@smake/eigen 1.0.2 → 1.1.1

package/eigen/Eigen/src/Core/arch/SSE/Complex.h

@@ -10,427 +10,389 @@
 #ifndef EIGEN_COMPLEX_SSE_H
 #define EIGEN_COMPLEX_SSE_H
 
+// IWYU pragma: private
+#include "../../InternalHeaderCheck.h"
+
 namespace Eigen {
 
 namespace internal {
 
 //---------- float ----------
-struct Packet2cf
-{
+struct Packet2cf {
   EIGEN_STRONG_INLINE Packet2cf() {}
   EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {}
-  __m128  v;
+  Packet4f v;
 };
 
 // Use the packet_traits defined in AVX/PacketMath.h instead if we're going
 // to leverage AVX instructions.
 #ifndef EIGEN_VECTORIZE_AVX
-template<> struct packet_traits<std::complex<float> >  : default_packet_traits
-{
+template <>
+struct packet_traits<std::complex<float> > : default_packet_traits {
   typedef Packet2cf type;
   typedef Packet2cf half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 1,
     size = 2,
-    HasHalfPacket = 0,
 
-    HasAdd    = 1,
-    HasSub    = 1,
-    HasMul    = 1,
-    HasDiv    = 1,
+    HasAdd = 1,
+    HasSub = 1,
+    HasMul = 1,
+    HasDiv = 1,
     HasNegate = 1,
-    HasAbs    = 0,
-    HasAbs2   = 0,
-    HasMin    = 0,
-    HasMax    = 0,
+    HasSqrt = 1,
+    HasLog = 1,
+    HasExp = 1,
+    HasAbs = 0,
+    HasAbs2 = 0,
+    HasMin = 0,
+    HasMax = 0,
     HasSetLinear = 0,
     HasBlend = 1
   };
 };
 #endif
 
-template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
-
-template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a)
-{
-  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
-  return Packet2cf(_mm_xor_ps(a.v,mask));
-}
-template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
-{
-  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
-  return Packet2cf(_mm_xor_ps(a.v,mask));
-}
-
-template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
-{
-  #ifdef EIGEN_VECTORIZE_SSE3
-  return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
-                                 _mm_mul_ps(_mm_movehdup_ps(a.v),
-                                            vec4f_swizzle1(b.v, 1, 0, 3, 2))));
-//   return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
-//                                  _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
-//                                             vec4f_swizzle1(b.v, 1, 0, 3, 2))));
-  #else
-  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000));
-  return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
-                              _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
-                                                    vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
-  #endif
-}
-
-template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); }
-
-template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); }
-template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); }
-
-template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
-{
-  Packet2cf res;
-#if EIGEN_GNUC_AT_MOST(4,2)
-  // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
-  res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
-#elif EIGEN_GNUC_AT_LEAST(4,6)
-  // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6
-  #pragma GCC diagnostic push
-  #pragma GCC diagnostic ignored "-Wuninitialized"
-  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
-  #pragma GCC diagnostic pop
-#else
-  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
-#endif
-  return Packet2cf(_mm_movelh_ps(res.v,res.v));
-}
-
-template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
-
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); }
+template <>
+struct unpacket_traits<Packet2cf> {
+  typedef std::complex<float> type;
+  typedef Packet2cf half;
+  typedef Packet4f as_real;
+  enum {
+    size = 2,
+    alignment = Aligned16,
+    vectorizable = true,
+    masked_load_available = false,
+    masked_store_available = false
+  };
+};
 
+template <>
+EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_add_ps(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_sub_ps(a.v, b.v));
+}
 
-template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
-{
-  return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]),
-                              std::imag(from[0*stride]), std::real(from[0*stride])));
+template <>
+EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) {
+  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000, 0x80000000, 0x80000000, 0x80000000));
+  return Packet2cf(_mm_xor_ps(a.v, mask));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) {
+  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000, 0x80000000, 0x00000000, 0x80000000));
+  return Packet2cf(_mm_xor_ps(a.v, mask));
 }
 
