$treeview $search $mathjax
Eigen
3.2.5
$projectbrief
|
$projectbrief
|
$searchbox |
00001 // This file is part of Eigen, a lightweight C++ template library 00002 // for linear algebra. 00003 // 00004 // Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr> 00005 // 00006 // This Source Code Form is subject to the terms of the Mozilla 00007 // Public License v. 2.0. If a copy of the MPL was not distributed 00008 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. 00009 00010 #ifndef EIGEN_PACKET_MATH_SSE_H 00011 #define EIGEN_PACKET_MATH_SSE_H 00012 00013 namespace Eigen { 00014 00015 namespace internal { 00016 00017 #ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 00018 #define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8 00019 #endif 00020 00021 #ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 00022 #define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*)) 00023 #endif 00024 00025 typedef __m128 Packet4f; 00026 typedef __m128i Packet4i; 00027 typedef __m128d Packet2d; 00028 00029 template<> struct is_arithmetic<__m128> { enum { value = true }; }; 00030 template<> struct is_arithmetic<__m128i> { enum { value = true }; }; 00031 template<> struct is_arithmetic<__m128d> { enum { value = true }; }; 00032 00033 #define vec4f_swizzle1(v,p,q,r,s) \ 00034 (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p))))) 00035 00036 #define vec4i_swizzle1(v,p,q,r,s) \ 00037 (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p)))) 00038 00039 #define vec2d_swizzle1(v,p,q) \ 00040 (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2))))) 00041 00042 #define vec4f_swizzle2(a,b,p,q,r,s) \ 00043 (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p)))) 00044 00045 #define vec4i_swizzle2(a,b,p,q,r,s) \ 00046 (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p)))))) 00047 00048 #define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \ 00049 const Packet4f p4f_##NAME = pset1<Packet4f>(X) 00050 00051 #define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \ 00052 const Packet2d p2d_##NAME = pset1<Packet2d>(X) 00053 00054 #define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \ 00055 const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X)) 00056 00057 #define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \ 00058 const Packet4i p4i_##NAME = pset1<Packet4i>(X) 00059 00060 00061 template<> struct packet_traits<float> : default_packet_traits 00062 { 00063 typedef Packet4f type; 00064 enum { 00065 Vectorizable = 1, 00066 AlignedOnScalar = 1, 00067 size=4, 00068 00069 HasDiv = 1, 00070 HasSin = EIGEN_FAST_MATH, 00071 HasCos = EIGEN_FAST_MATH, 00072 HasLog = 1, 00073 HasExp = 1, 00074 HasSqrt = 1 00075 }; 00076 }; 00077 template<> struct packet_traits<double> : default_packet_traits 00078 { 00079 typedef Packet2d type; 00080 enum { 00081 Vectorizable = 1, 00082 AlignedOnScalar = 1, 00083 size=2, 00084 00085 HasDiv = 1, 00086 HasExp = 1, 00087 HasSqrt = 1 00088 }; 00089 }; 00090 template<> struct packet_traits<int> : default_packet_traits 00091 { 00092 typedef Packet4i type; 00093 enum { 00094 // FIXME check the Has* 00095 Vectorizable = 1, 00096 AlignedOnScalar = 1, 00097 size=4 00098 }; 00099 }; 00100 00101 template<> struct unpacket_traits<Packet4f> { typedef float type; enum {size=4}; }; 00102 template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2}; }; 00103 template<> struct unpacket_traits<Packet4i> { typedef int type; enum {size=4}; }; 00104 00105 #if defined(_MSC_VER) && (_MSC_VER==1500) 00106 // Workaround MSVC 9 internal compiler error. 00107 // TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode 00108 // TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)). 00109 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); } 00110 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); } 00111 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); } 00112 #else 00113 template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set1_ps(from); } 00114 template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); } 00115 template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); } 00116 #endif 00117 00118 template<> EIGEN_STRONG_INLINE Packet4f plset<float>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); } 00119 template<> EIGEN_STRONG_INLINE Packet2d plset<double>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); } 00120 template<> EIGEN_STRONG_INLINE Packet4i plset<int>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); } 00121 00122 template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); } 00123 template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); } 00124 template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); } 00125 00126 template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); } 00127 template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); } 00128 template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); } 