Merged in suiyuan2009/eigen/fix_tanh_inconsistent_for_tensorflow (pull request PR-215)

Fix_tanh_inconsistent_for_tensorflow

1 files changed, 89 insertions, 28 deletions

diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index 04208c9fe..e2f3d869f 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -491,19 +491,62 @@ struct functor_traits<scalar_atan_op<Scalar> >
   };
 };
 
-
 /** \internal
   * \brief Template functor to compute the tanh of a scalar
   * \sa class CwiseUnaryOp, ArrayBase::tanh()
   */
-template<typename Scalar> struct scalar_tanh_op {
+template <typename Scalar>
+struct scalar_tanh_op {
   EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op)
-  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tanh(a); }
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const {
+    /** \internal \returns the hyperbolic tan of \a a (coeff-wise)
+          Doesn't do anything fancy, just a 13/6-degree rational interpolant
+       which
+          is accurate up to a couple of ulp in the range [-9, 9], outside of
+       which
+          the fl(tanh(x)) = +/-1. */
+
+    // Clamp the inputs to the range [-9, 9] since anything outside
+    // this range is +/-1.0f in single-precision.
+    const Scalar plus_9 = static_cast<Scalar>(9.0);
+    const Scalar minus_9 = static_cast<Scalar>(-9.0);
+    const Scalar x = numext::maxi(minus_9, numext::mini(plus_9, a));
+    // Scalarhe monomial coefficients of the numerator polynomial (odd).
+    const Scalar alpha_1 = static_cast<Scalar>(4.89352455891786e-03);
+    const Scalar alpha_3 = static_cast<Scalar>(6.37261928875436e-04);
+    const Scalar alpha_5 = static_cast<Scalar>(1.48572235717979e-05);
+    const Scalar alpha_7 = static_cast<Scalar>(5.12229709037114e-08);
+    const Scalar alpha_9 = static_cast<Scalar>(-8.60467152213735e-11);
+    const Scalar alpha_11 = static_cast<Scalar>(2.00018790482477e-13);
+    const Scalar alpha_13 = static_cast<Scalar>(-2.76076847742355e-16);
+    // Scalarhe monomial coefficients of the denominator polynomial (even).
+    const Scalar beta_0 = static_cast<Scalar>(4.89352518554385e-03);
+    const Scalar beta_2 = static_cast<Scalar>(2.26843463243900e-03);
+    const Scalar beta_4 = static_cast<Scalar>(1.18534705686654e-04);
+    const Scalar beta_6 = static_cast<Scalar>(1.19825839466702e-06);
+    // Since the polynomials are odd/even, we need x^2.
+    const Scalar x2 = x * x;
+    // Evaluate the numerator polynomial p.
+    Scalar p = x2 * alpha_13 + alpha_11;
+    p = x2 * p + alpha_9;
+    p = x2 * p + alpha_7;
+    p = x2 * p + alpha_5;
+    p = x2 * p + alpha_3;
+    p = x2 * p + alpha_1;
+    p = x * p;
+    // Evaluate the denominator polynomial p.
+    Scalar q = x2 * beta_6 + beta_4;
+    q = x2 * q + beta_2;
+    q = x2 * q + beta_0;
+    // Divide the numerator by the denominator.
+    return p / q;
+  }
   template <typename Packet>
   EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& _x) const {
     /** \internal \returns the hyperbolic tan of \a a (coeff-wise)
         Doesn't do anything fancy, just a 13/6-degree rational interpolant which
-        is accurate up to a couple of ulp in the range [-9, 9], outside of which the
+        is accurate up to a couple of ulp in the range [-9, 9], outside of which
+       the
         fl(tanh(x)) = +/-1. */
 
