Created by: chuanqi129

We found a FP32 latency regression of Image classification model especially on MobileNet from 3.04ms to 8.46ms based on SKX-8180 1 socket. This regression caused by PR#14707 and below commit ID:

commit 669191c9 Author: Yihua Xu yihua.xu@intel.com Date: Mon Dec 3 11:54:00 2018 +0800

Implement conv3d with mkldnn library (test=develop)

We can use CAPI Image classification application to reproduce it.

FLAGS_use_mkldnn=true FLAGS_paddle_num_threads=28 KMP_AFFINITY=compact,granularity=fine taskset -c 0-27 numactl -l  ./build/infer_image_classification \
        --infer_model=${infer_model} \
        --batch_size=1 \
        --profile \
        --skip_batch_num=10 \
        --iterations=1000  \
        --use_mkldnn \
        --paddle_num_threads=28  \
        --use_fake_data=true

You also can use python eval.py script to reproduce this latency regression, but it has lower regression ratio than CAPI caused by overhead. Please have a look into it. @luotao1 @yihuaxu @jianhang-liu