From e811c865677c80a5fce13d4bac7178bf0fa20d7b Mon Sep 17 00:00:00 2001
From: sweetsky0901 <work@yq01-idl-gpu-online20.yq01.baidu.com>
Date: Fri, 22 Dec 2017 18:36:40 +0800
Subject: [PATCH] for epsilon dataType

---
 paddle/operators/norm_op.cc | 10 ++++++----
 paddle/operators/norm_op.h  |  8 ++++----
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/paddle/operators/norm_op.cc b/paddle/operators/norm_op.cc
index 990a1504ea3..1d9b55d887e 100644
--- a/paddle/operators/norm_op.cc
+++ b/paddle/operators/norm_op.cc
@@ -16,6 +16,7 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
+template <typename AttrType>
 class NormOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   NormOpMaker(OpProto* proto, OpAttrChecker* op_checker)
@@ -28,9 +29,9 @@ class NormOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Scale",
              "(Tensor) The input tensor of norm operator. "
              "The format of input tensor is C * 1.");
-    AddAttr<float>("epsilon",
-                   "(float, default 1e-10) Constant "
-                   "for numerical stability.")
+    AddAttr<AttrType>("epsilon",
+                      "(float, default 1e-10) Constant "
+                      "for numerical stability.")
         .SetDefault(1.0e-10f);
     AddOutput("Out",
               "(Tensor) The output tensor of norm operator."
@@ -100,7 +101,8 @@ class NormOpGrad : public framework::OperatorWithKernel {
 }  // namespace paddle
 
 namespace ops = paddle::operators;
-REGISTER_OP(norm, ops::NormOp, ops::NormOpMaker, norm_grad, ops::NormOpGrad);
+REGISTER_OP(norm, ops::NormOp, ops::NormOpMaker<float>, norm_grad,
+            ops::NormOpGrad);
 REGISTER_OP_CPU_KERNEL(
     norm, ops::NormKernel<paddle::platform::CPUDeviceContext, float>,
     ops::NormKernel<paddle::platform::CPUDeviceContext, double>);
diff --git a/paddle/operators/norm_op.h b/paddle/operators/norm_op.h
index d3dcf483416..b22df373af1 100644
--- a/paddle/operators/norm_op.h
+++ b/paddle/operators/norm_op.h
@@ -26,14 +26,14 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class NormKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
     const framework::Tensor* in_x = context.Input<framework::Tensor>("X");
     const framework::Tensor* scale = context.Input<framework::Tensor>("Scale");
     auto* out = context.Output<framework::Tensor>("Out");
-    T epsilon = context.Attr<T>("epsilon");
+    auto epsilon = static_cast<T>(context.Attr<AttrType>("epsilon"));
     out->mutable_data<T>(context.GetPlace());
     int batch_size = in_x->dims()[0];
     int channels = in_x->dims()[1];
@@ -82,7 +82,7 @@ class NormKernel : public framework::OpKernel<T> {
     }
   }
 };
-template <typename DeviceContext, typename T>
+template <typename DeviceContext, typename T, typename AttrType = T>
 class NormGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
@@ -90,7 +90,7 @@ class NormGradKernel : public framework::OpKernel<T> {
     const framework::Tensor* scale = context.Input<framework::Tensor>("Scale");
     const framework::Tensor* out_grad =
         context.Input<framework::Tensor>(framework::GradVarName("Out"));
-    T epsilon = context.Attr<T>("epsilon");
+    auto epsilon = static_cast<T>(context.Attr<AttrType>("epsilon"));
     framework::Tensor* in_x_grad =
         context.Output<framework::Tensor>(framework::GradVarName("X"));
     in_x_grad->mutable_data<T>(context.GetPlace());
-- 
GitLab