From ee7b3ed09e699da191fe238ab409f32318637380 Mon Sep 17 00:00:00 2001
From: qiaolongfei <qiaolongfei@baidu.com>
Date: Wed, 4 Oct 2017 15:33:44 -0700
Subject: [PATCH 1/3] use EigenScalar to get learning_rate from GPU device

---
 paddle/operators/sgd_op.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h
index d72d333a9a03d..954fd4827252d 100644
--- a/paddle/operators/sgd_op.h
+++ b/paddle/operators/sgd_op.h
@@ -23,6 +23,9 @@ using Tensor = framework::Tensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
 
 template <typename Place, typename T>
 class SGDOpKernel : public framework::OpKernel<T> {
@@ -31,13 +34,14 @@ class SGDOpKernel : public framework::OpKernel<T> {
     auto param = ctx.Input<Tensor>("Param");
     auto grad = ctx.Input<Tensor>("Grad");
     auto param_out = ctx.Output<Tensor>("ParamOut");
-    float lr = ctx.Input<Tensor>("LearningRate")->data<float>()[0];
+    auto learning_rate = ctx.Input<Tensor>("LearningRate");
 
     param_out->mutable_data<T>(ctx.GetPlace());
 
     auto p = EigenVector<T>::Flatten(*param);
     auto g = EigenVector<T>::Flatten(*grad);
     auto o = EigenVector<T>::Flatten(*param_out);
+    auto lr = EigenScalar<T>::From(*learning_rate);
     auto place = ctx.GetEigenDevice<Place>();
 
     o.device(place) = p - lr * g;

From 775c60246b66469e06f01a50c89b7b39594a3b63 Mon Sep 17 00:00:00 2001
From: qiaolongfei <qiaolongfei@baidu.com>
Date: Wed, 4 Oct 2017 16:53:21 -0700
Subject: [PATCH 2/3] remove using in sgd header file

---
 paddle/operators/sgd_op.h | 27 ++++++++++-----------------
 1 file changed, 10 insertions(+), 17 deletions(-)

diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h
index 954fd4827252d..b501d244d7fc3 100644
--- a/paddle/operators/sgd_op.h
+++ b/paddle/operators/sgd_op.h
@@ -19,32 +19,25 @@ limitations under the License. */
 namespace paddle {
 namespace operators {
 
-using Tensor = framework::Tensor;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
-
 template <typename Place, typename T>
 class SGDOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto param = ctx.Input<Tensor>("Param");
-    auto grad = ctx.Input<Tensor>("Grad");
-    auto param_out = ctx.Output<Tensor>("ParamOut");
-    auto learning_rate = ctx.Input<Tensor>("LearningRate");
+    auto param = ctx.Input<framework::Tensor>("Param");
+    auto grad = ctx.Input<framework::Tensor>("Grad");
+    auto param_out = ctx.Output<framework::Tensor>("ParamOut");
+    auto learning_rate = ctx.Input<framework::Tensor>("LearningRate");
 
     param_out->mutable_data<T>(ctx.GetPlace());
 
-    auto p = EigenVector<T>::Flatten(*param);
-    auto g = EigenVector<T>::Flatten(*grad);
-    auto o = EigenVector<T>::Flatten(*param_out);
-    auto lr = EigenScalar<T>::From(*learning_rate);
+    auto p = framework::EigenVector<T>::Flatten(*param);
+    auto g = framework::EigenVector<T>::Flatten(*grad);
+    auto o = framework::EigenVector<T>::Flatten(*param_out);
+    auto lr = framework::EigenVector<T>::From(*learning_rate);
     auto place = ctx.GetEigenDevice<Place>();
 
-    o.device(place) = p - lr * g;
+    Eigen::DSizes<int, 2> grad_dsize(grad->dims()[0], grad->dims()[1]);
+    o.device(place) = p - lr.broadcast(grad_dsize) * g;
   }
 };
 

From 8ebc31d9358c919fdd6f50d502f4ee071a91d38e Mon Sep 17 00:00:00 2001
From: qiaolongfei <qiaolongfei@baidu.com>
Date: Wed, 4 Oct 2017 17:13:02 -0700
Subject: [PATCH 3/3] optimize the dsize

---
 paddle/operators/sgd_op.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/paddle/operators/sgd_op.h b/paddle/operators/sgd_op.h
index b501d244d7fc3..26f4012f25877 100644
--- a/paddle/operators/sgd_op.h
+++ b/paddle/operators/sgd_op.h
@@ -33,10 +33,10 @@ class SGDOpKernel : public framework::OpKernel<T> {
     auto p = framework::EigenVector<T>::Flatten(*param);
     auto g = framework::EigenVector<T>::Flatten(*grad);
     auto o = framework::EigenVector<T>::Flatten(*param_out);
-    auto lr = framework::EigenVector<T>::From(*learning_rate);
+    auto lr = framework::EigenVector<T>::Flatten(*learning_rate);
     auto place = ctx.GetEigenDevice<Place>();
 
-    Eigen::DSizes<int, 2> grad_dsize(grad->dims()[0], grad->dims()[1]);
+    Eigen::DSizes<int, 1> grad_dsize(grad->numel());
     o.device(place) = p - lr.broadcast(grad_dsize) * g;
   }
 };