diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc
index 5a874fe437d83e2ba795a0b063d7f1811afa04d8..fb99d2a2bc59f4c01ea16d8dbb5b8a20916527fc 100644
--- a/paddle/fluid/framework/operator.cc
+++ b/paddle/fluid/framework/operator.cc
@@ -921,57 +921,74 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  // check if op[type] has kernel registered.
-  auto& all_op_kernels = AllOpKernels();
-  auto kernels_iter = all_op_kernels.find(type_);
-  if (kernels_iter == all_op_kernels.end()) {
-    PADDLE_THROW(
-        "There are no kernels which are registered in the %s operator.", type_);
-  }
+  if (!kernel_type_) {
+    // LOG(INFO) << "1, kernel_type is not set.";
+    // check if op[type] has kernel registered.
+    auto& all_op_kernels = AllOpKernels();
+    auto kernels_iter = all_op_kernels.find(type_);
+    if (kernels_iter == all_op_kernels.end()) {
+      PADDLE_THROW(
+          "There are no kernels which are registered in the %s operator.",
+          type_);
+    }
 
-  OpKernelMap& kernels = kernels_iter->second;
+    OpKernelMap& kernels = kernels_iter->second;
 
-  auto expected_kernel_key = this->GetExpectedKernelType(
-      ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr));
-  VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
+    auto expected_kernel_key = this->GetExpectedKernelType(
+        ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr));
+    VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
-  auto kernel_iter = kernels.find(expected_kernel_key);
+    auto kernel_iter = kernels.find(expected_kernel_key);
 #ifdef PADDLE_WITH_MKLDNN
-  // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
-  if (kernel_iter == kernels.end() &&
-      expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
-    VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
-    expected_kernel_key.library_type_ = LibraryType::kPlain;
-    expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
-    kernel_iter = kernels.find(expected_kernel_key);
-  }
+    // workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
+    if (kernel_iter == kernels.end() &&
+        expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
+      VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
+      expected_kernel_key.library_type_ = LibraryType::kPlain;
+      expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
+      kernel_iter = kernels.find(expected_kernel_key);
+    }
 #endif
-  if (kernel_iter == kernels.end()) {
-    PADDLE_THROW("op %s does not have kernel for %s", type_,
-                 KernelTypeToString(expected_kernel_key));
+    if (kernel_iter == kernels.end()) {
+      PADDLE_THROW("op %s does not have kernel for %s", type_,
+                   KernelTypeToString(expected_kernel_key));
+    }
+
+    kernel_type_.reset(new OpKernelType(expected_kernel_key));
+    kernel_func_.reset(new OpKernelFunc(kernel_iter->second));
   }
 
+  // std::shared_ptr<OpKernelType> kernel_type = kernel_type_;
+  // std::shared_ptr<OpKernelFunc> kernel_func = kernel_func_;
+
   std::vector<KernelConfig>* kernel_configs =
-      GetKernelConfig(expected_kernel_key);
+      // GetKernelConfig(expected_kernel_key);
+      GetKernelConfig(*kernel_type_);
 
   // do data transformScope &transfer_scope;
   std::vector<std::string> transfered_inplace_vars;
   auto* transfer_scope =
-      PrepareData(scope, expected_kernel_key, &transfered_inplace_vars, &ctx);
+      // PrepareData(scope, expected_kernel_key, &transfered_inplace_vars,
+      // &ctx);
+      PrepareData(scope, *kernel_type_, &transfered_inplace_vars, &ctx);
 
   // exec scope is the scope that kernel actually executed on.
   const Scope& exec_scope =
       (transfer_scope == nullptr ? scope : *transfer_scope);
 
-  if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
-    dev_ctx = pool.Get(expected_kernel_key.place_);
+  // if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
+  // dev_ctx = pool.Get(expected_kernel_key.place_);
+  if (!(kernel_type_->place_ == dev_ctx->GetPlace())) {
+    dev_ctx = pool.Get(kernel_type_->place_);
   }
 
   RuntimeInferShapeContext infer_shape_ctx(*this, exec_scope, ctx);
   this->InferShape(&infer_shape_ctx);
   // TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
   // not Scope. Imperative mode only pass inputs and get outputs.
-  kernel_iter->second(
+  // kernel_iter->second(
+  //     ExecutionContext(*this, exec_scope, *dev_ctx, ctx, kernel_configs));
+  (*kernel_func_)(
       ExecutionContext(*this, exec_scope, *dev_ctx, ctx, kernel_configs));
 
   if (!transfered_inplace_vars.empty()) {
diff --git a/paddle/fluid/framework/operator.h b/paddle/fluid/framework/operator.h
index 8a86813e9362d7b82c2023428a35a1982adb0508..11ef729706b7c02c32ea5df934122e153f6afece 100644
--- a/paddle/fluid/framework/operator.h
+++ b/paddle/fluid/framework/operator.h
@@ -541,6 +541,8 @@ class OperatorWithKernel : public OperatorBase {
 
  protected:
   mutable OpKernelConfigsMap kernel_configs_map_;
+  mutable std::shared_ptr<OpKernelType> kernel_type_;
+  mutable std::shared_ptr<OpKernelFunc> kernel_func_;
 };
 
 extern bool OpSupportGPU(const std::string& op_type);