Update Eigen to commit:a30ecb7221a46824b85cad5f9016efe6e5871d69
CHANGELOG
=========
a30ecb722 - Don't use the fast implementation if EIGEN_GPU_CC, since integer_packet is not defined for float4 used by the GPU compiler (even on host).
5a0a165c0 - fix broken asserts
0b5873893 - Fix two corner cases in the new implementation of logistic sigmoid.
5d7ffe2ca - drop COPYING.GPL
a32e6a404 - Explicit type casting
8d81a2339 - Reduce usage of reserved names
c61b3cb0d - Fix IterativeSolverBase referring to itself as ConjugateGradient
80ccacc71 - Fix accuracy of logistic sigmoid
8b8125c57 - Make sure the scalar and vectorized path for array.exp() return consistent values.
c9df98b07 - Fix Gcc8.5 warning about missing base class initialisation (#2404)
47eac2107 - Make fixed-size Matrix and Array trivially copyable after C++20
c4b1dd2f6 - Add support for Cray, Fujitsu, and Intel ICX compilers
96dc37a03 - Some fixes/cleanups for numeric_limits & fix for related bug in psqrt
ed27d988c - Fixes #i2411
7b5a8b6bc - Improve plog: 20% speedup for float + handle denormals
a491c7f89 - Allow specifying inner & outer stride for CWiseUnaryView - fixes #2398
27a78e4f9 - Some serialization API changes were made in commit...
9210e71fb - ensure that eigen::internal::size is not found by ADL, rename to ssize and...
7244a74ab - Add bounds checking to Eigen serializer
ba91839d7 - Remove user survey from Doxygen header
a4098ac67 - Fix duplicate include guard *ALTIVEC_H -> *ZVECTOR_H
22a347b9d - Remove unused EIGEN_HAS_STATIC_ARRAY_TEMPLATE
d705eb5f8 - Revert "Select AVX2 even if the data size is not a multiple of 8"
8eab7b688 - Improve exp<float>(): Don't flush denormal results +4% speedup.
6e95c0cd9 - Add missing internal namespace
d3675b2e7 - Add vectorization_logic_1 test to list of CI smoketests
c06c3e52a - Include immintrin.h if F16C is available and vectorization is disabled
f7a056bf0 - Small fixes
2a6594de2 - Small cleanup of GDB pretty printer code
dee6428a7 - fixed clang warnings about alignment change and floating point precision
d0b4b75fb - Simplify logical_xor()
e93a07177 - Fix a bug introduced in !751.
PiperOrigin-RevId: 423164395
Change-Id: I8c7351eccfca04dccb07ca7ec6cfb0cbb3cd5a83
diff --git a/Eigen/src/Core/Array.h b/Eigen/src/Core/Array.h
index b652852..1242d19 100644
--- a/Eigen/src/Core/Array.h
+++ b/Eigen/src/Core/Array.h
@@ -21,6 +21,10 @@
typedef ArrayXpr XprKind;
typedef ArrayBase<Array<Scalar_, Rows_, Cols_, Options_, MaxRows_, MaxCols_> > XprBase;
};
+
+template<typename Scalar_, int Rows_, int Cols_, int Options_, int MaxRows_, int MaxCols_>
+struct has_trivially_copyable_storage<Array<Scalar_, Rows_, Cols_, Options_, MaxRows_, MaxCols_> >
+ : has_trivially_copyable_storage<Matrix<Scalar_, Rows_, Cols_, Options_, MaxRows_, MaxCols_> > {};
}
/** \class Array
@@ -120,6 +124,12 @@
return Base::_set(other);
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE Array& operator=(
+ const Array& other) requires internal::has_trivially_copyable_storage<Array>::value = default;
+#endif
+
/** Default constructor.
*
* For fixed-size matrices, does nothing.
@@ -159,6 +169,13 @@
return *this;
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC Array(Array&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Array>::value = default;
+ EIGEN_DEVICE_FUNC Array& operator=(Array&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Array>::value = default;
+#endif
+
/** \copydoc PlainObjectBase(const Scalar& a0, const Scalar& a1, const Scalar& a2, const Scalar& a3, const ArgTypes&... args)
*
* Example: \include Array_variadic_ctor_cxx11.cpp
@@ -266,6 +283,12 @@
: Base(other)
{ }
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE Array(const Array& other) requires internal::has_trivially_copyable_storage<Array>::value =
+ default;
+#endif
+
private:
struct PrivateType {};
public:
diff --git a/Eigen/src/Core/ArrayBase.h b/Eigen/src/Core/ArrayBase.h
index 28397e5..2d6fe42 100644
--- a/Eigen/src/Core/ArrayBase.h
+++ b/Eigen/src/Core/ArrayBase.h
@@ -112,6 +112,11 @@
return derived();
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayBase& operator=(
+ const ArrayBase& other) requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
/** Set all the entries to \a value.
* \sa DenseBase::setConstant(), DenseBase::fill() */
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
diff --git a/Eigen/src/Core/BandMatrix.h b/Eigen/src/Core/BandMatrix.h
index a8d8b19..913a967 100644
--- a/Eigen/src/Core/BandMatrix.h
+++ b/Eigen/src/Core/BandMatrix.h
@@ -235,17 +235,17 @@
internal::variable_if_dynamic<Index, Subs> m_subs;
};
-template<typename _CoefficientsType,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
+template<typename CoefficientsType_,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
class BandMatrixWrapper;
-template<typename _CoefficientsType,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
-struct traits<BandMatrixWrapper<_CoefficientsType,Rows_,Cols_,Supers_,Subs_,Options_> >
+template<typename CoefficientsType_,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
+struct traits<BandMatrixWrapper<CoefficientsType_,Rows_,Cols_,Supers_,Subs_,Options_> >
{
- typedef typename _CoefficientsType::Scalar Scalar;
- typedef typename _CoefficientsType::StorageKind StorageKind;
- typedef typename _CoefficientsType::StorageIndex StorageIndex;
+ typedef typename CoefficientsType_::Scalar Scalar;
+ typedef typename CoefficientsType_::StorageKind StorageKind;
+ typedef typename CoefficientsType_::StorageIndex StorageIndex;
enum {
- CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost,
+ CoeffReadCost = internal::traits<CoefficientsType_>::CoeffReadCost,
RowsAtCompileTime = Rows_,
ColsAtCompileTime = Cols_,
MaxRowsAtCompileTime = Rows_,
@@ -256,11 +256,11 @@
Options = Options_,
DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) ? 1 + Supers + Subs : Dynamic
};
- typedef _CoefficientsType CoefficientsType;
+ typedef CoefficientsType_ CoefficientsType;
};
-template<typename _CoefficientsType,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
-class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsType,Rows_,Cols_,Supers_,Subs_,Options_> >
+template<typename CoefficientsType_,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
+class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<CoefficientsType_,Rows_,Cols_,Supers_,Subs_,Options_> >
{
public:
@@ -339,9 +339,9 @@
typedef BandShape Shape;
};
-template<typename _CoefficientsType,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
-struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,Rows_,Cols_,Supers_,Subs_,Options_> >
- : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,Rows_,Cols_,Supers_,Subs_,Options_> >
+template<typename CoefficientsType_,int Rows_, int Cols_, int Supers_, int Subs_,int Options_>
+struct evaluator_traits<BandMatrixWrapper<CoefficientsType_,Rows_,Cols_,Supers_,Subs_,Options_> >
+ : public evaluator_traits_base<BandMatrixWrapper<CoefficientsType_,Rows_,Cols_,Supers_,Subs_,Options_> >
{
typedef BandShape Shape;
};
diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h
index 585323c..afbea86 100644
--- a/Eigen/src/Core/Block.h
+++ b/Eigen/src/Core/Block.h
@@ -23,7 +23,7 @@
typedef typename traits<XprType>::StorageKind StorageKind;
typedef typename traits<XprType>::XprKind XprKind;
typedef typename ref_selector<XprType>::type XprTypeNested;
- typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+ typedef typename remove_reference<XprTypeNested>::type XprTypeNested_;
enum{
MatrixRows = traits<XprType>::RowsAtCompileTime,
MatrixCols = traits<XprType>::ColsAtCompileTime,
diff --git a/Eigen/src/Core/CoreEvaluators.h b/Eigen/src/Core/CoreEvaluators.h
index 4a20312..ac8c8e5 100644
--- a/Eigen/src/Core/CoreEvaluators.h
+++ b/Eigen/src/Core/CoreEvaluators.h
@@ -812,11 +812,11 @@
// -------------------- CwiseUnaryView --------------------
-template<typename UnaryOp, typename ArgType>
-struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType>, IndexBased>
- : evaluator_base<CwiseUnaryView<UnaryOp, ArgType> >
+template<typename UnaryOp, typename ArgType, typename StrideType>
+struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType, StrideType>, IndexBased>
+ : evaluator_base<CwiseUnaryView<UnaryOp, ArgType, StrideType> >
{
- typedef CwiseUnaryView<UnaryOp, ArgType> XprType;
+ typedef CwiseUnaryView<UnaryOp, ArgType, StrideType> XprType;
enum {
CoeffReadCost = int(evaluator<ArgType>::CoeffReadCost) + int(functor_traits<UnaryOp>::Cost),
@@ -902,7 +902,8 @@
m_innerStride(map.innerStride()),
m_outerStride(map.outerStride())
{
- EIGEN_STATIC_ASSERT(check_implication(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1),
+ EIGEN_STATIC_ASSERT(check_implication((evaluator<Derived>::Flags & PacketAccessBit) != 0,
+ internal::inner_stride_at_compile_time<Derived>::ret == 1),
PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
}
diff --git a/Eigen/src/Core/CwiseBinaryOp.h b/Eigen/src/Core/CwiseBinaryOp.h
index a3828d1..ea491c6 100644
--- a/Eigen/src/Core/CwiseBinaryOp.h
+++ b/Eigen/src/Core/CwiseBinaryOp.h
@@ -45,10 +45,10 @@
typename traits<Rhs>::StorageIndex>::type StorageIndex;
typedef typename Lhs::Nested LhsNested;
typedef typename Rhs::Nested RhsNested;
- typedef typename remove_reference<LhsNested>::type _LhsNested;
- typedef typename remove_reference<RhsNested>::type _RhsNested;
+ typedef typename remove_reference<LhsNested>::type LhsNested_;
+ typedef typename remove_reference<RhsNested>::type RhsNested_;
enum {
- Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,_LhsNested::Flags & RowMajorBit,_RhsNested::Flags & RowMajorBit>::value
+ Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,LhsNested_::Flags & RowMajorBit,RhsNested_::Flags & RowMajorBit>::value
};
};
} // end namespace internal
@@ -102,8 +102,8 @@
typedef typename internal::ref_selector<LhsType>::type LhsNested;
typedef typename internal::ref_selector<RhsType>::type RhsNested;
- typedef typename internal::remove_reference<LhsNested>::type _LhsNested;
- typedef typename internal::remove_reference<RhsNested>::type _RhsNested;
+ typedef typename internal::remove_reference<LhsNested>::type LhsNested_;
+ typedef typename internal::remove_reference<RhsNested>::type RhsNested_;
#if EIGEN_COMP_MSVC
//Required for Visual Studio or the Copy constructor will probably not get inlined!
@@ -131,10 +131,10 @@
/** \returns the left hand side nested expression */
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- const _LhsNested& lhs() const { return m_lhs; }
+ const LhsNested_& lhs() const { return m_lhs; }
/** \returns the right hand side nested expression */
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- const _RhsNested& rhs() const { return m_rhs; }
+ const RhsNested_& rhs() const { return m_rhs; }
/** \returns the functor representing the binary operation */
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
const BinaryOp& functor() const { return m_functor; }
diff --git a/Eigen/src/Core/CwiseTernaryOp.h b/Eigen/src/Core/CwiseTernaryOp.h
index 52a0ae7..393279bc 100644
--- a/Eigen/src/Core/CwiseTernaryOp.h
+++ b/Eigen/src/Core/CwiseTernaryOp.h
@@ -43,10 +43,10 @@
typedef typename Arg1::Nested Arg1Nested;
typedef typename Arg2::Nested Arg2Nested;
typedef typename Arg3::Nested Arg3Nested;
- typedef typename remove_reference<Arg1Nested>::type _Arg1Nested;
- typedef typename remove_reference<Arg2Nested>::type _Arg2Nested;
- typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
- enum { Flags = _Arg1Nested::Flags & RowMajorBit };
+ typedef typename remove_reference<Arg1Nested>::type Arg1Nested_;
+ typedef typename remove_reference<Arg2Nested>::type Arg2Nested_;
+ typedef typename remove_reference<Arg3Nested>::type Arg3Nested_;
+ enum { Flags = Arg1Nested_::Flags & RowMajorBit };
};
} // end namespace internal
@@ -115,9 +115,9 @@
typedef typename internal::ref_selector<Arg1Type>::type Arg1Nested;
typedef typename internal::ref_selector<Arg2Type>::type Arg2Nested;
typedef typename internal::ref_selector<Arg3Type>::type Arg3Nested;
- typedef typename internal::remove_reference<Arg1Nested>::type _Arg1Nested;
- typedef typename internal::remove_reference<Arg2Nested>::type _Arg2Nested;
- typedef typename internal::remove_reference<Arg3Nested>::type _Arg3Nested;
+ typedef typename internal::remove_reference<Arg1Nested>::type Arg1Nested_;
+ typedef typename internal::remove_reference<Arg2Nested>::type Arg2Nested_;
+ typedef typename internal::remove_reference<Arg3Nested>::type Arg3Nested_;
EIGEN_DEVICE_FUNC
EIGEN_STRONG_INLINE CwiseTernaryOp(const Arg1& a1, const Arg2& a2,
@@ -165,13 +165,13 @@
/** \returns the first argument nested expression */
EIGEN_DEVICE_FUNC
- const _Arg1Nested& arg1() const { return m_arg1; }
+ const Arg1Nested_& arg1() const { return m_arg1; }
/** \returns the first argument nested expression */
EIGEN_DEVICE_FUNC
- const _Arg2Nested& arg2() const { return m_arg2; }
+ const Arg2Nested_& arg2() const { return m_arg2; }
/** \returns the third argument nested expression */
EIGEN_DEVICE_FUNC
- const _Arg3Nested& arg3() const { return m_arg3; }
+ const Arg3Nested_& arg3() const { return m_arg3; }
/** \returns the functor representing the ternary operation */
EIGEN_DEVICE_FUNC
const TernaryOp& functor() const { return m_functor; }
diff --git a/Eigen/src/Core/CwiseUnaryOp.h b/Eigen/src/Core/CwiseUnaryOp.h
index f2c6518..d9985c0 100644
--- a/Eigen/src/Core/CwiseUnaryOp.h
+++ b/Eigen/src/Core/CwiseUnaryOp.h
@@ -24,9 +24,9 @@
UnaryOp(const typename XprType::Scalar&)
>::type Scalar;
typedef typename XprType::Nested XprTypeNested;
- typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+ typedef typename remove_reference<XprTypeNested>::type XprTypeNested_;
enum {
- Flags = _XprTypeNested::Flags & RowMajorBit
+ Flags = XprTypeNested_::Flags & RowMajorBit
};
};
}
diff --git a/Eigen/src/Core/CwiseUnaryView.h b/Eigen/src/Core/CwiseUnaryView.h
index 9fc1dcd..fabb3f8 100644
--- a/Eigen/src/Core/CwiseUnaryView.h
+++ b/Eigen/src/Core/CwiseUnaryView.h
@@ -15,32 +15,37 @@
namespace Eigen {
namespace internal {
-template<typename ViewOp, typename MatrixType>
-struct traits<CwiseUnaryView<ViewOp, MatrixType> >
+template<typename ViewOp, typename MatrixType, typename StrideType>
+struct traits<CwiseUnaryView<ViewOp, MatrixType, StrideType> >
: traits<MatrixType>
{
typedef typename result_of<
ViewOp(const typename traits<MatrixType>::Scalar&)
>::type Scalar;
typedef typename MatrixType::Nested MatrixTypeNested;
- typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
+ typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNested_;
enum {
FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
- Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
+ Flags = traits<MatrixTypeNested_>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
MatrixTypeInnerStride = inner_stride_at_compile_time<MatrixType>::ret,
// need to cast the sizeof's from size_t to int explicitly, otherwise:
// "error: no integral type can represent all of the enumerator values
- InnerStrideAtCompileTime = MatrixTypeInnerStride == Dynamic
- ? int(Dynamic)
- : int(MatrixTypeInnerStride) * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)),
- OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret == Dynamic
- ? int(Dynamic)
- : outer_stride_at_compile_time<MatrixType>::ret * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar))
+ InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
+ ? (MatrixTypeInnerStride == Dynamic
+ ? int(Dynamic)
+ : int(MatrixTypeInnerStride) * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)))
+ : int(StrideType::InnerStrideAtCompileTime),
+
+ OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
+ ? (outer_stride_at_compile_time<MatrixType>::ret == Dynamic
+ ? int(Dynamic)
+ : outer_stride_at_compile_time<MatrixType>::ret * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)))
+ : int(StrideType::OuterStrideAtCompileTime)
};
};
}
-template<typename ViewOp, typename MatrixType, typename StorageKind>
+template<typename ViewOp, typename MatrixType, typename StrideType, typename StorageKind>
class CwiseUnaryViewImpl;
/** \class CwiseUnaryView
@@ -56,12 +61,12 @@
*
* \sa MatrixBase::unaryViewExpr(const CustomUnaryOp &) const, class CwiseUnaryOp
*/
-template<typename ViewOp, typename MatrixType>
-class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename internal::traits<MatrixType>::StorageKind>
+template<typename ViewOp, typename MatrixType, typename StrideType>
+class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, StrideType, typename internal::traits<MatrixType>::StorageKind>
{
public:
- typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
+ typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType, StrideType, typename internal::traits<MatrixType>::StorageKind>::Base Base;
EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView)
typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
typedef typename internal::remove_all<MatrixType>::type NestedExpression;
@@ -93,22 +98,22 @@
};
// Generic API dispatcher
-template<typename ViewOp, typename XprType, typename StorageKind>
+template<typename ViewOp, typename XprType, typename StrideType, typename StorageKind>
class CwiseUnaryViewImpl
- : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type
+ : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType, StrideType> >::type
{
public:
- typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base;
+ typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType, StrideType> >::type Base;
};
-template<typename ViewOp, typename MatrixType>
-class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
- : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type
+template<typename ViewOp, typename MatrixType, typename StrideType>
+class CwiseUnaryViewImpl<ViewOp,MatrixType,StrideType,Dense>
+ : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType, StrideType> >::type
{
public:
- typedef CwiseUnaryView<ViewOp, MatrixType> Derived;
- typedef typename internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type Base;
+ typedef CwiseUnaryView<ViewOp, MatrixType,StrideType> Derived;
+ typedef typename internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType,StrideType> >::type Base;
EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl)
@@ -118,12 +123,16 @@
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR inline Index innerStride() const
{
- return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
+ return StrideType::InnerStrideAtCompileTime != 0
+ ? int(StrideType::InnerStrideAtCompileTime)
+ : derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
}
EIGEN_DEVICE_FUNC EIGEN_CONSTEXPR inline Index outerStride() const
{
- return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
+ return StrideType::OuterStrideAtCompileTime != 0
+ ? int(StrideType::OuterStrideAtCompileTime)
+ : derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
}
protected:
EIGEN_DEFAULT_EMPTY_CONSTRUCTOR_AND_DESTRUCTOR(CwiseUnaryViewImpl)
diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index d62c851..6d5a100 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -275,6 +275,11 @@
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
Derived& operator=(const DenseBase& other);
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE DenseBase& operator=(
+ const DenseBase& other) requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
template<typename OtherDerived>
EIGEN_DEVICE_FUNC
Derived& operator=(const EigenBase<OtherDerived> &other);
diff --git a/Eigen/src/Core/Diagonal.h b/Eigen/src/Core/Diagonal.h
index 6d8df3d..2bf4527 100644
--- a/Eigen/src/Core/Diagonal.h
+++ b/Eigen/src/Core/Diagonal.h
@@ -20,8 +20,8 @@
*
* \brief Expression of a diagonal/subdiagonal/superdiagonal in a matrix
*
- * \param MatrixType the type of the object in which we are taking a sub/main/super diagonal
- * \param DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal.
