diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc
index 17f942571d0141537e992be9ab73847d2a794698..b29ac44699463312a1fdcea55e003daa75997302 100644
--- a/paddle/fluid/framework/op_desc.cc
+++ b/paddle/fluid/framework/op_desc.cc
@@ -50,6 +50,27 @@ class CompileTimeInferShapeContext : public InferShapeContext {
   const std::vector<std::string> &Outputs(
       const std::string &name) const override;
 
+  void ShareDim(const std::string &in, const std::string &out, size_t i = 0,
+                size_t j = 0) override {
+    PADDLE_ENFORCE_LT(i, Inputs(in).size());
+    PADDLE_ENFORCE_LT(j, Outputs(out).size());
+    const std::string &input_n = Inputs(in)[i];
+    const std::string &output_n = Outputs(out)[j];
+
+    PADDLE_ENFORCE(input_n != framework::kEmptyVarName, "The %s[%d] is @EMPTY@",
+                   in, i);
+    PADDLE_ENFORCE(output_n != framework::kEmptyVarName,
+                   "The %s[%d] is @EMPTY@", out, j);
+
+    auto *in_var = block_.FindVarRecursive(input_n);
+    auto *out_var = block_.FindVarRecursive(output_n);
+
+    PADDLE_ENFORCE(in_var->GetType() == out_var->GetType(),
+                   "The type of %s and %s is not the same.", input_n, output_n);
+
+    SetDim(output_n, GetDim(input_n));
+  }
+
   void ShareLoD(const std::string &in, const std::string &out, size_t i = 0,
                 size_t j = 0) const override {
     PADDLE_ENFORCE_LT(i, Inputs(in).size());
diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index a103be7191d02a96ee97d76f786f9364938c1c65..9aad9db3ca885507424989ff05cf5748ddb839a4 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -542,13 +542,45 @@ class RuntimeInferShapeContext : public InferShapeContext {
     return op_.Outputs(name);
   }
 
-  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
-                size_t j = 0) const override {
+  void ShareDim(const std::string& in, const std::string& out, size_t i = 0,
+                size_t j = 0) override {
     PADDLE_ENFORCE_LT(i, Inputs(in).size());
     PADDLE_ENFORCE_LT(j, Outputs(out).size());
-    Variable* in_var = scope_.FindVar(Inputs(in)[i]);
-    Variable* out_var = scope_.FindVar(Outputs(out)[j]);
+    const std::string& input_n = Inputs(in)[i];
+    const std::string& output_n = Outputs(out)[j];
+
+    Variable* in_var = scope_.FindVar(input_n);
+    Variable* out_var = scope_.FindVar(output_n);
+    PADDLE_ENFORCE(in_var->Type() == out_var->Type(),
+                   "The type of %s and %s is not the same.", output_n,
+                   GetDim(input_n));
+
+    if (in_var->IsType<framework::SelectedRows>()) {
+      auto& in_sele_rows = in_var->Get<framework::SelectedRows>();
+      auto out_sele_rows = out_var->GetMutable<framework::SelectedRows>();
+      out_sele_rows->mutable_value()->Resize(in_sele_rows.value().dims());
+      out_sele_rows->set_rows(in_sele_rows.rows());
+      out_sele_rows->set_height(in_sele_rows.height());
+    } else if (in_var->IsType<framework::LoDTensor>()) {
+      auto& in_lod_tensor = in_var->Get<framework::LoDTensor>();
+      auto* out_lod_tensor = out_var->GetMutable<framework::LoDTensor>();
+      out_lod_tensor->Resize(in_lod_tensor.dims());
+    } else {
+      PADDLE_THROW(
+          "Currently, the input type of ShareDim only can be LoDTensor "
+          "or SelectedRows.");
+    }
+  }
+
+  void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
+                size_t j = 0) const override {
+    const std::vector<std::string>& inputs = Inputs(in);
+    const std::vector<std::string>& outputs = Outputs(out);
+    PADDLE_ENFORCE_LT(i, inputs.size());
+    PADDLE_ENFORCE_LT(j, outputs.size());
+    Variable* in_var = scope_.FindVar(inputs.at(i));
     if (!in_var->IsType<LoDTensor>()) return;
+    Variable* out_var = scope_.FindVar(outputs.at(j));
     PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
                    "The %d-th output of Output(%s) must be LoDTensor.", j, out);
     auto in_tensor = in_var->Get<LoDTensor>();
diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h
index fd220d961af85dd55fe2031409180823d8f178fc..15ef5ce011a551a97c1d6fc2e6bf4477a4e09072 100644
--- a/paddle/fluid/framework/shape_inference.h
+++ b/paddle/fluid/framework/shape_inference.h
@@ -58,6 +58,9 @@ class InferShapeContext {
 