00129 00130 template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) 00131 { 00132 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000)); 00133 return _mm_xor_ps(a,mask); 00134 } 00135 template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) 00136 { 00137 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000)); 00138 return _mm_xor_pd(a,mask); 00139 } 00140 template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) 00141 { 00142 return psub(_mm_setr_epi32(0,0,0,0), a); 00143 } 00144 00145 template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; } 00146 template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } 00147 template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; } 00148 00149 template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); } 00150 template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); } 00151 template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) 00152 { 00153 #ifdef EIGEN_VECTORIZE_SSE4_1 00154 return _mm_mullo_epi32(a,b); 00155 #else 00156 // this version is slightly faster than 4 scalar products 00157 return vec4i_swizzle1( 00158 vec4i_swizzle2( 00159 _mm_mul_epu32(a,b), 00160 _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2), 00161 vec4i_swizzle1(b,1,0,3,2)), 00162 0,2,0,2), 00163 0,2,1,3); 00164 #endif 00165 } 00166 00167 template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); } 00168 template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); } 00169 template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) 00170 { eigen_assert(false && "packet integer division are not supported by SSE"); 00171 return pset1<Packet4i>(0); 00172 } 00173 00174 // for some weird raisons, it has to be overloaded for packet of integers 00175 template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); } 00176 00177 template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); } 00178 template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); } 00179 template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) 00180 { 00181 #ifdef EIGEN_VECTORIZE_SSE4_1 00182 return _mm_min_epi32(a,b); 00183 #else 00184 // after some bench, this version *is* faster than a scalar implementation 00185 Packet4i mask = _mm_cmplt_epi32(a,b); 00186 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); 00187 #endif 00188 } 00189 00190 template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); } 00191 template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); } 00192 template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) 00193 { 00194 #ifdef EIGEN_VECTORIZE_SSE4_1 00195 return _mm_max_epi32(a,b); 00196 #else 00197 // after some bench, this version *is* faster than a scalar implementation 00198 Packet4i mask = _mm_cmpgt_epi32(a,b); 00199 return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b)); 00200 #endif 00201 } 00202 00203 template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); } 00204 template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); } 00205 template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); } 00206 00207 template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); } 00208 template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); } 00209 template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); } 00210 00211 template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); } 00212 template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); } 00213 template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); } 00214 00215 template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); } 00216 template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); } 00217 template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); } 00218 00219 template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); } 00220 template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); } 00221 template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const Packet4i*>(from)); } 00222 00223 #if defined(_MSC_VER) 00224 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) { 00225 EIGEN_DEBUG_UNALIGNED_LOAD 00226 #if (_MSC_VER==1600) 00227 // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps 00228 // (i.e., it does not generate an unaligned load!! 00229 // TODO On most architectures this version should also be faster than a single _mm_loadu_ps 00230 // so we could also enable it for MSVC08 but first we have to make this later does not generate crap when doing so... 00231 __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from)); 00232 res = _mm_loadh_pi(res, (const __m64*)(from+2)); 00233 return res; 00234 #else 00235 return _mm_loadu_ps(from); 00236 #endif 00237 } 00238 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); } 00239 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); } 00240 #else 00241 // Fast unaligned loads. Note that here we cannot directly use intrinsics: this would 00242 // require pointer casting to incompatible pointer types and leads to invalid code 00243 // because of the strict aliasing rule. The "dummy" stuff are required to enforce 00244 // a correct instruction dependency. 