+ * \tparam MatrixType the type of the object in which we are taking a sub/main/super diagonal
+ * \tparam DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal.
* A positive value means a superdiagonal, a negative value means a subdiagonal.
* You can also use DynamicIndex so the index can be set at runtime.
*
@@ -40,7 +40,7 @@
: traits<MatrixType>
{
typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
- typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+ typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNested_;
typedef typename MatrixType::StorageKind StorageKind;
enum {
RowsAtCompileTime = (int(DiagIndex) == DynamicIndex || int(MatrixType::SizeAtCompileTime) == Dynamic) ? Dynamic
@@ -54,7 +54,7 @@
MatrixType::MaxColsAtCompileTime - plain_enum_max( DiagIndex, 0))),
MaxColsAtCompileTime = 1,
MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
- Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
+ Flags = (unsigned int)MatrixTypeNested_::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret,
InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1,
OuterStrideAtCompileTime = 0
diff --git a/Eigen/src/Core/DiagonalMatrix.h b/Eigen/src/Core/DiagonalMatrix.h
index d5053d6..453f69b 100644
--- a/Eigen/src/Core/DiagonalMatrix.h
+++ b/Eigen/src/Core/DiagonalMatrix.h
@@ -118,9 +118,9 @@
*
* \brief Represents a diagonal matrix with its storage
*
- * \param Scalar_ the type of coefficients
- * \param SizeAtCompileTime the dimension of the matrix, or Dynamic
- * \param MaxSizeAtCompileTime the dimension of the matrix, or Dynamic. This parameter is optional and defaults
+ * \tparam Scalar_ the type of coefficients
+ * \tparam SizeAtCompileTime the dimension of the matrix, or Dynamic
+ * \tparam MaxSizeAtCompileTime the dimension of the matrix, or Dynamic. This parameter is optional and defaults
* to SizeAtCompileTime. Most of the time, you do not need to specify it.
*
* \sa class DiagonalWrapper
@@ -261,7 +261,7 @@
*
* \brief Expression of a diagonal matrix
*
- * \param DiagonalVectorType_ the type of the vector of diagonal coefficients
+ * \tparam DiagonalVectorType_ the type of the vector of diagonal coefficients
*
* This class is an expression of a diagonal matrix, but not storing its own vector of diagonal coefficients,
* instead wrapping an existing vector expression. It is the return type of MatrixBase::asDiagonal()
diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h
index 387b6ce..a7c20f5 100644
--- a/Eigen/src/Core/Dot.h
+++ b/Eigen/src/Core/Dot.h
@@ -125,8 +125,8 @@
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::PlainObject
MatrixBase<Derived>::normalized() const
{
- typedef typename internal::nested_eval<Derived,2>::type _Nested;
- _Nested n(derived());
+ typedef typename internal::nested_eval<Derived,2>::type Nested_;
+ Nested_ n(derived());
RealScalar z = n.squaredNorm();
// NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
if(z>RealScalar(0))
@@ -168,8 +168,8 @@
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::PlainObject
MatrixBase<Derived>::stableNormalized() const
{
- typedef typename internal::nested_eval<Derived,3>::type _Nested;
- _Nested n(derived());
+ typedef typename internal::nested_eval<Derived,3>::type Nested_;
+ Nested_ n(derived());
RealScalar w = n.cwiseAbs().maxCoeff();
RealScalar z = (n/w).squaredNorm();
if(z>RealScalar(0))
diff --git a/Eigen/src/Core/GeneralProduct.h b/Eigen/src/Core/GeneralProduct.h
index 33b667d..0d24a00 100644
--- a/Eigen/src/Core/GeneralProduct.h
+++ b/Eigen/src/Core/GeneralProduct.h
@@ -52,17 +52,17 @@
template<typename Lhs, typename Rhs> struct product_type
{
- typedef typename remove_all<Lhs>::type _Lhs;
- typedef typename remove_all<Rhs>::type _Rhs;
+ typedef typename remove_all<Lhs>::type Lhs_;
+ typedef typename remove_all<Rhs>::type Rhs_;
enum {
- MaxRows = traits<_Lhs>::MaxRowsAtCompileTime,
- Rows = traits<_Lhs>::RowsAtCompileTime,
- MaxCols = traits<_Rhs>::MaxColsAtCompileTime,
- Cols = traits<_Rhs>::ColsAtCompileTime,
- MaxDepth = min_size_prefer_fixed(traits<_Lhs>::MaxColsAtCompileTime,
- traits<_Rhs>::MaxRowsAtCompileTime),
- Depth = min_size_prefer_fixed(traits<_Lhs>::ColsAtCompileTime,
- traits<_Rhs>::RowsAtCompileTime)
+ MaxRows = traits<Lhs_>::MaxRowsAtCompileTime,
+ Rows = traits<Lhs_>::RowsAtCompileTime,
+ MaxCols = traits<Rhs_>::MaxColsAtCompileTime,
+ Cols = traits<Rhs_>::ColsAtCompileTime,
+ MaxDepth = min_size_prefer_fixed(traits<Lhs_>::MaxColsAtCompileTime,
+ traits<Rhs_>::MaxRowsAtCompileTime),
+ Depth = min_size_prefer_fixed(traits<Lhs_>::ColsAtCompileTime,
+ traits<Rhs_>::RowsAtCompileTime)
};
// the splitting into different lines of code here, introducing the _select enums and the typedef below,
diff --git a/Eigen/src/Core/IndexedView.h b/Eigen/src/Core/IndexedView.h
index d27ec80..e7ca88b 100644
--- a/Eigen/src/Core/IndexedView.h
+++ b/Eigen/src/Core/IndexedView.h
@@ -124,10 +124,10 @@
{}
/** \returns number of rows */
- Index rows() const { return internal::size(m_rowIndices); }
+ Index rows() const { return internal::index_list_size(m_rowIndices); }
/** \returns number of columns */
- Index cols() const { return internal::size(m_colIndices); }
+ Index cols() const { return internal::index_list_size(m_colIndices); }
/** \returns the nested expression */
const typename internal::remove_all<XprType>::type&
diff --git a/Eigen/src/Core/Matrix.h b/Eigen/src/Core/Matrix.h
index 44c0e25..77051a0 100644
--- a/Eigen/src/Core/Matrix.h
+++ b/Eigen/src/Core/Matrix.h
@@ -52,6 +52,14 @@
Alignment = actual_alignment
};
};
+
+template<typename Scalar_, int Rows_, int Cols_, int Options_, int MaxRows_, int MaxCols_>
+struct has_trivially_copyable_storage<Matrix<Scalar_, Rows_, Cols_, Options_, MaxRows_, MaxCols_> >
+{
+ // Must be identical to the type of PlainObjectBase::m_storage.
+ typedef DenseStorage<Scalar_, internal::size_at_compile_time<MaxRows_, MaxCols_>::ret, Rows_, Cols_, Options_> Storage;
+ static const bool value = std::is_trivially_copyable<Storage>::value;
+};
}
/** \class Matrix
@@ -210,6 +218,12 @@
return Base::_set(other);
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE Matrix& operator=(
+ const Matrix& other) requires internal::has_trivially_copyable_storage<Matrix>::value = default;
+#endif
+
/** \internal
* \brief Copies the value of the expression \a other into \c *this with automatic resizing.
*
@@ -279,6 +293,13 @@
return *this;
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix(Matrix&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Matrix>::value = default;
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Matrix>::value = default;
+#endif
+
/** \copydoc PlainObjectBase(const Scalar&, const Scalar&, const Scalar&, const Scalar&, const ArgTypes&... args)
*
* Example: \include Matrix_variadic_ctor_cxx11.cpp
@@ -404,6 +425,12 @@
EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other)
{ }
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE Matrix(const Matrix& other) requires internal::has_trivially_copyable_storage<Matrix>::value =
+ default;
+#endif
+
/** \brief Copy constructor for generic expressions.
* \sa MatrixBase::operator=(const EigenBase<OtherDerived>&)
*/
diff --git a/Eigen/src/Core/MatrixBase.h b/Eigen/src/Core/MatrixBase.h
index 70d0cf7..89a4110 100644
--- a/Eigen/src/Core/MatrixBase.h
+++ b/Eigen/src/Core/MatrixBase.h
@@ -140,6 +140,11 @@
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
Derived& operator=(const MatrixBase& other);
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE MatrixBase& operator=(
+ const MatrixBase& other) requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
// We cannot inherit here via Base::operator= since it is causing
// trouble with MSVC.
diff --git a/Eigen/src/Core/PartialReduxEvaluator.h b/Eigen/src/Core/PartialReduxEvaluator.h
index a179688..b6f31f9 100644
--- a/Eigen/src/Core/PartialReduxEvaluator.h
+++ b/Eigen/src/Core/PartialReduxEvaluator.h
@@ -154,16 +154,16 @@
: TraversalSize==0 ? 1
: int(TraversalSize) * int(evaluator<ArgType>::CoeffReadCost) + int(CostOpType::value),
- _ArgFlags = evaluator<ArgType>::Flags,
+ ArgFlags_ = evaluator<ArgType>::Flags,
- _Vectorizable = bool(int(_ArgFlags)&PacketAccessBit)
+ Vectorizable_ = bool(int(ArgFlags_)&PacketAccessBit)
&& bool(MemberOp::Vectorizable)
- && (Direction==int(Vertical) ? bool(_ArgFlags&RowMajorBit) : (_ArgFlags&RowMajorBit)==0)
+ && (Direction==int(Vertical) ? bool(ArgFlags_&RowMajorBit) : (ArgFlags_&RowMajorBit)==0)
&& (TraversalSize!=0),
Flags = (traits<XprType>::Flags&RowMajorBit)
| (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit)))
- | (_Vectorizable ? PacketAccessBit : 0)
+ | (Vectorizable_ ? PacketAccessBit : 0)
| LinearAccessBit,
Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized
diff --git a/Eigen/src/Core/PermutationMatrix.h b/Eigen/src/Core/PermutationMatrix.h
index a6910e2..1b4195a 100644
--- a/Eigen/src/Core/PermutationMatrix.h
+++ b/Eigen/src/Core/PermutationMatrix.h
@@ -391,20 +391,20 @@
namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int _PacketAccess>
-struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int PacketAccess_>
+struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,PacketAccess_> >
: traits<Matrix<StorageIndex_,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
{
typedef PermutationStorage StorageKind;
- typedef Map<const Matrix<StorageIndex_, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+ typedef Map<const Matrix<StorageIndex_, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, PacketAccess_> IndicesType;
typedef StorageIndex_ StorageIndex;
typedef void Scalar;
};
}
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int _PacketAccess>
-class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,_PacketAccess>
- : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int PacketAccess_>
+class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,PacketAccess_>
+ : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, StorageIndex_>,PacketAccess_> >
{
typedef PermutationBase<Map> Base;
typedef internal::traits<Map> Traits;
diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index 4367ea5..33ae507 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -463,6 +463,12 @@
return _set(other);
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE PlainObjectBase& operator=(
+ const PlainObjectBase& other) requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
/** \sa MatrixBase::lazyAssign() */
template<typename OtherDerived>
EIGEN_DEVICE_FUNC
@@ -514,10 +520,27 @@
return *this;
}
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ PlainObjectBase(PlainObjectBase&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Derived>::value = default;
+
+ EIGEN_DEVICE_FUNC
+ PlainObjectBase& operator=(PlainObjectBase&& other) EIGEN_NOEXCEPT
+ requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
/** Copy constructor */
EIGEN_DEVICE_FUNC
EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
: Base(), m_storage(other.m_storage) { }
+
+#if EIGEN_COMP_HAS_P0848R3
+ EIGEN_DEVICE_FUNC
+ EIGEN_STRONG_INLINE PlainObjectBase(
+ const PlainObjectBase& other) requires internal::has_trivially_copyable_storage<Derived>::value = default;
+#endif
+
EIGEN_DEVICE_FUNC
EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols)
: m_storage(size, rows, cols)
diff --git a/Eigen/src/Core/ProductEvaluators.h b/Eigen/src/Core/ProductEvaluators.h
index aff3572..5f9ffe8 100644
--- a/Eigen/src/Core/ProductEvaluators.h
+++ b/Eigen/src/Core/ProductEvaluators.h
@@ -841,19 +841,19 @@
StorageOrder_ = (Derived::MaxRowsAtCompileTime==1 && Derived::MaxColsAtCompileTime!=1) ? RowMajor
: (Derived::MaxColsAtCompileTime==1 && Derived::MaxRowsAtCompileTime!=1) ? ColMajor
: MatrixFlags & RowMajorBit ? RowMajor : ColMajor,
- _SameStorageOrder = StorageOrder_ == (MatrixFlags & RowMajorBit ? RowMajor : ColMajor),
+ SameStorageOrder_ = StorageOrder_ == (MatrixFlags & RowMajorBit ? RowMajor : ColMajor),
- _ScalarAccessOnDiag = !((int(StorageOrder_) == ColMajor && int(ProductOrder) == OnTheLeft)
+ ScalarAccessOnDiag_ = !((int(StorageOrder_) == ColMajor && int(ProductOrder) == OnTheLeft)
||(int(StorageOrder_) == RowMajor && int(ProductOrder) == OnTheRight)),
- _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
+ SameTypes_ = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
// FIXME currently we need same types, but in the future the next rule should be the one
- //_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))),
- _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit)
- && _SameTypes
- && (_SameStorageOrder || (MatrixFlags&LinearAccessBit)==LinearAccessBit)
- && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))),
- _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
- Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0),
+ //Vectorizable_ = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (SameTypes_ && bool(int(DiagFlags)&PacketAccessBit))),
+ Vectorizable_ = bool(int(MatrixFlags)&PacketAccessBit)
+ && SameTypes_
+ && (SameStorageOrder_ || (MatrixFlags&LinearAccessBit)==LinearAccessBit)
+ && (ScalarAccessOnDiag_ || (bool(int(DiagFlags)&PacketAccessBit))),
+ LinearAccessMask_ = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
+ Flags = ((HereditaryBits|LinearAccessMask_) & (unsigned int)(MatrixFlags)) | (Vectorizable_ ? PacketAccessBit : 0),
Alignment = evaluator<MatrixType>::Alignment,
AsScalarProduct = (DiagonalType::SizeAtCompileTime==1)
diff --git a/Eigen/src/Core/Ref.h b/Eigen/src/Core/Ref.h
index 3e1d99c..c31dfcc 100644
--- a/Eigen/src/Core/Ref.h
+++ b/Eigen/src/Core/Ref.h
@@ -16,16 +16,16 @@
namespace internal {
-template<typename _PlainObjectType, int Options_, typename _StrideType>
-struct traits<Ref<_PlainObjectType, Options_, _StrideType> >
- : public traits<Map<_PlainObjectType, Options_, _StrideType> >
+template<typename PlainObjectType_, int Options_, typename StrideType_>
+struct traits<Ref<PlainObjectType_, Options_, StrideType_> >
+ : public traits<Map<PlainObjectType_, Options_, StrideType_> >
{
- typedef _PlainObjectType PlainObjectType;
- typedef _StrideType StrideType;
+ typedef PlainObjectType_ PlainObjectType;
+ typedef StrideType_ StrideType;
enum {
Options = Options_,
- Flags = traits<Map<_PlainObjectType, Options_, _StrideType> >::Flags | NestByRefBit,
- Alignment = traits<Map<_PlainObjectType, Options_, _StrideType> >::Alignment
+ Flags = traits<Map<PlainObjectType_, Options_, StrideType_> >::Flags | NestByRefBit,
+ Alignment = traits<Map<PlainObjectType_, Options_, StrideType_> >::Alignment
};
template<typename Derived> struct match {
diff --git a/Eigen/src/Core/Replicate.h b/Eigen/src/Core/Replicate.h
index 0ee9a10..6b5f9fe 100644
--- a/Eigen/src/Core/Replicate.h
+++ b/Eigen/src/Core/Replicate.h
@@ -23,7 +23,7 @@
typedef typename traits<MatrixType>::StorageKind StorageKind;
typedef typename traits<MatrixType>::XprKind XprKind;
typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
- typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+ typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNested_;
enum {
RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic
? Dynamic
@@ -64,7 +64,7 @@
: public internal::dense_xpr_base< Replicate<MatrixType,RowFactor,ColFactor> >::type
{
typedef typename internal::traits<Replicate>::MatrixTypeNested MatrixTypeNested;
- typedef typename internal::traits<Replicate>::_MatrixTypeNested _MatrixTypeNested;
+ typedef typename internal::traits<Replicate>::MatrixTypeNested_ MatrixTypeNested_;
public:
typedef typename internal::dense_xpr_base<Replicate>::type Base;
@@ -96,7 +96,7 @@
inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); }
EIGEN_DEVICE_FUNC
- const _MatrixTypeNested& nestedExpression() const
+ const MatrixTypeNested_& nestedExpression() const
{
return m_matrix;
}
diff --git a/Eigen/src/Core/Reverse.h b/Eigen/src/Core/Reverse.h
index 7ad5023..dbc6929 100644
--- a/Eigen/src/Core/Reverse.h
+++ b/Eigen/src/Core/Reverse.h
@@ -26,13 +26,13 @@
typedef typename traits<MatrixType>::StorageKind StorageKind;
typedef typename traits<MatrixType>::XprKind XprKind;
typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
- typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+ typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNested_;
enum {
RowsAtCompileTime = MatrixType::RowsAtCompileTime,
ColsAtCompileTime = MatrixType::ColsAtCompileTime,
MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
- Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit)
+ Flags = MatrixTypeNested_::Flags & (RowMajorBit | LvalueBit)
};
};
diff --git a/Eigen/src/Core/Select.h b/Eigen/src/Core/Select.h
index f2a2a03..d9ed2b2 100644
--- a/Eigen/src/Core/Select.h
+++ b/Eigen/src/Core/Select.h
@@ -19,9 +19,9 @@
*
* \brief Expression of a coefficient wise version of the C++ ternary operator ?:
*
- * \param ConditionMatrixType the type of the \em condition expression which must be a boolean matrix
- * \param ThenMatrixType the type of the \em then expression
- * \param ElseMatrixType the type of the \em else expression
+ * \tparam ConditionMatrixType the type of the \em condition expression which must be a boolean matrix
+ * \tparam ThenMatrixType the type of the \em then expression
+ * \tparam ElseMatrixType the type of the \em else expression
*
* This class represents an expression of a coefficient wise version of the C++ ternary operator ?:.