   void ShareLoDs(const std::string &in, const std::string &out) const;
 
+  virtual void ShareDim(const std::string &in, const std::string &out,
+                        size_t i = 0, size_t j = 0) = 0;
+
   virtual void ShareLoD(const std::string &in, const std::string &out,
                         size_t i = 0, size_t j = 0) const = 0;
 
diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc
index 286b03d7b7d11a50f33f0190c1a5b9097ed0f4a2..677fcdcd81f18271c2836d4e0c106b396c946da5 100644
--- a/paddle/fluid/operators/activation_op.cc
+++ b/paddle/fluid/operators/activation_op.cc
@@ -79,7 +79,7 @@ class ActivationOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 
@@ -90,12 +90,26 @@ class ActivationOp : public framework::OperatorWithKernel {
   }
 };
 
+class ActivationOpInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDesc& op_desc,
+                  framework::BlockDesc* block) const override {
+    auto x_name = op_desc.Input("X")[0];
+    auto out_name = op_desc.Output("Out")[0];
+    auto& x = block->FindRecursiveOrCreateVar(x_name);
+    auto& out = block->FindRecursiveOrCreateVar(out_name);
+    out.SetType(x.GetType());
+    out.SetDataType(x.GetDataType());
+  }
+};
+
 class ActivationOpGrad : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
 
   void InferShape(framework::InferShapeContext* ctx) const override {
-    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
+    ctx->ShareDim("Out", framework::GradVarName("X"));
+    ctx->ShareLoD("Out", framework::GradVarName("X"));
   }
 
  protected:
@@ -524,12 +538,14 @@ namespace ops = paddle::operators;
 #define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)        \
   REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \
                     ::paddle::operators::OP_NAME##OpMaker,          \
+                    ::paddle::operators::ActivationOpInferVarType,  \
                     ::paddle::operators::OP_NAME##GradMaker);       \
   REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad)
 
 #define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE)                    \
   REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp,     \
                     ::paddle::operators::OP_NAME##OpMaker,              \
+                    ::paddle::operators::ActivationOpInferVarType,      \
                     ::paddle::framework::DefaultGradOpDescMaker<true>); \
   REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad)
 
diff --git a/paddle/fluid/operators/argsort_op.cc b/paddle/fluid/operators/argsort_op.cc
index a2f5a2545701991263c1ef842e9275b1edbfd2ca..d25160f4232b5a621d16b9f469f56bd5aa7c88e3 100644
--- a/paddle/fluid/operators/argsort_op.cc
+++ b/paddle/fluid/operators/argsort_op.cc
@@ -42,8 +42,8 @@ class ArgsortOp : public framework::OperatorWithKernel {
                    "-rank(Input(X)) (%d).",
                    axis, num_dims);
 
