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


acado
Author(s): Milan Vukov, Rien Quirynen
autogenerated on Thu Aug 27 2015 11:59:23