* It is the return type of DenseBase::select() and most of the time this is the only way it is used.
diff --git a/Eigen/src/Core/SelfAdjointView.h b/Eigen/src/Core/SelfAdjointView.h
index 6b11247..7096058 100644
--- a/Eigen/src/Core/SelfAdjointView.h
+++ b/Eigen/src/Core/SelfAdjointView.h
@@ -20,8 +20,8 @@
*
* \brief Expression of a selfadjoint matrix from a triangular part of a dense matrix
*
- * \param MatrixType the type of the dense matrix storing the coefficients
- * \param TriangularPart can be either \c #Lower or \c #Upper
+ * \tparam MatrixType the type of the dense matrix storing the coefficients
+ * \tparam TriangularPart can be either \c #Lower or \c #Upper
*
* This class is an expression of a sefladjoint matrix from a triangular part of a matrix
* with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView()
diff --git a/Eigen/src/Core/Transpositions.h b/Eigen/src/Core/Transpositions.h
index 6996ca2..84a9773 100644
--- a/Eigen/src/Core/Transpositions.h
+++ b/Eigen/src/Core/Transpositions.h
@@ -201,11 +201,11 @@
namespace internal {
-template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int _PacketAccess>
-struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,StorageIndex_>,_PacketAccess> >
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename StorageIndex_, int PacketAccess_>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,StorageIndex_>,PacketAccess_> >
: traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,StorageIndex_> >
{
- typedef Map<const Matrix<StorageIndex_,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+ typedef Map<const Matrix<StorageIndex_,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, PacketAccess_> IndicesType;
typedef StorageIndex_ StorageIndex;
typedef TranspositionsStorage StorageKind;
};
diff --git a/Eigen/src/Core/TriangularMatrix.h b/Eigen/src/Core/TriangularMatrix.h
index 9376a79..28f515f 100644
--- a/Eigen/src/Core/TriangularMatrix.h
+++ b/Eigen/src/Core/TriangularMatrix.h
@@ -155,8 +155,8 @@
*
* \brief Expression of a triangular part in a matrix
*
- * \param MatrixType the type of the object in which we are taking the triangular part
- * \param Mode the kind of triangular matrix expression to construct. Can be #Upper,
+ * \tparam MatrixType the type of the object in which we are taking the triangular part
+ * \tparam Mode the kind of triangular matrix expression to construct. Can be #Upper,
* #Lower, #UnitUpper, #UnitLower, #StrictlyUpper, or #StrictlyLower.
* This is in fact a bit field; it must have either #Upper or #Lower,
* and additionally it may have #UnitDiag or #ZeroDiag or neither.
diff --git a/Eigen/src/Core/arch/AVX/MathFunctions.h b/Eigen/src/Core/arch/AVX/MathFunctions.h
index 17b9d0b..d517dff 100644
--- a/Eigen/src/Core/arch/AVX/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX/MathFunctions.h
@@ -101,17 +101,22 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet8f psqrt<Packet8f>(const Packet8f& _x) {
- Packet8f minus_half_x = pmul(_x, pset1<Packet8f>(-0.5f));
- Packet8f denormal_mask = pandnot(
- pcmp_lt(_x, pset1<Packet8f>((std::numeric_limits<float>::min)())),
- pcmp_lt(_x, pzero(_x)));
+ const Packet8f minus_half_x = pmul(_x, pset1<Packet8f>(-0.5f));
+ const Packet8f negative_mask = pcmp_lt(_x, pzero(_x));
+ const Packet8f denormal_mask =
+ pandnot(pcmp_lt(_x, pset1<Packet8f>((std::numeric_limits<float>::min)())),
+ negative_mask);
// Compute approximate reciprocal sqrt.
- Packet8f x = _mm256_rsqrt_ps(_x);
+ Packet8f rs = _mm256_rsqrt_ps(_x);
+ // Flush negative arguments to zero. This is a workaround which ensures
+ // that sqrt of a negative denormal returns -NaN, despite _mm256_rsqrt_ps
+ // returning -Inf for such values.
+ const Packet8f x_flushed = pandnot(_x, negative_mask);
// Do a single step of Newton's iteration.
- x = pmul(x, pmadd(minus_half_x, pmul(x,x), pset1<Packet8f>(1.5f)));
+ rs = pmul(rs, pmadd(minus_half_x, pmul(rs,rs), pset1<Packet8f>(1.5f)));
// Flush results for denormals to zero.
- return pandnot(pmul(_x,x), denormal_mask);
+ return pandnot(pmul(x_flushed, rs), denormal_mask);
}
#else
@@ -131,10 +136,10 @@
#if EIGEN_FAST_MATH
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
- _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
- _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
- _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
- _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
+ EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
+ EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+ EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+ EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
Packet8f neg_half = pmul(_x, p8f_minus_half);
@@ -164,14 +169,14 @@
#else
template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
- _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+ EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(_x));
}
#endif
template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4d prsqrt<Packet4d>(const Packet4d& _x) {
- _EIGEN_DECLARE_CONST_Packet4d(one, 1.0);
+ EIGEN_DECLARE_CONST_Packet4d(one, 1.0);
return _mm256_div_pd(p4d_one, _mm256_sqrt_pd(_x));
}
diff --git a/Eigen/src/Core/arch/AVX/PacketMath.h b/Eigen/src/Core/arch/AVX/PacketMath.h
index fdf1afb..ef73d35 100644
--- a/Eigen/src/Core/arch/AVX/PacketMath.h
+++ b/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -42,16 +42,16 @@
template<> struct is_arithmetic<Packet8h> { enum { value = true }; };
template<> struct is_arithmetic<Packet8bf> { enum { value = true }; };
-#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet8f(NAME,X) \
const Packet8f p8f_##NAME = pset1<Packet8f>(X)
-#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4d(NAME,X) \
const Packet4d p4d_##NAME = pset1<Packet4d>(X)
-#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X))
-#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
const Packet8i p8i_##NAME = pset1<Packet8i>(X)
// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going
diff --git a/Eigen/src/Core/arch/AVX512/MathFunctions.h b/Eigen/src/Core/arch/AVX512/MathFunctions.h
index 26ab168..54be6cf 100644
--- a/Eigen/src/Core/arch/AVX512/MathFunctions.h
+++ b/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -19,22 +19,22 @@
// Disable the code for older versions of gcc that don't support many of the required avx512 instrinsics.
#if EIGEN_GNUC_AT_LEAST(5, 3) || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC >= 1923
-#define _EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
const Packet16f p16f_##NAME = pset1<Packet16f>(X)
-#define _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \
const Packet16f p16f_##NAME = preinterpret<Packet16f,Packet16i>(pset1<Packet16i>(X))
-#define _EIGEN_DECLARE_CONST_Packet8d(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet8d(NAME, X) \
const Packet8d p8d_##NAME = pset1<Packet8d>(X)
-#define _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \
const Packet8d p8d_##NAME = _mm512_castsi512_pd(_mm512_set1_epi64(X))
-#define _EIGEN_DECLARE_CONST_Packet16bf(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet16bf(NAME, X) \
const Packet16bf p16bf_##NAME = pset1<Packet16bf>(X)
-#define _EIGEN_DECLARE_CONST_Packet16bf_FROM_INT(NAME, X) \
+#define EIGEN_DECLARE_CONST_Packet16bf_FROM_INT(NAME, X) \
const Packet16bf p16bf_##NAME = preinterpret<Packet16bf,Packet16i>(pset1<Packet16i>(X))
template <>
@@ -73,21 +73,21 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
pexp<Packet16f>(const Packet16f& _x) {
- _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
- _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
- _EIGEN_DECLARE_CONST_Packet16f(127, 127.0f);
+ EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+ EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+ EIGEN_DECLARE_CONST_Packet16f(127, 127.0f);
- _EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f);
- _EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f);
+ EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f);
+ EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f);
- _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f);
+ EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f);
// Clamp x.
Packet16f x = pmax(pmin(_x, p16f_exp_hi), p16f_exp_lo);
@@ -98,7 +98,7 @@
// Get r = x - m*ln(2). Note that we can do this without losing more than one
// ulp precision due to the FMA instruction.
- _EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f);
+ EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f);
Packet16f r = _mm512_fmadd_ps(m, p16f_nln2, x);
Packet16f r2 = pmul(r, r);
Packet16f r3 = pmul(r2, r);
@@ -227,9 +227,9 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
prsqrt<Packet16f>(const Packet16f& _x) {
- _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000);
- _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
- _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+ EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000);
+ EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+ EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
Packet16f neg_half = pmul(_x, p16f_minus_half);
@@ -257,7 +257,7 @@
template <>
EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
- _EIGEN_DECLARE_CONST_Packet16f(one, 1.0f);
+ EIGEN_DECLARE_CONST_Packet16f(one, 1.0f);
return _mm512_div_ps(p16f_one, _mm512_sqrt_ps(x));
}
#endif
@@ -270,9 +270,9 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
prsqrt<Packet8d>(const Packet8d& _x) {
- _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
- _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
- _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL);
+ EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+ EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+ EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL);
Packet8d neg_half = pmul(_x, p8d_minus_half);
@@ -307,7 +307,7 @@
#else
template <>
EIGEN_STRONG_INLINE Packet8d prsqrt<Packet8d>(const Packet8d& x) {
- _EIGEN_DECLARE_CONST_Packet8d(one, 1.0f);
+ EIGEN_DECLARE_CONST_Packet8d(one, 1.0f);
return _mm512_div_pd(p8d_one, _mm512_sqrt_pd(x));
}
#endif
diff --git a/Eigen/src/Core/arch/AltiVec/PacketMath.h b/Eigen/src/Core/arch/AltiVec/PacketMath.h
index c6b632f..efb0242 100755
--- a/Eigen/src/Core/arch/AltiVec/PacketMath.h
+++ b/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -41,34 +41,34 @@
// We don't want to write the same code all the time, but we need to reuse the constants
// and it doesn't really work to declare them global, so we define macros instead
-#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
Packet4f p4f_##NAME = {X, X, X, X}
-#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
Packet4i p4i_##NAME = vec_splat_s32(X)
-#define _EIGEN_DECLARE_CONST_FAST_Packet4ui(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet4ui(NAME,X) \
Packet4ui p4ui_##NAME = {X, X, X, X}
-#define _EIGEN_DECLARE_CONST_FAST_Packet8us(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet8us(NAME,X) \
Packet8us p8us_##NAME = {X, X, X, X, X, X, X, X}
-#define _EIGEN_DECLARE_CONST_FAST_Packet16uc(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet16uc(NAME,X) \
Packet16uc p16uc_##NAME = {X, X, X, X, X, X, X, X, X, X, X, X, X, X, X, X}
-#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
Packet4f p4f_##NAME = pset1<Packet4f>(X)
-#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
Packet4i p4i_##NAME = pset1<Packet4i>(X)
-#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
Packet2d p2d_##NAME = pset1<Packet2d>(X)
-#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
Packet2l p2l_##NAME = pset1<Packet2l>(X)
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
#define DST_CHAN 1
@@ -76,15 +76,15 @@
#define __UNPACK_TYPE__(PACKETNAME) typename unpacket_traits<PACKETNAME>::type
// These constants are endian-agnostic
-static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
-static _EIGEN_DECLARE_CONST_FAST_Packet4ui(SIGN, 0x80000000u);
-static _EIGEN_DECLARE_CONST_FAST_Packet4ui(PREV0DOT5, 0x3EFFFFFFu);
-static _EIGEN_DECLARE_CONST_FAST_Packet8us(ONE,1); //{ 1, 1, 1, 1, 1, 1, 1, 1}
-static _EIGEN_DECLARE_CONST_FAST_Packet16uc(ONE,1);
+static EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
+static EIGEN_DECLARE_CONST_FAST_Packet4ui(SIGN, 0x80000000u);
+static EIGEN_DECLARE_CONST_FAST_Packet4ui(PREV0DOT5, 0x3EFFFFFFu);
+static EIGEN_DECLARE_CONST_FAST_Packet8us(ONE,1); //{ 1, 1, 1, 1, 1, 1, 1, 1}
+static EIGEN_DECLARE_CONST_FAST_Packet16uc(ONE,1);
static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
#ifndef __VSX__
static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
@@ -1182,7 +1182,7 @@
template<> EIGEN_STRONG_INLINE Packet16c pabs(const Packet16c& a) { return vec_abs(a); }
template<> EIGEN_STRONG_INLINE Packet16uc pabs(const Packet16uc& a) { return a; }
template<> EIGEN_STRONG_INLINE Packet8bf pabs(const Packet8bf& a) {
- _EIGEN_DECLARE_CONST_FAST_Packet8us(abs_mask,0x7FFF);
+ EIGEN_DECLARE_CONST_FAST_Packet8us(abs_mask,0x7FFF);
return pand<Packet8us>(p8us_abs_mask, a);
}
@@ -1194,38 +1194,38 @@
{ return vec_sl(a,reinterpret_cast<Packet4ui>(pset1<Packet4i>(N))); }
template<int N> EIGEN_STRONG_INLINE Packet4f plogical_shift_left(const Packet4f& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
Packet4ui r = vec_sl(reinterpret_cast<Packet4ui>(a), p4ui_mask);
return reinterpret_cast<Packet4f>(r);
}
template<int N> EIGEN_STRONG_INLINE Packet4f plogical_shift_right(const Packet4f& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
Packet4ui r = vec_sr(reinterpret_cast<Packet4ui>(a), p4ui_mask);
return reinterpret_cast<Packet4f>(r);
}
template<int N> EIGEN_STRONG_INLINE Packet4ui plogical_shift_right(const Packet4ui& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
return vec_sr(a, p4ui_mask);
}
template<int N> EIGEN_STRONG_INLINE Packet4ui plogical_shift_left(const Packet4ui& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(mask, N);
return vec_sl(a, p4ui_mask);
}
template<int N> EIGEN_STRONG_INLINE Packet8us plogical_shift_left(const Packet8us& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet8us(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet8us(mask, N);
return vec_sl(a, p8us_mask);
}
template<int N> EIGEN_STRONG_INLINE Packet8us plogical_shift_right(const Packet8us& a)
{
- const _EIGEN_DECLARE_CONST_FAST_Packet8us(mask, N);
+ const EIGEN_DECLARE_CONST_FAST_Packet8us(mask, N);
return vec_sr(a, p8us_mask);
}
@@ -1234,7 +1234,7 @@
}
EIGEN_STRONG_INLINE Packet4f Bf16ToF32Odd(const Packet8bf& bf){
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(high_mask, 0xFFFF0000);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(high_mask, 0xFFFF0000);
return pand<Packet4f>(
reinterpret_cast<Packet4f>(bf.m_val),
reinterpret_cast<Packet4f>(p4ui_high_mask)
@@ -1244,7 +1244,7 @@
// Simple interleaving of bool masks, prevents true values from being
// converted to NaNs.
