Add SSE vectorization of Quaternion::conjugate. Significant speed-up when combined with products like q1*q2.conjugate()

Eigen/src/Geometry/Quaternion.h

@@ -441,7 +441,7 @@ QuaternionBase<Derived>::operator* (const QuaternionBase<OtherDerived>& other) c
    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
   return internal::quat_product<Architecture::Target, Derived, OtherDerived,
                          typename internal::traits<Derived>::Scalar,
-                         internal::traits<Derived>::IsAligned && internal::traits<OtherDerived>::IsAligned>::run(*this, other);
+                         (internal::traits<Derived>::IsAligned && internal::traits<OtherDerived>::IsAligned)?Aligned:Unaligned>::run(*this, other);
 }
 
 /** \sa operator*(Quaternion) */
@@ -646,6 +646,16 @@ inline Quaternion<typename internal::traits<Derived>::Scalar> QuaternionBase<Der
   }
 }
 
+// Generic conjugate of a Quaternion
+namespace internal {
+template<int Arch, class Derived, typename Scalar, int _Options> struct quat_conj
+{
+  static EIGEN_STRONG_INLINE Quaternion<Scalar> run(const QuaternionBase<Derived>& q){
+    return Quaternion<Scalar>(q.w(),-q.x(),-q.y(),-q.z());
+  }
+};
+}
+                         
 /** \returns the conjugate of the \c *this which is equal to the multiplicative inverse
   * if the quaternion is normalized.
   * The conjugate of a quaternion represents the opposite rotation.
@@ -656,7 +666,10 @@ template <class Derived>
 inline Quaternion<typename internal::traits<Derived>::Scalar>
 QuaternionBase<Derived>::conjugate() const
 {
-  return Quaternion<Scalar>(this->w(),-this->x(),-this->y(),-this->z());
+  return internal::quat_conj<Architecture::Target, Derived,
+                         typename internal::traits<Derived>::Scalar,
+                         internal::traits<Derived>::IsAligned?Aligned:Unaligned>::run(*this);
+                         
 }
 
 /** \returns the angle (in radian) between two rotations

Eigen/src/Geometry/arch/Geometry_SSE.h

@@ -20,23 +20,35 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, float, Aligned>
 {
   static inline Quaternion<float> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
   {
-    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0,0,0,0x80000000));
     Quaternion<float> res;
+    const __m128 mask = _mm_setr_ps(0.f,0.f,0.f,-0.f);
     __m128 a = _a.coeffs().template packet<Aligned>(0);
     __m128 b = _b.coeffs().template packet<Aligned>(0);
-    __m128 flip1 = _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),
+    __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
-                                         vec4f_swizzle1(b,2,0,1,2)),mask);
+    __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
-    __m128 flip2 = _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),
-                                         vec4f_swizzle1(b,0,1,2,1)),mask);
     pstore(&res.x(),
               _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
                                     _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0),
                                                vec4f_swizzle1(b,1,2,0,0))),
-                         _mm_add_ps(flip1,flip2)));
+                         _mm_xor_ps(mask,_mm_add_ps(s1,s2))));
+    
     return res;
   }
 };
 
+template<class Derived, int Alignment>
+struct quat_conj<Architecture::SSE, Derived, float, Alignment>
+{
+  static inline Quaternion<float> run(const QuaternionBase<Derived>& q)
+  {
+    Quaternion<float> res;
+    const __m128 mask = _mm_setr_ps(-0.f,-0.f,-0.f,0.f);
+    pstore(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet<Alignment>(0)));
+    return res;
+  }
+};
+
+
 template<typename VectorLhs,typename VectorRhs>
 struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
 {
@@ -56,8 +68,8 @@ struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
 
 
 
-template<class Derived, class OtherDerived>
+template<class Derived, class OtherDerived, int Alignment>
-struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Aligned>
+struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Alignment>
 {
   static inline Quaternion<double> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
   {
@@ -66,8 +78,8 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Aligned>
   Quaternion<double> res;
 
   const double* a = _a.coeffs().data();
-  Packet2d b_xy = _b.coeffs().template packet<Aligned>(0);
+  Packet2d b_xy = _b.coeffs().template packet<Alignment>(0);
-  Packet2d b_zw = _b.coeffs().template packet<Aligned>(2);
+  Packet2d b_zw = _b.coeffs().template packet<Alignment>(2);
   Packet2d a_xx = pset1<Packet2d>(a[0]);
   Packet2d a_yy = pset1<Packet2d>(a[1]);
   Packet2d a_zz = pset1<Packet2d>(a[2]);
@@ -108,6 +120,20 @@ struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Aligned>
 }
 };
 
+template<class Derived, int Alignment>
+struct quat_conj<Architecture::SSE, Derived, double, Alignment>
+{
+  static inline Quaternion<double> run(const QuaternionBase<Derived>& q)
+  {
+    Quaternion<double> res;
+    const __m128d mask0 = _mm_setr_pd(-0.,-0.);
+    const __m128d mask2 = _mm_setr_pd(-0.,0.);
+    pstore(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet<Alignment>(0)));
+    pstore(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet<Alignment>(2)));
+    return res;
+  }
+};
+
 } // end namespace internal
 
 } // end namespace Eigen