diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index f147cc5a6e233ff88208f4f966ab93e8737b6f13..66840a2e037e7ca0fd1eacc64421865b170b47f8 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -402,19 +402,28 @@ void OperatorWithKernel::Run(const Scope& scope,
   OpKernelMap& kernels = kernels_iter->second;
 
   ExecutionContext ctx(*this, scope, *dev_ctx);
-  auto kernel_key = GetKernelType(ctx);
-  auto kernel_iter = kernels.find(kernel_key);
+  auto actual_kernel_key = GetActualKernelType(ctx);
+  auto expected_kernel_key = GetExpectedKernelType(actual_kernel_key);
+  auto kernel_iter = kernels.find(expected_kernel_key);
 
   if (kernel_iter == kernels.end()) {
-    PADDLE_THROW("The operator %s does not support %s", type_, kernel_key);
+    PADDLE_THROW("The operator %s does not support %s", type_,
+                 expected_kernel_key);
   }
 
   kernel_iter->second->Compute(ctx);
 }
-OpKernelType OperatorWithKernel::GetKernelType(
+
+OpKernelType OperatorWithKernel::GetActualKernelType(
     const ExecutionContext& ctx) const {
   return OpKernelType(IndicateDataType(ctx), ctx.GetPlace());
 }
+
+OpKernelType OperatorWithKernel::GetExpectedKernelType(
+    const OpKernelType& actual_kernel_type) const {
+  return actual_kernel_type;
+}
+
 proto::DataType OperatorWithKernel::IndicateDataType(
     const ExecutionContext& ctx) const {
   auto& scope = ctx.scope();
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index b592eea1b96113a8cbcb0e137890927cfcc22670..55eed57e6665515aec36dab4be8028dc75dbf7f3 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -52,6 +52,11 @@ constexpr char kGradVarSuffix[] = "@GRAD";
 /// Variables with this suffix are supposed to be filled up with zeros.
 constexpr char kZeroVarSuffix[] = "@ZERO";
 
+// define some kernel hint
+const std::string kUseCPU = "use_cpu";
+const std::string kUseCUDNN = "use_cudnn";
+const std::string kUseMKLDNN = "use_mkldnn";
+
 inline std::string GradVarName(const std::string& var_name) {
   return var_name + kGradVarSuffix;
 }
@@ -373,7 +378,9 @@ class OperatorWithKernel : public OperatorBase {
   }
 
  protected:
-  virtual OpKernelType GetKernelType(const ExecutionContext& ctx) const;
+  virtual OpKernelType GetActualKernelType(const ExecutionContext& ctx) const;
+  virtual OpKernelType GetExpectedKernelType(
+      const OpKernelType& actual_kernel_type) const;
 
  private:
   // indicate kernel DataType by input data. Defaultly all input data must be
diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc
index fbca45b59dc5446e93e79599f471d80a06ea3661..4d38a7ada91af834aa1a19b49e36d606ebe786ba 100644
--- a/paddle/framework/operator_test.cc
+++ b/paddle/framework/operator_test.cc
@@ -114,7 +114,7 @@ class OpWithKernelTest : public OperatorWithKernel {
 