-    ctx->SetOutputDim("Out", in_dims);
-    ctx->SetOutputDim("Indices", in_dims);
+    ctx->ShareDim("X", "Out");
+    ctx->ShareDim("X", "Indices");
     ctx->ShareLoD("X", "Out");
     ctx->ShareLoD("X", "Indices");
   }
diff --git a/paddle/fluid/operators/conv_shift_op.cc b/paddle/fluid/operators/conv_shift_op.cc
index f2549e814d6f3b5674fe2eec1139f1c3dc6fa0b4..08506ddd18ed35831702814e70962cb36ec958b1 100644
--- a/paddle/fluid/operators/conv_shift_op.cc
+++ b/paddle/fluid/operators/conv_shift_op.cc
@@ -44,7 +44,7 @@ class ConvShiftOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
                       "The 2nd dimension of Input(Y) should be less than or "
                       "equal to the 2nd dimension of Input(X).");
-    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h
index 94df11bee70dec44f19ee9ffff04ca92d5990ee8..7e5975ead64ab39a9c618a33e300c4fce55a5b22 100644
--- a/paddle/fluid/operators/elementwise_op.h
+++ b/paddle/fluid/operators/elementwise_op.h
@@ -41,7 +41,8 @@ class ElementwiseOp : public framework::OperatorWithKernel {
     auto y_dim = ctx->GetInputDim("Y");
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
                       "Rank of first input must >= rank of second input.");
-    ctx->SetOutputDim("Out", x_dim);
+
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 
@@ -70,6 +71,7 @@ class ElementwiseOpInferVarType : public framework::VarTypeInference {
     auto& x = block->FindRecursiveOrCreateVar(x_name);
     auto& out = block->FindRecursiveOrCreateVar(out_name);
     out.SetType(x.GetType());
+    out.SetDataType(x.GetDataType());
   }
 };
 
@@ -157,10 +159,12 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
     auto x_grad_name = framework::GradVarName("X");
     auto y_grad_name = framework::GradVarName("Y");
     if (ctx->HasOutput(x_grad_name)) {
-      ctx->SetOutputDim(x_grad_name, x_dims);
+      ctx->ShareDim("X", /*->*/ x_grad_name);
+      ctx->ShareLoD("X", /*->*/ x_grad_name);
     }
     if (ctx->HasOutput(y_grad_name)) {
-      ctx->SetOutputDim(y_grad_name, y_dims);
+      ctx->ShareDim("Y", /*->*/ y_grad_name);
+      ctx->ShareLoD("Y", /*->*/ y_grad_name);
     }
   }
 
@@ -193,14 +197,15 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad {
 
     auto x_grad_name = framework::GradVarName("X");
     if (ctx->HasOutput(x_grad_name)) {
-      auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
-      ctx->SetOutputDim(x_grad_name, out_dims);
+      ctx->ShareDim(framework::GradVarName("Out"), /*->*/ x_grad_name);
+      ctx->ShareLoD(framework::GradVarName("Out"), /*->*/ x_grad_name);
     }
     auto y_grad_name = framework::GradVarName("Y");
     if (ctx->HasOutput(y_grad_name)) {
       PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
-      auto y_dims = ctx->GetInputDim("Y");
-      ctx->SetOutputDim(y_grad_name, y_dims);
+
+      ctx->ShareDim("Y", /*->*/ y_grad_name);
+      ctx->ShareLoD("Y", /*->*/ y_grad_name);
     }
   }
 };
diff --git a/paddle/fluid/operators/fake_dequantize_op.cc b/paddle/fluid/operators/fake_dequantize_op.cc
index 2008e7027524ffd1f80a6eede015801b8a0b0254..5d6488c67e0db440c8d4609736523643dd666dcc 100644
--- a/paddle/fluid/operators/fake_dequantize_op.cc
+++ b/paddle/fluid/operators/fake_dequantize_op.cc
@@ -48,7 +48,8 @@ class FakeDequantizeMaxAbsOp : public framework::OperatorWithKernel {
                    "Input(X) of FakeDequantizeMaxAbsOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of FakeDequantizeMaxAbsOp should not be null.");
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
diff --git a/paddle/fluid/operators/prelu_op.cc b/paddle/fluid/operators/prelu_op.cc
index e0c4c81bdd5b5d0af3bafe632a2fa033efd08050..58cfbb76e93a1c15c9b7cf9f9e596066c29b7ebb 100644
--- a/paddle/fluid/operators/prelu_op.cc
+++ b/paddle/fluid/operators/prelu_op.cc
@@ -49,7 +49,7 @@ class PReluOp : public framework::OperatorWithKernel {
     } else {
       PADDLE_THROW("Unkown mode %s", mode);
     }
-    ctx->SetOutputDim("Out", x_dim);
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 
diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc
index 13df1d4b4bb6c240610f96ccc8f223fc984d63f7..0fb7776fd9dbf437673820c7cf9411644272626c 100644
--- a/paddle/fluid/operators/rnn_memory_helper_op.cc
+++ b/paddle/fluid/operators/rnn_memory_helper_op.cc
@@ -54,7 +54,7 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase {
                    "Input(X) of rnn_memory_helper op should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output of rnn_memory_helper op should not be null.");
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
diff --git a/paddle/fluid/operators/sequence_conv_op.cc b/paddle/fluid/operators/sequence_conv_op.cc
index ec6cb24350ae276724aae339590d40be1e9ea400..95a21a5d3ee6d8037431083edc25d1cddf05dedb 100644
--- a/paddle/fluid/operators/sequence_conv_op.cc
+++ b/paddle/fluid/operators/sequence_conv_op.cc
@@ -90,8 +90,8 @@ class SequenceConvGradOp : public framework::OperatorWithKernel {
                         ctx->GetInputDim("PaddingData"));
     }
     if (ctx->HasOutput(framework::GradVarName("X"))) {
-      ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
-      ctx->ShareLoD("X", framework::GradVarName("X"));
+      ctx->ShareDim("X", /*->*/ framework::GradVarName("X"));
+      ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
     }
     if (ctx->HasOutput(framework::GradVarName("Filter"))) {
       ctx->SetOutputDim(framework::GradVarName("Filter"),
diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc
index 5c6fd13d42e43e3502a1cab85a56e019420c708d..15d3f064eb7b025dc9a85b2aabad24186061cbd4 100644
--- a/paddle/fluid/operators/sequence_pool_op.cc
+++ b/paddle/fluid/operators/sequence_pool_op.cc
@@ -102,8 +102,9 @@ class SequencePoolGradOp : public framework::OperatorWithKernel {
     for (int64_t i = 1; i < og_dims.size(); ++i) {
       PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch.");
     }
-    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
-    ctx->ShareLoD("X", framework::GradVarName("X"));
+
+    ctx->ShareDim("X", /*->*/ framework::GradVarName("X"));
+    ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
   }
 
