Complete elementwise_max_op

paddle/operators/elementwise_max_op.cc

@@ -42,4 +42,4 @@ REGISTER_OP_CPU_KERNEL(
     ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, float>,
     ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, double>,
     ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int>,
-    ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
+    ops::ElementwiseMaxGradKernel<paddle::platform::CPUDeviceContext, int64_t>);

paddle/operators/elementwise_max_op.cu

@@ -0,0 +1,32 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/elementwise_max_op.h"
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_CUDA_KERNEL(
+    elementwise_max,
+    ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMaxKernel<paddle::platform::CUDADeviceContext, int64_t>);
+REGISTER_OP_CUDA_KERNEL(
+    elementwise_max_grad,
+    ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMaxGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);

paddle/operators/elementwise_max_op.h

@@ -65,43 +65,88 @@ class ElementwiseMaxKernel : public framework::OpKernel<T> {
 };
 
 template <typename T>
-struct ElementwiseSubGradFunctor {
+struct ElementwiseMaxGradFunctor {
   template <typename Device, typename X, typename Y, typename Z, typename dX,
             typename dY, typename dZ>
   void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
-    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
     auto x_e = framework::EigenVector<T>::Flatten(*x);
     auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
 
     if (dx) {
       auto dx_e = framework::EigenVector<T>::Flatten(*dx);
-      dx_e.device(d) = (x_e > y_e) * dz_e;
+      dx_e.device(d) = (x_e > y_e).template cast<T>() * dz_e;
     }
     if (dy) {
       auto dy_e = framework::EigenVector<T>::Flatten(*dy);
-      dy_e.device(d) = (y_e >= x_e) * dz_e;
+      dy_e.device(d) = (y_e >= x_e).template cast<T>() * dz_e;
     }
   }
 };
 
 template <typename T>
-struct ElementwiseSubOneGradFunctor {
+struct ElementwiseMaxBroadCastGradFunctor {
   template <typename Device, typename X, typename Y, typename Z, typename dX,
-            typename dY, typename dZ>
-  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
+            typename dY, typename dZ, typename Pre, typename N>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
+    auto x_e = framework::EigenVector<T>::Flatten(*x);
+    auto y_e = framework::EigenVector<T>::Flatten(*y);
     auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
+    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
+                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
+
+    if (dx) {
+      auto dx_e = framework::EigenVector<T>::Flatten(*dx);
+      dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
+    }
+
+    if (dy) {
+      auto dy_e = framework::EigenVector<T>::Flatten(*dy);
+      dy_e.device(d) = ((y_e_bcast >= x_e).template cast<T>() * dz_e)
+                           .reshape(Eigen::DSizes<int, 2>(pre, n))
+                           .sum(Eigen::array<int, 1>{{0}});
+    }
+  }
+};
+
+template <typename T>
+struct ElementwiseMaxBroadCast2GradFunctor {
+  template <typename Device, typename X, typename Y, typename Z, typename dX,
+            typename dY, typename dZ, typename Pre, typename N, typename Post>
+  void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
+                  Post post) {
     auto x_e = framework::EigenVector<T>::Flatten(*x);
     auto y_e = framework::EigenVector<T>::Flatten(*y);
+    auto dz_e = framework::EigenVector<T>::Flatten(*dz);
+
+    auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
+                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
+                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));
     if (dx) {
       auto dx_e = framework::EigenVector<T>::Flatten(*dx);
-      dx_e.device(d) = dz_e;
+      dx_e.device(d) = (x_e > y_e_bcast).template cast<T>() * dz_e;
     }
+
     if (dy) {
       auto dy_e = framework::EigenVector<T>::Flatten(*dy);
-      dy_e.device(d) = (-1.0) * dz_e.sum();
+      dy_e.device(d) = ((y_e_bcast >= x_e).template cast<T>() * dz_e)
+                           .reshape(Eigen::DSizes<int, 3>(pre, n, post))
+                           .sum(Eigen::array<int, 2>{{0, 2}});
     }
   }
 };
 
+template <typename DeviceContext, typename T>
+class ElementwiseMaxGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    ElementwiseGradCompute<DeviceContext, T, ElementwiseMaxGradFunctor<T>,
+                           ElementwiseMaxBroadCastGradFunctor<T>,
+                           ElementwiseMaxBroadCast2GradFunctor<T>>(ctx);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle