GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Add PacketConv implementation for non-vectorizable src expressions

Browse Source
This commit is contained in:
Eugene Zhulenev 2019-02-08 15:47:25 -08:00
parent 1e36166ed1
commit 21eb97d3e0
1 changed files with 26 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
View File

@ -230,8 +230,21 @@ struct PacketConv<SrcPacket, TargetPacket, LoadMode, true, IsSameT> {
} }
}; };
template <typename SrcPacket, typename TargetPacket, int LoadMode, bool ActuallyVectorize> template <typename SrcPacket, typename TargetPacket, int LoadMode>
struct PacketConv<SrcPacket, TargetPacket, LoadMode, ActuallyVectorize, true> { struct PacketConv<SrcPacket, TargetPacket, LoadMode, /*ActuallyVectorize=*/false, /*IsSameT=*/true> {
typedef typename internal::unpacket_traits<TargetPacket>::type TargetType;
static const int PacketSize = internal::unpacket_traits<TargetPacket>::size;
template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
EIGEN_ALIGN_MAX typename internal::remove_const<TargetType>::type values[PacketSize];
for (int i = 0; i < PacketSize; ++i) values[i] = impl.coeff(index+i);
return internal::pload<TargetPacket>(values);
}
};
template <typename SrcPacket, typename TargetPacket, int LoadMode>
struct PacketConv<SrcPacket, TargetPacket, LoadMode, /*ActuallyVectorize=*/true, /*IsSameT=*/true> {
template <typename ArgType, typename Device> template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) { static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
return impl.template packet<LoadMode>(index); return impl.template packet<LoadMode>(index);
@ -287,10 +300,17 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
} }
template<int LoadMode> template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType
{ packet(Index index) const {
const bool Vectorizable = TensorEvaluator<ArgType, Device>::PacketAccess & // If we are not going to do the cast, we just need to check that base
internal::type_casting_traits<SrcType, TargetType>::VectorizedCast; // TensorEvaluator has packet access. Otherwise we also need to make sure,
// that we have an implementation of vectorized cast.
const bool Vectorizable =
IsSameType
? TensorEvaluator<ArgType, Device>::PacketAccess
: TensorEvaluator<ArgType, Device>::PacketAccess &
internal::type_casting_traits<SrcType, TargetType>::VectorizedCast;
return internal::PacketConv<PacketSourceType, PacketReturnType, LoadMode, return internal::PacketConv<PacketSourceType, PacketReturnType, LoadMode,
Vectorizable, IsSameType>::run(m_impl, index); Vectorizable, IsSameType>::run(m_impl, index);
} }