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) 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#ifndef EIGEN_VECTORIZE_AVX 413template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 414template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 415template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 416#endif 417 418#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64 419// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 420// Direct of the struct members fixed bug #62. 421template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; } 422template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; } 423template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } 424#elif EIGEN_COMP_MSVC_STRICT 425// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 426template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; } 427template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; } 428template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } 429#else 430template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); } 431template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); } 432template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); } 433#endif 434 435template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) 436{ return _mm_shuffle_ps(a,a,0x1B); } 437template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) 438{ return _mm_shuffle_pd(a,a,0x1); } 439template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) 440{ return _mm_shuffle_epi32(a,0x1B); } 441 442template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) 443{ 444 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); 445 return _mm_and_ps(a,mask); 446} 447template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) 448{ 449 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); 450 return _mm_and_pd(a,mask); 451} 452template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) 453{ 454 #ifdef EIGEN_VECTORIZE_SSSE3 455 return _mm_abs_epi32(a); 456 #else 457 Packet4i aux = _mm_srai_epi32(a,31); 458 return _mm_sub_epi32(_mm_xor_si128(a,aux),aux); 459 #endif 460} 461 462// with AVX, the default implementations based on pload1 are faster 463#ifndef __AVX__ 464template<> EIGEN_STRONG_INLINE void 465pbroadcast4<Packet4f>(const float *a, 466 Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3) 467{ 468 a3 = pload<Packet4f>(a); 469 a0 = vec4f_swizzle1(a3, 0,0,0,0); 470 a1 = vec4f_swizzle1(a3, 1,1,1,1); 471 a2 = vec4f_swizzle1(a3, 2,2,2,2); 472 a3 = vec4f_swizzle1(a3, 3,3,3,3); 473} 474template<> EIGEN_STRONG_INLINE void 475pbroadcast4<Packet2d>(const double *a, 476 Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3) 477{ 478#ifdef EIGEN_VECTORIZE_SSE3 479 a0 = _mm_loaddup_pd(a+0); 480 a1 = _mm_loaddup_pd(a+1); 481 a2 = _mm_loaddup_pd(a+2); 482 a3 = _mm_loaddup_pd(a+3); 483#else 484 a1 = pload<Packet2d>(a); 485 a0 = vec2d_swizzle1(a1, 0,0); 486 a1 = vec2d_swizzle1(a1, 1,1); 487 a3 = pload<Packet2d>(a+2); 488 a2 = vec2d_swizzle1(a3, 0,0); 489 a3 = vec2d_swizzle1(a3, 1,1); 490#endif 491} 492#endif 493 494EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) 495{ 496 vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); 497 vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA)); 498 vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF)); 499 vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00)); 500} 501 502#ifdef EIGEN_VECTORIZE_SSE3 503template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) 504{ 505 return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3])); 506} 507 508template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) 509{ 510 return _mm_hadd_pd(vecs[0], vecs[1]); 511} 512 513#else 514template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) 515{ 516 Packet4f tmp0, tmp1, tmp2; 517 tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]); 518 tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]); 519 tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]); 520 tmp0 = _mm_add_ps(tmp0, tmp1); 521 tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]); 522 tmp1 = _mm_add_ps(tmp1, tmp2); 523 tmp2 = _mm_movehl_ps(tmp1, tmp0); 524 tmp0 = _mm_movelh_ps(tmp0, tmp1); 525 return _mm_add_ps(tmp0, tmp2); 526} 527 528template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) 529{ 530 return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1])); 531} 532#endif // SSE3 533 534template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) 535{ 536 // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures 537 // (from Nehalem to Haswell) 538// #ifdef EIGEN_VECTORIZE_SSE3 539// Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3)); 540// return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp)); 541// #else 542 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); 543 return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 544// #endif 545} 546 547template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) 548{ 549 // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures 550 // (from Nehalem to Haswell) 551// #ifdef EIGEN_VECTORIZE_SSE3 552// return pfirst<Packet2d>(_mm_hadd_pd(a, a)); 