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.

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);
+  Packet not_nan_mask_a = pcmp_eq(a, a);
-  return pselect(not_nan_mask, pmin(a, b), b);
+  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.*/
+    If either \a a or \a b are NaN, NaN is returned. */
 template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
-pfmax(const Packet& a, const Packet& b) {
+pfmax_nan(const Packet& a, const Packet& b) {
-  Packet not_nan_mask = pcmp_eq(a, a);
+  Packet not_nan_mask_a = pcmp_eq(a, a);
-  return pselect(not_nan_mask, pmax(a, b), b);
+  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
 ***************************************************************************/

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>
+template<typename LhsScalar,typename RhsScalar, int NaNPropagation>
-struct scalar_max_op  : binary_op_base<LhsScalar,RhsScalar>
+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

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:
  *

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, int NaNPropagation=PropagateFast> 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_max_op;
 template<typename Scalar> struct scalar_opposite_op;
 template<typename Scalar> struct scalar_conjugate_op;
 template<typename Scalar> struct scalar_real_op;

test/packetmath.cpp

@@ -763,6 +763,20 @@ void packetmath_real<bfloat16, typename internal::packet_traits<bfloat16>::type>
 
 }
 
+template <typename Scalar>
+Scalar propagate_nan_max(const Scalar& a, const Scalar& b) {
+  if ((std::isnan)(a)) return a;
+  if ((std::isnan)(b)) return b;
+  return (std::max)(a,b);
+}
+
+template <typename Scalar>
+Scalar propagate_nan_min(const Scalar& a, const Scalar& b) {
+  if ((std::isnan)(a)) return a;
+  if ((std::isnan)(b)) return b;
+  return (std::min)(a,b);
+}
+
 template <typename Scalar, typename Packet>
 void packetmath_notcomplex() {
   typedef internal::packet_traits<Scalar> PacketTraits;
@@ -829,12 +843,12 @@ void packetmath_notcomplex() {
     data1[i] = internal::random<bool>() ? std::numeric_limits<Scalar>::quiet_NaN() : Scalar(0);
     data1[i + PacketSize] = internal::random<bool>() ? std::numeric_limits<Scalar>::quiet_NaN() : Scalar(0);
   }
-  // Test NaN propagation for pmin and pmax. It should be equivalent to std::min.
-  CHECK_CWISE2_IF(PacketTraits::HasMin, (std::min), internal::pmin);
-  CHECK_CWISE2_IF(PacketTraits::HasMax, (std::max), internal::pmax);
   // Test NaN propagation for pfmin and pfmax. It should be equivalent to std::fmin.
+  // Note: NaN propagation is implementation defined for pmin/pmax, so we do not test it here.
   CHECK_CWISE2_IF(PacketTraits::HasMin, fmin, internal::pfmin);
   CHECK_CWISE2_IF(PacketTraits::HasMax, fmax, internal::pfmax);
+  CHECK_CWISE2_IF(PacketTraits::HasMin, propagate_nan_min, internal::pfmin_nan);
+  CHECK_CWISE2_IF(PacketTraits::HasMax, propagate_nan_max, internal::pfmax_nan);
 }
 
 template <>

unsupported/Eigen/CXX11/src/Tensor/TensorBase.h

@@ -395,16 +395,18 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return unaryExpr(internal::scalar_mod_op<Scalar>(rhs));
     }
 
+    template <int NanPropagation=PropagateFast>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+        EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar,NanPropagation>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
     cwiseMax(Scalar threshold) const {
-      return cwiseMax(constant(threshold));
+      return cwiseMax<NanPropagation>(constant(threshold));
     }
 
+    template <int NanPropagation=PropagateFast>
     EIGEN_DEVICE_FUNC
-    EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
+        EIGEN_STRONG_INLINE const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar,NanPropagation>, const Derived, const TensorCwiseNullaryOp<internal::scalar_constant_op<Scalar>, const Derived> >
     cwiseMin(Scalar threshold) const {
-      return cwiseMin(constant(threshold));
+      return cwiseMin<NanPropagation>(constant(threshold));
     }
 
     template<typename NewType>
@@ -472,16 +474,16 @@ class TensorBase<Derived, ReadOnlyAccessors>
       return binaryExpr(other.derived(), internal::scalar_quotient_op<Scalar>());
     }
 
-    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  template<int NaNPropagation=PropagateFast, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar>, const Derived, const OtherDerived>
+      const TensorCwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar, NaNPropagation>, const Derived, const OtherDerived>
     cwiseMax(const OtherDerived& other) const {
-      return binaryExpr(other.derived(), internal::scalar_max_op<Scalar>());
+    return binaryExpr(other.derived(), internal::scalar_max_op<Scalar,Scalar, NaNPropagation>());
     }
 
