Created by: bingyanghuang

This is our internal branch ResNet50 optimization integration , you can use this branch to reproduce the latest MKLDNN ResNet50 performance. This PR is only for temporary test, not for merge.

Here is my benchmark. Test Platform: Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz Environment Config: OMP_NUM_THREADS=1, unset KMP_AFFINITY, Enable Turbo boost Command line: ./paddle/fluid/inference/tests/api/test_analyzer_image_classification --use_fake_data --infer_model=third_party/inference_demo/resnet50/model --iterations=100 --height=318 --width=318 --skip_batch_num=5 --batch_size= Test with MKL: export FLAGS_use_mkldnn=False Test with MKLDNN: export FLAGS_use_mkldnn=True Input image size: 318x318