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

Fix tensor casts for large packets and casts to/from std::complex

Browse Source 
The original tensor casts were only defined for
`SrcCoeffRatio`:`TgtCoeffRatio` 1:1, 1:2, 2:1, 4:1. Here we add the
missing 1:N and 8:1.

We also add casting `Eigen::half` to/from `std::complex<T>`, which
was missing to make it consistent with `Eigen:bfloat16`, and
generalize the overload to work for any complex type.

Tests were added to `basicstuff`, `packetmath`, and
`cxx11_tensor_casts` to test all cast configurations.
This commit is contained in:
Antonio Sanchez
2020-06-25 14:31:16 -07:00
committed by Antonio Sánchez
parent 145e51516f
commit 9cb8771e9c
7 changed files with 388 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
View File

@@ -51,7 +51,10 @@ struct nested<TensorConversionOp<TargetType, XprType>, 1, typename eval<TensorCo
template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket, int SrcCoeffRatio, int TgtCoeffRatio>
struct PacketConverter {
struct PacketConverter;
template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, 1> {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
PacketConverter(const TensorEvaluator& impl)
: m_impl(impl) {}
@@ -109,7 +112,33 @@ struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 4, 1> {
};
template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket>
struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, 2> {
struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 8, 1> {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
PacketConverter(const TensorEvaluator& impl)
: m_impl(impl) {}
template<int LoadMode, typename Index>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TgtPacket packet(Index index) const {
const int SrcPacketSize = internal::unpacket_traits<SrcPacket>::size;
SrcPacket src1 = m_impl.template packet<LoadMode>(index);
SrcPacket src2 = m_impl.template packet<LoadMode>(index + 1 * SrcPacketSize);
SrcPacket src3 = m_impl.template packet<LoadMode>(index + 2 * SrcPacketSize);
SrcPacket src4 = m_impl.template packet<LoadMode>(index + 3 * SrcPacketSize);
SrcPacket src5 = m_impl.template packet<LoadMode>(index + 4 * SrcPacketSize);
SrcPacket src6 = m_impl.template packet<LoadMode>(index + 5 * SrcPacketSize);
SrcPacket src7 = m_impl.template packet<LoadMode>(index + 6 * SrcPacketSize);
SrcPacket src8 = m_impl.template packet<LoadMode>(index + 7 * SrcPacketSize);
TgtPacket result = internal::pcast<SrcPacket, TgtPacket>(src1, src2, src3, src4, src5, src6, src7, src8);
return result;
}
private:
const TensorEvaluator& m_impl;
};
template <typename TensorEvaluator, typename SrcPacket, typename TgtPacket, int TgtCoeffRatio>
struct PacketConverter<TensorEvaluator, SrcPacket, TgtPacket, 1, TgtCoeffRatio> {
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
PacketConverter(const TensorEvaluator& impl)
: m_impl(impl), m_maxIndex(impl.dimensions().TotalSize()) {}