EIGEN_STRONG_INLINE Packet8bf F32ToBf16Bool(Packet4f even, Packet4f odd) {
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(high_mask, 0xFFFF0000);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(high_mask, 0xFFFF0000);
Packet4f bf_odd, bf_even;
bf_odd = pand(reinterpret_cast<Packet4f>(p4ui_high_mask), odd);
bf_even = plogical_shift_right<16>(even);
@@ -1256,18 +1256,18 @@
Packet4ui lsb = plogical_shift_right<16>(input);
lsb = pand<Packet4ui>(lsb, reinterpret_cast<Packet4ui>(p4i_ONE));
- _EIGEN_DECLARE_CONST_FAST_Packet4ui(BIAS,0x7FFFu);
+ EIGEN_DECLARE_CONST_FAST_Packet4ui(BIAS,0x7FFFu);
Packet4ui rounding_bias = padd<Packet4ui>(lsb, p4ui_BIAS);
input = padd<Packet4ui>(input, rounding_bias);
//Test NaN and Subnormal - Begin
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(exp_mask, 0x7F800000);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(exp_mask, 0x7F800000);
Packet4ui exp = pand<Packet4ui>(p4ui_exp_mask, reinterpret_cast<Packet4ui>(p4f));
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(mantissa_mask, 0x7FFFFF);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(mantissa_mask, 0x7FFFFF);
Packet4ui mantissa = pand<Packet4ui>(p4ui_mantissa_mask, reinterpret_cast<Packet4ui>(p4f));
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(max_exp, 0x7F800000);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(max_exp, 0x7F800000);
Packet4bi is_max_exp = vec_cmpeq(exp, p4ui_max_exp);
Packet4bi is_zero_exp = vec_cmpeq(exp, reinterpret_cast<Packet4ui>(p4i_ZERO));
@@ -1282,7 +1282,7 @@
reinterpret_cast<Packet4ui>(is_mant_zero)
);
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(nan, 0x7FC00000);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(nan, 0x7FC00000);
input = vec_sel(input, p4ui_nan, nan_selector);
input = vec_sel(input, reinterpret_cast<Packet4ui>(p4f), subnormal_selector);
//Test NaN and Subnormal - End
@@ -2210,7 +2210,7 @@
Packet4f float_odd = Bf16ToF32Odd(a);
Packet4ui int_even = pcast<Packet4f, Packet4ui>(float_even);
Packet4ui int_odd = pcast<Packet4f, Packet4ui>(float_odd);
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(low_mask, 0x0000FFFF);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(low_mask, 0x0000FFFF);
Packet4ui low_even = pand<Packet4ui>(int_even, p4ui_low_mask);
Packet4ui low_odd = pand<Packet4ui>(int_odd, p4ui_low_mask);
@@ -2233,7 +2233,7 @@
template<> EIGEN_STRONG_INLINE Packet8bf pcast<Packet8us, Packet8bf>(const Packet8us& a) {
//short -> int -> float -> bfloat16
- const _EIGEN_DECLARE_CONST_FAST_Packet4ui(low_mask, 0x0000FFFF);
+ const EIGEN_DECLARE_CONST_FAST_Packet4ui(low_mask, 0x0000FFFF);
Packet4ui int_cast = reinterpret_cast<Packet4ui>(a);
Packet4ui int_even = pand<Packet4ui>(int_cast, p4ui_low_mask);
Packet4ui int_odd = plogical_shift_right<16>(int_cast);
diff --git a/Eigen/src/Core/arch/Default/BFloat16.h b/Eigen/src/Core/arch/Default/BFloat16.h
index 36df320..fb99d86 100644
--- a/Eigen/src/Core/arch/Default/BFloat16.h
+++ b/Eigen/src/Core/arch/Default/BFloat16.h
@@ -137,21 +137,25 @@
static EIGEN_CONSTEXPR const bool has_infinity = true;
static EIGEN_CONSTEXPR const bool has_quiet_NaN = true;
static EIGEN_CONSTEXPR const bool has_signaling_NaN = true;
- static EIGEN_CONSTEXPR const float_denorm_style has_denorm = std::denorm_absent;
+ static EIGEN_CONSTEXPR const float_denorm_style has_denorm = std::denorm_present;
static EIGEN_CONSTEXPR const bool has_denorm_loss = false;
static EIGEN_CONSTEXPR const std::float_round_style round_style = numeric_limits<float>::round_style;
- static EIGEN_CONSTEXPR const bool is_iec559 = false;
+ static EIGEN_CONSTEXPR const bool is_iec559 = true;
+ // The C++ standard defines this as "true if the set of values representable
+ // by the type is finite." BFloat16 has finite precision.
static EIGEN_CONSTEXPR const bool is_bounded = true;
static EIGEN_CONSTEXPR const bool is_modulo = false;
static EIGEN_CONSTEXPR const int digits = 8;
static EIGEN_CONSTEXPR const int digits10 = 2;
static EIGEN_CONSTEXPR const int max_digits10 = 4;
- static EIGEN_CONSTEXPR const int radix = 2;
+ static EIGEN_CONSTEXPR const int radix = numeric_limits<float>::radix;
static EIGEN_CONSTEXPR const int min_exponent = numeric_limits<float>::min_exponent;
static EIGEN_CONSTEXPR const int min_exponent10 = numeric_limits<float>::min_exponent10;
static EIGEN_CONSTEXPR const int max_exponent = numeric_limits<float>::max_exponent;
static EIGEN_CONSTEXPR const int max_exponent10 = numeric_limits<float>::max_exponent10;
static EIGEN_CONSTEXPR const bool traps = numeric_limits<float>::traps;
+ // IEEE754: "The implementer shall choose how tininess is detected, but shall
+ // detect tininess in the same way for all operations in radix two"
static EIGEN_CONSTEXPR const bool tinyness_before = numeric_limits<float>::tinyness_before;
static EIGEN_CONSTEXPR Eigen::bfloat16 (min)() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x0080); }
@@ -161,7 +165,7 @@
static EIGEN_CONSTEXPR Eigen::bfloat16 round_error() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x3f00); }
static EIGEN_CONSTEXPR Eigen::bfloat16 infinity() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7f80); }
static EIGEN_CONSTEXPR Eigen::bfloat16 quiet_NaN() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7fc0); }
- static EIGEN_CONSTEXPR Eigen::bfloat16 signaling_NaN() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7f81); }
+ static EIGEN_CONSTEXPR Eigen::bfloat16 signaling_NaN() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x7fa0); }
static EIGEN_CONSTEXPR Eigen::bfloat16 denorm_min() { return Eigen::bfloat16_impl::raw_uint16_to_bfloat16(0x0001); }
};
diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
index 5f2a130..137daa8 100644
--- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
+++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
@@ -170,33 +170,14 @@
EIGEN_UNUSED
Packet plog_impl_float(const Packet _x)
{
- Packet x = _x;
-
const Packet cst_1 = pset1<Packet>(1.0f);
- const Packet cst_neg_half = pset1<Packet>(-0.5f);
- // The smallest non denormalized float number.
- const Packet cst_min_norm_pos = pset1frombits<Packet>( 0x00800000u);
const Packet cst_minus_inf = pset1frombits<Packet>( 0xff800000u);
const Packet cst_pos_inf = pset1frombits<Packet>( 0x7f800000u);
- // Polynomial coefficients.
const Packet cst_cephes_SQRTHF = pset1<Packet>(0.707106781186547524f);
- const Packet cst_cephes_log_p0 = pset1<Packet>(7.0376836292E-2f);
- const Packet cst_cephes_log_p1 = pset1<Packet>(-1.1514610310E-1f);
- const Packet cst_cephes_log_p2 = pset1<Packet>(1.1676998740E-1f);
- const Packet cst_cephes_log_p3 = pset1<Packet>(-1.2420140846E-1f);
- const Packet cst_cephes_log_p4 = pset1<Packet>(+1.4249322787E-1f);
- const Packet cst_cephes_log_p5 = pset1<Packet>(-1.6668057665E-1f);
- const Packet cst_cephes_log_p6 = pset1<Packet>(+2.0000714765E-1f);
- const Packet cst_cephes_log_p7 = pset1<Packet>(-2.4999993993E-1f);
- const Packet cst_cephes_log_p8 = pset1<Packet>(+3.3333331174E-1f);
-
- // Truncate input values to the minimum positive normal.
- x = pmax(x, cst_min_norm_pos);
-
- Packet e;
+ Packet e, x;
// extract significant in the range [0.5,1) and exponent
- x = pfrexp(x,e);
+ x = pfrexp(_x,e);
// part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
// and shift by -1. The values are then centered around 0, which improves
@@ -211,24 +192,22 @@
e = psub(e, pand(cst_1, mask));
x = padd(x, tmp);
- Packet x2 = pmul(x, x);
- Packet x3 = pmul(x2, x);
+ // Polynomial coefficients for rational (3,3) r(x) = p(x)/q(x)
+ // approximating log(1+x) on [sqrt(0.5)-1;sqrt(2)-1].
+ const Packet cst_p1 = pset1<Packet>(1.0000000190281136f);
+ const Packet cst_p2 = pset1<Packet>(1.0000000190281063f);
+ const Packet cst_p3 = pset1<Packet>(0.18256296349849254f);
+ const Packet cst_q1 = pset1<Packet>(1.4999999999999927f);
+ const Packet cst_q2 = pset1<Packet>(0.59923249590823520f);
+ const Packet cst_q3 = pset1<Packet>(0.049616247954120038f);
- // Evaluate the polynomial approximant of degree 8 in three parts, probably
- // to improve instruction-level parallelism.
- Packet y, y1, y2;
- y = pmadd(cst_cephes_log_p0, x, cst_cephes_log_p1);
- y1 = pmadd(cst_cephes_log_p3, x, cst_cephes_log_p4);
- y2 = pmadd(cst_cephes_log_p6, x, cst_cephes_log_p7);
- y = pmadd(y, x, cst_cephes_log_p2);
- y1 = pmadd(y1, x, cst_cephes_log_p5);
- y2 = pmadd(y2, x, cst_cephes_log_p8);
- y = pmadd(y, x3, y1);
- y = pmadd(y, x3, y2);
- y = pmul(y, x3);
-
- y = pmadd(cst_neg_half, x2, y);
- x = padd(x, y);
+ Packet p = pmadd(x, cst_p3, cst_p2);
+ p = pmadd(x, p, cst_p1);
+ p = pmul(x, p);
+ Packet q = pmadd(x, cst_q3, cst_q2);
+ q = pmadd(x, q, cst_q1);
+ q = pmadd(x, q, cst_1);
+ x = pdiv(p, q);
// Add the logarithm of the exponent back to the result of the interpolation.
if (base2) {
@@ -284,8 +263,6 @@
const Packet cst_1 = pset1<Packet>(1.0);
const Packet cst_neg_half = pset1<Packet>(-0.5);
- // The smallest non denormalized double.
- const Packet cst_min_norm_pos = pset1frombits<Packet>( static_cast<uint64_t>(0x0010000000000000ull));
const Packet cst_minus_inf = pset1frombits<Packet>( static_cast<uint64_t>(0xfff0000000000000ull));
const Packet cst_pos_inf = pset1frombits<Packet>( static_cast<uint64_t>(0x7ff0000000000000ull));
@@ -307,9 +284,6 @@
const Packet cst_cephes_log_q4 = pset1<Packet>(7.11544750618563894466E1);
const Packet cst_cephes_log_q5 = pset1<Packet>(2.31251620126765340583E1);
- // Truncate input values to the minimum positive normal.
- x = pmax(x, cst_min_norm_pos);
-
Packet e;
// extract significant in the range [0.5,1) and exponent
x = pfrexp(x,e);
@@ -435,24 +409,24 @@
// Exponential function. Works by writing "x = m*log(2) + r" where
// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+// exp(r) is computed using a 6th order minimax polynomial approximation.
template <typename Packet>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
EIGEN_UNUSED
Packet pexp_float(const Packet _x)
{
const Packet cst_zero = pset1<Packet>(0.0f);
- const Packet cst_1 = pset1<Packet>(1.0f);
+ const Packet cst_one = pset1<Packet>(1.0f);
const Packet cst_half = pset1<Packet>(0.5f);
- const Packet cst_exp_hi = pset1<Packet>( 88.723f);
- const Packet cst_exp_lo = pset1<Packet>(-88.723f);
+ const Packet cst_exp_hi = pset1<Packet>(88.723f);
+ const Packet cst_exp_lo = pset1<Packet>(-104.f);
const Packet cst_cephes_LOG2EF = pset1<Packet>(1.44269504088896341f);
- const Packet cst_cephes_exp_p0 = pset1<Packet>(1.9875691500E-4f);
- const Packet cst_cephes_exp_p1 = pset1<Packet>(1.3981999507E-3f);
- const Packet cst_cephes_exp_p2 = pset1<Packet>(8.3334519073E-3f);
- const Packet cst_cephes_exp_p3 = pset1<Packet>(4.1665795894E-2f);
- const Packet cst_cephes_exp_p4 = pset1<Packet>(1.6666665459E-1f);
- const Packet cst_cephes_exp_p5 = pset1<Packet>(5.0000001201E-1f);
+ const Packet cst_p2 = pset1<Packet>(0.49999988079071044921875f);
+ const Packet cst_p3 = pset1<Packet>(0.16666518151760101318359375f);
+ const Packet cst_p4 = pset1<Packet>(4.166965186595916748046875e-2f);
+ const Packet cst_p5 = pset1<Packet>(8.36894474923610687255859375e-3f);
+ const Packet cst_p6 = pset1<Packet>(1.37449637986719608306884765625e-3f);
// Clamp x.
Packet zero_mask = pcmp_lt(_x, cst_exp_lo);
@@ -470,18 +444,15 @@
Packet r = pmadd(m, cst_cephes_exp_C1, x);
r = pmadd(m, cst_cephes_exp_C2, r);
- Packet r2 = pmul(r, r);
- Packet r3 = pmul(r2, r);
-
- // Evaluate the polynomial approximant,improved by instruction-level parallelism.
- Packet y, y1, y2;
- y = pmadd(cst_cephes_exp_p0, r, cst_cephes_exp_p1);
- y1 = pmadd(cst_cephes_exp_p3, r, cst_cephes_exp_p4);
- y2 = padd(r, cst_1);
- y = pmadd(y, r, cst_cephes_exp_p2);
- y1 = pmadd(y1, r, cst_cephes_exp_p5);
- y = pmadd(y, r3, y1);
- y = pmadd(y, r2, y2);
+ // Evaluate the 6th order polynomial approximation to exp(r)
+ // with r in the interval [-ln(2)/2;ln(2)/2].
