mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-01-30 17:40:05 +08:00
Code Issues Packages Projects Releases Wiki Activity

Created the ptranspose packet primitive that can transpose an array of N packets, where N is the number of words in each packet. This primitive will be used to complete the vectorization of the gemm_pack_lhs and gemm_pack_rhs functions.

Browse Source 
Implemented the primitive using SSE instructions.
This commit is contained in:
Benoit Steiner 2014-03-26 19:03:07 -07:00
parent 14bc4b9704
commit a419cea4a0
4 changed files with 61 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
Eigen/src/Core/GenericPacketMath.h
View File

@ -386,9 +386,22 @@ template<> inline std::complex<double> pmul(const std::complex<double>& a, const
#endif
/***************************************************************************
* Kernel, that is a collection of N packets where N is the number of words
* in the packet.
***************************************************************************/
template <typename Packet> struct Kernel {
Packet packet[unpacket_traits<Packet>::size];
};
template<typename Packet> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet>& /*kernel*/) {
// Nothing to do in the scalar case, i.e. a 1x1 matrix.
}
} // end namespace internal
} // end namespace Eigen
#endif // EIGEN_GENERIC_PACKET_MATH_H

10
Eigen/src/Core/arch/SSE/Complex.h
View File

@ -435,6 +435,16 @@ EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
return Packet1cd(preverse(x.v));
}
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet2cf>& kernel) {
__m128d w1 = _mm_castps_pd(kernel.packet[0].v);
__m128d w2 = _mm_castps_pd(kernel.packet[1].v);
__m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
kernel.packet[1].v = tmp;
}
} // end namespace internal
} // end namespace Eigen

25
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -707,6 +707,31 @@ struct palign_impl<Offset,Packet2d>
};
#endif
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet4f>& kernel) {
_MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
}
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet2d>& kernel) {
__m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
kernel.packet[1] = tmp;
}
template<> EIGEN_DEVICE_FUNC inline void
ptranspose(Kernel<Packet4i>& kernel) {
__m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
__m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
__m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
__m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
}
} // end namespace internal
} // end namespace Eigen

12
test/packetmath.cpp
View File

@ -208,6 +208,18 @@ template<typename Scalar> void packetmath()
ref[i] = data1[PacketSize-i-1];
internal::pstore(data2, internal::preverse(internal::pload<Packet>(data1)));
VERIFY(areApprox(ref, data2, PacketSize) && "internal::preverse");
internal::Kernel<Packet> kernel;
for (int i=0; i<PacketSize; ++i) {
kernel.packet[i] = internal::pload<Packet>(data1+i*PacketSize);
}
ptranspose(kernel);
for (int i=0; i<PacketSize; ++i) {
internal::pstore(data2, kernel.packet[i]);
for (int j = 0; j < PacketSize; ++j) {
VERIFY(isApproxAbs(data2[j], data1[i+j*PacketSize], refvalue));
}
}
}
template<typename Scalar> void packetmath_real()