diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc
index 60c761c5281e2f535aab0200c93fb738addcdb87..987da18116cc6f4902bd66ae317f2470a8bc5057 100644
--- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc
+++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc
@@ -21,6 +21,7 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model.");
 DEFINE_int32(batch_size, 1, "Batch size of input data");
 DEFINE_int32(repeat, 1, "Running the inference program repeat times");
 DEFINE_bool(skip_cpu, false, "Skip the cpu test");
+DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
 
 TEST(inference, image_classification) {
   if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
@@ -58,8 +59,10 @@ TEST(inference, image_classification) {
     // Run inference on CPU
     LOG(INFO) << "--- CPU Runs: ---";
     LOG(INFO) << "Batch size is " << FLAGS_batch_size;
+    LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn;
     TestInference<paddle::platform::CPUPlace, false, true>(
-        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined);
+        dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined,
+        FLAGS_use_mkldnn);
     LOG(INFO) << output1.dims();
   }
 
diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h
index b02e5c99f00eaf03c3753e43575cbc67e834774e..61334a10dea1019ac12dfad23dcc146c4af937a4 100644
--- a/paddle/fluid/inference/tests/test_helper.h
+++ b/paddle/fluid/inference/tests/test_helper.h
@@ -133,11 +133,24 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
   return feed_target_shapes;
 }
 
+void EnableMKLDNN(
+    const std::unique_ptr<paddle::framework::ProgramDesc>& program) {
+  for (size_t bid = 0; bid < program->Size(); ++bid) {
+    auto* block = program->MutableBlock(bid);
+    for (auto* op : block->AllOps()) {
+      if (op->HasAttr("use_mkldnn")) {
+        op->SetAttr("use_mkldnn", true);
+      }
+    }
+  }
+}
+
 template <typename Place, bool CreateVars = true, bool PrepareContext = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
-                   const int repeat = 1, const bool is_combined = false) {
+                   const int repeat = 1, const bool is_combined = false,
+                   const bool use_mkldnn = false) {
   // 1. Define place, executor, scope
   auto place = Place();
   auto executor = paddle::framework::Executor(place);
@@ -169,6 +182,9 @@ void TestInference(const std::string& dirname,
         "init_program",
         paddle::platform::DeviceContextPool::Instance().Get(place));
     inference_program = InitProgram(&executor, scope, dirname, is_combined);
+    if (use_mkldnn) {
+      EnableMKLDNN(inference_program);
+    }
   }
   // Disable the profiler and print the timing information
   paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,