Created by: ddokupil

System information

• PaddlePaddle version: 234060f8 / develop branch
• CPU: including CLX with MKLDNN
• OS Platform Ubuntu 16.04
• Python version: 2.7.12
• Cmake orders

 -DCMAKE_BUILD_TYPE=Release -DWITH_GPU=OFF -DWITH_MKLDNN=ON -DWITH_TESTING=ON -DWITH_PROFILER=ON -DWITH_STYLE_CHECK=OFF -DON_INFER=ON -DWITH_INFERENCE_API_TEST=ON

• C++version: gcc (Ubuntu 5.4.0-6ubuntu1~16.04.11) 5.4.0 20160609
• API information

To Reproduce Build with above cmake and execute with:

/paddle_paddle/build/paddle/fluid/inference/tests/api/test_analyzer_bert
--gtest_filter=Analyzer_bert.profile
--infer_data=/paddle_paddle/build/third_party/inference_demo/bert_emb128/data.txt
--infer_model=/paddle_paddle/build/third_party/inference_demo/bert_emb128/model
--paddle_num_threads=1
--batch_size=1
--repeat=1000

Describe your current behavior

• Performance on 234060f8 is 3-4% lower than was on 6cfcbe05 (a committ before)

Code to reproduce the issue Other info / logs