553// #else 554 return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); 555// #endif 556} 557 558#ifdef EIGEN_VECTORIZE_SSSE3 559template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) 560{ 561 return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); 562} 563template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) 564{ 565 Packet4i tmp0 = _mm_hadd_epi32(a,a); 566 return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0)); 567} 568#else 569template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) 570{ 571 Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); 572 return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1)); 573} 574 575template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) 576{ 577 Packet4i tmp0, tmp1, tmp2; 578 tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]); 579 tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]); 580 tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]); 581 tmp0 = _mm_add_epi32(tmp0, tmp1); 582 tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]); 583 tmp1 = _mm_add_epi32(tmp1, tmp2); 584 tmp2 = _mm_unpacklo_epi64(tmp0, tmp1); 585 tmp0 = _mm_unpackhi_epi64(tmp0, tmp1); 586 return _mm_add_epi32(tmp0, tmp2); 587} 588#endif 589// Other reduction functions: 590 591// mul 592template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) 593{ 594 Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); 595 return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 596} 597template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) 598{ 599 return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); 600} 601template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) 602{ 603 // after some experiments, it is seems this is the fastest way to implement it 604 // for GCC (eg., reusing pmul is very slow !) 605 // TODO try to call _mm_mul_epu32 directly 606 EIGEN_ALIGN16 int aux[4]; 607 pstore(aux, a); 608 return (aux[0] * aux[1]) * (aux[2] * aux[3]);; 609} 610 611// min 612template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) 613{ 614 Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); 615 return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 616} 617template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) 618{ 619 return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); 620} 621template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) 622{ 623#ifdef EIGEN_VECTORIZE_SSE4_1 624 Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); 625 return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); 626#else 627 // after some experiments, it is seems this is the fastest way to implement it 628 // for GCC (eg., it does not like using std::min after the pstore !!) 629 EIGEN_ALIGN16 int aux[4]; 630 pstore(aux, a); 631 int aux0 = aux[0]<aux[1] ? aux[0] : aux[1]; 632 int aux2 = aux[2]<aux[3] ? aux[2] : aux[3]; 633 return aux0<aux2 ? aux0 : aux2; 634#endif // EIGEN_VECTORIZE_SSE4_1 635} 636 637// max 638template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) 639{ 640 Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); 641 return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 642} 643template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) 644{ 645 return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); 646} 647template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) 648{ 649#ifdef EIGEN_VECTORIZE_SSE4_1 650 Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2))); 651 return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1))); 652#else 653 // after some experiments, it is seems this is the fastest way to implement it 654 // for GCC (eg., it does not like using std::min after the pstore !!) 655 EIGEN_ALIGN16 int aux[4]; 656 pstore(aux, a); 657 int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; 658 int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; 659 return aux0>aux2 ? aux0 : aux2; 660#endif // EIGEN_VECTORIZE_SSE4_1 661} 662 663#if EIGEN_COMP_GNUC 664// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) 665// { 666// Packet4f res = b; 667// asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c)); 668// return res; 669// } 670// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i) 671// { 672// Packet4i res = a; 673// asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i)); 674// return res; 675// } 676#endif 677 678#ifdef EIGEN_VECTORIZE_SSSE3 679// SSSE3 versions 680template<int Offset> 681struct palign_impl<Offset,Packet4f> 682{ 683 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) 684 { 685 if (Offset!=0) 686 first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4)); 687 } 688}; 689 690template<int Offset> 691struct palign_impl<Offset,Packet4i> 692{ 693 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) 694 { 695 if (Offset!