+ const Packet r2 = pmul(r, r);
+ Packet p_even = pmadd(r2, cst_p6, cst_p4);
+ const Packet p_odd = pmadd(r2, cst_p5, cst_p3);
+ p_even = pmadd(r2, p_even, cst_p2);
+ const Packet p_low = padd(r, cst_one);
+ Packet y = pmadd(r, p_odd, p_even);
+ y = pmadd(r2, y, p_low);
// Return 2^m * exp(r).
// TODO: replace pldexp with faster implementation since y in [-1, 1).
@@ -494,7 +465,7 @@
Packet pexp_double(const Packet _x)
{
Packet x = _x;
- const Packet cst_zero = pset1<Packet>(0.0f);
+ const Packet cst_zero = pset1<Packet>(0.0);
const Packet cst_1 = pset1<Packet>(1.0);
const Packet cst_2 = pset1<Packet>(2.0);
const Packet cst_half = pset1<Packet>(0.5);
diff --git a/Eigen/src/Core/arch/Default/Half.h b/Eigen/src/Core/arch/Default/Half.h
index 4155335..876045e 100644
--- a/Eigen/src/Core/arch/Default/Half.h
+++ b/Eigen/src/Core/arch/Default/Half.h
@@ -218,29 +218,33 @@
static EIGEN_CONSTEXPR const float_denorm_style has_denorm = denorm_present;
static EIGEN_CONSTEXPR const bool has_denorm_loss = false;
static EIGEN_CONSTEXPR const std::float_round_style round_style = std::round_to_nearest;
- static EIGEN_CONSTEXPR const bool is_iec559 = false;
- static EIGEN_CONSTEXPR const bool is_bounded = false;
+ static EIGEN_CONSTEXPR const bool is_iec559 = true;
+ // The C++ standard defines this as "true if the set of values representable
+ // by the type is finite." Half has finite precision.
+ static EIGEN_CONSTEXPR const bool is_bounded = true;
static EIGEN_CONSTEXPR const bool is_modulo = false;
static EIGEN_CONSTEXPR const int digits = 11;
static EIGEN_CONSTEXPR const int digits10 = 3; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
static EIGEN_CONSTEXPR const int max_digits10 = 5; // according to http://half.sourceforge.net/structstd_1_1numeric__limits_3_01half__float_1_1half_01_4.html
- static EIGEN_CONSTEXPR const int radix = 2;
+ static EIGEN_CONSTEXPR const int radix = numeric_limits<float>::radix;
static EIGEN_CONSTEXPR const int min_exponent = -13;
static EIGEN_CONSTEXPR const int min_exponent10 = -4;
static EIGEN_CONSTEXPR const int max_exponent = 16;
static EIGEN_CONSTEXPR const int max_exponent10 = 4;
- static EIGEN_CONSTEXPR const bool traps = true;
- static EIGEN_CONSTEXPR const bool tinyness_before = false;
+ static EIGEN_CONSTEXPR const bool traps = numeric_limits<float>::traps;
+ // IEEE754: "The implementer shall choose how tininess is detected, but shall
+ // detect tininess in the same way for all operations in radix two"
+ static EIGEN_CONSTEXPR const bool tinyness_before = std::numeric_limits<float>::tinyness_before;
- static EIGEN_CONSTEXPR Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
+ static EIGEN_CONSTEXPR Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x0400); }
static EIGEN_CONSTEXPR Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
static EIGEN_CONSTEXPR Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
- static EIGEN_CONSTEXPR Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
+ static EIGEN_CONSTEXPR Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x1400); }
static EIGEN_CONSTEXPR Eigen::half round_error() { return Eigen::half_impl::raw_uint16_to_half(0x3800); }
static EIGEN_CONSTEXPR Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
static EIGEN_CONSTEXPR Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
static EIGEN_CONSTEXPR Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7d00); }
- static EIGEN_CONSTEXPR Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
+ static EIGEN_CONSTEXPR Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x0001); }
};
// If std::numeric_limits<T> is specialized, should also specialize
diff --git a/Eigen/src/Core/arch/MSA/MathFunctions.h b/Eigen/src/Core/arch/MSA/MathFunctions.h
index bbebde5..c3295a51 100644
--- a/Eigen/src/Core/arch/MSA/MathFunctions.h
+++ b/Eigen/src/Core/arch/MSA/MathFunctions.h
@@ -35,20 +35,20 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
plog<Packet4f>(const Packet4f& _x) {
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292e-2f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
- static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
- static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292e-2f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, -1.1514610310e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, -1.2420140846e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, +1.4249322787e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, -1.6668057665e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, +2.0000714765e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, -2.4999993993e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, +3.3333331174e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+ static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+ static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
// Convert negative argument into NAN (quiet negative, to be specific).
Packet4f zero = (Packet4f)__builtin_msa_ldi_w(0);
@@ -125,19 +125,19 @@
pexp<Packet4f>(const Packet4f& _x) {
// Limiting single-precision pexp's argument to [-128, +128] lets pexp
// reach 0 and INFINITY naturally.
- static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -128.0f);
- static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, +128.0f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894e-2f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
- static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(exp_lo, -128.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(exp_hi, +128.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894e-2f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+ static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
Packet4f x = _x;
@@ -176,21 +176,21 @@
template <>
EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
ptanh<Packet4f>(const Packet4f& _x) {
- static _EIGEN_DECLARE_CONST_Packet4f(tanh_tiny, 1e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(tanh_hi, 9.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(tanh_tiny, 1e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(tanh_hi, 9.0f);
// The monomial coefficients of the numerator polynomial (odd).
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-5f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-8f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
- static _EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_1, 4.89352455891786e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_3, 6.37261928875436e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_5, 1.48572235717979e-5f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_7, 5.12229709037114e-8f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_9, -8.60467152213735e-11f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_11, 2.00018790482477e-13f);
+ static EIGEN_DECLARE_CONST_Packet4f(alpha_13, -2.76076847742355e-16f);
// The monomial coefficients of the denominator polynomial (even).
- static _EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-6f);
+ static EIGEN_DECLARE_CONST_Packet4f(beta_0, 4.89352518554385e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(beta_2, 2.26843463243900e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(beta_4, 1.18534705686654e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(beta_6, 1.19825839466702e-6f);
Packet4f x = pabs(_x);
Packet4i tiny_mask = __builtin_msa_fclt_w(x, p4f_tanh_tiny);
@@ -231,19 +231,19 @@
template <bool sine>
Packet4f psincos_inner_msa_float(const Packet4f& _x) {
- static _EIGEN_DECLARE_CONST_Packet4f(sincos_max_arg, 13176795.0f); // Approx. (2**24) / (4/Pi).
- static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1, -0.78515625f);
- static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
- static _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891e-4f);
- static _EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611e-1f);
- static _EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948e-5f);
- static _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765e-3f);
- static _EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827e-2f);
- static _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4/Pi.
- static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
- static _EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
+ static EIGEN_DECLARE_CONST_Packet4f(sincos_max_arg, 13176795.0f); // Approx. (2**24) / (4/Pi).
+ static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1, -0.78515625f);
+ static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
+ static EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891e-4f);
+ static EIGEN_DECLARE_CONST_Packet4f(sincof_p1, 8.3321608736e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611e-1f);
+ static EIGEN_DECLARE_CONST_Packet4f(coscof_p0, 2.443315711809948e-5f);
+ static EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765e-3f);
+ static EIGEN_DECLARE_CONST_Packet4f(coscof_p2, 4.166664568298827e-2f);
+ static EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4/Pi.
+ static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+ static EIGEN_DECLARE_CONST_Packet4f(1, 1.0f);
Packet4f x = pabs(_x);
@@ -328,21 +328,21 @@
pexp<Packet2d>(const Packet2d& _x) {
// Limiting double-precision pexp's argument to [-1024, +1024] lets pexp
// reach 0 and INFINITY naturally.
- static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -1024.0);
- static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, +1024.0);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
- static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
- static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
- static _EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
- static _EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
+ static EIGEN_DECLARE_CONST_Packet2d(exp_lo, -1024.0);
+ static EIGEN_DECLARE_CONST_Packet2d(exp_hi, +1024.0);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+ static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+ static EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+ static EIGEN_DECLARE_CONST_Packet2d(1, 1.0);
+ static EIGEN_DECLARE_CONST_Packet2d(2, 2.0);
Packet2d x = _x;
diff --git a/Eigen/src/Core/arch/MSA/PacketMath.h b/Eigen/src/Core/arch/MSA/PacketMath.h
index 3dc95a1..f03dbed 100644
--- a/Eigen/src/Core/arch/MSA/PacketMath.h
+++ b/Eigen/src/Core/arch/MSA/PacketMath.h
@@ -53,9 +53,9 @@
typedef v4i32 Packet4i;
typedef v4u32 Packet4ui;
-#define _EIGEN_DECLARE_CONST_Packet4f(NAME, X) const Packet4f p4f_##NAME = { X, X, X, X }
-#define _EIGEN_DECLARE_CONST_Packet4i(NAME, X) const Packet4i p4i_##NAME = { X, X, X, X }
-#define _EIGEN_DECLARE_CONST_Packet4ui(NAME, X) const Packet4ui p4ui_##NAME = { X, X, X, X }
+#define EIGEN_DECLARE_CONST_Packet4f(NAME, X) const Packet4f p4f_##NAME = { X, X, X, X }
+#define EIGEN_DECLARE_CONST_Packet4i(NAME, X) const Packet4i p4i_##NAME = { X, X, X, X }
+#define EIGEN_DECLARE_CONST_Packet4ui(NAME, X) const Packet4ui p4ui_##NAME = { X, X, X, X }
inline std::ostream& operator<<(std::ostream& os, const Packet4f& value) {
os << "[ " << value[0] << ", " << value[1] << ", " << value[2] << ", " << value[3] << " ]";
@@ -823,9 +823,9 @@
typedef v2i64 Packet2l;
typedef v2u64 Packet2ul;
-#define _EIGEN_DECLARE_CONST_Packet2d(NAME, X) const Packet2d p2d_##NAME = { X, X }
-#define _EIGEN_DECLARE_CONST_Packet2l(NAME, X) const Packet2l p2l_##NAME = { X, X }
-#define _EIGEN_DECLARE_CONST_Packet2ul(NAME, X) const Packet2ul p2ul_##NAME = { X, X }
+#define EIGEN_DECLARE_CONST_Packet2d(NAME, X) const Packet2d p2d_##NAME = { X, X }
+#define EIGEN_DECLARE_CONST_Packet2l(NAME, X) const Packet2l p2l_##NAME = { X, X }
+#define EIGEN_DECLARE_CONST_Packet2ul(NAME, X) const Packet2ul p2ul_##NAME = { X, X }
inline std::ostream& operator<<(std::ostream& os, const Packet2d& value) {
os << "[ " << value[0] << ", " << value[1] << " ]";
diff --git a/Eigen/src/Core/arch/NEON/Complex.h b/Eigen/src/Core/arch/NEON/Complex.h
index fd6b14b..008dd7a 100644
--- a/Eigen/src/Core/arch/NEON/Complex.h
+++ b/Eigen/src/Core/arch/NEON/Complex.h
@@ -376,7 +376,7 @@
#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
// See bug 1325, clang fails to call vld1q_u64.
-#if EIGEN_COMP_CLANG || EIGEN_COMP_CASTXML
+#if EIGEN_COMP_CLANG || EIGEN_COMP_CASTXML || EIGEN_COMP_CPE
static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000};
#else
const uint64_t p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h
index 707f7d7..dda0a1e 100644
--- a/Eigen/src/Core/arch/NEON/PacketMath.h
+++ b/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -139,13 +139,13 @@
#define vec4f_duplane(a, p) \
vdupq_lane_f32(vget_low_f32(a), p)
-#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
const Packet4f p4f_##NAME = pset1<Packet4f>(X)
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int32_t>(X))
-#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
const Packet4i p4i_##NAME = pset1<Packet4i>(X)
#if EIGEN_ARCH_ARM64
diff --git a/Eigen/src/Core/arch/SSE/Complex.h b/Eigen/src/Core/arch/SSE/Complex.h
index 61e9406..60308ce 100644
--- a/Eigen/src/Core/arch/SSE/Complex.h
+++ b/Eigen/src/Core/arch/SSE/Complex.h
@@ -137,7 +137,7 @@
template<> EIGEN_STRONG_INLINE std::complex<float> pfirst<Packet2cf>(const Packet2cf& a)
{
- std::complex<float> res;
+ alignas(alignof(__m64)) std::complex<float> res;
_mm_storel_pi((__m64*)&res, a.v);
return res;
}
diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h
index 4bdb9af..5a063d3 100644
--- a/Eigen/src/Core/arch/SSE/MathFunctions.h
+++ b/Eigen/src/Core/arch/SSE/MathFunctions.h
@@ -75,8 +75,6 @@
return pcos_float(_x);
}
-#if EIGEN_FAST_MATH
-
// Functions for sqrt.
// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
@@ -85,11 +83,13 @@
// it can be inlined and pipelined with other computations, further reducing its
// effective latency. This is similar to Quake3's fast inverse square root.
// For detail see here: http://www.beyond3d.com/content/articles/8/
-template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+#if EIGEN_FAST_MATH
+template<>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4f psqrt<Packet4f>(const Packet4f& _x)
{
- Packet4f minus_half_x = pmul(_x, pset1<Packet4f>(-0.5f));
- Packet4f denormal_mask = pandnot(
+ const Packet4f minus_half_x = pmul(_x, pset1<Packet4f>(-0.5f));
+ const Packet4f denormal_mask = pandnot(
pcmp_lt(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())),
pcmp_lt(_x, pzero(_x)));
@@ -118,10 +118,10 @@
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
- _EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
- _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
- _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000u);
- _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000u);
+ EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
+ EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
+ EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000u);
+ EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000u);
Packet4f neg_half = pmul(_x, p4f_minus_half);
diff --git a/Eigen/src/Core/arch/SSE/PacketMath.h b/Eigen/src/Core/arch/SSE/PacketMath.h
index 45e219c..a843226 100755
--- a/Eigen/src/Core/arch/SSE/PacketMath.h
+++ b/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -108,16 +108,16 @@
#define vec2d_duplane(a,p) \
vec2d_swizzle2(a,a,(p<<1)|p)
-#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
const Packet4f p4f_##NAME = pset1<Packet4f>(X)
-#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
const Packet2d p2d_##NAME = pset1<Packet2d>(X)
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = pset1frombits<Packet4f>(X)
-#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
const Packet4i p4i_##NAME = pset1<Packet4i>(X)
diff --git a/Eigen/src/Core/arch/ZVector/Complex.h b/Eigen/src/Core/arch/ZVector/Complex.h
index acb04fb..df5c8d4 100644
--- a/Eigen/src/Core/arch/ZVector/Complex.h
+++ b/Eigen/src/Core/arch/ZVector/Complex.h
@@ -8,8 +8,8 @@
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
-#ifndef EIGEN_COMPLEX32_ALTIVEC_H
-#define EIGEN_COMPLEX32_ALTIVEC_H
+#ifndef EIGEN_COMPLEX32_ZVECTOR_H
+#define EIGEN_COMPLEX32_ZVECTOR_H
#include "../../InternalHeaderCheck.h"
@@ -425,4 +425,4 @@
} // end namespace Eigen
-#endif // EIGEN_COMPLEX32_ALTIVEC_H
+#endif // EIGEN_COMPLEX32_ZVECTOR_H
diff --git a/Eigen/src/Core/arch/ZVector/MathFunctions.h b/Eigen/src/Core/arch/ZVector/MathFunctions.h
index c86e185..7ee14ab 100644
--- a/Eigen/src/Core/arch/ZVector/MathFunctions.h
+++ b/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -13,8 +13,8 @@
* Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
*/
-#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
-#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#ifndef EIGEN_MATH_FUNCTIONS_ZVECTOR_H
+#define EIGEN_MATH_FUNCTIONS_ZVECTOR_H
#include "../../InternalHeaderCheck.h"
@@ -23,69 +23,69 @@
namespace internal {
#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
-static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
-static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
-static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
-static _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+static EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+static EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+static EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+static EIGEN_DECLARE_CONST_Packet4i(23, 23);
-static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
/* the smallest non denormalized float number */
-static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000);
-static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000); // -1.f/0.f
-static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan, 0xffffffff);
+static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos, 0x00800000);
+static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf, 0xff800000); // -1.f/0.f
+static EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan, 0xffffffff);
/* natural logarithm computed for 4 simultaneous float
return NaN for x <= 0
*/
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
-static _EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f);
-static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+static EIGEN_DECLARE_CONST_Packet4f(exp_hi, 88.3762626647950f);
+static EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
-static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+static EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
#endif
-static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
-static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
-static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+static EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
-static _EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437);
-static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+static EIGEN_DECLARE_CONST_Packet2d(exp_hi, 709.437);
+static EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
-static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
Packet2d pexp<Packet2d>(const Packet2d& _x)
@@ -232,4 +232,4 @@
} // end namespace Eigen
-#endif // EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#endif // EIGEN_MATH_FUNCTIONS_ZVECTOR_H
diff --git a/Eigen/src/Core/arch/ZVector/PacketMath.h b/Eigen/src/Core/arch/ZVector/PacketMath.h
index 3d81aed..26b6f0d 100755
--- a/Eigen/src/Core/arch/ZVector/PacketMath.h
+++ b/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -66,48 +66,48 @@
// We don't want to write the same code all the time, but we need to reuse the constants
// and it doesn't really work to declare them global, so we define macros instead
-#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
Packet4i p4i_##NAME = reinterpret_cast<Packet4i>(vec_splat_s32(X))
-#define _EIGEN_DECLARE_CONST_FAST_Packet2d(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet2d(NAME,X) \
Packet2d p2d_##NAME = reinterpret_cast<Packet2d>(vec_splat_s64(X))
-#define _EIGEN_DECLARE_CONST_FAST_Packet2l(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet2l(NAME,X) \
Packet2l p2l_##NAME = reinterpret_cast<Packet2l>(vec_splat_s64(X))
-#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
Packet4i p4i_##NAME = pset1<Packet4i>(X)
-#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
Packet2d p2d_##NAME = pset1<Packet2d>(X)
-#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
Packet2l p2l_##NAME = pset1<Packet2l>(X)
// These constants are endian-agnostic
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
-static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
-static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
-static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
+static EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
+static EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
+static EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
static Packet2d p2d_ONE = { 1.0, 1.0 };
static Packet2d p2d_ZERO_ = { numext::bit_cast<double>(0x8000000000000000ull),
numext::bit_cast<double>(0x8000000000000000ull) };
#if !defined(__ARCH__) || (defined(__ARCH__) && __ARCH__ >= 12)
-#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
-#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
Packet4f p4f_##NAME = pset1<Packet4f>(X)
-#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+#define EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
-static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
-static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
+static EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
+static EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
static Packet4f p4f_MZERO = { 0x80000000, 0x80000000, 0x80000000, 0x80000000};
#endif
@@ -119,9 +119,9 @@
static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
// Mask alignment
-#define _EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0
+#define EIGEN_MASK_ALIGNMENT 0xfffffffffffffff0
-#define _EIGEN_ALIGNED_PTR(x) ((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+#define EIGEN_ALIGNED_PTR(x) ((std::ptrdiff_t)(x) & EIGEN_MASK_ALIGNMENT)
// Handle endianness properly while loading constants
// Define global static constants:
diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index 044f7dd..d56aae5 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -278,7 +278,7 @@
*/
template<typename Scalar> struct scalar_exp_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::exp(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return internal::pexp(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
};
@@ -1026,87 +1026,105 @@
template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
Packet packetOp(const Packet& x) const {
const Packet one = pset1<Packet>(T(1));
- return pdiv(one, padd(one, pexp(pnegate(x))));
+ const Packet inf = pset1<Packet>(NumTraits<T>::infinity());
+ const Packet e = pexp(x);
+ const Packet inf_mask = pcmp_eq(e, inf);
+ return pselect(inf_mask, one, pdiv(e, padd(one, e)));
}
};
-#ifndef EIGEN_GPU_COMPILE_PHASE
+// TODO(rmlarsen): Enable the following on host when integer_packet is defined
+// for the relevant packet types.
