MKLDNN error for ppyolov2 model

[xiegegege]

Paddle Test Environment

---

Paddle version: 0.0.0
Paddle With CUDA: True

OS: Ubuntu 16.04
Python version: 3.7.0

CUDA version: 10.1.243
cuDNN version: 7.6
Nvidia driver version: None
CPU: Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz

---

install paddle develop (commit id:1e62c239d323354eccfc974d4e2e6496f93d848e)

git clone https://github.com/PaddlePaddle/PaddleDetection.git 
cd PaddleDetection
git checkout develop

python -m pip install -r requirments.txt
python tools/export_model.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --output_dir=inference_model -o weights=https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams
python deploy/python/infer.py --model_dir=./inference_model/ppyolov2_r50vd_dcn_365e_coco/ --image_file=demo/000000570688.jpg --device=CPU --enable_mkldnn=True

I have tested this case and it works well in paddle2.1.0 and paddle2.1.1， so this should be the new problem caused by paddle develop. The model ttfnet_darknet53_1x_coco has the same problem.

ppyolov2_cpu_mkldnnn_log.txt

[paddle-bot-old]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[lidanqing-vv]

1. 6148 is capable with AVX512, You had this issue because you used some virtual machine ? @xiegegege

[xiegegege]

@lidanqing-intel Yes, it's a virtual machine. I tested in some other physical machine, and it works well.

[jakpiase]

Hi @xiegegege
#33944 will fix your issue. Your virtual machine probably uses AVX2 and we recommend switching to AVX512, because our performance improvements with oneDNN library are made mainly for that instruction set.

[xiegegege]

@jakpiase OK, got it, thank you.

[lidanqing-vv]

@xiegegege
#33944 will solve this error on AVX2 (your virtual machine). But if you test this model on AVX2 environment, it may not be speedup as well as on AVX512 (6148 itself suppots AVX512). Because oneDNN improve well on AVX512. But in AVX2, some oneDNN op will call reference version (slower).

Anyway, to achieve expected performance, please use avx512. This fix will only guarantee no error on AVX2, but could not guarantee performance on AVX2.

[xiegegege]

@lidanqing-intel OK，I will use avx512 in subsequent test.

[lidanqing-vv]

@xiegegege @OliverLPH Since the PR has been merged, if you have have verified ok, let me know, I will close this issue.