=0) 696 first = _mm_alignr_epi8(second,first, Offset*4); 697 } 698}; 699 700template<int Offset> 701struct palign_impl<Offset,Packet2d> 702{ 703 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) 704 { 705 if (Offset==1) 706 first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8)); 707 } 708}; 709#else 710// SSE2 versions 711template<int Offset> 712struct palign_impl<Offset,Packet4f> 713{ 714 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) 715 { 716 if (Offset==1) 717 { 718 first = _mm_move_ss(first,second); 719 first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39)); 720 } 721 else if (Offset==2) 722 { 723 first = _mm_movehl_ps(first,first); 724 first = _mm_movelh_ps(first,second); 725 } 726 else if (Offset==3) 727 { 728 first = _mm_move_ss(first,second); 729 first = _mm_shuffle_ps(first,second,0x93); 730 } 731 } 732}; 733 734template<int Offset> 735struct palign_impl<Offset,Packet4i> 736{ 737 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) 738 { 739 if (Offset==1) 740 { 741 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 742 first = _mm_shuffle_epi32(first,0x39); 743 } 744 else if (Offset==2) 745 { 746 first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first))); 747 first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 748 } 749 else if (Offset==3) 750 { 751 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 752 first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93)); 753 } 754 } 755}; 756 757template<int Offset> 758struct palign_impl<Offset,Packet2d> 759{ 760 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) 761 { 762 if (Offset==1) 763 { 764 first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first))); 765 first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second))); 766 } 767 } 768}; 769#endif 770 771EIGEN_DEVICE_FUNC inline void 772ptranspose(PacketBlock<Packet4f,4>& kernel) { 773 _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]); 774} 775 776EIGEN_DEVICE_FUNC inline void 777ptranspose(PacketBlock<Packet2d,2>& kernel) { 778 __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]); 779 kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]); 780 kernel.packet[1] = tmp; 781} 782 783EIGEN_DEVICE_FUNC inline void 784ptranspose(PacketBlock<Packet4i,4>& kernel) { 785 __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]); 786 __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]); 787 __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]); 788 __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]); 789 790 kernel.packet[0] = _mm_unpacklo_epi64(T0, T1); 791 kernel.packet[1] = _mm_unpackhi_epi64(T0, T1); 792 kernel.packet[2] = _mm_unpacklo_epi64(T2, T3); 793 kernel.packet[3] = _mm_unpackhi_epi64(T2, T3); 794} 795 796template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) { 797 const __m128i zero = _mm_setzero_si128(); 798 const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); 799 __m128i false_mask = _mm_cmpeq_epi32(select, zero); 800#ifdef EIGEN_VECTORIZE_SSE4_1 801 return _mm_blendv_epi8(thenPacket, elsePacket, false_mask); 802#else 803 return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket)); 804#endif 805} 806template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) { 807 const __m128 zero = _mm_setzero_ps(); 808 const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]); 809 __m128 false_mask = _mm_cmpeq_ps(select, zero); 810#ifdef EIGEN_VECTORIZE_SSE4_1 811 return _mm_blendv_ps(thenPacket, elsePacket, false_mask); 812#else 813 return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket)); 814#endif 815} 816template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) { 817 const __m128d zero = _mm_setzero_pd(); 818 const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]); 819 __m128d false_mask = _mm_cmpeq_pd(select, zero); 820#ifdef EIGEN_VECTORIZE_SSE4_1 821 return _mm_blendv_pd(thenPacket, elsePacket, false_mask); 822#else 823 return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket)); 824#endif 825} 826 827template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b) 828{ 829#ifdef EIGEN_VECTORIZE_SSE4_1 830 return _mm_blend_ps(a,pset1<Packet4f>(b),1); 831#else 832 return _mm_move_ss(a, _mm_load_ss(&b)); 833#endif 834} 835 836template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b) 837{ 838#ifdef EIGEN_VECTORIZE_SSE4_1 839 return _mm_blend_pd(a,pset1<Packet2d>(b),1); 840#else 841 return _mm_move_sd(a, _mm_load_sd(&b)); 842#endif 843} 844 845template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b) 846{ 847#ifdef EIGEN_VECTORIZE_SSE4_1 848 return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3)); 849#else 850 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF)); 851 return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b))); 852#endif 853} 854 855template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b) 856{ 857#ifdef EIGEN_VECTORIZE_SSE4_1 858 return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1)); 859#else 860 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF)); 861 return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b))); 862#endif 863} 864 865// Scalar path for pmadd with FMA to ensure consistency with vectorized path. 866#ifdef __FMA__ 867template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) { 868 return ::fmaf(a,b,c); 869} 870template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) { 871 return ::fma(a,b,c); 872} 873#endif 874 875} // end namespace internal 876 877} // end namespace Eigen 878 879#endif // EIGEN_PACKET_MATH_SSE_H 880