GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-08 17:59:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Created a NEON version of the ptranspose packet primitives

Browse Source
This commit is contained in:
Benoit Steiner 2014-04-23 18:22:10 -07:00
parent 82b09fcb91
commit ccb4dec719
2 changed files with 30 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
Eigen/src/Core/arch/NEON/Complex.h
View File

@ -263,6 +263,14 @@ template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, con
return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s))); return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s)));
} }
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet2cf>& kernel) {
float32x4_t tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v));
kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v));
kernel.packet[1].v = tmp;
}
} // end namespace internal } // end namespace internal
} // end namespace Eigen } // end namespace Eigen

23
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@ -447,9 +447,30 @@ PALIGN_NEON(0,Packet4i,vextq_s32)
PALIGN_NEON(1,Packet4i,vextq_s32) PALIGN_NEON(1,Packet4i,vextq_s32)
PALIGN_NEON(2,Packet4i,vextq_s32) PALIGN_NEON(2,Packet4i,vextq_s32)
PALIGN_NEON(3,Packet4i,vextq_s32) PALIGN_NEON(3,Packet4i,vextq_s32)
#undef PALIGN_NEON #undef PALIGN_NEON
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet4f>& kernel) {
float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]);
float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]);
kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0]));
kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0]));
kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1]));
kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1]));
}
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet4i>& kernel) {
int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]);
int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]);
kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0]));
kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0]));
kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1]));
kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1]));
}
} // end namespace internal } // end namespace internal
} // end namespace Eigen } // end namespace Eigen