diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index 2a9a7aed480e76edbac4d5ba6d7bc3b8b2dc5006..cd52114713ba85cf5ced617a18fe901eb3c0cf86 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -77,6 +77,9 @@ bool AnalysisPredictor::Init(
 
   OptimizeInferenceProgram();
   ctx_ = executor_->Prepare(*inference_program_, 0);
+  if (config_.use_mkldnn) {
+    executor_->EnableMKLDNN(*inference_program_);
+  }
 
   VLOG(5) << "to create variables";
   PADDLE_ENFORCE(scope_.get());
diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc
index 6fe13ed027de403bdc21882c26225bcd4cc7e49a..c6cb09667e0e58962c073dbbb7affebf76cc185d 100644
--- a/paddle/fluid/inference/api/api_impl.cc
+++ b/paddle/fluid/inference/api/api_impl.cc
@@ -106,6 +106,9 @@ bool NativePaddlePredictor::Init(
   }
 
   ctx_ = executor_->Prepare(*inference_program_, 0);
+  if (config_.use_mkldnn) {
+    executor_->EnableMKLDNN(*inference_program_);
+  }
   executor_->CreateVariables(*inference_program_,
                              sub_scope_ ? sub_scope_ : scope_.get(), 0);
 
diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h
index 995da11e4a30eca72a91a53d3293aa8b033b012b..e8d51bb72cd08cb443b809d196c3e95f6050f5be 100644
--- a/paddle/fluid/inference/api/paddle_inference_api.h
+++ b/paddle/fluid/inference/api/paddle_inference_api.h
@@ -45,7 +45,7 @@ class PaddleBuf {
   PaddleBuf(void* data, size_t length)
       : data_(data), length_(length), memory_owned_{false} {}
   // Own memory.
-  PaddleBuf(size_t length)
+  explicit PaddleBuf(size_t length)
       : data_(new char[length]), length_(length), memory_owned_(true) {}
   // Resize to `length` bytes.
   void Resize(size_t length);
@@ -121,6 +121,8 @@ struct NativeConfig : public PaddlePredictor::Config {
   bool use_gpu{false};
   int device{0};
   float fraction_of_gpu_memory{-1.f};  // Negative to notify initialization.
+  // MKLDNN related fields.
+  bool use_mkldnn{false};
   // Specify the variable's name of each input.
   bool specify_input_name{false};
 
diff --git a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc
index 67bde723042f24326909dea1ed4305ede21dc824..135a81a85c2d86e1cb06c5b1adf02c23d91857f6 100644
--- a/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc
+++ b/paddle/fluid/inference/tests/api/analyzer_vis_tester.cc
@@ -66,12 +66,13 @@ Record ProcessALine(const std::string &line) {
  * Use the native and analysis fluid engine to inference the demo.
  * ocr, mobilenet and se_resnext50
  */
-void TestVisualPrediction() {
+void TestVisualPrediction(bool use_mkldnn) {
   std::unique_ptr<PaddlePredictor> predictor;
   AnalysisConfig cfg;
   cfg.param_file = FLAGS_infer_model + "/__params__";
   cfg.prog_file = FLAGS_infer_model + "/__model__";
   cfg.use_gpu = false;
+  cfg.use_mkldnn = use_mkldnn;
   cfg.device = 0;
   cfg.enable_ir_optim = true;
   cfg.ir_passes.push_back("fc_gru_fuse_pass");
@@ -163,7 +164,10 @@ void TestVisualPrediction() {
   }
 }
 
-TEST(Analyzer_vis, analysis) { TestVisualPrediction(); }
+TEST(Analyzer_vis, analysis) { TestVisualPrediction(/*use_mkldnn*/ false); }
+TEST(Analyzer_vis, analysis_mkldnn) {
+  TestVisualPrediction(/*use_mkldnn*/ true);
+}
 
 }  // namespace analysis
 }  // namespace inference