1 | /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. |
2 | |
3 | Licensed under the Apache License, Version 2.0 (the "License"); |
4 | you may not use this file except in compliance with the License. |
5 | You may obtain a copy of the License at |
6 | |
7 | http://www.apache.org/licenses/LICENSE-2.0 |
8 | |
9 | Unless required by applicable law or agreed to in writing, software |
10 | distributed under the License is distributed on an "AS IS" BASIS, |
11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | See the License for the specific language governing permissions and |
13 | limitations under the License. |
14 | ==============================================================================*/ |
15 | |
16 | #ifndef TENSORFLOW_CORE_KERNELS_CWISE_OPS_GRADIENTS_H_ |
17 | #define TENSORFLOW_CORE_KERNELS_CWISE_OPS_GRADIENTS_H_ |
18 | |
19 | #define EIGEN_USE_THREADS |
20 | #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" |
21 | #include "tensorflow/core/kernels/cwise_ops.h" |
22 | |
23 | namespace Eigen { |
24 | namespace internal { |
25 | |
26 | // Gradient for the tanh function |
27 | template <typename T> |
28 | struct scalar_tanh_gradient_op { |
29 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T |
30 | operator()(const T& output, const T& output_gradient) const { |
31 | return output_gradient * (T(1) - output * output); |
32 | } |
33 | template <typename Packet> |
34 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet |
35 | packetOp(const Packet& output, const Packet& output_gradient) const { |
36 | return pmul(output_gradient, |
37 | psub(pset1<Packet>(T(1)), pmul(output, output))); |
38 | } |
39 | }; |
40 | template <typename T> |
41 | struct functor_traits<scalar_tanh_gradient_op<T>> { |
42 | enum { |
43 | Cost = NumTraits<T>::AddCost + 2 * NumTraits<T>::MulCost, |
44 | PacketAccess = packet_traits<T>::HasSub && packet_traits<T>::HasMul, |
45 | }; |
46 | }; |
47 | |
48 | // Gradient for the sigmoid function |
49 | template <typename T> |
50 | struct scalar_sigmoid_gradient_op { |
51 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T |
52 | operator()(const T& output, const T& output_gradient) const { |
53 | return output_gradient * output * (T(1) - output); |
54 | } |
55 | template <typename Packet> |
56 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet |
57 | packetOp(const Packet& output, const Packet& output_gradient) const { |
58 | return pmul(output_gradient, |
59 | pmul(output, psub(pset1<Packet>(T(1)), output))); |
60 | } |
61 | }; |
62 | template <typename T> |
63 | struct functor_traits<scalar_sigmoid_gradient_op<T>> { |
64 | enum { |
65 | Cost = NumTraits<T>::AddCost + 2 * NumTraits<T>::MulCost, |
66 | PacketAccess = packet_traits<T>::HasSub && packet_traits<T>::HasMul, |
67 | }; |
68 | }; |
69 | |
70 | // Gradient for the inverse function |
71 | template <typename T> |
72 | struct scalar_inverse_gradient_op { |
73 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T |
74 | operator()(const T& output, const T& output_gradient) const { |
75 | if (output_gradient == T(0)) { |
76 | return T(0); |
77 | } else { |
78 | const T out_conj = numext::conj(output); |
79 | return -out_conj * out_conj * output_gradient; |
80 | } |
81 | } |
82 | template <typename Packet> |
83 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet |
84 | packetOp(const Packet& output, const Packet& output_gradient) const { |
85 | const Packet out_conj = pconj(output); |
86 | return mul_no_nan_op<T>().packetOp(pnegate(pmul(out_conj, out_conj)), |
87 | output_gradient); |
88 | } |
89 | }; |
90 | template <typename T> |
91 | struct functor_traits<scalar_inverse_gradient_op<T>> { |
92 | enum { |
93 | Cost = NumTraits<T>::AddCost + 2 * NumTraits<T>::MulCost, |
94 | PacketAccess = packet_traits<T>::HasMul, |
95 | }; |
96 | }; |
97 | |
98 | // Gradient for the sqrt function |
99 | template <typename T> |
100 | struct scalar_sqrt_gradient_op { |
101 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T |
102 | operator()(const T& output, const T& output_gradient) const { |
