mirror of
https://gitlab.com/libeigen/eigen.git
synced 2026-04-10 11:34:33 +08:00
Implement ptranspose on altivec and fix pgather/pscatter
This commit is contained in:
@@ -146,6 +146,7 @@ inline std::ostream & operator <<(std::ostream & s, const Packetbi & v)
|
||||
return s;
|
||||
}
|
||||
*/
|
||||
|
||||
template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) {
|
||||
// Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
|
||||
float EIGEN_ALIGN16 af[4];
|
||||
@@ -533,6 +534,32 @@ struct palign_impl<Offset,Packet4i>
|
||||
}
|
||||
};
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline void
|
||||
ptranspose(Kernel<Packet4f>& kernel) {
|
||||
Packet4f t0, t1, t2, t3;
|
||||
t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
|
||||
t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
|
||||
t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
|
||||
t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
|
||||
kernel.packet[0] = vec_mergeh(t0, t2);
|
||||
kernel.packet[1] = vec_mergel(t0, t2);
|
||||
kernel.packet[2] = vec_mergeh(t1, t3);
|
||||
kernel.packet[3] = vec_mergel(t1, t3);
|
||||
}
|
||||
|
||||
template<> EIGEN_DEVICE_FUNC inline void
|
||||
ptranspose(Kernel<Packet4i>& kernel) {
|
||||
Packet4i t0, t1, t2, t3;
|
||||
t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
|
||||
t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
|
||||
t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
|
||||
t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
|
||||
kernel.packet[0] = vec_mergeh(t0, t2);
|
||||
kernel.packet[1] = vec_mergel(t0, t2);
|
||||
kernel.packet[2] = vec_mergeh(t1, t3);
|
||||
kernel.packet[3] = vec_mergel(t1, t3);
|
||||
}
|
||||
|
||||
} // end namespace internal
|
||||
|
||||
} // end namespace Eigen
|
||||
|
||||
Reference in New Issue
Block a user