  protected:
   void InferShape(framework::InferShapeContext* ctx) const override {}
-  OpKernelType GetKernelType(const ExecutionContext& ctx) const override {
+  OpKernelType GetActualKernelType(const ExecutionContext& ctx) const override {
     return OpKernelType(proto::DataType::FP32, ctx.GetPlace());
   }
 };
diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc
index b8ed93f4eb549fbd76bf360d4b843c1fa9635b40..d7baa6e90538484b400f32587a052d394a8d10d5 100644
--- a/paddle/operators/accuracy_op.cc
+++ b/paddle/operators/accuracy_op.cc
@@ -53,7 +53,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),
diff --git a/paddle/operators/auc_op.cc b/paddle/operators/auc_op.cc
index 811c487089fcf4044f129ad6bf95b46535d4fcd6..c16bc11931e6733d567107913521eafc34a30066 100644
--- a/paddle/operators/auc_op.cc
+++ b/paddle/operators/auc_op.cc
@@ -39,7 +39,7 @@ class AucOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Out")->type()),
diff --git a/paddle/operators/batch_norm_op.cc b/paddle/operators/batch_norm_op.cc
index 1c14acbe11fbad9654bd0309f5674176ebdb5e6f..49cb0fa4d9d8ada649105a9b14ad960f6cda1a7e 100644
--- a/paddle/operators/batch_norm_op.cc
+++ b/paddle/operators/batch_norm_op.cc
@@ -304,7 +304,7 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     const auto *var = ctx.InputVar(framework::GradVarName("Y"));
     if (var == nullptr) {
diff --git a/paddle/operators/chunk_eval_op.cc b/paddle/operators/chunk_eval_op.cc
index f1f274a7af079d68c7c1bcd8ec07962e18b0ea60..a040404266c3cd44230b141cfed1aaede3f05187 100644
--- a/paddle/operators/chunk_eval_op.cc
+++ b/paddle/operators/chunk_eval_op.cc
@@ -55,7 +55,7 @@ class ChunkEvalOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(framework::proto::DataType::FP32,
                                    ctx.device_context());
diff --git a/paddle/operators/compare_op.cc b/paddle/operators/compare_op.cc
index 1148172f3a2cc9b3f849ee04cefc19f16742d3eb..10bf3d4bbc38a0809a071acc8a154a63515dc42d 100644
--- a/paddle/operators/compare_op.cc
+++ b/paddle/operators/compare_op.cc
@@ -66,9 +66,9 @@ class CompareOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
     // CompareOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;
diff --git a/paddle/operators/crf_decoding_op.cc b/paddle/operators/crf_decoding_op.cc
index 27d0871f82beed4ceb3a4439be097a580631d4c6..024e1d061a5b2eabc27110d5379cb8226a104079 100644
--- a/paddle/operators/crf_decoding_op.cc
+++ b/paddle/operators/crf_decoding_op.cc
@@ -120,12 +120,18 @@ class CRFDecodingOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
         ctx.device_context());
   }
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::OpKernelType& actual_kernel_type) const override {
+    return framework::OpKernelType(actual_kernel_type.data_type_,
+                                   platform::CPUPlace());
+  }
 };
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/cross_entropy_op.cc b/paddle/operators/cross_entropy_op.cc
index 1ab7c0a06f85f332b290cb6cac82d0cfbe8f3242..a9c5c7046f8642072f2fdef14e1800ce7b05915b 100644
--- a/paddle/operators/cross_entropy_op.cc
+++ b/paddle/operators/cross_entropy_op.cc
@@ -51,7 +51,7 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -101,7 +101,7 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of cross_entropy
   // is determined by its input "X".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
diff --git a/paddle/operators/fill_constant_batch_size_like_op.cc b/paddle/operators/fill_constant_batch_size_like_op.cc
index 7a7e280e78309582a627087bdbdfea358c37b9eb..852ecdfe45e7f4737a505c1f722d25457ad6ad32 100644
--- a/paddle/operators/fill_constant_batch_size_like_op.cc
+++ b/paddle/operators/fill_constant_batch_size_like_op.cc
@@ -49,7 +49,7 @@ class FillConstantBatchSizeLikeOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),
diff --git a/paddle/operators/gather_op.cc b/paddle/operators/gather_op.cc
index 47af222314c40a2c77ee422ccc70602078b3f1fb..45e9d8df702403e66f9100e4edaf5c17470eb20d 100644
--- a/paddle/operators/gather_op.cc
+++ b/paddle/operators/gather_op.cc
@@ -40,7 +40,7 @@ class GatherOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
@@ -57,7 +57,7 @@ class GatherGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc
index 5eab1d5f4ee067db602ab81a9df1854bcfaf78a8..da4d281081531ccbf09b262db8e0a0b6a22f57c7 100644
--- a/paddle/operators/gaussian_random_op.cc
+++ b/paddle/operators/gaussian_random_op.cc
@@ -57,7 +57,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),
diff --git a/paddle/operators/linear_chain_crf_op.cc b/paddle/operators/linear_chain_crf_op.cc
index ad15e8ebd2b323929a4448e98a18c5cad6f5ed12..666207ea07628ca5f2a8313fa3f5febda140a294 100644
--- a/paddle/operators/linear_chain_crf_op.cc
+++ b/paddle/operators/linear_chain_crf_op.cc
@@ -183,7 +183,7 @@ class LinearChainCRFOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of computation kernel of linear_chain_crf
   // is determined by its input "Emission".
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("Emission")->type()),
@@ -242,7 +242,7 @@ class LinearChainCRFGradOp : public framework::OperatorWithKernel {
  protected:
   // Explicitly set that the data type of output of the linear_chain_crf_grad
   // operator is determined by its input: gradients of LogLikelihood.
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(
diff --git a/paddle/operators/lod_reset_op.cc b/paddle/operators/lod_reset_op.cc
index ccb87258c6b8629cd18d08185bfcc84c247070dd..f33874bd7b7d680f2b620c1bf32fdcad7a45c1ee 100644
--- a/paddle/operators/lod_reset_op.cc
+++ b/paddle/operators/lod_reset_op.cc
@@ -38,7 +38,7 @@ class LoDResetOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
@@ -97,7 +97,7 @@ class LoDResetGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
diff --git a/paddle/operators/logical_op.cc b/paddle/operators/logical_op.cc
index 2bd6c6efae38d6d8d49cc9f3fd97cf316fbbdd0a..ee8e4dd2adb31e3c80f9edbea8825796f82ed043 100644
--- a/paddle/operators/logical_op.cc
+++ b/paddle/operators/logical_op.cc
@@ -99,9 +99,9 @@ class LogicalOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
-    framework::OpKernelType kt = OperatorWithKernel::GetKernelType(ctx);
+    framework::OpKernelType kt = OperatorWithKernel::GetActualKernelType(ctx);
     // LogicalOp kernel's device type is decided by input tensor place
     kt.place_ = ctx.Input<framework::LoDTensor>("X")->place();
     return kt;
diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc
index 0a9defa8c50453abf3eefdcb89126b1349d6d756..73b746492905e0db63ec38eb3b7ba2d7b01ebbcf 100644
--- a/paddle/operators/lookup_table_op.cc
+++ b/paddle/operators/lookup_table_op.cc
@@ -41,7 +41,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),
@@ -98,7 +98,7 @@ class LookupTableOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<LoDTensor>("W")->type()),
diff --git a/paddle/operators/lstm_op.cc b/paddle/operators/lstm_op.cc
index f82156170e672b5e590ddb8e0e6e8a2a24ea6868..b8fcec0f29b46e838f91ad1ee0fded8e42f27bd5 100644
--- a/paddle/operators/lstm_op.cc
+++ b/paddle/operators/lstm_op.cc
@@ -92,7 +92,7 @@ class LSTMOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),
@@ -260,7 +260,7 @@ class LSTMGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("Input")->type()),
diff --git a/paddle/operators/multiplex_op.cc b/paddle/operators/multiplex_op.cc
index f524de60dbb3c652aa2a74478af6c0e38fb3cb43..d25e4c269c7647582915294338768e7088c14ab2 100644
--- a/paddle/operators/multiplex_op.cc
+++ b/paddle/operators/multiplex_op.cc
@@ -51,7 +51,7 @@ class MultiplexOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),
@@ -102,7 +102,7 @@ class MultiplexGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.MultiInput<Tensor>("X")[0]->type()),
diff --git a/paddle/operators/nce_op.cc b/paddle/operators/nce_op.cc
index 6dd457f7a2e410b65680004599ab753acbb34f71..d39ca87d53518963f652f7b8c8cb289a6fef70fd 100644
--- a/paddle/operators/nce_op.cc
+++ b/paddle/operators/nce_op.cc
@@ -63,7 +63,7 @@ class NCEOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
@@ -166,7 +166,7 @@ class NCEOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Input")->type()),
diff --git a/paddle/operators/pool_with_index_op.cc b/paddle/operators/pool_with_index_op.cc
index 980e9dc08b2ac160e6e06dfb11ff8f3e1279be46..76c5123527c3ff5b7e6c7eec39f4eb1d612759d3 100644