-template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
-{
-  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
-                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
-  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
-                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
+template <>
+EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) {
+#ifdef EIGEN_VECTORIZE_SSE3
+  __m128 tmp1 = _mm_mul_ps(_mm_movehdup_ps(a.v), vec4f_swizzle1(b.v, 1, 0, 3, 2));
+  __m128 tmp2 = _mm_moveldup_ps(a.v);
+#else
+  __m128 tmp1 = _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3), vec4f_swizzle1(b.v, 1, 0, 3, 2));
+  __m128 tmp2 = vec4f_swizzle1(a.v, 0, 0, 2, 2);
+#endif
+#ifdef EIGEN_VECTORIZE_FMA
+  __m128 result = _mm_fmaddsub_ps(tmp2, b.v, tmp1);
+#else
+#ifdef EIGEN_VECTORIZE_SSE3
+  __m128 result = _mm_addsub_ps(_mm_mul_ps(tmp2, b.v), tmp1);
+#else
+  const __m128 mask = _mm_setr_ps(-0.0f, 0.0f, -0.0f, 0.0f);
+  __m128 result = _mm_add_ps(_mm_mul_ps(tmp2, b.v), _mm_xor_ps(tmp1, mask));
+#endif
+#endif
+  return Packet2cf(result);
 }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template <>
+EIGEN_STRONG_INLINE Packet2cf ptrue<Packet2cf>(const Packet2cf& a) {
+  return Packet2cf(ptrue(Packet4f(a.v)));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pand<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_and_ps(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf por<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_or_ps(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pxor<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_xor_ps(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return Packet2cf(_mm_andnot_ps(b.v, a.v));
+}
 
-template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
-{
-  #if EIGEN_GNUC_AT_MOST(4,3)
-  // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares...
-  // This workaround also fix invalid code generation with gcc 4.3
-  EIGEN_ALIGN16 std::complex<float> res[2];
-  _mm_store_ps((float*)res, a.v);
-  return res[0];
-  #else
-  std::complex<float> res;
-  _mm_storel_pi((__m64*)&res, a.v);
-  return res;
-  #endif
+template <>
+EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(_mm_load_ps(&numext::real_ref(*from)));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(_mm_loadu_ps(&numext::real_ref(*from)));
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); }
+template <>
+EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>& from) {
+  const float re = std::real(from);
+  const float im = std::imag(from);
+  return Packet2cf(_mm_set_ps(im, re, im, re));
+}
 
-template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
-{
-  return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
+template <>
+EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) {
+  return pset1<Packet2cf>(*from);
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
-{
-  return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
+template <>
+EIGEN_STRONG_INLINE void pstore<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(&numext::real_ref(*to), from.v);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet2cf& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(&numext::real_ref(*to), from.v);
 }
 
-template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
-{
-  return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
+template <>
+EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from,
+                                                                           Index stride) {
+  return Packet2cf(_mm_set_ps(std::imag(from[1 * stride]), std::real(from[1 * stride]), std::imag(from[0 * stride]),
+                              std::real(from[0 * stride])));
 }
 
-template<int Offset>
-struct palign_impl<Offset,Packet2cf>
-{
-  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
-  {
-    if (Offset==1)
-    {
-      first.v = _mm_movehl_ps(first.v, first.v);
-      first.v = _mm_movelh_ps(first.v, second.v);
-    }
-  }
-};
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from,
+                                                                       Index stride) {
+  to[stride * 0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
+                                       _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
+  to[stride * 1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
+                                       _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
+}
 
-template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return internal::pmul(a, pconj(b));
-    #else
-    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
-    return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
-                                _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
-                                           vec4f_swizzle1(b.v, 1, 0, 3, 2))));
-    #endif
-  }
-};
+template <>
+EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float>* addr) {
+  _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
+}
 
-template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return internal::pmul(pconj(a), b);
-    #else
-    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
-    return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
-                                _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
-                                                      vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
-    #endif
-  }
-};
+template <>
+EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a) {
+  alignas(alignof(__m64)) std::complex<float> res;
+  _mm_storel_pi((__m64*)&res, a.v);
+  return res;
+}
 
-template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
-{
-  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return pconj(internal::pmul(a, b));
-    #else
-    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
-    return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
-                                _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
-                                           vec4f_swizzle1(b.v, 1, 0, 3, 2))));
-    #endif
-  }
-};
+template <>
+EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) {
+  return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v)))));
+}
 
-EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf,Packet4f)
+template <>
+EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a) {
+  return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v, a.v))));
+}
 
-template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
-{
-  // TODO optimize it for SSE3 and 4
-  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
-  __m128 s = _mm_mul_ps(b.v,b.v);
-  return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
+template <>
+EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a) {
+  return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v, a.v))));
 }
 
-EIGEN_STRONG_INLINE Packet2cf pcplxflip/* <Packet2cf> */(const Packet2cf& x)
-{
+EIGEN_STRONG_INLINE Packet2cf pcplxflip /* <Packet2cf> */ (const Packet2cf& x) {
   return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
 }
 
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet2cf, Packet4f)
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b) {
+  return pdiv_complex(a, b);
+}
 
 //---------- double ----------
-struct Packet1cd
-{
+struct Packet1cd {
   EIGEN_STRONG_INLINE Packet1cd() {}
   EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
-  __m128d  v;
+  Packet2d v;
 };
 
 // Use the packet_traits defined in AVX/PacketMath.h instead if we're going
 // to leverage AVX instructions.
 #ifndef EIGEN_VECTORIZE_AVX
-template<> struct packet_traits<std::complex<double> >  : default_packet_traits
-{
+template <>
+struct packet_traits<std::complex<double> > : default_packet_traits {
   typedef Packet1cd type;
   typedef Packet1cd half;
   enum {
     Vectorizable = 1,
     AlignedOnScalar = 0,
     size = 1,
-    HasHalfPacket = 0,
 
-    HasAdd    = 1,
-    HasSub    = 1,
-    HasMul    = 1,
-    HasDiv    = 1,
+    HasAdd = 1,
+    HasSub = 1,
+    HasMul = 1,
+    HasDiv = 1,
     HasNegate = 1,
-    HasAbs    = 0,
-    HasAbs2   = 0,
-    HasMin    = 0,
-    HasMax    = 0,
+    HasSqrt = 1,
+    HasLog = 1,
+    HasAbs = 0,
+    HasAbs2 = 0,
+    HasMin = 0,
+    HasMax = 0,
     HasSetLinear = 0
   };
 };
 #endif
 
-template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+template <>
+struct unpacket_traits<Packet1cd> {
+  typedef std::complex<double> type;
+  typedef Packet1cd half;
+  typedef Packet2d as_real;
+  enum {
+    size = 1,
+    alignment = Aligned16,
+    vectorizable = true,
+    masked_load_available = false,
+    masked_store_available = false
+  };
+};
 
-template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
-template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
-{
-  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
-  return Packet1cd(_mm_xor_pd(a.v,mask));
+template <>
+EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_add_pd(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_sub_pd(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) {
+  return Packet1cd(pnegate(Packet2d(a.v)));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) {
+  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000, 0x0, 0x0, 0x0));
+  return Packet1cd(_mm_xor_pd(a.v, mask));
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
-{
-  #ifdef EIGEN_VECTORIZE_SSE3
-  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v),
-                                 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
-                                            vec2d_swizzle1(b.v, 1, 0))));
-  #else
-  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
-  return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
-                              _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
-                                                    vec2d_swizzle1(b.v, 1, 0)), mask)));
-  #endif
+template <>
+EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) {
+  __m128d tmp1 = _mm_mul_pd(_mm_unpackhi_pd(a.v, a.v), vec2d_swizzle1(b.v, 1, 0));
+#ifdef EIGEN_VECTORIZE_SSE3
+  __m128d tmp2 = _mm_movedup_pd(a.v);
+#else
+  __m128d tmp2 = _mm_unpacklo_pd(a.v, a.v);
+#endif
+#ifdef EIGEN_VECTORIZE_FMA
+  __m128d result = _mm_fmaddsub_pd(tmp2, b.v, tmp1);
+#else
+#ifdef EIGEN_VECTORIZE_SSE3
+  __m128d result = _mm_addsub_pd(_mm_mul_pd(tmp2, b.v), tmp1);
+#else
+  const __m128d mask = _mm_setr_pd(-0.0, 0.0);
+  __m128d result = _mm_add_pd(_mm_mul_pd(tmp2, b.v), _mm_xor_pd(tmp1, mask));
+#endif
+#endif
+  return Packet1cd(result);
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
+template <>
+EIGEN_STRONG_INLINE Packet1cd ptrue<Packet1cd>(const Packet1cd& a) {
+  return Packet1cd(ptrue(Packet2d(a.v)));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pand<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_and_pd(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd por<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_or_pd(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pxor<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_xor_pd(a.v, b.v));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return Packet1cd(_mm_andnot_pd(b.v, a.v));
+}
 