  protected:
diff --git a/paddle/fluid/operators/sequence_reshape_op.cc b/paddle/fluid/operators/sequence_reshape_op.cc
index ef5e6f3210234d59298fcf04c812390643c693d0..31d28d723498892f287246ba228df757d5b9f6c8 100644
--- a/paddle/fluid/operators/sequence_reshape_op.cc
+++ b/paddle/fluid/operators/sequence_reshape_op.cc
@@ -92,7 +92,7 @@ class SequenceReshapeGradOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(ctx->HasInput("X"),
                    "Input(X) of SequenceReshapeGradOp should  not be null.");
 
-    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
+    ctx->ShareDim("X", /*->*/ framework::GradVarName("X"));
     ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
   }
 };
diff --git a/paddle/fluid/operators/sequence_softmax_op.cc b/paddle/fluid/operators/sequence_softmax_op.cc
index c44f8206eb5079fef969e3e527552512eebd0f1a..ada3e0c8dbba38729c2b9c8b02335327835f2ef4 100644
--- a/paddle/fluid/operators/sequence_softmax_op.cc
+++ b/paddle/fluid/operators/sequence_softmax_op.cc
@@ -27,7 +27,8 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel {
                    "Input(X) of SequenceSoftmaxOp should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Out"),
                    "Output(Out) of SequenceSoftmaxOp should not be null.");
-    ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
+
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 
diff --git a/paddle/fluid/operators/shrink_rnn_memory_op.cc b/paddle/fluid/operators/shrink_rnn_memory_op.cc
index 29d2fb989754f5621222768a279a1c898ea1c355..e1c74c3a2f89235ba92c396d1a548271bb7d939d 100644
--- a/paddle/fluid/operators/shrink_rnn_memory_op.cc
+++ b/paddle/fluid/operators/shrink_rnn_memory_op.cc
@@ -151,9 +151,9 @@ class ShrinkRNNMemoryGradInferShape : public framework::InferShapeBase {
   void operator()(framework::InferShapeContext *context) const override {
     PADDLE_ENFORCE(context->HasInput("X"));
     PADDLE_ENFORCE(context->HasOutput(framework::GradVarName("X")));
-    context->SetOutputDim(framework::GradVarName("X"),
-                          context->GetInputDim("X"));
-    context->ShareLoD("X", framework::GradVarName("X"));
+
+    context->ShareDim("X", /*->*/ framework::GradVarName("X"));
+    context->ShareLoD("X", /*->*/ framework::GradVarName("X"));
   }
 };
 
diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
index c3b0fe32098cb4b41ccc155db58809ef9f1bf46b..193de05422bb78572c0e5eaf4cd46744c3bcb113 100644
--- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
+++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
@@ -40,7 +40,7 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
                       "The 2nd dimension of Input(X) and Input(Label) should "
                       "be equal.");
 