     // Clamp the inputs to the range [-9, 9] since anything outside
@@ -511,7 +554,7 @@ template<typename Scalar> struct scalar_tanh_op {
     const Packet plus_9 = pset1<Packet>(9.0);
     const Packet minus_9 = pset1<Packet>(-9.0);
     const Packet x = pmax(minus_9, pmin(plus_9, _x));
-    
+
     // The monomial coefficients of the numerator polynomial (odd).
     const Packet alpha_1 = pset1<Packet>(4.89352455891786e-03);
     const Packet alpha_3 = pset1<Packet>(6.37261928875436e-04);
@@ -520,17 +563,17 @@ template<typename Scalar> struct scalar_tanh_op {
     const Packet alpha_9 = pset1<Packet>(-8.60467152213735e-11);
     const Packet alpha_11 = pset1<Packet>(2.00018790482477e-13);
     const Packet alpha_13 = pset1<Packet>(-2.76076847742355e-16);
-    
+
     // The monomial coefficients of the denominator polynomial (even).
     const Packet beta_0 = pset1<Packet>(4.89352518554385e-03);
     const Packet beta_2 = pset1<Packet>(2.26843463243900e-03);
     const Packet beta_4 = pset1<Packet>(1.18534705686654e-04);
     const Packet beta_6 = pset1<Packet>(1.19825839466702e-06);
-    
+
     // Since the polynomials are odd/even, we need x^2.
     const Packet x2 = pmul(x, x);
-  
-  // Evaluate the numerator polynomial p.
+
+    // Evaluate the numerator polynomial p.
     Packet p = pmadd(x2, alpha_13, alpha_11);
     p = pmadd(x2, p, alpha_9);
     p = pmadd(x2, p, alpha_7);
@@ -538,38 +581,56 @@ template<typename Scalar> struct scalar_tanh_op {
     p = pmadd(x2, p, alpha_3);
     p = pmadd(x2, p, alpha_1);
     p = pmul(x, p);
-    
+
     // Evaluate the denominator polynomial p.
     Packet q = pmadd(x2, beta_6, beta_4);
     q = pmadd(x2, q, beta_2);
     q = pmadd(x2, q, beta_0);
-    
+
     // Divide the numerator by the denominator.
     return pdiv(p, q);
   }
 };
-template<typename Scalar>
-struct functor_traits<scalar_tanh_op<Scalar> >
-{
+template <>
+struct scalar_tanh_op<std::complex<double> > {
+  EIGEN_DEVICE_FUNC inline const std::complex<double> operator()(
+      const std::complex<double>& a) const {
+    return numext::tanh(a);
+  }
+};
+template <>
+struct scalar_tanh_op<std::complex<float> > {
+  EIGEN_DEVICE_FUNC inline const std::complex<float> operator()(
+      const std::complex<float>& a) const {
+    return numext::tanh(a);
+  }
+};
+template <typename Scalar>
+struct functor_traits<scalar_tanh_op<Scalar> > {
   enum {
     PacketAccess = packet_traits<Scalar>::HasTanh,
-    Cost =
-    (PacketAccess
-     // The following numbers are based on the AVX implementation,
+    Cost = (PacketAccess && (!is_same<Scalar, std::complex<float> >::value) &&
+                    (!is_same<Scalar, std::complex<double> >::value)
+// The following numbers are based on the AVX implementation,
 #ifdef EIGEN_VECTORIZE_FMA
-     // Haswell can issue 2 add/mul/madd per cycle.
-     // 9 pmadd, 2 pmul, 1 div, 2 other
-     ? (2 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost +
-     NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)
+                // Haswell can issue 2 add/mul/madd per cycle.
+                // 9 pmadd, 2 pmul, 1 div, 2 other
+                ? (2 * NumTraits<Scalar>::AddCost +
+                   6 * NumTraits<Scalar>::MulCost +
+                   NumTraits<Scalar>::template Div<
+                       packet_traits<Scalar>::HasDiv>::Cost)
 #else
-     ? (11 * NumTraits<Scalar>::AddCost +
-        11 * NumTraits<Scalar>::MulCost +
-        NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost)
+                ? (11 * NumTraits<Scalar>::AddCost +
+                   11 * NumTraits<Scalar>::MulCost +
+                   NumTraits<Scalar>::template Div<
+                       packet_traits<Scalar>::HasDiv>::Cost)
 #endif
-     // This number assumes a naive implementation of tanh
-     : (6 * NumTraits<Scalar>::AddCost + 3 * NumTraits<Scalar>::MulCost +
-        2 * NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost +
-        functor_traits<scalar_exp_op<Scalar> >::Cost))
+                // This number assumes a naive implementation of tanh
+                : (6 * NumTraits<Scalar>::AddCost +
+                   3 * NumTraits<Scalar>::MulCost +
+                   2 * NumTraits<Scalar>::template Div<
+                           packet_traits<Scalar>::HasDiv>::Cost +
+                   functor_traits<scalar_exp_op<Scalar> >::Cost))
   };
 };