00245 // TODO: do the same for MSVC (ICC is compatible) 00246 // NOTE: with the code below, MSVC's compiler crashes! 00247 00248 #if defined(__GNUC__) && defined(__i386__) 00249 // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd 00250 #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 00251 #elif defined(__clang__) 00252 // bug 201: Segfaults in __mm_loadh_pd with clang 2.8 00253 #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1 00254 #else 00255 #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0 00256 #endif 00257 00258 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) 00259 { 00260 EIGEN_DEBUG_UNALIGNED_LOAD 00261 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 00262 return _mm_loadu_ps(from); 00263 #else 00264 __m128d res; 00265 res = _mm_load_sd((const double*)(from)) ; 00266 res = _mm_loadh_pd(res, (const double*)(from+2)) ; 00267 return _mm_castpd_ps(res); 00268 #endif 00269 } 00270 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) 00271 { 00272 EIGEN_DEBUG_UNALIGNED_LOAD 00273 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 00274 return _mm_loadu_pd(from); 00275 #else 00276 __m128d res; 00277 res = _mm_load_sd(from) ; 00278 res = _mm_loadh_pd(res,from+1); 00279 return res; 00280 #endif 00281 } 00282 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from) 00283 { 00284 EIGEN_DEBUG_UNALIGNED_LOAD 00285 #if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 00286 return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); 00287 #else 00288 __m128d res; 00289 res = _mm_load_sd((const double*)(from)) ; 00290 res = _mm_loadh_pd(res, (const double*)(from+2)) ; 00291 return _mm_castpd_si128(res); 00292 #endif 00293 } 00294 #endif 00295 00296 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from) 00297 { 00298 return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1); 00299 } 00300 template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from) 00301 { return pset1<Packet2d>(from[0]); } 00302 template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from) 00303 { 00304 Packet4i tmp; 00305 tmp = _mm_loadl_epi64(reinterpret_cast<const Packet4i*>(from)); 00306 return vec4i_swizzle1(tmp, 0, 0, 1, 1); 00307 } 00308 00309 template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); } 00310 template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); } 00311 template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<Packet4i*>(to), from); } 00312 00313 template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { 00314 EIGEN_DEBUG_UNALIGNED_STORE 00315 _mm_storel_pd((to), from); 00316 _mm_storeh_pd((to+1), from); 00317 } 00318 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castps_pd(from)); } 00319 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(reinterpret_cast<double*>(to), _mm_castsi128_pd(from)); } 00320 00321 // some compilers might be tempted to perform multiple moves instead of using a vector path. 00322 template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a) 00323 { 00324 Packet4f pa = _mm_set_ss(a); 00325 pstore(to, vec4f_swizzle1(pa,0,0,0,0)); 00326 } 00327 // some compilers might be tempted to perform multiple moves instead of using a vector path. 00328 template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a) 00329 { 00330 Packet2d pa = _mm_set_sd(a); 00331 pstore(to, vec2d_swizzle1(pa,0,0)); 00332 } 00333 00334 template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 00335 template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 00336 template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); } 00337 00338 #if defined(_MSC_VER) && defined(_WIN64) && !defined(__INTEL_COMPILER) 00339 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 00340 // Direct of the struct members fixed bug #62. 00341 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; } 00342 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; } 00343 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } 00344 #elif defined(_MSC_VER) && !defined(__INTEL_COMPILER) 00345 // The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010 00346 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; } 00347 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; } 00348 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; } 00349 #else 00350 template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); } 00351 template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); } 00352 template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); } 00353 #endif 00354 00355 template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) 00356 { return _mm_shuffle_ps(a,a,0x1B); } 