GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-12 19:59:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

VSX supports vec_div, implement where appropriate (float/doubles)

Browse Source
This commit is contained in:
Konstantinos Margaritis 2014-09-21 08:12:22 +00:00
parent 10f8aabb61
commit c0205ca4af
1 changed files with 6 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
Eigen/src/Core/arch/AltiVec/PacketMath.h
View File

@ -336,6 +336,7 @@ template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const
*/ */
template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
{ {
#ifndef __VSX__ // VSX actually provides a div instruction
Packet4f t, y_0, y_1, res; Packet4f t, y_0, y_1, res;
// Altivec does not offer a divide instruction, we have to do a reciprocal approximation // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
@ -345,8 +346,10 @@ template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const
t = vec_nmsub(y_0, b, p4f_ONE); t = vec_nmsub(y_0, b, p4f_ONE);
y_1 = vec_madd(y_0, t, y_0); y_1 = vec_madd(y_0, t, y_0);
res = vec_madd(a, y_1, p4f_ZERO); return vec_madd(a, y_1, p4f_ZERO);
return res; #else
return vec_div(a, b);
#endif
} }
template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/) template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
@ -801,20 +804,7 @@ template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return psub
template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; } template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); } template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_ZERO); }
template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
{
Packet2d t, y_0, y_1, res;
// Altivec does not offer a divide instruction, we have to do a reciprocal approximation
y_0 = vec_re(b);
// Do one Newton-Raphson iteration to get the needed accuracy
t = vec_nmsub(y_0, b, p2d_ONE);
y_1 = vec_madd(y_0, t, y_0);
res = vec_madd(a, y_1, p2d_ZERO);
return res;
}
// for some weird raisons, it has to be overloaded for packet of integers // for some weird raisons, it has to be overloaded for packet of integers
template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); } template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }