Merge pull request #4632 from luotao1/reduce

Unify Reduce functions and simplify register code

paddle/operators/activation_op.cc

@@ -285,11 +285,9 @@ REGISTER_OP(stanh, ops::ActivationOp, ops::STanhOpMaker<float>, stanh_grad,
 #define REGISTER_ACTIVATION_CPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_CPU_KERNEL(                                                      \
       act_type,                                                                \
-      paddle::operators::ActivationKernel<paddle::platform::CPUPlace,          \
-                                          paddle::operators::functor<float>>); \
+      ops::ActivationKernel<paddle::platform::CPUPlace, ops::functor<float>>); \
   REGISTER_OP_CPU_KERNEL(act_type##_grad,                                      \
-                         paddle::operators::ActivationGradKernel<              \
-                             paddle::platform::CPUPlace,                       \
-                             paddle::operators::grad_functor<float>>);
+                         ops::ActivationGradKernel<paddle::platform::CPUPlace, \
+                                                   ops::grad_functor<float>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CPU_KERNEL);

paddle/operators/activation_op.cu

@@ -15,14 +15,14 @@
 #define EIGEN_USE_GPU
 #include "paddle/operators/activation_op.h"
 
+namespace ops = paddle::operators;
+
 #define REGISTER_ACTIVATION_GPU_KERNEL(act_type, functor, grad_functor)        \
   REGISTER_OP_GPU_KERNEL(                                                      \
       act_type,                                                                \
-      paddle::operators::ActivationKernel<paddle::platform::GPUPlace,          \
-                                          paddle::operators::functor<float>>); \
+      ops::ActivationKernel<paddle::platform::GPUPlace, ops::functor<float>>); \
   REGISTER_OP_GPU_KERNEL(act_type##_grad,                                      \
-                         paddle::operators::ActivationGradKernel<              \
-                             paddle::platform::GPUPlace,                       \
-                             paddle::operators::grad_functor<float>>);
+                         ops::ActivationGradKernel<paddle::platform::GPUPlace, \
+                                                   ops::grad_functor<float>>);
 
 FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_GPU_KERNEL);

paddle/operators/reduce_op.cc

@@ -168,36 +168,22 @@ namespace ops = paddle::operators;
 
 REGISTER_OP(reduce_sum, ops::ReduceOp, ops::ReduceSumOpMaker, reduce_sum_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_sum,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::SumFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_sum_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::SumGradFunctor>);
 
 REGISTER_OP(reduce_mean, ops::ReduceOp, ops::ReduceMeanOpMaker,
             reduce_mean_grad, ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_mean,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MeanFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_mean_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MeanGradFunctor>);
 
 REGISTER_OP(reduce_max, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_max_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_max,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MaxFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_max_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
-
-REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMaxOpMaker, reduce_min_grad,
+
+REGISTER_OP(reduce_min, ops::ReduceOp, ops::ReduceMinOpMaker, reduce_min_grad,
             ops::ReduceGradOp);
-REGISTER_OP_CPU_KERNEL(
-    reduce_min,
-    ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::MinFunctor>);
-REGISTER_OP_CPU_KERNEL(reduce_min_grad,
-                       ops::ReduceGradKernel<paddle::platform::CPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
+
+#define REGISTER_REDUCE_CPU_KERNEL(reduce_type, functor, grad_functor)     \
+  REGISTER_OP_CPU_KERNEL(                                                  \
+      reduce_type,                                                         \
+      ops::ReduceKernel<paddle::platform::CPUPlace, float, ops::functor>); \
+  REGISTER_OP_CPU_KERNEL(reduce_type##_grad,                               \
+                         ops::ReduceGradKernel<paddle::platform::CPUPlace, \
+                                               float, ops::grad_functor>);
+
+FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_CPU_KERNEL);

paddle/operators/reduce_op.cu

@@ -17,30 +17,12 @@
 
 namespace ops = paddle::operators;
 
-REGISTER_OP_GPU_KERNEL(
-    reduce_sum,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::SumFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_sum_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::SumGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_mean,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MeanFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_mean_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MeanGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_max,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MaxFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_max_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
-
-REGISTER_OP_GPU_KERNEL(
-    reduce_min,
-    ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::MinFunctor>);
-REGISTER_OP_GPU_KERNEL(reduce_min_grad,
-                       ops::ReduceGradKernel<paddle::platform::GPUPlace, float,
-                                             ops::MaxOrMinGradFunctor>);
+#define REGISTER_REDUCE_GPU_KERNEL(reduce_type, functor, grad_functor)     \
+  REGISTER_OP_GPU_KERNEL(                                                  \
+      reduce_type,                                                         \
+      ops::ReduceKernel<paddle::platform::GPUPlace, float, ops::functor>); \
+  REGISTER_OP_GPU_KERNEL(reduce_type##_grad,                               \
+                         ops::ReduceGradKernel<paddle::platform::GPUPlace, \
+                                               float, ops::grad_functor>);
+
+FOR_EACH_KERNEL_FUNCTOR(REGISTER_REDUCE_GPU_KERNEL);

paddle/operators/reduce_op.h

@@ -198,3 +198,9 @@ class ReduceGradKernel : public framework::OpKernel<T> {
 
 }  // namespace operators
 }  // namespace paddle
+
+#define FOR_EACH_KERNEL_FUNCTOR(__macro)                \
+  __macro(reduce_sum, SumFunctor, SumGradFunctor);      \
+  __macro(reduce_mean, MeanFunctor, MeanGradFunctor);   \
+  __macro(reduce_max, MaxFunctor, MaxOrMinGradFunctor); \
+  __macro(reduce_min, MinFunctor, MaxOrMinGradFunctor);