103 | if (output_gradient == T(0)) { |
104 | return T(0); |
105 | } else { |
106 | const T out_conj = numext::conj(output); |
107 | return (static_cast<T>(0.5) * output_gradient) / out_conj; |
108 | } |
109 | } |
110 | template <typename Packet> |
111 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet |
112 | packetOp(const Packet& output, const Packet& output_gradient) const { |
113 | const Packet const_half = pset1<Packet>(static_cast<T>(0.5)); |
114 | const Packet out_conj = pconj(output); |
115 | return mul_no_nan_op<T>().packetOp(pdiv(const_half, out_conj), |
116 | output_gradient); |
117 | } |
118 | }; |
119 | template <typename T> |
120 | struct functor_traits<scalar_sqrt_gradient_op<T>> { |
121 | enum { |
122 | PacketAccess = packet_traits<T>::HasMul & packet_traits<T>::HasDiv, |
123 | Cost = NumTraits<T>::MulCost + scalar_div_cost<T, PacketAccess>::value, |
124 | }; |
125 | }; |
126 | |
127 | // Gradient for the rsqrt function |
128 | template <typename T> |
129 | struct scalar_rsqrt_gradient_op { |
130 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const T |
131 | operator()(const T& output, const T& output_gradient) const { |
132 | if (output_gradient == T(0)) { |
133 | return T(0); |
134 | } else { |
135 | const T out_conj = numext::conj(output); |
136 | return static_cast<T>(-0.5) * (output_gradient * out_conj) * |
137 | (out_conj * out_conj); |
138 | } |
139 | } |
140 | template <typename Packet> |
141 | EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet |
142 | packetOp(const Packet& output, const Packet& output_gradient) const { |
143 | const Packet const_half = pset1<Packet>(static_cast<T>(-0.5)); |
144 | const Packet out_conj = pconj(output); |
145 | auto safe_pmul = [](const Packet& a, const Packet& b) { |
146 | return mul_no_nan_op<T>().packetOp(a, b); |
147 | }; |
148 | return safe_pmul(pmul(const_half, pmul(out_conj, out_conj)), |
149 | safe_pmul(out_conj, output_gradient)); |
150 | } |
151 | }; |
152 | template <typename T> |
153 | struct functor_traits<scalar_rsqrt_gradient_op<T>> { |
154 | enum { |
155 | Cost = 4 * NumTraits<T>::MulCost, |
156 | PacketAccess = packet_traits<T>::HasMul, |
157 | }; |
158 | }; |
159 | |
160 | } // end namespace internal |
161 | } // end namespace Eigen |
162 | |
163 | namespace tensorflow { |
164 | |
165 | namespace functor { |
166 | |
167 | template <typename Device, typename Functor> |
168 | struct SimpleBinaryFunctor { |
169 | void operator()(const Device& d, typename Functor::tout_type out, |
170 | typename Functor::tin_type in0, |
171 | typename Functor::tin_type in1); |
172 | }; |
173 | |
174 | // Partial specialization of BinaryFunctor for CPU devices |
175 | typedef Eigen::ThreadPoolDevice CPUDevice; |
176 | |
177 | template <typename Functor> |
178 | struct SimpleBinaryFunctor<CPUDevice, Functor> { |
179 | void operator()(const CPUDevice& d, typename Functor::tout_type out, |
180 | typename Functor::tin_type in0, |
181 | typename Functor::tin_type in1) { |
182 | out.device(d) = in0.binaryExpr(in1, typename Functor::func()); |
183 | } |
184 | }; |
185 | |
186 | |
187 | template <typename T> |
188 | struct tanh_grad : base<T, Eigen::internal::scalar_tanh_gradient_op<T>> {}; |
189 | |
190 | template <typename T> |
191 | struct sigmoid_grad : base<T, Eigen::internal::scalar_sigmoid_gradient_op<T>> { |
192 | }; |
193 | |
194 | template <typename T> |
195 | struct inverse_grad : base<T, Eigen::internal::scalar_inverse_gradient_op<T>> { |
196 | }; |
197 | |
198 | template <typename T> |
199 | struct sqrt_grad : base<T, Eigen::internal::scalar_sqrt_gradient_op<T>> {}; |
200 | |
201 | template <typename T> |
202 | struct rsqrt_grad : base<T, Eigen::internal::scalar_rsqrt_gradient_op<T>> {}; |
203 | |
204 | template <typename T> |
205 | struct igamma_grad_a : base<T, Eigen::internal::scalar_igamma_der_a_op<T>> {}; |
206 | |
207 | } // end namespace functor |
208 | |
209 | } // end namespace tensorflow |
210 | #endif // TENSORFLOW_CORE_KERNELS_CWISE_OPS_GRADIENTS_H_ |
211 | |