aboutsummaryrefslogtreecommitdiffhomepage
path: root/Eigen/src/Core/functors/UnaryFunctors.h
diff options
context:
space:
mode:
Diffstat (limited to 'Eigen/src/Core/functors/UnaryFunctors.h')
-rw-r--r--Eigen/src/Core/functors/UnaryFunctors.h106
1 files changed, 85 insertions, 21 deletions
diff --git a/Eigen/src/Core/functors/UnaryFunctors.h b/Eigen/src/Core/functors/UnaryFunctors.h
index 7ba0abedc..5baba1494 100644
--- a/Eigen/src/Core/functors/UnaryFunctors.h
+++ b/Eigen/src/Core/functors/UnaryFunctors.h
@@ -234,9 +234,33 @@ template<typename Scalar> struct scalar_exp_op {
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
};
-template<typename Scalar>
-struct functor_traits<scalar_exp_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasExp }; };
+template <typename Scalar>
+struct functor_traits<scalar_exp_op<Scalar> > {
+ enum {
+ PacketAccess = packet_traits<Scalar>::HasExp,
+ // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+ // Haswell can issue 2 add/mul/madd per cycle.
+ Cost =
+ (sizeof(Scalar) == 4
+ // float: 8 pmadd, 4 pmul, 2 padd/psub, 6 other
+ ? (8 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost)
+ // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other
+ : (14 * NumTraits<Scalar>::AddCost +
+ 6 * NumTraits<Scalar>::MulCost +
+ NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost))
+#else
+ Cost =
+ (sizeof(Scalar) == 4
+ // float: 7 pmadd, 6 pmul, 4 padd/psub, 10 other
+ ? (21 * NumTraits<Scalar>::AddCost + 13 * NumTraits<Scalar>::MulCost)
+ // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div, 13 other
+ : (23 * NumTraits<Scalar>::AddCost +
+ 12 * NumTraits<Scalar>::MulCost +
+ NumTraits<Scalar>::template Div<packet_traits<Scalar>::HasDiv>::Cost))
+#endif
+ };
+};
/** \internal
*
@@ -250,9 +274,24 @@ template<typename Scalar> struct scalar_log_op {
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
};
-template<typename Scalar>
-struct functor_traits<scalar_log_op<Scalar> >
-{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog }; };
+template <typename Scalar>
+struct functor_traits<scalar_log_op<Scalar> > {
+ enum {
+ PacketAccess = packet_traits<Scalar>::HasLog,
+ Cost =
+ (PacketAccess
+ // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+ // 8 pmadd, 6 pmul, 8 padd/psub, 16 other, can issue 2 add/mul/madd per cycle.
+ ? (20 * NumTraits<Scalar>::AddCost + 7 * NumTraits<Scalar>::MulCost)
+#else
+ // 8 pmadd, 6 pmul, 8 padd/psub, 20 other
+ ? (36 * NumTraits<Scalar>::AddCost + 14 * NumTraits<Scalar>::MulCost)
+#endif
+ // Measured cost of std::log.
+ : sizeof(Scalar)==4 ? 40 : 85)
+ };
+};
/** \internal
*
@@ -280,10 +319,19 @@ template<typename Scalar> struct scalar_sqrt_op {
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
};
-template<typename Scalar>
-struct functor_traits<scalar_sqrt_op<Scalar> >
-{ enum {
- Cost = 5 * NumTraits<Scalar>::MulCost,
+template <typename Scalar>
+struct functor_traits<scalar_sqrt_op<Scalar> > {
+ enum {
+#if EIGEN_FAST_MATH
+ // The following numbers are based on the AVX implementation.
+ Cost = (sizeof(Scalar) == 8 ? 28
+ // 4 pmul, 1 pmadd, 3 other
+ : (3 * NumTraits<Scalar>::AddCost +
+ 5 * NumTraits<Scalar>::MulCost)),
+#else
+ // The following numbers are based on min VSQRT throughput on Haswell.
+ Cost = (sizeof(Scalar) == 8 ? 28 : 14),
+#endif
PacketAccess = packet_traits<Scalar>::HasSqrt
};
};
@@ -313,7 +361,7 @@ struct functor_traits<scalar_rsqrt_op<Scalar> >
*/
template<typename Scalar> struct scalar_cos_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
- EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { using std::cos; return cos(a); }
+ EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return numext::cos(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
};
@@ -332,7 +380,7 @@ struct functor_traits<scalar_cos_op<Scalar> >
*/
template<typename Scalar> struct scalar_sin_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sin; return sin(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sin(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
};
@@ -352,7 +400,7 @@ struct functor_traits<scalar_sin_op<Scalar> >
*/
template<typename Scalar> struct scalar_tan_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::tan; return tan(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tan(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
};
@@ -371,7 +419,7 @@ struct functor_traits<scalar_tan_op<Scalar> >
*/
template<typename Scalar> struct scalar_acos_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::acos; return acos(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::acos(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
};
@@ -390,7 +438,7 @@ struct functor_traits<scalar_acos_op<Scalar> >
*/
template<typename Scalar> struct scalar_asin_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::asin; return asin(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::asin(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
};
@@ -546,7 +594,7 @@ struct functor_traits<scalar_erfc_op<Scalar> >
*/
template<typename Scalar> struct scalar_atan_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::atan; return atan(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::atan(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::patan(a); }
};
@@ -566,7 +614,7 @@ struct functor_traits<scalar_atan_op<Scalar> >
*/
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 { using std::tanh; return tanh(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tanh(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptanh(a); }
};
@@ -574,8 +622,24 @@ template<typename Scalar>
struct functor_traits<scalar_tanh_op<Scalar> >
{
enum {
- Cost = 5 * NumTraits<Scalar>::MulCost,
- PacketAccess = packet_traits<Scalar>::HasTanh
+ PacketAccess = packet_traits<Scalar>::HasTanh,
+ Cost =
+ (PacketAccess
+ // 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)
+#else
+ ? (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))
};
};
@@ -585,7 +649,7 @@ struct functor_traits<scalar_tanh_op<Scalar> >
*/
template<typename Scalar> struct scalar_sinh_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_sinh_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::sinh; return sinh(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sinh(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psinh(a); }
};
@@ -604,7 +668,7 @@ struct functor_traits<scalar_sinh_op<Scalar> >
*/
template<typename Scalar> struct scalar_cosh_op {
EIGEN_EMPTY_STRUCT_CTOR(scalar_cosh_op)
- EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { using std::cosh; return cosh(a); }
+ EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::cosh(a); }
template <typename Packet>
EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); }
};
generated by cgit v1.2.3 (git 2.39.1) at 2024-05-23 05:54:22 +0000