--- a/paddle/operators/pool_with_index_op.cc
+++ b/paddle/operators/pool_with_index_op.cc
@@ -69,7 +69,7 @@ class MaxPoolWithIndexOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -90,7 +90,7 @@ class MaxPoolWithIndexOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
diff --git a/paddle/operators/positive_negative_pair_op.cc b/paddle/operators/positive_negative_pair_op.cc
index c607c93a15609e51a3019b91b67aa328abf6a054..a6b23c995b8b9104f2da9d3d29ceb3eb88e7da63 100644
--- a/paddle/operators/positive_negative_pair_op.cc
+++ b/paddle/operators/positive_negative_pair_op.cc
@@ -85,7 +85,7 @@ class PositiveNegativePairOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Score")->type()),
diff --git a/paddle/operators/precision_recall_op.cc b/paddle/operators/precision_recall_op.cc
index 21dcd28c67bb5eb1d3af0ac8ba16f1d5df1958a8..c5753147effd17c012683e1058e34af46288f366 100644
--- a/paddle/operators/precision_recall_op.cc
+++ b/paddle/operators/precision_recall_op.cc
@@ -80,7 +80,7 @@ class PrecisionRecallOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext &ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("MaxProbs")->type()),
diff --git a/paddle/operators/roi_pool_op.cc b/paddle/operators/roi_pool_op.cc
index 85b6a8e15160d0c259a270f5e12ca9e67a6508ab..ef1804d9762200686ac8537140af046c21443779 100644
--- a/paddle/operators/roi_pool_op.cc
+++ b/paddle/operators/roi_pool_op.cc
@@ -68,7 +68,7 @@ class ROIPoolOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -89,7 +89,7 @@ class ROIPoolGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
diff --git a/paddle/operators/scatter_op.cc b/paddle/operators/scatter_op.cc
index 173c9582557eb4e020824d5830731e3e2312dc3c..806dccc6ca78bf64da828fe13e08e043097bd939 100644
--- a/paddle/operators/scatter_op.cc
+++ b/paddle/operators/scatter_op.cc
@@ -49,7 +49,7 @@ class ScatterOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Ref")->type()),
@@ -68,7 +68,7 @@ class ScatterGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Ref")->type()),
diff --git a/paddle/operators/sequence_pool_op.cc b/paddle/operators/sequence_pool_op.cc
index 0eb675caaddf1274a941cbfe29017cb9ea11f40f..47f5bd891f3a5c661111d1f6fb6dbc13292f702b 100644
--- a/paddle/operators/sequence_pool_op.cc
+++ b/paddle/operators/sequence_pool_op.cc
@@ -107,7 +107,7 @@ class SequencePoolGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("X")->type()),
diff --git a/paddle/operators/sequence_slice_op.cc b/paddle/operators/sequence_slice_op.cc
index 309ee1f3a82c35104db74084c4ef761bd4b06695..98bd8854903e5abf6d27432a2af0aaae980c0b1d 100644
--- a/paddle/operators/sequence_slice_op.cc
+++ b/paddle/operators/sequence_slice_op.cc
@@ -48,7 +48,7 @@ class SequenceSliceOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
@@ -69,7 +69,7 @@ class SequenceSliceGradOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
diff --git a/paddle/operators/softmax_with_cross_entropy_op.cc b/paddle/operators/softmax_with_cross_entropy_op.cc
index d9911a6901447d8900c3881a60c7a0852dcbf429..13266d394d161f9cf073e2aeab33ad7420de49af 100644
--- a/paddle/operators/softmax_with_cross_entropy_op.cc
+++ b/paddle/operators/softmax_with_cross_entropy_op.cc
@@ -118,7 +118,7 @@ class SoftmaxWithCrossEntropyOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<Tensor>("Logits")->type()),
@@ -159,7 +159,7 @@ class SoftmaxWithCrossEntropyOpGrad : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(
diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc
index 891839bf9cd991e15d96b86e24ea61b09e35a7c7..b86e8266425ca094a51d224fd39ce33700057f13 100644
--- a/paddle/operators/sum_op.cc
+++ b/paddle/operators/sum_op.cc
@@ -53,7 +53,7 @@ class SumOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     auto x_vars = ctx.MultiInputVar("X");
     if (x_vars[0]->IsType<framework::LoDTensor>()) {
diff --git a/paddle/operators/uniform_random_op.cc b/paddle/operators/uniform_random_op.cc
index 3c705cb3396f68f88882388675ab145660e13070..e985e491e9b7343eb96494e15d28a0c3438a4807 100644
--- a/paddle/operators/uniform_random_op.cc
+++ b/paddle/operators/uniform_random_op.cc
@@ -63,7 +63,7 @@ class UniformRandomOp : public framework::OperatorWithKernel {
   }
 
