)
751 {
752 first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first)));
753 first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
754 }
755 else if (Offset==3)
756 {
757 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
758 first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93));
759 }
760 }
761};
762
763template<int Offset>
764struct palign_impl<Offset,Packet2d>
765{
766 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
767 {
768 if (Offset==1)
769 {
770 first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first)));
771 first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second)));
772 }
773 }
774};
775#endif
776
777EIGEN_DEVICE_FUNC inline void
778ptranspose(PacketBlock<Packet4f,4>& kernel) {
779 _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
780}
781
782EIGEN_DEVICE_FUNC inline void
783ptranspose(PacketBlock<Packet2d,2>& kernel) {
784 __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
785 kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
786 kernel.packet[1] = tmp;
787}
788
789EIGEN_DEVICE_FUNC inline void
790ptranspose(PacketBlock<Packet4i,4>& kernel) {
791 __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
792 __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
793 __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
794 __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
795
796 kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
797 kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
798 kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
799 kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
800}
801
802template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
803 const __m128i zero = _mm_setzero_si128();
804 const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
805 __m128i false_mask = _mm_cmpeq_epi32(select, zero);
806#ifdef EIGEN_VECTORIZE_SSE4_1
807 return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
808#else
809 return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
810#endif
811}
812template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
813 const __m128 zero = _mm_setzero_ps();
814 const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
815 __m128 false_mask = _mm_cmpeq_ps(select, zero);
816#ifdef EIGEN_VECTORIZE_SSE4_1
817 return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
818#else
819 return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
820#endif
821}
822template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
823 const __m128d zero = _mm_setzero_pd();
824 const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
825 __m128d false_mask = _mm_cmpeq_pd(select, zero);
826#ifdef EIGEN_VECTORIZE_SSE4_1
827 return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
828#else
829 return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
830#endif
831}
832
833template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b)
834{
835#ifdef EIGEN_VECTORIZE_SSE4_1
836 return _mm_blend_ps(a,pset1<Packet4f>(b),1);
837#else
838 return _mm_move_ss(a, _mm_load_ss(&b));
839#endif
840}
841
842template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b)
843{
844#ifdef EIGEN_VECTORIZE_SSE4_1
845 return _mm_blend_pd(a,pset1<Packet2d>(b),1);
846#else
847 return _mm_move_sd(a, _mm_load_sd(&b));
848#endif
849}
850
851template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b)
852{
853#ifdef EIGEN_VECTORIZE_SSE4_1
854 return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3));
855#else
856 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF));
857 return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b)));
858#endif
859}
860
861template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b)
862{
863#ifdef EIGEN_VECTORIZE_SSE4_1
864 return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1));
865#else
866 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF));
867 return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b)));
868#endif
869}
870
871// Scalar path for pmadd with FMA to ensure consistency with vectorized path.
872#ifdef __FMA__
873template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) {
874 return ::fmaf(a,b,c);
875}
876template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) {
877 return ::fma(a,b,c);
878}
879#endif
880
881} // end namespace internal
882
883} // end namespace Eigen
884
885#if EIGEN_COMP_PGI
886// PGI++ does not define the following intrinsics in C++ mode.
887static inline __m128 _mm_castpd_ps (__m128d x) { return reinterpret_cast<__m128&>(x); }
888static inline __m128i _mm_castpd_si128(__m128d x) { return reinterpret_cast<__m128i&>(x); }
889static inline __m128d _mm_castps_pd (__m128 x) { return reinterpret_cast<__m128d&>(x); }
890static inline __m128i _mm_castps_si128(__m128 x) { return reinterpret_cast<__m128i&>(x); }
891static inline __m128 _mm_castsi128_ps(__m128i x) { return reinterpret_cast<__m128&>(x); }
892static inline __m128d _mm_castsi128_pd(__m128i x) { return reinterpret_cast<__m128d&>(x); }
893#endif
894
895#endif // EIGEN_PACKET_MATH_SSE_H
896