00357 template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) 00358 { return _mm_shuffle_pd(a,a,0x1); } 00359 template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) 00360 { return _mm_shuffle_epi32(a,0x1B); } 00361 00362 00363 template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) 00364 { 00365 const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF)); 00366 return _mm_and_ps(a,mask); 00367 } 00368 template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) 00369 { 00370 const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF)); 00371 return _mm_and_pd(a,mask); 00372 } 00373 template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) 00374 { 00375 #ifdef EIGEN_VECTORIZE_SSSE3 00376 return _mm_abs_epi32(a); 00377 #else 00378 Packet4i aux = _mm_srai_epi32(a,31); 00379 return _mm_sub_epi32(_mm_xor_si128(a,aux),aux); 00380 #endif 00381 } 00382 00383 EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs) 00384 { 00385 vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55)); 00386 vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA)); 00387 vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF)); 00388 vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00)); 00389 } 00390 00391 #ifdef EIGEN_VECTORIZE_SSE3 00392 // TODO implement SSE2 versions as well as integer versions 00393 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) 00394 { 00395 return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3])); 00396 } 00397 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) 00398 { 00399 return _mm_hadd_pd(vecs[0], vecs[1]); 00400 } 00401 // SSSE3 version: 00402 // EIGEN_STRONG_INLINE Packet4i preduxp(const Packet4i* vecs) 00403 // { 00404 // return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3])); 00405 // } 00406 00407 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) 00408 { 00409 Packet4f tmp0 = _mm_hadd_ps(a,a); 00410 return pfirst(_mm_hadd_ps(tmp0, tmp0)); 00411 } 00412 00413 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return pfirst(_mm_hadd_pd(a, a)); } 00414 00415 // SSSE3 version: 00416 // EIGEN_STRONG_INLINE float predux(const Packet4i& a) 00417 // { 00418 // Packet4i tmp0 = _mm_hadd_epi32(a,a); 00419 // return pfirst(_mm_hadd_epi32(tmp0, tmp0)); 00420 // } 00421 #else 00422 // SSE2 versions 00423 template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a) 00424 { 00425 Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a)); 00426 return pfirst(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 00427 } 00428 template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) 00429 { 00430 return pfirst(_mm_add_sd(a, _mm_unpackhi_pd(a,a))); 00431 } 00432 00433 template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs) 00434 { 00435 Packet4f tmp0, tmp1, tmp2; 00436 tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]); 00437 tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]); 00438 tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]); 00439 tmp0 = _mm_add_ps(tmp0, tmp1); 00440 tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]); 00441 tmp1 = _mm_add_ps(tmp1, tmp2); 00442 tmp2 = _mm_movehl_ps(tmp1, tmp0); 00443 tmp0 = _mm_movelh_ps(tmp0, tmp1); 00444 return _mm_add_ps(tmp0, tmp2); 00445 } 00446 00447 template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs) 00448 { 00449 return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1])); 00450 } 00451 #endif // SSE3 00452 00453 template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a) 00454 { 00455 Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a)); 00456 return pfirst(tmp) + pfirst(_mm_shuffle_epi32(tmp, 1)); 00457 } 00458 00459 template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs) 00460 { 00461 Packet4i tmp0, tmp1, tmp2; 00462 tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]); 00463 tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]); 00464 tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]); 00465 tmp0 = _mm_add_epi32(tmp0, tmp1); 00466 tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]); 00467 tmp1 = _mm_add_epi32(tmp1, tmp2); 00468 tmp2 = _mm_unpacklo_epi64(tmp0, tmp1); 00469 tmp0 = _mm_unpackhi_epi64(tmp0, tmp1); 00470 return _mm_add_epi32(tmp0, tmp2); 00471 } 00472 00473 // Other reduction functions: 00474 00475 // mul 00476 template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a) 00477 { 00478 Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a)); 00479 return pfirst(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 00480 } 00481 template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) 00482 { 00483 return pfirst(_mm_mul_sd(a, _mm_unpackhi_pd(a,a))); 00484 } 00485 template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a) 00486 { 00487 // after some experiments, it is seems this is the fastest way to implement it 00488 // for GCC (eg., reusing pmul is very slow !) 