-    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareDim("X", /*->*/ "Out");
     ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cc b/paddle/fluid/operators/truncated_gaussian_random_op.cc
index d854e2803975543b51c50ea2bc173322d3c3ca5e..1e8708f2648d7dd3c10319bd0a4be193d2458d53 100644
--- a/paddle/fluid/operators/truncated_gaussian_random_op.cc
+++ b/paddle/fluid/operators/truncated_gaussian_random_op.cc
@@ -148,7 +148,7 @@ struct TruncatedNormal {
 
   T operator()(T value) const {
     auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value;
-    return (std::sqrt(2.0) * Erfinv(2 * p - 1) + mean) * std;
+    return std::sqrt(2.0) * Erfinv(2 * p - 1) * std + mean;
   }
 };
 
diff --git a/paddle/fluid/operators/truncated_gaussian_random_op.cu b/paddle/fluid/operators/truncated_gaussian_random_op.cu
index ad2a9021bfe344d838dff2040b3fb9371274e218..2b1403db2ee8995d5679f481444f23d56a52e209 100644
--- a/paddle/fluid/operators/truncated_gaussian_random_op.cu
+++ b/paddle/fluid/operators/truncated_gaussian_random_op.cu
@@ -42,7 +42,7 @@ struct TruncatedNormal {
     rng.discard(n);
     T value = dist(rng);
     auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value;
-    return (std::sqrt(2.0) * erfinvf(2 * p - 1) + mean) * std;
+    return std::sqrt(2.0) * erfinvf(2 * p - 1) * std + mean;
   }
 };
 
diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h
index c7c533bd42859c374c4783d43ec4cdd34a6a994a..4ea0cd7283b55649dbdbbf97f81f10c69ac6a1d2 100644
--- a/paddle/fluid/platform/dynload/cublas.h
+++ b/paddle/fluid/platform/dynload/cublas.h
@@ -55,7 +55,7 @@ extern void *cublas_dso_handle;
   struct DynLoad__##__name {                         \
     template <typename... Args>                      \
     inline cublasStatus_t operator()(Args... args) { \
-      return __name(args...);                        \
+      return ::__name(args...);                      \
     }                                                \
   };                                                 \
   extern DynLoad__##__name __name
diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h
index 0103e7a3accf88f3c83f109298010c3c9af3d549..e6353f67ef118072a2d8e49111e8ecc486589998 100644
--- a/paddle/fluid/platform/dynload/cudnn.h
+++ b/paddle/fluid/platform/dynload/cudnn.h
@@ -13,6 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#define GLOG_NO_ABBREVIATED_SEVERITIES
+#define GOOGLE_GLOG_DLL_DECL
+#include <glog/logging.h>
 
 #include <cudnn.h>
 #include <mutex>  // NOLINT
@@ -47,13 +50,13 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
 
 #else
 
-#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)                  \
-  struct DynLoad__##__name {                                     \
-    template <typename... Args>                                  \
-    auto operator()(Args... args) -> decltype(__name(args...)) { \
-      return __name(args...);                                    \
-    }                                                            \
-  };                                                             \
+#define DECLARE_DYNAMIC_LOAD_CUDNN_WRAP(__name)     \
+  struct DynLoad__##__name {                        \
+    template <typename... Args>                     \
+    inline cudnnStatus_t operator()(Args... args) { \
+      return ::__name(args...);                     \
+    }                                               \
+  };                                                \
   extern DynLoad__##__name __name
 
