diff options
author | Rasmus Munk Larsen <rmlarsen@google.com> | 2020-10-07 19:05:18 +0000 |
---|---|---|
committer | Rasmus Munk Larsen <rmlarsen@google.com> | 2020-10-07 19:05:18 +0000 |
commit | b43102440489df9d0175c88e602dfa425b574a94 (patch) | |
tree | 9325c3401de7047451d4a59ad343cdf1c5a83679 /Eigen/src | |
parent | f66f3393e3d567e5c8b138fbad69b316214a4ce9 (diff) |
Don't make assumptions about NaN-propagation for pmin/pmax - it various across platforms.
Change test to only test for NaN-propagation for pfmin/pfmax.
Diffstat (limited to 'Eigen/src')
-rw-r--r-- | Eigen/src/Core/GenericPacketMath.h | 55 | ||||
-rw-r--r-- | Eigen/src/Core/functors/BinaryFunctors.h | 58 | ||||
-rw-r--r-- | Eigen/src/Core/util/Constants.h | 15 | ||||
-rw-r--r-- | Eigen/src/Core/util/ForwardDeclarations.h | 4 |
4 files changed, 104 insertions, 28 deletions
diff --git a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h index 075d18aa6..b5eb1cf99 100644 --- a/Eigen/src/Core/GenericPacketMath.h +++ b/Eigen/src/Core/GenericPacketMath.h @@ -216,12 +216,12 @@ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pdiv(const Packet& a, const Packet& b) { return a/b; } /** \internal \returns the min of \a a and \a b (coeff-wise). -Equivalent to std::min(a, b), so if either a or b is NaN, a is returned. */ + If \a a or \b b is NaN, the return value is implementation defined. */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pmin(const Packet& a, const Packet& b) { return numext::mini(a, b); } /** \internal \returns the max of \a a and \a b (coeff-wise) -Equivalent to std::max(a, b), so if either a or b is NaN, a is returned.*/ + If \a a or \b b is NaN, the return value is implementation defined. */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pmax(const Packet& a, const Packet& b) { return numext::maxi(a, b); } @@ -635,23 +635,54 @@ Packet print(const Packet& a) { using numext::rint; return rint(a); } template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); } -/** \internal \returns the min of \a a and \a b (coeff-wise) - Equivalent to std::fmin(a, b). Only if both a and b are NaN is NaN returned. -*/ + +/** \internal \returns the max of \a a and \a b (coeff-wise) + If both \a a and \a b are NaN, NaN is returned. + Equivalent to std::fmax(a, b). */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pfmax(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmax(a, b), a), + b); +} + +/** \internal \returns the min of \a a and \a b (coeff-wise) + If both \a a and \a b are NaN, NaN is returned. + Equivalent to std::fmin(a, b). */ template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pfmin(const Packet& a, const Packet& b) { - Packet not_nan_mask = pcmp_eq(a, a); - return pselect(not_nan_mask, pmin(a, b), b); + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmin(a, b), a), + b); } /** \internal \returns the max of \a a and \a b (coeff-wise) - Equivalent to std::fmax(a, b). Only if both a and b are NaN is NaN returned.*/ -template<typename Packet> EIGEN_DEVICE_FUNC inline Packet -pfmax(const Packet& a, const Packet& b) { - Packet not_nan_mask = pcmp_eq(a, a); - return pselect(not_nan_mask, pmax(a, b), b); + If either \a a or \a b are NaN, NaN is returned. */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pfmax_nan(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmax(a, b), b), + a); +} + +/** \internal \returns the min of \a a and \a b (coeff-wise) + If either \a a or \a b are NaN, NaN is returned. */ +template<typename Packet> EIGEN_DEVICE_FUNC inline Packet +pfmin_nan(const Packet& a, const Packet& b) { + Packet not_nan_mask_a = pcmp_eq(a, a); + Packet not_nan_mask_b = pcmp_eq(b, b); + return pselect(not_nan_mask_a, + pselect(not_nan_mask_b, pmin(a, b), b), + a); } + /*************************************************************************** * The following functions might not have to be overwritten for vectorized types ***************************************************************************/ diff --git a/Eigen/src/Core/functors/BinaryFunctors.h b/Eigen/src/Core/functors/BinaryFunctors.h index d8b7b1eba..55650bb8d 100644 --- a/Eigen/src/Core/functors/BinaryFunctors.h +++ b/Eigen/src/Core/functors/BinaryFunctors.h @@ -134,21 +134,39 @@ struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > { * * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff() */ -template<typename LhsScalar,typename RhsScalar> +template<typename LhsScalar,typename RhsScalar, int NaNPropagation> struct scalar_min_op : binary_op_base<LhsScalar,RhsScalar> { typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_min_op>::ReturnType result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::mini(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const LhsScalar& a, const RhsScalar& b) const { + if (NaNPropagation == PropagateFast) { + return numext::mini(a, b); + } else if (NaNPropagation == PropagateNumbers) { + return internal::pfmin(a,b); + } else if (NaNPropagation == PropagateNaN) { + return internal::pfmin_nan(a,b); + } + } template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pmin(a,b); } + { + if (NaNPropagation == PropagateFast) { + return internal::pmin(a,b); + } else if (NaNPropagation == PropagateNumbers) { + return internal::pfmin(a,b); + } else if (NaNPropagation == PropagateNaN) { + return internal::pfmin_nan(a,b); + } + } + // TODO(rmlarsen): Handle all NaN propagation semantics reductions. template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type predux(const Packet& a) const { return internal::predux_min(a); } }; -template<typename LhsScalar,typename RhsScalar> -struct functor_traits<scalar_min_op<LhsScalar,RhsScalar> > { + +template<typename LhsScalar,typename RhsScalar, int NaNPropagation> +struct functor_traits<scalar_min_op<LhsScalar,RhsScalar, NaNPropagation> > { enum { Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMin @@ -160,21 +178,39 @@ struct functor_traits<scalar_min_op<LhsScalar,RhsScalar> > { * * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff() */ -template<typename LhsScalar,typename RhsScalar> -struct scalar_max_op : binary_op_base<LhsScalar,RhsScalar> +template<typename LhsScalar,typename RhsScalar, int NaNPropagation> +struct scalar_max_op : binary_op_base<LhsScalar,RhsScalar> { typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_max_op>::ReturnType result_type; EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op) - EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::maxi(a, b); } + EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const LhsScalar& a, const RhsScalar& b) const { + if (NaNPropagation == PropagateFast) { + return numext::maxi(a, b); + } else if (NaNPropagation == PropagateNumbers) { + return internal::pfmax(a,b); + } else if (NaNPropagation == PropagateNaN) { + return internal::pfmax_nan(a,b); + } + } template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet packetOp(const Packet& a, const Packet& b) const - { return internal::pmax(a,b); } + { + if (NaNPropagation == PropagateFast) { + return internal::pmax(a,b); + } else if (NaNPropagation == PropagateNumbers) { + return internal::pfmax(a,b); + } else if (NaNPropagation == PropagateNaN) { + return internal::pfmax_nan(a,b); + } + } + // TODO(rmlarsen): Handle all NaN propagation semantics reductions. template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type predux(const Packet& a) const { return internal::predux_max(a); } }; -template<typename LhsScalar,typename RhsScalar> -struct functor_traits<scalar_max_op<LhsScalar,RhsScalar> > { + +template<typename LhsScalar,typename RhsScalar, int NaNPropagation> +struct functor_traits<scalar_max_op<LhsScalar,RhsScalar, NaNPropagation> > { enum { Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMax diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h index 7ada82195..ad9af5727 100644 --- a/Eigen/src/Core/util/Constants.h +++ b/Eigen/src/Core/util/Constants.h @@ -328,12 +328,21 @@ enum StorageOptions { * Enum for specifying whether to apply or solve on the left or right. */ enum SideType { /** Apply transformation on the left. */ - OnTheLeft = 1, + OnTheLeft = 1, /** Apply transformation on the right. */ - OnTheRight = 2 + OnTheRight = 2 }; - +/** \ingroup enums + * Enum for specifying NaN-propagation behavior, e.g. for coeff-wise min/max. */ +enum NaNPropagationOptions { + /** Implementation defined behavior if NaNs are present. */ + PropagateFast = 0, + /** Always propagate NaNs. */ + PropagateNaN, + /** Always propagate not-NaNs. */ + PropagateNumbers +}; /* the following used to be written as: * diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h index 208b96c9c..2f9cc4491 100644 --- a/Eigen/src/Core/util/ForwardDeclarations.h +++ b/Eigen/src/Core/util/ForwardDeclarations.h @@ -180,8 +180,8 @@ template<typename LhsScalar, typename RhsScalar, bool ConjLhs=false, bool ConjRh template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_sum_op; template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_difference_op; template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_conj_product_op; -template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_min_op; -template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_max_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar, int NaNPropagation=PropagateFast> struct scalar_min_op; +template<typename LhsScalar,typename RhsScalar=LhsScalar, int NaNPropagation=PropagateFast> struct scalar_max_op; template<typename Scalar> struct scalar_opposite_op; template<typename Scalar> struct scalar_conjugate_op; template<typename Scalar> struct scalar_real_op; |