FC+elementwise_add (residual connection)

[Silv3S]

PR types

Performance optimization

PR changes

OPs

Describe

• Fuse FC with elementwise_add so residual data is added inside FC operator,
• add unit tests with multiple cases for this fuse pass,
• download VIT-OCR model and perform accuracy test.

Benchmarks

Test were conducted with oneDNN 2.5.4 , fc_mkldnn_pass and fc_act_mkldnn_pass were enabled:

• VIT-OCR is performing 4% faster (12 fc+residual fusions),
• BERT is performing 8% faster (24 fc+residual fusions).

[paddle-bot-old]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zhiqiu]

LGTM

[Silv3S]

Hi @XieYunshen @chalsliu could you please approve setting TIMEOUT for newly added unit test?

[XieYunshen]

LGTM for set_tests_properties(test_mkldnn_fc_elementwise_add_fuse_pass PROPERTIES TIMEOUT 120)

[paddle-bot-old]

你的PR已合入Paddle库，请关注后续测试结果。
Your PR has been merged into the repository. An official integration test will be conducted later. Stay tuned.

[lidanqing-intel]

#40834 has been reviewed by many developers and performance and model UT was given. This PR is the same as #40834. Hence I merge it.

[baoachun]

Hi @Silv3S, this pr will cause the test_analyzer_vit_ocr single test execution of the Windows task to fail, could you plsease check it? https://xly.bce.baidu.com/paddlepaddle/paddle/newipipe/detail/5400245/job/14068509

[lidanqing-intel]

@Silv3S Hi, this new test failed on windows, please investigate thank you.

[Silv3S]

@baoachun @lidanqing-intel Thank you for letting me know. As I can see there is error with downloading model and data for test. That's unexpected because all windows tests were ok on PR CI and test_analyzer_vit_ocr works locally for me. I will investigate more and fix it

[lidanqing-intel]

Hi, @baoachun . Siwek checked it seems models downloading issue and the link address is rarely used in API tests before. Please move vit_ocr.tgz and datavit.txt to more frequently used link and combine them into one tgz.

original model link
https://paddle-qa.bj.bcebos.com/inference_model/2.1.1/ocr/vit_ocr.tgz
original data link
https://paddle-qa.bj.bcebos.com/inference_model/2.1.1/ocr/datavit.txt

Move the two files into one http://paddle-inference-dist.bj.bcebos.com/ocr/vit_ocr.tgz.
Its distracted files should look like

- vit_ocr/
  - datavit.txt
  - model/
    - inference.pdiparams
    - inference.pdmodel  

Thanks.

[Silv3S]

I changed config, and now VIT-OCR model and data will be downloaded from same source as other models #42041