 #endif
diff --git a/paddle/fluid/platform/dynload/curand.h b/paddle/fluid/platform/dynload/curand.h
index 2daf1b4215ce1f7f771bbac72bfe103b0b941976..0bb300ec33076d9ddfaf69190f14131279cc888e 100644
--- a/paddle/fluid/platform/dynload/curand.h
+++ b/paddle/fluid/platform/dynload/curand.h
@@ -44,7 +44,7 @@ extern void *curand_dso_handle;
   struct DynLoad__##__name {                     \
     template <typename... Args>                  \
     curandStatus_t operator()(Args... args) {    \
-      return __name(args...);                    \
+      return ::__name(args...);                  \
     }                                            \
   };                                             \
   extern DynLoad__##__name __name
diff --git a/paddle/fluid/platform/dynload/dynamic_loader.cc b/paddle/fluid/platform/dynload/dynamic_loader.cc
index 6a3ad2151081504fda2a3818c5f99ad47039d91d..cc5cda6106c188f3156d33480b5d3641eed32556 100644
--- a/paddle/fluid/platform/dynload/dynamic_loader.cc
+++ b/paddle/fluid/platform/dynload/dynamic_loader.cc
@@ -107,7 +107,11 @@ static inline void* GetDsoHandleFromDefaultPath(const std::string& dso_path,
 static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
                                                const std::string& dso_name,
                                                bool throw_on_error = true) {
+#if !defined(_WIN32)
   int dynload_flags = RTLD_LAZY | RTLD_LOCAL;
+#else
+  int dynload_flags = 0;
+#endif  // !_WIN32
   void* dso_handle = nullptr;
 
   std::string dlPath = dso_name;
@@ -117,10 +121,15 @@ static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
     // search xxx.so from custom path
     dlPath = join(search_root, dso_name);
     dso_handle = dlopen(dlPath.c_str(), dynload_flags);
+#if !defined(_WIN32)
+    auto errorno = dlerror();
+#else
+    auto errorno = GetLastError();
+#endif  // !_WIN32
     // if not found, search from default path
     if (nullptr == dso_handle) {
       LOG(WARNING) << "Failed to find dynamic library: " << dlPath << " ("
-                   << dlerror() << ")";
+                   << errorno << ")";
       if (dlPath.find("nccl") != std::string::npos) {
         std::cout
             << "You may need to install 'nccl2' from NVIDIA official website: "
@@ -139,10 +148,15 @@ static inline void* GetDsoHandleFromSearchPath(const std::string& search_root,
       "export LD_LIBRARY_PATH=... \n Note: After Mac OS 10.11, "
       "using the DYLD_LIBRARY_PATH is impossible unless System "
       "Integrity Protection (SIP) is disabled.";
+#if !defined(_WIN32)
+  auto errorno = dlerror();
+#else
+  auto errorno = GetLastError();
+#endif  // !_WIN32
   if (throw_on_error) {
-    PADDLE_ENFORCE(nullptr != dso_handle, error_msg, dlPath, dlerror());
+    PADDLE_ENFORCE(nullptr != dso_handle, error_msg, dlPath, errorno);
   } else if (nullptr == dso_handle) {
-    LOG(WARNING) << string::Sprintf(error_msg, dlPath, dlerror());
+    LOG(WARNING) << string::Sprintf(error_msg, dlPath, errorno);
   }
 
   return dso_handle;
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index f3ecc1083a8f7c9d8efd870e1e6fbf4731f6f336..bb288384e8501caf1cb6f9cd186cc09985a5de55 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -600,7 +600,7 @@ EOF
     if [[ ${WITH_GPU} == "ON"  ]]; then
         NCCL_DEPS="apt-get install -y --allow-downgrades libnccl2=2.2.13-1+cuda${CUDA_MAJOR} libnccl-dev=2.2.13-1+cuda${CUDA_MAJOR} || true"
     else
-        NCCL_DEPS=""
+        NCCL_DEPS="true"
     fi
 
     if [[ ${WITH_FLUID_ONLY:-OFF} == "OFF" ]]; then
diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt
index 9cdcb87df5dd1669066c204c86c269973df506f1..aad94479a9e4fbf5c4f70fb5e763b147a0e7590b 100644
--- a/python/CMakeLists.txt
+++ b/python/CMakeLists.txt
@@ -52,7 +52,6 @@ add_custom_command(OUTPUT ${FLUID_CORE}
         DEPENDS paddle_pybind)
 add_custom_target(copy_paddle_pybind ALL DEPENDS ${FLUID_CORE})
 
