Home | History | Annotate | Download | only in kernels
      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 // See docs in ../ops/nn_ops.cc.
     17 
     18 #define EIGEN_USE_THREADS
     19 
     20 #include "tensorflow/core/kernels/softsign_op.h"
     21 #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
     22 #include "tensorflow/core/framework/numeric_op.h"
     23 #include "tensorflow/core/framework/op_kernel.h"
     24 #include "tensorflow/core/framework/register_types.h"
     25 #include "tensorflow/core/framework/tensor.h"
     26 #include "tensorflow/core/kernels/warn_about_ints.h"
     27 #include "tensorflow/core/lib/core/errors.h"
     28 
     29 namespace tensorflow {
     30 
     31 typedef Eigen::ThreadPoolDevice CPUDevice;
     32 typedef Eigen::GpuDevice GPUDevice;
     33 
     34 template <typename Device, typename T>
     35 class SoftsignOp : public UnaryElementWiseOp<T, SoftsignOp<Device, T>> {
     36  public:
     37   explicit SoftsignOp(OpKernelConstruction* context)
     38       : UnaryElementWiseOp<T, SoftsignOp<Device, T>>(context) {
     39     WarnAboutInts(context);
     40   }
     41 
     42   void Operate(OpKernelContext* context, const Tensor& input, Tensor* output) {
     43     functor::Softsign<Device, T> functor;
     44     functor(context->eigen_device<Device>(), input.flat<T>(),
     45             output->flat<T>());
     46   }
     47 };
     48 
     49 template <typename Device, typename T>
     50 class SoftsignGradOp
     51     : public BinaryElementWiseOp<T, SoftsignGradOp<Device, T>> {
     52  public:
     53   explicit SoftsignGradOp(OpKernelConstruction* context)
     54       : BinaryElementWiseOp<T, SoftsignGradOp<Device, T>>(context) {
     55     WarnAboutInts(context);
     56   }
     57 
     58   void OperateNoTemplate(OpKernelContext* context, const Tensor& g,
     59                          const Tensor& a, Tensor* output);
     60 
     61   // INPUTS:
     62   //   g (gradients): backpropagated gradients
     63   //   a (inputs): inputs that were passed to SoftsignOp()
     64   // OUTPUT:
     65   //   gradients to backprop
     66   template <int NDIMS>
     67   void Operate(OpKernelContext* context, const Tensor& g, const Tensor& a,
     68                Tensor* output) {
     69     OperateNoTemplate(context, g, a, output);
     70   }
     71 };
     72 
     73 template <typename Device, typename T>
     74 void SoftsignGradOp<Device, T>::OperateNoTemplate(OpKernelContext* context,
     75                                                   const Tensor& g,
     76                                                   const Tensor& a,
     77                                                   Tensor* output) {
     78   OP_REQUIRES(context, a.IsSameSize(g),
     79               errors::InvalidArgument("g and a must be the same size"));
     80   functor::SoftsignGrad<Device, T> functor;
     81   functor(context->eigen_device<Device>(), g.flat<T>(), a.flat<T>(),
     82           output->flat<T>());
     83 }
     84 
     85 #define REGISTER_KERNELS(type)                                           \
     86   REGISTER_KERNEL_BUILDER(                                               \
     87       Name("Softsign").Device(DEVICE_CPU).TypeConstraint<type>("T"),     \
     88       SoftsignOp<CPUDevice, type>);                                      \
     89   REGISTER_KERNEL_BUILDER(                                               \
     90       Name("SoftsignGrad").Device(DEVICE_CPU).TypeConstraint<type>("T"), \
     91       SoftsignGradOp<CPUDevice, type>);
     92 
     93 TF_CALL_REAL_NUMBER_TYPES(REGISTER_KERNELS);
     94 #undef REGISTER_KERNELS
     95 
     96 #if GOOGLE_CUDA
     97 // Forward declarations of the functor specializations for GPU.
     98 namespace functor {
     99 #define DECLARE_GPU_SPEC(T)                                          \
    100   template <>                                                        \
    101   void Softsign<GPUDevice, T>::operator()(                           \
    102       const GPUDevice& d, typename TTypes<T>::ConstTensor features,  \
    103       typename TTypes<T>::Tensor activations);                       \
    104   extern template struct Softsign<GPUDevice, T>;                     \
    105                                                                      \
    106   template <>                                                        \
    107   void SoftsignGrad<GPUDevice, T>::operator()(                       \
    108       const GPUDevice& d, typename TTypes<T>::ConstTensor gradients, \
    109       typename TTypes<T>::ConstTensor features,                      \
    110       typename TTypes<T>::Tensor backprops);                         \
    111   extern template struct SoftsignGrad<GPUDevice, T>;
    112 
    113 TF_CALL_GPU_NUMBER_TYPES(DECLARE_GPU_SPEC);
    114 }  // namespace functor
    115 
    116 // Registration of the GPU implementations.
    117 #define REGISTER_GPU_KERNELS(type)                                       \
    118   REGISTER_KERNEL_BUILDER(                                               \
    119       Name("Softsign").Device(DEVICE_GPU).TypeConstraint<type>("T"),     \
    120       SoftsignOp<GPUDevice, type>);                                      \
    121   REGISTER_KERNEL_BUILDER(                                               \
    122       Name("SoftsignGrad").Device(DEVICE_GPU).TypeConstraint<type>("T"), \
    123       SoftsignGradOp<GPUDevice, type>);
    124 
    125 TF_CALL_GPU_NUMBER_TYPES(REGISTER_GPU_KERNELS);
    126 #undef REGISTER_GPU_KERNELS
    127 
    128 #endif  // GOOGLE_CUDA
    129 
    130 }  // namespace tensorflow
    131