00489 // TODO try to call _mm_mul_epu32 directly 00490 EIGEN_ALIGN16 int aux[4]; 00491 pstore(aux, a); 00492 return (aux[0] * aux[1]) * (aux[2] * aux[3]);; 00493 } 00494 00495 // min 00496 template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a) 00497 { 00498 Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a)); 00499 return pfirst(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 00500 } 00501 template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) 00502 { 00503 return pfirst(_mm_min_sd(a, _mm_unpackhi_pd(a,a))); 00504 } 00505 template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a) 00506 { 00507 // after some experiments, it is seems this is the fastest way to implement it 00508 // for GCC (eg., it does not like using std::min after the pstore !!) 00509 EIGEN_ALIGN16 int aux[4]; 00510 pstore(aux, a); 00511 int aux0 = aux[0]<aux[1] ? aux[0] : aux[1]; 00512 int aux2 = aux[2]<aux[3] ? aux[2] : aux[3]; 00513 return aux0<aux2 ? aux0 : aux2; 00514 } 00515 00516 // max 00517 template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a) 00518 { 00519 Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a)); 00520 return pfirst(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1))); 00521 } 00522 template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) 00523 { 00524 return pfirst(_mm_max_sd(a, _mm_unpackhi_pd(a,a))); 00525 } 00526 template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a) 00527 { 00528 // after some experiments, it is seems this is the fastest way to implement it 00529 // for GCC (eg., it does not like using std::min after the pstore !!) 00530 EIGEN_ALIGN16 int aux[4]; 00531 pstore(aux, a); 00532 int aux0 = aux[0]>aux[1] ? aux[0] : aux[1]; 00533 int aux2 = aux[2]>aux[3] ? aux[2] : aux[3]; 00534 return aux0>aux2 ? aux0 : aux2; 00535 } 00536 00537 #if (defined __GNUC__) 00538 // template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) 00539 // { 00540 // Packet4f res = b; 00541 // asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c)); 00542 // return res; 00543 // } 00544 // EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i) 00545 // { 00546 // Packet4i res = a; 00547 // asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i)); 00548 // return res; 00549 // } 00550 #endif 00551 00552 #ifdef EIGEN_VECTORIZE_SSSE3 00553 // SSSE3 versions 00554 template<int Offset> 00555 struct palign_impl<Offset,Packet4f> 00556 { 00557 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) 00558 { 00559 if (Offset!=0) 00560 first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4)); 00561 } 00562 }; 00563 00564 template<int Offset> 00565 struct palign_impl<Offset,Packet4i> 00566 { 00567 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) 00568 { 00569 if (Offset!=0) 00570 first = _mm_alignr_epi8(second,first, Offset*4); 00571 } 00572 }; 00573 00574 template<int Offset> 00575 struct palign_impl<Offset,Packet2d> 00576 { 00577 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) 00578 { 00579 if (Offset==1) 00580 first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8)); 00581 } 00582 }; 00583 #else 00584 // SSE2 versions 00585 template<int Offset> 00586 struct palign_impl<Offset,Packet4f> 00587 { 00588 static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second) 00589 { 00590 if (Offset==1) 00591 { 00592 first = _mm_move_ss(first,second); 00593 first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39)); 00594 } 00595 else if (Offset==2) 00596 { 00597 first = _mm_movehl_ps(first,first); 00598 first = _mm_movelh_ps(first,second); 00599 } 00600 else if (Offset==3) 00601 { 00602 first = _mm_move_ss(first,second); 00603 first = _mm_shuffle_ps(first,second,0x93); 00604 } 00605 } 00606 }; 00607 00608 template<int Offset> 00609 struct palign_impl<Offset,Packet4i> 00610 { 00611 static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second) 00612 { 00613 if (Offset==1) 00614 { 00615 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 00616 first = _mm_shuffle_epi32(first,0x39); 00617 } 00618 else if (Offset==2) 00619 { 00620 first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first))); 00621 first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 00622 } 00623 else if (Offset==3) 00624 { 00625 first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second))); 00626 first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93)); 00627 } 00628 } 00629 }; 00630 00631 template<int Offset> 00632 struct palign_impl<Offset,Packet2d> 00633 { 00634 static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second) 00635 { 00636 if (Offset==1) 00637 { 00638 first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first))); 00639 first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second))); 00640 } 00641 } 00642 }; 00643 #endif 00644 00645 } // end namespace internal 00646 00647 } // end namespace Eigen 00648 00649 #endif // EIGEN_PACKET_MATH_SSE_H