devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source
This commit is contained in:
Gael Guennebaud
2015-03-02 20:09:33 +01:00
parent ef09ce4552
commit 3109f0e74e
2 changed files with 51 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
Eigen/src/Geometry/arch/Geometry_SSE.h
View File

@@ -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),
vec4f_swizzle1(b,2,0,1,2)),mask);
__m128 flip2 = _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),
vec4f_swizzle1(b,0,1,2,1)),mask);
__m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
__m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
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>
struct quat_product<Architecture::SSE, Derived, OtherDerived, double, Aligned>
template<class Derived, class OtherDerived, int Alignment>
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_zw = _b.coeffs().template packet<Aligned>(2);
Packet2d b_xy = _b.coeffs().template packet<Alignment>(0);
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