-    template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  template<int NaNPropagation=PropagateFast, typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
-    const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar>, const Derived, const OtherDerived>
+    const TensorCwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar, NaNPropagation>, const Derived, const OtherDerived>
     cwiseMin(const OtherDerived& other) const {
-      return binaryExpr(other.derived(), internal::scalar_min_op<Scalar>());
+      return binaryExpr(other.derived(), internal::scalar_min_op<Scalar,Scalar, NaNPropagation>());
     }
 
     template<typename OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE

unsupported/test/cxx11_tensor_expr.cpp

@@ -303,40 +303,79 @@ template <typename Scalar>
 void test_minmax_nan_propagation_templ() {
   for (int size = 1; size < 17; ++size) {
     const Scalar kNan = std::numeric_limits<Scalar>::quiet_NaN();
+    const Scalar kZero(0);
     Tensor<Scalar, 1> vec_nan(size);
     Tensor<Scalar, 1> vec_zero(size);
-    Tensor<Scalar, 1> vec_res(size);
     vec_nan.setConstant(kNan);
     vec_zero.setZero();
-    vec_res.setZero();
 
-    // Test that we propagate NaNs in the tensor when applying the
+    auto verify_all_nan = [&](const Tensor<Scalar, 1>& v) {
-    // cwiseMax(scalar) operator, which is used for the Relu operator.
+      for (int i = 0; i < size; ++i) {
-    vec_res = vec_nan.cwiseMax(Scalar(0));
+        VERIFY((numext::isnan)(v(i)));
-    for (int i = 0; i < size; ++i) {
+      }
-      VERIFY((numext::isnan)(vec_res(i)));
+    };
-    }
 
-    // Test that NaNs do not propagate if we reverse the arguments.
+    auto verify_all_zero = [&](const Tensor<Scalar, 1>& v) {
-    vec_res = vec_zero.cwiseMax(kNan);
+      for (int i = 0; i < size; ++i) {
-    for (int i = 0; i < size; ++i) {
+        VERIFY_IS_EQUAL(v(i), Scalar(0));
-      VERIFY_IS_EQUAL(vec_res(i), Scalar(0));
+      }
-    }
+    };
 
-    // Test that we propagate NaNs in the tensor when applying the
+    // Test NaN propagating max.
-    // cwiseMin(scalar) operator.
+    // max(nan, nan) = nan
-    vec_res.setZero();
+    // max(nan, 0) = nan
-    vec_res = vec_nan.cwiseMin(Scalar(0));
+    // max(0, nan) = nan
-    for (int i = 0; i < size; ++i) {
+    // max(0, 0) = 0
-      VERIFY((numext::isnan)(vec_res(i)));
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNaN>(kNan));
-    }
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNaN>(vec_nan));
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNaN>(kZero));
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNaN>(vec_zero));
+    verify_all_nan(vec_zero.template cwiseMax<PropagateNaN>(kNan));
+    verify_all_nan(vec_zero.template cwiseMax<PropagateNaN>(vec_nan));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNaN>(kZero));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNaN>(vec_zero));
 
+    // Test number propagating max.
+    // max(nan, nan) = nan
+    // max(nan, 0) = 0
+    // max(0, nan) = 0
+    // max(0, 0) = 0
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNumbers>(kNan));
+    verify_all_nan(vec_nan.template cwiseMax<PropagateNumbers>(vec_nan));
+    verify_all_zero(vec_nan.template cwiseMax<PropagateNumbers>(kZero));
+    verify_all_zero(vec_nan.template cwiseMax<PropagateNumbers>(vec_zero));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNumbers>(kNan));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNumbers>(vec_nan));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNumbers>(kZero));
+    verify_all_zero(vec_zero.template cwiseMax<PropagateNumbers>(vec_zero));
 
-    // Test that NaNs do not propagate if we reverse the arguments.
+    // Test NaN propagating min.
-    vec_res = vec_zero.cwiseMin(kNan);
+    // min(nan, nan) = nan
-    for (int i = 0; i < size; ++i) {
+    // min(nan, 0) = nan
-      VERIFY_IS_EQUAL(vec_res(i), Scalar(0));
+    // min(0, nan) = nan
-    }
+    // min(0, 0) = 0
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNaN>(kNan));
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNaN>(vec_nan));
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNaN>(kZero));
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNaN>(vec_zero));
+    verify_all_nan(vec_zero.template cwiseMin<PropagateNaN>(kNan));
+    verify_all_nan(vec_zero.template cwiseMin<PropagateNaN>(vec_nan));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNaN>(kZero));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNaN>(vec_zero));
+
+    // Test number propagating min.
+    // min(nan, nan) = nan
+    // min(nan, 0) = 0
+    // min(0, nan) = 0
+    // min(0, 0) = 0
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNumbers>(kNan));
+    verify_all_nan(vec_nan.template cwiseMin<PropagateNumbers>(vec_nan));
+    verify_all_zero(vec_nan.template cwiseMin<PropagateNumbers>(kZero));
+    verify_all_zero(vec_nan.template cwiseMin<PropagateNumbers>(vec_zero));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNumbers>(kNan));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNumbers>(vec_nan));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNumbers>(kZero));
+    verify_all_zero(vec_zero.template cwiseMin<PropagateNumbers>(vec_zero));
   }
 }