+#ifdef EIGEN_GPU_CC
+
/** \internal
* \brief Template specialization of the logistic function for float.
- *
- * Uses just a 9/10-degree rational interpolant which
- * interpolates 1/(1+exp(-x)) - 0.5 up to a couple of ulps in the range
- * [-9, 18]. Below -9 we use the more accurate approximation
- * 1/(1+exp(-x)) ~= exp(x), and above 18 the logistic function is 1 within
- * one ulp. The shifted logistic is interpolated because it was easier to
- * make the fit converge.
- *
+ * Computes S(x) = exp(x) / (1 + exp(x)), where exp(x) is implemented
+ * using an algorithm partly adopted from the implementation of
+ * pexp_float. See the individual steps described in the code below.
+ * Note that compared to pexp, we use an additional outer multiplicative
+ * range reduction step using the identity exp(x) = exp(x/2)^2.
+ * This prevert us from having to call ldexp on values that could produce
+ * a denormal result, which allows us to call the faster implementation in
+ * pldexp_fast_impl<Packet>::run(p, m).
+ * The final squaring, however, doubles the error bound on the final
+ * approximation. Exhaustive testing shows that we have a worst case error
+ * of 4.5 ulps (compared to computing S(x) in double precision), which is
+ * acceptable.
*/
template <>
struct scalar_logistic_op<float> {
EIGEN_EMPTY_STRUCT_CTOR(scalar_logistic_op)
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator()(const float& x) const {
- return packetOp(x);
+ // Truncate at the first point where the interpolant is exactly one.
+ const float cst_exp_hi = 16.6355324f;
+ const float e = numext::exp(numext::mini(x, cst_exp_hi));
+ return e / (1.0f + e);
}
- template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- Packet packetOp(const Packet& _x) const {
- const Packet cutoff_lower = pset1<Packet>(-9.f);
- const Packet lt_mask = pcmp_lt<Packet>(_x, cutoff_lower);
- const bool any_small = predux_any(lt_mask);
+ template <typename Packet>
+ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
+ packetOp(const Packet& _x) const {
+ const Packet cst_zero = pset1<Packet>(0.0f);
+ const Packet cst_one = pset1<Packet>(1.0f);
+ const Packet cst_half = pset1<Packet>(0.5f);
+ // Truncate at the first point where the interpolant is exactly one.
+ const Packet cst_exp_hi = pset1<Packet>(16.6355324f);
+ const Packet cst_exp_lo = pset1<Packet>(-104.f);
- // The upper cut-off is the smallest x for which the rational approximation evaluates to 1.
- // Choosing this value saves us a few instructions clamping the results at the end.
-#ifdef EIGEN_VECTORIZE_FMA
- const Packet cutoff_upper = pset1<Packet>(15.7243833541870117f);
-#else
- const Packet cutoff_upper = pset1<Packet>(15.6437711715698242f);
-#endif
- const Packet x = pmin(_x, cutoff_upper);
+ // Clamp x to the non-trivial range where S(x). Outside this
+ // interval the correctly rounded value of S(x) is either zero
+ // or one.
+ Packet zero_mask = pcmp_lt(_x, cst_exp_lo);
+ Packet x = pmin(_x, cst_exp_hi);
- // The monomial coefficients of the numerator polynomial (odd).
- const Packet alpha_1 = pset1<Packet>(2.48287947061529e-01f);
- const Packet alpha_3 = pset1<Packet>(8.51377133304701e-03f);
- const Packet alpha_5 = pset1<Packet>(6.08574864600143e-05f);
- const Packet alpha_7 = pset1<Packet>(1.15627324459942e-07f);
- const Packet alpha_9 = pset1<Packet>(4.37031012579801e-11f);
+ // 1. Multiplicative range reduction:
+ // Reduce the range of x by a factor of 2. This avoids having
+ // to compute exp(x) accurately where the result is a denormalized
+ // value.
+ x = pmul(x, cst_half);
- // The monomial coefficients of the denominator polynomial (even).
- const Packet beta_0 = pset1<Packet>(9.93151921023180e-01f);
- const Packet beta_2 = pset1<Packet>(1.16817656904453e-01f);
- const Packet beta_4 = pset1<Packet>(1.70198817374094e-03f);
- const Packet beta_6 = pset1<Packet>(6.29106785017040e-06f);
- const Packet beta_8 = pset1<Packet>(5.76102136993427e-09f);
- const Packet beta_10 = pset1<Packet>(6.10247389755681e-13f);
+ // 2. Subtractive range reduction:
+ // Express exp(x) as exp(m*ln(2) + r) = 2^m*exp(r), start by extracting
+ // m = floor(x/ln(2) + 0.5), such that x = m*ln(2) + r.
+ const Packet cst_cephes_LOG2EF = pset1<Packet>(1.44269504088896341f);
+ Packet m = pfloor(pmadd(x, cst_cephes_LOG2EF, cst_half));
+ // Get r = x - m*ln(2). We use a trick from Cephes where the term
+ // m*ln(2) is subtracted out in two parts, m*C1+m*C2 = m*ln(2),
+ // to avoid accumulating truncation errors.
+ const Packet cst_cephes_exp_C1 = pset1<Packet>(-0.693359375f);
+ const Packet cst_cephes_exp_C2 = pset1<Packet>(2.12194440e-4f);
+ Packet r = pmadd(m, cst_cephes_exp_C1, x);
+ r = pmadd(m, cst_cephes_exp_C2, r);
- // Since the polynomials are odd/even, we need x^2.
- const Packet x2 = pmul(x, x);
+ // 3. Compute an approximation to exp(r) using a degree 5 minimax polynomial.
+ // We compute even and odd terms separately to increase instruction level
+ // parallelism.
+ Packet r2 = pmul(r, r);
+ const Packet cst_p2 = pset1<Packet>(0.49999141693115234375f);
+ const Packet cst_p3 = pset1<Packet>(0.16666877269744873046875f);
+ const Packet cst_p4 = pset1<Packet>(4.1898667812347412109375e-2f);
+ const Packet cst_p5 = pset1<Packet>(8.33471305668354034423828125e-3f);
- // Evaluate the numerator polynomial p.
- Packet p = pmadd(x2, alpha_9, alpha_7);
- p = pmadd(x2, p, alpha_5);
- p = pmadd(x2, p, alpha_3);
- p = pmadd(x2, p, alpha_1);
- p = pmul(x, p);
+ const Packet p_even = pmadd(r2, cst_p4, cst_p2);
+ const Packet p_odd = pmadd(r2, cst_p5, cst_p3);
+ const Packet p_low = padd(r, cst_one);
+ Packet p = pmadd(r, p_odd, p_even);
+ p = pmadd(r2, p, p_low);
- // Evaluate the denominator polynomial q.
- Packet q = pmadd(x2, beta_10, beta_8);
- q = pmadd(x2, q, beta_6);
- q = pmadd(x2, q, beta_4);
- q = pmadd(x2, q, beta_2);
- q = pmadd(x2, q, beta_0);
- // Divide the numerator by the denominator and shift it up.
- const Packet logistic = padd(pdiv(p, q), pset1<Packet>(0.5f));
- if (EIGEN_PREDICT_FALSE(any_small)) {
- const Packet exponential = pexp(_x);
- return pselect(lt_mask, exponential, logistic);
- } else {
- return logistic;
- }
+ // 4. Undo subtractive range reduction exp(m*ln(2) + r) = 2^m * exp(r).
+ Packet e = pldexp_fast_impl<Packet>::run(p, m);
+
+ // 5. Undo multiplicative range reduction by using exp(r) = exp(r/2)^2.
+ e = pmul(e, e);
+
+ // Return exp(x) / (1 + exp(x))
+ return pselect(zero_mask, cst_zero, pdiv(e, padd(cst_one, e)));
}
};
#endif // #ifndef EIGEN_GPU_COMPILE_PHASE
+
template <typename T>
struct functor_traits<scalar_logistic_op<T> > {
enum {
diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index 938f5fb..7580887 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -378,12 +378,12 @@
template <typename T1, typename T2, typename T3>
struct packet_conditional<GEBPPacketHalf, T1, T2, T3> { typedef T2 type; };
-#define PACKET_DECL_COND_PREFIX(prefix, name, packet_size) \
+#define PACKET_DECL_COND_POSTFIX(postfix, name, packet_size) \
typedef typename packet_conditional<packet_size, \
typename packet_traits<name ## Scalar>::type, \
typename packet_traits<name ## Scalar>::half, \
typename unpacket_traits<typename packet_traits<name ## Scalar>::half>::half>::type \
- prefix ## name ## Packet
+ name ## Packet ## postfix
#define PACKET_DECL_COND(name, packet_size) \
typedef typename packet_conditional<packet_size, \
@@ -392,12 +392,12 @@
typename unpacket_traits<typename packet_traits<name ## Scalar>::half>::half>::type \
name ## Packet
-#define PACKET_DECL_COND_SCALAR_PREFIX(prefix, packet_size) \
+#define PACKET_DECL_COND_SCALAR_POSTFIX(postfix, packet_size) \
typedef typename packet_conditional<packet_size, \
typename packet_traits<Scalar>::type, \
typename packet_traits<Scalar>::half, \
typename unpacket_traits<typename packet_traits<Scalar>::half>::half>::type \
- prefix ## ScalarPacket
+ ScalarPacket ## postfix
#define PACKET_DECL_COND_SCALAR(packet_size) \
typedef typename packet_conditional<packet_size, \
@@ -424,17 +424,17 @@
typedef RhsScalar_ RhsScalar;
typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
- PACKET_DECL_COND_PREFIX(_, Lhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Rhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Res, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Lhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Rhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Res, PacketSize_);
enum {
ConjLhs = ConjLhs_,
ConjRhs = ConjRhs_,
- Vectorizable = unpacket_traits<_LhsPacket>::vectorizable && unpacket_traits<_RhsPacket>::vectorizable,
- LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1,
- RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1,
- ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1,
+ Vectorizable = unpacket_traits<LhsPacket_>::vectorizable && unpacket_traits<RhsPacket_>::vectorizable,
+ LhsPacketSize = Vectorizable ? unpacket_traits<LhsPacket_>::size : 1,
+ RhsPacketSize = Vectorizable ? unpacket_traits<RhsPacket_>::size : 1,
+ ResPacketSize = Vectorizable ? unpacket_traits<ResPacket_>::size : 1,
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
@@ -459,9 +459,9 @@
};
- typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
- typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
- typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+ typedef typename conditional<Vectorizable,LhsPacket_,LhsScalar>::type LhsPacket;
+ typedef typename conditional<Vectorizable,RhsPacket_,RhsScalar>::type RhsPacket;
+ typedef typename conditional<Vectorizable,ResPacket_,ResScalar>::type ResPacket;
typedef LhsPacket LhsPacket4Packing;
typedef QuadPacket<RhsPacket> RhsPacketx4;
@@ -553,17 +553,17 @@
typedef RealScalar RhsScalar;
typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
- PACKET_DECL_COND_PREFIX(_, Lhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Rhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Res, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Lhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Rhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Res, PacketSize_);
enum {
ConjLhs = ConjLhs_,
ConjRhs = false,
- Vectorizable = unpacket_traits<_LhsPacket>::vectorizable && unpacket_traits<_RhsPacket>::vectorizable,
- LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1,
- RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1,
- ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1,
+ Vectorizable = unpacket_traits<LhsPacket_>::vectorizable && unpacket_traits<RhsPacket_>::vectorizable,
+ LhsPacketSize = Vectorizable ? unpacket_traits<LhsPacket_>::size : 1,
+ RhsPacketSize = Vectorizable ? unpacket_traits<RhsPacket_>::size : 1,
+ ResPacketSize = Vectorizable ? unpacket_traits<ResPacket_>::size : 1,
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
nr = 4,
@@ -578,9 +578,9 @@
RhsProgress = 1
};
- typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
- typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
- typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+ typedef typename conditional<Vectorizable,LhsPacket_,LhsScalar>::type LhsPacket;
+ typedef typename conditional<Vectorizable,RhsPacket_,RhsScalar>::type RhsPacket;
+ typedef typename conditional<Vectorizable,ResPacket_,ResScalar>::type ResPacket;
typedef LhsPacket LhsPacket4Packing;
typedef QuadPacket<RhsPacket> RhsPacketx4;
@@ -765,9 +765,9 @@
typedef std::complex<RealScalar> RhsScalar;
typedef std::complex<RealScalar> ResScalar;
- PACKET_DECL_COND_PREFIX(_, Lhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Rhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Res, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Lhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Rhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Res, PacketSize_);
PACKET_DECL_COND(Real, PacketSize_);
PACKET_DECL_COND_SCALAR(PacketSize_);
@@ -776,8 +776,8 @@
ConjRhs = ConjRhs_,
Vectorizable = unpacket_traits<RealPacket>::vectorizable
&& unpacket_traits<ScalarPacket>::vectorizable,
- ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1,
- LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1,
+ ResPacketSize = Vectorizable ? unpacket_traits<ResPacket_>::size : 1,
+ LhsPacketSize = Vectorizable ? unpacket_traits<LhsPacket_>::size : 1,
RhsPacketSize = Vectorizable ? unpacket_traits<RhsScalar>::size : 1,
RealPacketSize = Vectorizable ? unpacket_traits<RealPacket>::size : 1,
@@ -931,25 +931,25 @@
typedef Scalar RhsScalar;
typedef Scalar ResScalar;
- PACKET_DECL_COND_PREFIX(_, Lhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Rhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Res, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Real, PacketSize_);
- PACKET_DECL_COND_SCALAR_PREFIX(_, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Lhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Rhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Res, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Real, PacketSize_);
+ PACKET_DECL_COND_SCALAR_POSTFIX(_, PacketSize_);
-#undef PACKET_DECL_COND_SCALAR_PREFIX
-#undef PACKET_DECL_COND_PREFIX
+#undef PACKET_DECL_COND_SCALAR_POSTFIX
+#undef PACKET_DECL_COND_POSTFIX
#undef PACKET_DECL_COND_SCALAR
#undef PACKET_DECL_COND
enum {
ConjLhs = false,
ConjRhs = ConjRhs_,
- Vectorizable = unpacket_traits<_RealPacket>::vectorizable
- && unpacket_traits<_ScalarPacket>::vectorizable,
- LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1,
- RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1,
- ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1,
+ Vectorizable = unpacket_traits<RealPacket_>::vectorizable
+ && unpacket_traits<ScalarPacket_>::vectorizable,
+ LhsPacketSize = Vectorizable ? unpacket_traits<LhsPacket_>::size : 1,
+ RhsPacketSize = Vectorizable ? unpacket_traits<RhsPacket_>::size : 1,
+ ResPacketSize = Vectorizable ? unpacket_traits<ResPacket_>::size : 1,
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
// FIXME: should depend on NumberOfRegisters
@@ -960,9 +960,9 @@
RhsProgress = 1
};
- typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
- typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
- typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+ typedef typename conditional<Vectorizable,LhsPacket_,LhsScalar>::type LhsPacket;
+ typedef typename conditional<Vectorizable,RhsPacket_,RhsScalar>::type RhsPacket;
+ typedef typename conditional<Vectorizable,ResPacket_,ResScalar>::type ResPacket;
typedef LhsPacket LhsPacket4Packing;
typedef QuadPacket<RhsPacket> RhsPacketx4;
typedef ResPacket AccPacket;
diff --git a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
index 465294b..9728f30 100644
--- a/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
+++ b/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -213,13 +213,13 @@
typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
typedef internal::blas_traits<Lhs> LhsBlasTraits;
typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
- typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+ typedef typename internal::remove_all<ActualLhs>::type ActualLhs_;
typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
typedef internal::blas_traits<Rhs> RhsBlasTraits;
typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
- typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+ typedef typename internal::remove_all<ActualRhs>::type ActualRhs_;
typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
@@ -229,19 +229,19 @@
enum {
StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
- UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1,
- UseRhsDirectly = _ActualRhs::InnerStrideAtCompileTime==1
+ UseLhsDirectly = ActualLhs_::InnerStrideAtCompileTime==1,
+ UseRhsDirectly = ActualRhs_::InnerStrideAtCompileTime==1
};
internal::gemv_static_vector_if<Scalar,Lhs::SizeAtCompileTime,Lhs::MaxSizeAtCompileTime,!UseLhsDirectly> static_lhs;
ei_declare_aligned_stack_constructed_variable(Scalar, actualLhsPtr, actualLhs.size(),
(UseLhsDirectly ? const_cast<Scalar*>(actualLhs.data()) : static_lhs.data()));
- if(!UseLhsDirectly) Map<typename _ActualLhs::PlainObject>(actualLhsPtr, actualLhs.size()) = actualLhs;
+ if(!UseLhsDirectly) Map<typename ActualLhs_::PlainObject>(actualLhsPtr, actualLhs.size()) = actualLhs;
internal::gemv_static_vector_if<Scalar,Rhs::SizeAtCompileTime,Rhs::MaxSizeAtCompileTime,!UseRhsDirectly> static_rhs;
ei_declare_aligned_stack_constructed_variable(Scalar, actualRhsPtr, actualRhs.size(),
(UseRhsDirectly ? const_cast<Scalar*>(actualRhs.data()) : static_rhs.data()));
- if(!UseRhsDirectly) Map<typename _ActualRhs::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+ if(!UseRhsDirectly) Map<typename ActualRhs_::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
@@ -259,13 +259,13 @@
typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
typedef internal::blas_traits<Lhs> LhsBlasTraits;
typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
- typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+ typedef typename internal::remove_all<ActualLhs>::type ActualLhs_;
typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
typedef internal::blas_traits<Rhs> RhsBlasTraits;
typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
- typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+ typedef typename internal::remove_all<ActualRhs>::type ActualRhs_;
typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
@@ -275,8 +275,8 @@
enum {
IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
- LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0,
- RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0,
+ LhsIsRowMajor = ActualLhs_::Flags&RowMajorBit ? 1 : 0,
+ RhsIsRowMajor = ActualRhs_::Flags&RowMajorBit ? 1 : 0,
SkipDiag = (UpLo&(UnitDiag|ZeroDiag))!=0
};
@@ -286,7 +286,7 @@
Index depth = actualLhs.cols();
typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,typename Lhs::Scalar,typename Rhs::Scalar,
- MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualRhs::MaxColsAtCompileTime> BlockingType;
+ MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, ActualRhs_::MaxColsAtCompileTime> BlockingType;
BlockingType blocking(size, size, depth, 1, false);
diff --git a/Eigen/src/Core/products/GeneralMatrixVector.h b/Eigen/src/Core/products/GeneralMatrixVector.h
index b775dbe..e7dd61d 100644
--- a/Eigen/src/Core/products/GeneralMatrixVector.h
+++ b/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -36,31 +36,31 @@
{
typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
-#define PACKET_DECL_COND_PREFIX(prefix, name, packet_size) \
+#define PACKET_DECL_COND_POSTFIX(postfix, name, packet_size) \
typedef typename gemv_packet_cond<packet_size, \
typename packet_traits<name ## Scalar>::type, \
typename packet_traits<name ## Scalar>::half, \
typename unpacket_traits<typename packet_traits<name ## Scalar>::half>::half>::type \
- prefix ## name ## Packet
+ name ## Packet ## postfix
- PACKET_DECL_COND_PREFIX(_, Lhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Rhs, PacketSize_);
- PACKET_DECL_COND_PREFIX(_, Res, PacketSize_);
-#undef PACKET_DECL_COND_PREFIX
+ PACKET_DECL_COND_POSTFIX(_, Lhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Rhs, PacketSize_);
+ PACKET_DECL_COND_POSTFIX(_, Res, PacketSize_);
+#undef PACKET_DECL_COND_POSTFIX
public:
enum {
- Vectorizable = unpacket_traits<_LhsPacket>::vectorizable &&
- unpacket_traits<_RhsPacket>::vectorizable &&
- int(unpacket_traits<_LhsPacket>::size)==int(unpacket_traits<_RhsPacket>::size),
- LhsPacketSize = Vectorizable ? unpacket_traits<_LhsPacket>::size : 1,
- RhsPacketSize = Vectorizable ? unpacket_traits<_RhsPacket>::size : 1,
- ResPacketSize = Vectorizable ? unpacket_traits<_ResPacket>::size : 1
+ Vectorizable = unpacket_traits<LhsPacket_>::vectorizable &&
+ unpacket_traits<RhsPacket_>::vectorizable &&
+ int(unpacket_traits<LhsPacket_>::size)==int(unpacket_traits<RhsPacket_>::size),
+ LhsPacketSize = Vectorizable ? unpacket_traits<LhsPacket_>::size : 1,
+ RhsPacketSize = Vectorizable ? unpacket_traits<RhsPacket_>::size : 1,
+ ResPacketSize = Vectorizable ? unpacket_traits<ResPacket_>::size : 1
};
- typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
- typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
- typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+ typedef typename conditional<Vectorizable,LhsPacket_,LhsScalar>::type LhsPacket;
+ typedef typename conditional<Vectorizable,RhsPacket_,RhsScalar>::type RhsPacket;
+ typedef typename conditional<Vectorizable,ResPacket_,ResScalar>::type ResPacket;
};
diff --git a/Eigen/src/Core/products/SelfadjointProduct.h b/Eigen/src/Core/products/SelfadjointProduct.h
index 629ce1d..026bc19 100644
--- a/Eigen/src/Core/products/SelfadjointProduct.h
+++ b/Eigen/src/Core/products/SelfadjointProduct.h
@@ -57,14 +57,14 @@
typedef typename MatrixType::Scalar Scalar;
typedef internal::blas_traits<OtherType> OtherBlasTraits;
typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
- typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+ typedef typename internal::remove_all<ActualOtherType>::type ActualOtherType_;
typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
enum {
StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
- UseOtherDirectly = _ActualOtherType::InnerStrideAtCompileTime==1
+ UseOtherDirectly = ActualOtherType_::InnerStrideAtCompileTime==1
};
internal::gemv_static_vector_if<Scalar,OtherType::SizeAtCompileTime,OtherType::MaxSizeAtCompileTime,!UseOtherDirectly> static_other;
@@ -72,7 +72,7 @@
(UseOtherDirectly ? const_cast<Scalar*>(actualOther.data()) : static_other.data()));
if(!UseOtherDirectly)
- Map<typename _ActualOtherType::PlainObject>(actualOtherPtr, actualOther.size()) = actualOther;
+ Map<typename ActualOtherType_::PlainObject>(actualOtherPtr, actualOther.size()) = actualOther;
selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
OtherBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex,
@@ -89,21 +89,21 @@
typedef typename MatrixType::Scalar Scalar;
typedef internal::blas_traits<OtherType> OtherBlasTraits;
typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
- typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+ typedef typename internal::remove_all<ActualOtherType>::type ActualOtherType_;
typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
enum {
IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
- OtherIsRowMajor = _ActualOtherType::Flags&RowMajorBit ? 1 : 0
+ OtherIsRowMajor = ActualOtherType_::Flags&RowMajorBit ? 1 : 0
};
Index size = mat.cols();
Index depth = actualOther.cols();
typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,Scalar,Scalar,
- MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualOtherType::MaxColsAtCompileTime> BlockingType;
+ MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, ActualOtherType_::MaxColsAtCompileTime> BlockingType;
BlockingType blocking(size, size, depth, 1, false);
diff --git a/Eigen/src/Core/products/SelfadjointRank2Update.h b/Eigen/src/Core/products/SelfadjointRank2Update.h
index 3153009..44c3381 100644
--- a/Eigen/src/Core/products/SelfadjointRank2Update.h
+++ b/Eigen/src/Core/products/SelfadjointRank2Update.h
@@ -65,12 +65,12 @@
{
typedef internal::blas_traits<DerivedU> UBlasTraits;
typedef typename UBlasTraits::DirectLinearAccessType ActualUType;
- typedef typename internal::remove_all<ActualUType>::type _ActualUType;
+ typedef typename internal::remove_all<ActualUType>::type ActualUType_;
typename internal::add_const_on_value_type<ActualUType>::type actualU = UBlasTraits::extract(u.derived());
typedef internal::blas_traits<DerivedV> VBlasTraits;
typedef typename VBlasTraits::DirectLinearAccessType ActualVType;
- typedef typename internal::remove_all<ActualVType>::type _ActualVType;
+ typedef typename internal::remove_all<ActualVType>::type ActualVType_;
typename internal::add_const_on_value_type<ActualVType>::type actualV = VBlasTraits::extract(v.derived());
// If MatrixType is row major, then we use the routine for lower triangular in the upper triangular case and
@@ -82,8 +82,8 @@
if (IsRowMajor)
actualAlpha = numext::conj(actualAlpha);
- typedef typename internal::remove_all<typename internal::conj_expr_if<int(IsRowMajor) ^ int(UBlasTraits::NeedToConjugate), _ActualUType>::type>::type UType;
- typedef typename internal::remove_all<typename internal::conj_expr_if<int(IsRowMajor) ^ int(VBlasTraits::NeedToConjugate), _ActualVType>::type>::type VType;
+ typedef typename internal::remove_all<typename internal::conj_expr_if<int(IsRowMajor) ^ int(UBlasTraits::NeedToConjugate), ActualUType_>::type>::type UType;
+ typedef typename internal::remove_all<typename internal::conj_expr_if<int(IsRowMajor) ^ int(VBlasTraits::NeedToConjugate), ActualVType_>::type>::type VType;
internal::selfadjoint_rank2_update_selector<Scalar, Index, UType, VType,
(IsRowMajor ? int(UpLo==Upper ? Lower : Upper) : UpLo)>
::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha);
diff --git a/Eigen/src/Core/util/BlasUtil.h b/Eigen/src/Core/util/BlasUtil.h
index 2268593..d2dd702 100755
--- a/Eigen/src/Core/util/BlasUtil.h
+++ b/Eigen/src/Core/util/BlasUtil.h
@@ -405,7 +405,7 @@
{
typedef typename traits<XprType>::Scalar Scalar;
typedef const XprType& ExtractType;
- typedef XprType _ExtractType;
+ typedef XprType ExtractType_;
enum {
IsComplex = NumTraits<Scalar>::IsComplex,
IsTransposed = false,
@@ -418,7 +418,7 @@
};
typedef typename conditional<bool(HasUsableDirectAccess),
ExtractType,
- typename _ExtractType::PlainObject
+ typename ExtractType_::PlainObject
>::type DirectLinearAccessType;
static inline EIGEN_DEVICE_FUNC ExtractType extract(const XprType& x) { return x; }
static inline EIGEN_DEVICE_FUNC const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
@@ -500,8 +500,8 @@
typedef typename NestedXpr::Scalar Scalar;
typedef blas_traits<NestedXpr> Base;
typedef Transpose<NestedXpr> XprType;
- typedef Transpose<const typename Base::_ExtractType> ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS
- typedef Transpose<const typename Base::_ExtractType> _ExtractType;
+ typedef Transpose<const typename Base::ExtractType_> ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS
+ typedef Transpose<const typename Base::ExtractType_> ExtractType_;
typedef typename conditional<bool(Base::HasUsableDirectAccess),
ExtractType,
typename ExtractType::PlainObject
diff --git a/Eigen/src/Core/util/ConfigureVectorization.h b/Eigen/src/Core/util/ConfigureVectorization.h
index ba2049b..4ba1370 100644
--- a/Eigen/src/Core/util/ConfigureVectorization.h
+++ b/Eigen/src/Core/util/ConfigureVectorization.h
@@ -420,9 +420,11 @@
// We can use the optimized fp16 to float and float to fp16 conversion routines
#define EIGEN_HAS_FP16_C
- #if EIGEN_COMP_CLANG
- // Workaround for clang: The FP16C intrinsics for clang are included by
- // immintrin.h, as opposed to emmintrin.h as suggested by Intel:
+ #if EIGEN_COMP_GNUC
+ // Make sure immintrin.h is included, even if e.g. vectorization is
+ // explicitly disabled (see also issue #2395).