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         static_cast<framework::proto::DataType>(ctx.Attr<int>("dtype")),
diff --git a/paddle/operators/unpool_op.cc b/paddle/operators/unpool_op.cc
index 1b682d5c72b003eccb3ea411048b6bcdf737bc59..aeed9679b2a3cce744189809c223a5b0d581ccdd 100644
--- a/paddle/operators/unpool_op.cc
+++ b/paddle/operators/unpool_op.cc
@@ -71,7 +71,7 @@ int OutputSize(int input_size, int ksize, int padding, int stride) {
 
 class UnpoolOp : public framework::OperatorWithKernel {
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
@@ -110,7 +110,7 @@ class UnpoolOp : public framework::OperatorWithKernel {
 
 class UnpoolOpGrad : public framework::OperatorWithKernel {
  protected:
-  framework::OpKernelType GetKernelType(
+  framework::OpKernelType GetActualKernelType(
       const framework::ExecutionContext& ctx) const override {
     return framework::OpKernelType(
         framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
diff --git a/paddle/pybind/const_value.cc b/paddle/pybind/const_value.cc
index b13ad42ea29453354798d88bff8ef47339d1a614..761635aa5e5eac445c2ec8331b0dc37ffd11248c 100644
--- a/paddle/pybind/const_value.cc
+++ b/paddle/pybind/const_value.cc
@@ -23,6 +23,11 @@ void BindConstValue(pybind11::module& m) {
   m.def("kTempVarName", [] { return framework::kTempVarName; });
   m.def("kGradVarSuffix", [] { return framework::kGradVarSuffix; });
   m.def("kZeroVarSuffix", [] { return framework::kZeroVarSuffix; });
+
+  // for kernel_hint key
+  m.def("kUseCPU", [] { return framework::kUseCPU; });
+  m.def("kUseCUDNN", [] { return framework::kUseCUDNN; });
+  m.def("kUseMKLDNN", [] { return framework::kUseMKLDNN; });
 }
 
 }  // namespace pybind
diff --git a/python/paddle/v2/fluid/framework.py b/python/paddle/v2/fluid/framework.py
index 7b65fe80aed6a450c7aea1ef5e0fcf03a2a26686..add854306ea7fa527943de871d2716cd2aa9f530 100644
--- a/python/paddle/v2/fluid/framework.py
+++ b/python/paddle/v2/fluid/framework.py
@@ -17,6 +17,10 @@ TEMP_VAR_NAME = core.kTempVarName()
 GRAD_VAR_SUFFIX = core.kGradVarSuffix()
 ZERO_VAR_SUFFIX = core.kZeroVarSuffix()
 
+USE_CPU = core.kUseCPU()
+USE_CUDNN = core.kUseMKLDNN()
+USE_MKLDNN = core.kUseMKLDNN()
+
 
 def grad_var_name(var_name):
     """