mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-12-21 07:19:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

Optimize TensorConversion evaluator: do not convert same type

Browse Source
This commit is contained in:
Eugene Zhulenev 2019-02-08 15:13:24 -08:00
parent 953ca5ba2f
commit 1e36166ed1
1 changed files with 69 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

99
unsupported/Eigen/CXX11/src/Tensor/TensorConversion.h
View File

@ -32,7 +32,7 @@ struct traits<TensorConversionOp<TargetType, XprType> >
static const int NumDimensions = traits<XprType>::NumDimensions;
static const int Layout = traits<XprType>::Layout;
enum { Flags = 0 };
typedef typename TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
typedef typename TypeConversion<Scalar, typename traits<XprType>::PointerType>::type PointerType;
};
template<typename TargetType, typename XprType>
@ -177,6 +177,68 @@ template <typename Eval, typename Scalar> struct ConversionSubExprEval<true, Eva
}
};
namespace internal {
template <typename SrcType, typename TargetType, bool IsSameT>
struct CoeffConv {
template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
internal::scalar_cast_op<SrcType, TargetType> converter;
return converter(impl.coeff(index));
}
};
template <typename SrcType, typename TargetType>
struct CoeffConv<SrcType, TargetType, true> {
template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
return impl.coeff(index);
}
};
template <typename SrcPacket, typename TargetPacket, int LoadMode, bool ActuallyVectorize, bool IsSameT>
struct PacketConv {
typedef typename internal::unpacket_traits<SrcPacket>::type SrcType;
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) {
internal::scalar_cast_op<SrcType, TargetType> converter;
EIGEN_ALIGN_MAX typename internal::remove_const<TargetType>::type values[PacketSize];
for (int i = 0; i < PacketSize; ++i) {
values[i] = converter(impl.coeff(index+i));
}
TargetPacket rslt = internal::pload<TargetPacket>(values);
return rslt;
}
};
template <typename SrcPacket, typename TargetPacket, int LoadMode, bool IsSameT>
struct PacketConv<SrcPacket, TargetPacket, LoadMode, true, IsSameT> {
typedef typename internal::unpacket_traits<SrcPacket>::type SrcType;
typedef typename internal::unpacket_traits<TargetPacket>::type TargetType;
template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
const int SrcCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::SrcCoeffRatio;
const int TgtCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::TgtCoeffRatio;
PacketConverter<TensorEvaluator<ArgType, Device>, SrcPacket, TargetPacket,
SrcCoeffRatio, TgtCoeffRatio> converter(impl);
return converter.template packet<LoadMode>(index);
}
};
template <typename SrcPacket, typename TargetPacket, int LoadMode, bool ActuallyVectorize>
struct PacketConv<SrcPacket, TargetPacket, LoadMode, ActuallyVectorize, true> {
template <typename ArgType, typename Device>
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TargetPacket run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
return impl.template packet<LoadMode>(index);
}
};
} // namespace internal
// Eval as rvalue
template<typename TargetType, typename ArgType, typename Device>
@ -191,6 +253,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
typedef typename PacketType<SrcType, Device>::type PacketSourceType;
static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
static const bool IsSameType = internal::is_same<TargetType, SrcType>::value;
enum {
IsAligned = false,
@ -210,7 +273,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* data)
{
return ConversionSubExprEval<internal::is_same<TargetType, SrcType>::value, TensorEvaluator<ArgType, Device>, Scalar>::run(m_impl, data);
return ConversionSubExprEval<IsSameType, TensorEvaluator<ArgType, Device>, Scalar>::run(m_impl, data);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup()
@ -220,8 +283,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE CoeffReturnType coeff(Index index) const
{
internal::scalar_cast_op<SrcType, TargetType> converter;
return converter(m_impl.coeff(index));
return internal::CoeffConv<SrcType, TargetType, IsSameType>::run(m_impl,index);
}
template<int LoadMode>
@ -229,7 +291,8 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
{
const bool Vectorizable = TensorEvaluator<ArgType, Device>::PacketAccess &
internal::type_casting_traits<SrcType, TargetType>::VectorizedCast;
return PacketConv<LoadMode, Vectorizable>::run(m_impl, index);
return internal::PacketConv<PacketSourceType, PacketReturnType, LoadMode,
Vectorizable, IsSameType>::run(m_impl, index);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
@ -252,31 +315,7 @@ struct TensorEvaluator<const TensorConversionOp<TargetType, ArgType>, Device>
/// required by sycl in order to extract the sycl accessor
const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
protected:
template <int LoadMode, bool ActuallyVectorize>
struct PacketConv {
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
internal::scalar_cast_op<SrcType, TargetType> converter;
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
for (int i = 0; i < PacketSize; ++i) {
values[i] = converter(impl.coeff(index+i));
}
PacketReturnType rslt = internal::pload<PacketReturnType>(values);
return rslt;
}
};
template <int LoadMode>
struct PacketConv<LoadMode, true> {
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType run(const TensorEvaluator<ArgType, Device>& impl, Index index) {
const int SrcCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::SrcCoeffRatio;
const int TgtCoeffRatio = internal::type_casting_traits<SrcType, TargetType>::TgtCoeffRatio;
PacketConverter<TensorEvaluator<ArgType, Device>, PacketSourceType, PacketReturnType,
SrcCoeffRatio, TgtCoeffRatio> converter(impl);
return converter.template packet<LoadMode>(index);
}
};
protected:
TensorEvaluator<ArgType, Device> m_impl;
};