-
 add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
     COMMAND touch stub.cc
     COMMAND cp -r ${PADDLE_SOURCE_DIR}/python/paddle ${PADDLE_BINARY_DIR}/python
@@ -60,7 +59,7 @@ add_custom_command(OUTPUT ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
     COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
     COMMAND ${CMAKE_COMMAND} -E touch ${PADDLE_PYTHON_BUILD_DIR}/.timestamp
     COMMAND ${CMAKE_COMMAND} -E remove_directory ${PADDLE_PYTHON_BUILD_DIR}/lib-python
-    COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
+    COMMAND ${CMAKE_COMMAND} -E copy_directory ${PADDLE_PYTHON_BUILD_DIR}/lib.* ${PADDLE_PYTHON_BUILD_DIR}/lib-python
     DEPENDS gen_proto_py copy_paddle_pybind ${FLUID_CORE} framework_py_proto profiler_py_proto ${PY_FILES} ${external_project_dependencies} ${COPY_PADDLE_MASTER})
 
 set(paddle_python_deps ${PADDLE_PYTHON_BUILD_DIR}/.timestamp ${MKL_DEPENDS})
diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
index 775c2253ab3b27708b745b85fc007fcb504d1aed..6a129b6df9bf1830fdf5eb5cb9ae0c5e4f7bb4ec 100644
--- a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
+++ b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py
@@ -16,6 +16,8 @@ from __future__ import print_function
 import unittest
 import numpy as np
 from op_test import OpTest
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
 
 
 class ElementwiseMulOp(OpTest):
@@ -115,5 +117,56 @@ class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp):
         }
 
 
+class TestElementWiseMulSelectedRows(OpTest):
+    def setUp(self):
+        self.rows = [0, 1, 2, 3, 4, 5, 6]
+        self.feature = 12
+        self.height = 100
+        self.input_shape = (len(self.rows), self.feature)
+
+    def prepare_input(self, scope, place):
+        self.input = {
+            "X": np.random.random(self.input_shape).astype("float32"),
+            "Y": np.random.random(self.input_shape).astype("float32")
+        }
+
+        def init_input(in_name):
+            x_selected_rows = scope.var(in_name).get_selected_rows()
+            x_selected_rows.set_height(self.height)
+            x_selected_rows.set_rows(self.rows)
+            x_array = self.input[in_name]
+            x_tensor = x_selected_rows.get_tensor()
+            x_tensor.set(x_array, place)
+
+        init_input("X")
+        init_input("Y")
+
+    def create_out_selected_row(self, scope):
+        return scope.var('Out').get_selected_rows()
+
+    def check_result(self, out_selected_rows):
+        assert out_selected_rows.height() == self.height
+        assert out_selected_rows.rows() == self.rows
+        out_tensor = np.array(out_selected_rows.get_tensor())
+        assert out_tensor.shape == self.input_shape
+
+    def check_with_place(self, place):
+        scope = core.Scope()
+        self.prepare_input(scope, place)
+
+        out_selected_rows = self.create_out_selected_row(scope)
+        out_selected_rows.set_height(0)
+        out_selected_rows.set_rows([])
+
+        elementwise_mul = Operator("elementwise_mul", X='X', Y='Y', Out='Out')
+        elementwise_mul.run(scope, place)
+        self.check_result(out_selected_rows)
+
+    def test_elewisemul_with_selected_rows_input(self):
+        places = [core.CPUPlace()]
+        for place in places:
+            self.check_with_place(place)
+
+
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/setup.py.in b/python/setup.py.in
index 786c9f2e39880b68700b8acb94b3d35a48323958..6e439d28ac61a28e8c096f7cbdd9f225f9966a66 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -27,7 +27,7 @@ def _get_version_detail(idx):
     if re.match('@TAG_VERSION_REGEX@', '@PADDLE_VERSION@'):
         version_details = '@PADDLE_VERSION@'.split('.')
 
-        if len(version_details) == 3:
+        if len(version_details) >= 3:
             return version_details[idx]
 
     return 0