 // FIXME force unaligned load, this is a temporary fix
-template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from)
-{ EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
-template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
-{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
-template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
-{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+template <>
+EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(_mm_load_pd((const double*)from));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(_mm_loadu_pd((const double*)from));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd
+pset1<Packet1cd>(const std::complex<double>& from) { /* here we really have to use unaligned loads :( */
+  return ploadu<Packet1cd>(&from);
+}
 
-template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
+template <>
+EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) {
+  return pset1<Packet1cd>(*from);
+}
 
 // FIXME force unaligned store, this is a temporary fix
-template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
-template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
+template <>
+EIGEN_STRONG_INLINE void pstore<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd((double*)to, from.v);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double>* to, const Packet1cd& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd((double*)to, from.v);
+}
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
+template <>
+EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double>* addr) {
+  _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0);
+}
 
-template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
-{
+template <>
+EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet1cd>(const Packet1cd& a) {
   EIGEN_ALIGN16 double res[2];
   _mm_store_pd(res, a.v);
-  return std::complex<double>(res[0],res[1]);
+  return std::complex<double>(res[0], res[1]);
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
-
-template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
-{
-  return pfirst(a);
+template <>
+EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) {
+  return a;
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
-{
-  return vecs[0];
+template <>
+EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) {
+  return pfirst(a);
 }
 
-template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
-{
+template <>
+EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) {
   return pfirst(a);
 }
 
-template<int Offset>
-struct palign_impl<Offset,Packet1cd>
-{
-  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
-  {
-    // FIXME is it sure we never have to align a Packet1cd?
-    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
-  }
-};
-
-template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return internal::pmul(a, pconj(b));
-    #else
-    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
-    return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
-                                _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
-                                           vec2d_swizzle1(b.v, 1, 0))));
-    #endif
-  }
-};
-
-template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return internal::pmul(pconj(a), b);
-    #else
-    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
-    return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
-                                _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
-                                                      vec2d_swizzle1(b.v, 1, 0)), mask)));
-    #endif
-  }
-};
-
-template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
-{
-  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
-  { return padd(pmul(x,y),c); }
-
-  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
-  {
-    #ifdef EIGEN_VECTORIZE_SSE3
-    return pconj(internal::pmul(a, b));
-    #else
-    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
-    return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
-                                _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
-                                           vec2d_swizzle1(b.v, 1, 0))));
-    #endif
-  }
-};
-
-EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd,Packet2d)
+EIGEN_MAKE_CONJ_HELPER_CPLX_REAL(Packet1cd, Packet2d)
 
-template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
-{
-  // TODO optimize it for SSE3 and 4
-  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
-  __m128d s = _mm_mul_pd(b.v,b.v);
-  return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1))));
+template <>
+EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b) {
+  return pdiv_complex(a, b);
 }
 
-EIGEN_STRONG_INLINE Packet1cd pcplxflip/* <Packet1cd> */(const Packet1cd& x)
-{
+EIGEN_STRONG_INLINE Packet1cd pcplxflip /* <Packet1cd> */ (const Packet1cd& x) {
   return Packet1cd(preverse(Packet2d(x.v)));
 }
 
-EIGEN_DEVICE_FUNC inline void
-ptranspose(PacketBlock<Packet2cf,2>& kernel) {
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet2cf, 2>& kernel) {
   __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
   __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
 
@@ -439,33 +401,103 @@ ptranspose(PacketBlock<Packet2cf,2>& kernel) {
   kernel.packet[1].v = tmp;
 }
 