+ // Note that FP16C intrinsics for gcc and clang are included by immintrin.h,
+ // as opposed to emmintrin.h as suggested by Intel:
// https://software.intel.com/sites/landingpage/IntrinsicsGuide/#othertechs=FP16C&expand=1711
#include <immintrin.h>
#endif
diff --git a/Eigen/src/Core/util/DisableStupidWarnings.h b/Eigen/src/Core/util/DisableStupidWarnings.h
index d7e8a13..7e1c76b 100755
--- a/Eigen/src/Core/util/DisableStupidWarnings.h
+++ b/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -54,7 +54,7 @@
#pragma clang diagnostic ignored "-Wc11-extensions"
#endif
-#elif defined __GNUC__
+#elif defined __GNUC__ && !defined(__FUJITSU)
#if (!defined(EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS)) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6))
#pragma GCC diagnostic push
diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h
index 2f33599..f3df372 100644
--- a/Eigen/src/Core/util/ForwardDeclarations.h
+++ b/Eigen/src/Core/util/ForwardDeclarations.h
@@ -42,6 +42,11 @@
template< typename T> struct evaluator;
+template<typename T> struct has_trivially_copyable_storage
+{
+ static const bool value = false;
+};
+
} // end namespace internal
template<typename T> struct NumTraits;
@@ -78,7 +83,6 @@
template<typename MatrixType> class Conjugate;
template<typename NullaryOp, typename MatrixType> class CwiseNullaryOp;
template<typename UnaryOp, typename MatrixType> class CwiseUnaryOp;
-template<typename ViewOp, typename MatrixType> class CwiseUnaryView;
template<typename BinaryOp, typename Lhs, typename Rhs> class CwiseBinaryOp;
template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3> class CwiseTernaryOp;
template<typename Decomposition, typename Rhstype> class Solve;
@@ -87,7 +91,7 @@
template<typename Lhs, typename Rhs, int Option = DefaultProduct> class Product;
template<typename Derived> class DiagonalBase;
-template<typename _DiagonalVectorType> class DiagonalWrapper;
+template<typename DiagonalVectorType_> class DiagonalWrapper;
template<typename Scalar_, int SizeAtCompileTime, int MaxSizeAtCompileTime=SizeAtCompileTime> class DiagonalMatrix;
template<typename MatrixType, typename DiagonalType, int ProductOrder> class DiagonalProduct;
template<typename MatrixType, int Index = 0> class Diagonal;
@@ -108,6 +112,7 @@
template<typename Derived> class RefBase;
template<typename PlainObjectType, int Options = 0,
typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
+template<typename ViewOp, typename MatrixType, typename StrideType = Stride<0,0>> class CwiseUnaryView;
template<typename Derived> class TriangularBase;
template<typename MatrixType, unsigned int Mode> class TriangularView;
@@ -272,8 +277,8 @@
template<typename Scalar,int Dim> class AlignedBox;
template<typename Scalar, int Options = AutoAlign> class Quaternion;
template<typename Scalar,int Dim,int Mode,int Options_=AutoAlign> class Transform;
-template <typename Scalar_, int _AmbientDim, int Options=AutoAlign> class ParametrizedLine;
-template <typename Scalar_, int _AmbientDim, int Options=AutoAlign> class Hyperplane;
+template <typename Scalar_, int AmbientDim_, int Options=AutoAlign> class ParametrizedLine;
+template <typename Scalar_, int AmbientDim_, int Options=AutoAlign> class Hyperplane;
template<typename Scalar> class UniformScaling;
template<typename MatrixType,int Direction> class Homogeneous;
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 0aafb9e..d0ac15a 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -91,13 +91,20 @@
#define EIGEN_COMP_LLVM 0
#endif
-/// \internal EIGEN_COMP_ICC set to __INTEL_COMPILER if the compiler is Intel compiler, 0 otherwise
+/// \internal EIGEN_COMP_ICC set to __INTEL_COMPILER if the compiler is Intel icc compiler, 0 otherwise
#if defined(__INTEL_COMPILER)
#define EIGEN_COMP_ICC __INTEL_COMPILER
#else
#define EIGEN_COMP_ICC 0
#endif
+/// \internal EIGEN_COMP_CLANGICC set to __INTEL_CLANG_COMPILER if the compiler is Intel icx compiler, 0 otherwise
+#if defined(__INTEL_CLANG_COMPILER)
+ #define EIGEN_COMP_CLANGICC __INTEL_CLANG_COMPILER
+#else
+ #define EIGEN_COMP_CLANGICC 0
+#endif
+
/// \internal EIGEN_COMP_MINGW set to 1 if the compiler is mingw
#if defined(__MINGW32__)
#define EIGEN_COMP_MINGW 1
@@ -197,9 +204,45 @@
#define EIGEN_COMP_EMSCRIPTEN 0
#endif
+/// \internal EIGEN_COMP_FCC set to FCC version if the compiler is Fujitsu Compiler (traditional mode)
+/// \note The Fujitsu C/C++ compiler uses the traditional mode based
+/// on EDG g++ 6.1 by default or if envoked with the -Nnoclang flag
+#if defined(__FUJITSU)
+ #define EIGEN_COMP_FCC (__FCC_major__*100+__FCC_minor__*10+__FCC_patchlevel__)
+#else
+ #define EIGEN_COMP_FCC 0
+#endif
+
+/// \internal EIGEN_COMP_CLANGFCC set to FCC version if the compiler is Fujitsu Compiler (Clang mode)
+/// \note The Fujitsu C/C++ compiler uses the non-traditional mode
+/// based on Clang 7.1.0 if envoked with the -Nclang flag
+#if defined(__CLANG_FUJITSU)
+ #define EIGEN_COMP_CLANGFCC (__FCC_major__*100+__FCC_minor__*10+__FCC_patchlevel__)
+#else
+ #define EIGEN_COMP_CLANGFCC 0
+#endif
+
+/// \internal EIGEN_COMP_CPE set to CPE version if the compiler is HPE Cray Compiler (GCC based)
+/// \note This is the SVE-enabled C/C++ compiler from the HPE Cray
+/// Programming Environment (CPE) based on Cray GCC 8.1
+#if defined(_CRAYC) && !defined(__clang__)
+ #define EIGEN_COMP_CPE (_RELEASE_MAJOR*100+_RELEASE_MINOR*10+_RELEASE_PATCHLEVEL)
+#else
+ #define EIGEN_COMP_CPE 0
+#endif
+
+/// \internal EIGEN_COMP_CLANGCPE set to CPE version if the compiler is HPE Cray Compiler (Clang based)
+/// \note This is the C/C++ compiler from the HPE Cray Programming
+/// Environment (CPE) based on Cray Clang 11.0 without SVE-support
+#if defined(_CRAYC) && defined(__clang__)
+ #define EIGEN_COMP_CLANGCPE (_RELEASE_MAJOR*100+_RELEASE_MINOR*10+_RELEASE_PATCHLEVEL)
+#else
+ #define EIGEN_COMP_CLANGCPE 0
+#endif
+
/// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.)
-#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM || EIGEN_COMP_EMSCRIPTEN)
+#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_CLANGICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM || EIGEN_COMP_EMSCRIPTEN || EIGEN_COMP_FCC || EIGEN_COMP_CLANGFCC || EIGEN_COMP_CPE || EIGEN_COMP_CLANGCPE)
#define EIGEN_COMP_GNUC_STRICT 1
#else
#define EIGEN_COMP_GNUC_STRICT 0
@@ -616,6 +659,7 @@
#define EIGEN_COMP_CXXVER 03
#endif
+
// The macros EIGEN_HAS_CXX?? defines a rough estimate of available c++ features
// but in practice we should not rely on them but rather on the availability of
// individual features as defined later.
@@ -1044,6 +1088,16 @@
}
#endif
+/**
+ * \internal
+ * \brief Macro for conditionally trivial special member functions for supporting trivially copyable types.
+ * This feature is officially known as C++20's P0848R3 and is enabled on supported compilers.
+ */
+#if (EIGEN_COMP_CXXVER >= 20) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MSVC_STRICT)
+#define EIGEN_COMP_HAS_P0848R3 1
+#else
+#define EIGEN_COMP_HAS_P0848R3 0
+#endif
/**
* \internal
diff --git a/Eigen/src/Core/util/Meta.h b/Eigen/src/Core/util/Meta.h
index b641f97..c0232c2 100755
--- a/Eigen/src/Core/util/Meta.h
+++ b/Eigen/src/Core/util/Meta.h
@@ -212,20 +212,33 @@
enum { value = N };
};
+
/** \internal
- * Analogue of the std::size free function.
- * It returns the size of the container or view \a x of type \c T
+ * Analogue of the std::ssize free function.
+ * It returns the signed size of the container or view \a x of type \c T
*
* It currently supports:
* - any types T defining a member T::size() const
* - plain C arrays as T[N]
*
+ * For C++20, this function just forwards to `std::ssize`, or any ADL discoverable `ssize` function.
*/
-template<typename T>
-EIGEN_CONSTEXPR Index size(const T& x) { return x.size(); }
+#if EIGEN_COMP_CXXVER < 20
+template <typename T>
+EIGEN_CONSTEXPR auto index_list_size(const T& x) {
+ using R = std::common_type_t<std::ptrdiff_t, std::make_signed_t<decltype(x.size())>>;
+ return static_cast<R>(x.size());
+}
-template<typename T,std::size_t N>
-EIGEN_CONSTEXPR Index size(const T (&) [N]) { return N; }
+template<typename T, std::ptrdiff_t N>
+EIGEN_CONSTEXPR std::ptrdiff_t index_list_size(const T (&)[N]) { return N; }
+#else
+template <typename T>
+EIGEN_CONSTEXPR auto index_list_size(T&& x) {
+ using std::ssize;
+ return ssize(std::forward<T>(x));
+}
+#endif // EIGEN_COMP_CXXVER
/** \internal
* Convenient struct to get the result type of a nullary, unary, binary, or
@@ -646,7 +659,7 @@
/// \internal Calculate logical XOR at compile time
inline constexpr bool logical_xor(bool a, bool b) {
- return (a || b) && !(a && b);
+ return a != b;
}
/// \internal Calculate logical IMPLIES at compile time
diff --git a/Eigen/src/Core/util/ReenableStupidWarnings.h b/Eigen/src/Core/util/ReenableStupidWarnings.h
index f830de1..7021e6d 100644
--- a/Eigen/src/Core/util/ReenableStupidWarnings.h
+++ b/Eigen/src/Core/util/ReenableStupidWarnings.h
@@ -12,7 +12,7 @@
#pragma warning pop
#elif defined __clang__
#pragma clang diagnostic pop
- #elif defined __GNUC__ && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6))
+ #elif defined __GNUC__ && !defined(__FUJITSU) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6))
#pragma GCC diagnostic pop
#endif
diff --git a/Eigen/src/Core/util/Serializer.h b/Eigen/src/Core/util/Serializer.h
index 7ec5eed..b77c5de 100644
--- a/Eigen/src/Core/util/Serializer.h
+++ b/Eigen/src/Core/util/Serializer.h
@@ -45,11 +45,14 @@
/**
* Serializes a value to a byte buffer.
- * \param dest the destination buffer.
- * \param T the value to serialize.
+ * \param dest the destination buffer; if this is nullptr, does nothing.
+ * \param end the end of the destination buffer.
+ * \param value the value to serialize.
* \return the next memory address past the end of the serialized data.
*/
- EIGEN_DEVICE_FUNC uint8_t* serialize(uint8_t* dest, const T& value) {
+ EIGEN_DEVICE_FUNC uint8_t* serialize(uint8_t* dest, uint8_t* end, const T& value) {
+ if (EIGEN_PREDICT_FALSE(dest == nullptr)) return nullptr;
+ if (EIGEN_PREDICT_FALSE(dest + sizeof(value) > end)) return nullptr;
EIGEN_USING_STD(memcpy)
memcpy(dest, &value, sizeof(value));
return dest + sizeof(value);
@@ -57,11 +60,14 @@
/**
* Deserializes a value from a byte buffer.
- * \param src the source buffer.
+ * \param src the source buffer; if this is nullptr, does nothing.
+ * \param end the end of the source buffer.
* \param value the value to populate.
- * \return the next unprocessed memory address.
+ * \return the next unprocessed memory address; nullptr if parsing errors are detected.
*/
- EIGEN_DEVICE_FUNC uint8_t* deserialize(uint8_t* src, T& value) const {
+ EIGEN_DEVICE_FUNC const uint8_t* deserialize(const uint8_t* src, const uint8_t* end, T& value) const {
+ if (EIGEN_PREDICT_FALSE(src == nullptr)) return nullptr;
+ if (EIGEN_PREDICT_FALSE(src + sizeof(value) > end)) return nullptr;
EIGEN_USING_STD(memcpy)
memcpy(&value, src, sizeof(value));
return src + sizeof(value);
@@ -84,7 +90,9 @@
return sizeof(Header) + sizeof(Scalar) * value.size();
}
- EIGEN_DEVICE_FUNC uint8_t* serialize(uint8_t* dest, const Derived& value) {
+ EIGEN_DEVICE_FUNC uint8_t* serialize(uint8_t* dest, uint8_t* end, const Derived& value) {
+ if (EIGEN_PREDICT_FALSE(dest == nullptr)) return nullptr;
+ if (EIGEN_PREDICT_FALSE(dest + size(value) > end)) return nullptr;
const size_t header_bytes = sizeof(Header);
const size_t data_bytes = sizeof(Scalar) * value.size();
Header header = {value.rows(), value.cols()};
@@ -95,14 +103,17 @@
return dest + data_bytes;
}
- EIGEN_DEVICE_FUNC uint8_t* deserialize(uint8_t* src, Derived& value) const {
+ EIGEN_DEVICE_FUNC const uint8_t* deserialize(const uint8_t* src, const uint8_t* end, Derived& value) const {
+ if (EIGEN_PREDICT_FALSE(src == nullptr)) return nullptr;
+ if (EIGEN_PREDICT_FALSE(src + sizeof(Header) > end)) return nullptr;
const size_t header_bytes = sizeof(Header);
Header header;
EIGEN_USING_STD(memcpy)
memcpy(&header, src, header_bytes);
src += header_bytes;
- value.resize(header.rows, header.cols);
const size_t data_bytes = sizeof(Scalar) * header.rows * header.cols;
+ if (EIGEN_PREDICT_FALSE(src + data_bytes > end)) return nullptr;
+ value.resize(header.rows, header.cols);
memcpy(value.data(), src, data_bytes);
return src + data_bytes;
}
@@ -134,17 +145,17 @@
}
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- uint8_t* serialize(uint8_t* dest, const T1& value, const Ts&... args) {
+ uint8_t* serialize(uint8_t* dest, uint8_t* end, const T1& value, const Ts&... args) {
Serializer serializer;
- dest = serializer.serialize(dest, value);
- return serialize_impl<N-1, Ts...>::serialize(dest, args...);
+ dest = serializer.serialize(dest, end, value);
+ return serialize_impl<N-1, Ts...>::serialize(dest, end, args...);
}
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- uint8_t* deserialize(uint8_t* src, T1& value, Ts&... args) {
+ const uint8_t* deserialize(const uint8_t* src, const uint8_t* end, T1& value, Ts&... args) {
Serializer serializer;
- src = serializer.deserialize(src, value);
- return serialize_impl<N-1, Ts...>::deserialize(src, args...);
+ src = serializer.deserialize(src, end, value);
+ return serialize_impl<N-1, Ts...>::deserialize(src, end, args...);
}
};
@@ -155,10 +166,10 @@
size_t serialize_size() { return 0; }
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- uint8_t* serialize(uint8_t* dest) { return dest; }
+ uint8_t* serialize(uint8_t* dest, uint8_t* /*end*/) { return dest; }
static EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
- uint8_t* deserialize(uint8_t* src) { return src; }
+ const uint8_t* deserialize(const uint8_t* src, const uint8_t* /*end*/) { return src; }
};
} // namespace internal
@@ -179,27 +190,29 @@
/**
* Serialize a set of values to the byte buffer.
*
- * \param dest output byte buffer.
+ * \param dest output byte buffer; if this is nullptr, does nothing.
+ * \param end the end of the output byte buffer.
* \param args ... arguments to serialize in sequence.
* \return the next address after all serialized values.
*/
template<typename... Args>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-uint8_t* serialize(uint8_t* dest, const Args&... args) {
- return internal::serialize_impl<sizeof...(args), Args...>::serialize(dest, args...);
+uint8_t* serialize(uint8_t* dest, uint8_t* end, const Args&... args) {
+ return internal::serialize_impl<sizeof...(args), Args...>::serialize(dest, end, args...);
}
/**
* Deserialize a set of values from the byte buffer.
*
- * \param src input byte buffer.
+ * \param src input byte buffer; if this is nullptr, does nothing.
+ * \param end the end of input byte buffer.
* \param args ... arguments to deserialize in sequence.
- * \return the next address after all parsed values.
+ * \return the next address after all parsed values; nullptr if parsing errors are detected.
*/
template<typename... Args>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-uint8_t* deserialize(uint8_t* src, Args&... args) {
- return internal::serialize_impl<sizeof...(args), Args...>::deserialize(src, args...);
+const uint8_t* deserialize(const uint8_t* src, const uint8_t* end, Args&... args) {
+ return internal::serialize_impl<sizeof...(args), Args...>::deserialize(src, end, args...);
}
} // namespace Eigen
diff --git a/Eigen/src/Core/util/XprHelper.h b/Eigen/src/Core/util/XprHelper.h
index a1314e9..73bad6c 100644
--- a/Eigen/src/Core/util/XprHelper.h
+++ b/Eigen/src/Core/util/XprHelper.h
@@ -180,32 +180,8 @@
template<typename T> struct unpacket_traits;
-// If we vectorize regardless of alignment, pick the full-sized packet if:
-//
-// * The size is large enough;
-// * Picking it will result in less operations than picking the half size.
-// Consider the case where the size is 12, the full packet is 8, and the
-// half packet is 4. If we pick the full packet we'd have 1 + 4 operations,
-// but only 3 operations if we pick the half-packet.
-//
-// The reason why we only do this with EIGEN_UNALIGNED_VECTORIZE is that if
-// we chose packets which do not divide the data size exactly we're going to
-// be left with some possibly unaligned data at the end.
-#if EIGEN_UNALIGNED_VECTORIZE
-template<int Size, typename PacketType,
- bool Stop =
- Size==Dynamic ||
- (Size >= unpacket_traits<PacketType>::size &&
- // If the packet size is 1 we're always good -- it will always divide things perfectly.
- // We have this check since otherwise 1/2 would be 0 in the division below.
- (unpacket_traits<PacketType>::size == 1 ||
- (Size/unpacket_traits<PacketType>::size + Size%unpacket_traits<PacketType>::size) <=
- (Size/(unpacket_traits<PacketType>::size/2) + Size%(unpacket_traits<PacketType>::size/2)))) ||
- is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
-#else
template<int Size, typename PacketType,
bool Stop = Size==Dynamic || (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
-#endif
struct find_best_packet_helper;
template< int Size, typename PacketType>
@@ -527,8 +503,8 @@
template<typename XprType, typename CastType> struct cast_return_type
{
typedef typename XprType::Scalar CurrentScalarType;
- typedef typename remove_all<CastType>::type _CastType;
- typedef typename _CastType::Scalar NewScalarType;
+ typedef typename remove_all<CastType>::type CastType_;
+ typedef typename CastType_::Scalar NewScalarType;
typedef typename conditional<is_same<CurrentScalarType,NewScalarType>::value,
const XprType&,CastType>::type type;
};