-template<>  EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+template <>
+EIGEN_STRONG_INLINE Packet2cf pcmp_eq(const Packet2cf& a, const Packet2cf& b) {
+  __m128 eq = _mm_cmpeq_ps(a.v, b.v);
+  return Packet2cf(pand<Packet4f>(eq, vec4f_swizzle1(eq, 1, 0, 3, 2)));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet1cd pcmp_eq(const Packet1cd& a, const Packet1cd& b) {
+  __m128d eq = _mm_cmpeq_pd(a.v, b.v);
+  return Packet1cd(pand<Packet2d>(eq, vec2d_swizzle1(eq, 1, 0)));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket,
+                                     const Packet2cf& elsePacket) {
   __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
   return Packet2cf(_mm_castpd_ps(result));
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b)
-{
-  return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+template <>
+EIGEN_STRONG_INLINE Packet1cd psqrt<Packet1cd>(const Packet1cd& a) {
+  return psqrt_complex<Packet1cd>(a);
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b)
-{
-  return pset1<Packet1cd>(b);
+template <>
+EIGEN_STRONG_INLINE Packet2cf psqrt<Packet2cf>(const Packet2cf& a) {
+  return psqrt_complex<Packet2cf>(a);
 }
 
-template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b)
-{
-  return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+template <>
+EIGEN_STRONG_INLINE Packet1cd plog<Packet1cd>(const Packet1cd& a) {
+  return plog_complex<Packet1cd>(a);
 }
 
-template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b)
-{
-  return pset1<Packet1cd>(b);
+template <>
+EIGEN_STRONG_INLINE Packet2cf plog<Packet2cf>(const Packet2cf& a) {
+  return plog_complex<Packet2cf>(a);
 }
 
-} // end namespace internal
+template <>
+EIGEN_STRONG_INLINE Packet2cf pexp<Packet2cf>(const Packet2cf& a) {
+  return pexp_complex<Packet2cf>(a);
+}
 
-} // end namespace Eigen
+#ifdef EIGEN_VECTORIZE_FMA
+// std::complex<float>
+template <>
+EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
+  __m128 a_odd = _mm_movehdup_ps(a.v);
+  __m128 a_even = _mm_moveldup_ps(a.v);
+  __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
+  __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmaddsub_ps(a_odd, b_swap, c.v));
+  return Packet2cf(result);
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pmsub(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
+  __m128 a_odd = _mm_movehdup_ps(a.v);
+  __m128 a_even = _mm_moveldup_ps(a.v);
+  __m128 b_swap = _mm_permute_ps(b.v, _MM_SHUFFLE(2, 3, 0, 1));
+  __m128 result = _mm_fmaddsub_ps(a_even, b.v, _mm_fmsubadd_ps(a_odd, b_swap, c.v));
+  return Packet2cf(result);
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pnmadd(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
+  return pnegate(pmsub(a, b, c));
+}
+template <>
+EIGEN_STRONG_INLINE Packet2cf pnmsub(const Packet2cf& a, const Packet2cf& b, const Packet2cf& c) {
+  return pnegate(pmadd(a, b, c));
+}
+// std::complex<double>
+template <>
+EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
+  __m128d a_odd = _mm_permute_pd(a.v, 0x3);
+  __m128d a_even = _mm_movedup_pd(a.v);
+  __m128d b_swap = _mm_permute_pd(b.v, 0x1);
+  __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmaddsub_pd(a_odd, b_swap, c.v));
+  return Packet1cd(result);
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pmsub(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
+  __m128d a_odd = _mm_permute_pd(a.v, 0x3);
+  __m128d a_even = _mm_movedup_pd(a.v);
+  __m128d b_swap = _mm_permute_pd(b.v, 0x1);
+  __m128d result = _mm_fmaddsub_pd(a_even, b.v, _mm_fmsubadd_pd(a_odd, b_swap, c.v));
+  return Packet1cd(result);
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pnmadd(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
+  return pnegate(pmsub(a, b, c));
+}
+template <>
+EIGEN_STRONG_INLINE Packet1cd pnmsub(const Packet1cd& a, const Packet1cd& b, const Packet1cd& c) {
+  return pnegate(pmadd(a, b, c));
+}
+#endif
+}  // end namespace internal
+}  // end namespace Eigen
 
-#endif // EIGEN_COMPLEX_SSE_H
+#endif  // EIGEN_COMPLEX_SSE_H