config. SetMkldnnCacheCapacity is useless

[juncaipeng]

Ubuntu
C++
MKLDNN

Download the demo (Link:https://dubox.com/s/1RVo80hVAC0zr5iZ2EuG0xQ Password:3v5m ) and refer to the readme.

[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-intel]

Reproduced

[lidanqing-intel]

I just tested again, From #32664 there are memory leak @jczaja Please fix this issue, thank you very much!

[jczaja]

@luotao1 ,
While working on this issue I had remove(#33174, still testing) some code you added at #18580 . My question is do you remember why restoring oneDNN caching was needed as done here:

Paddle/paddle/fluid/inference/api/analysis_predictor.cc

Lines 332 to 335 in 481ee79

	paddle::platform::MKLDNNDeviceContext::tls().set_cur_mkldnn_session_id(
	platform::MKLDNNDeviceContextThreadLocals::kMKLDNNSessionID_Default);
	platform::MKLDNNDeviceContext::tls().set_cur_input_shape_cache_capacity(0);
	platform::MKLDNNDeviceContext::tls().set_cur_input_shape_str("");

[luotao1]

I remember it is used to fix a detect model #18081 (comment) which input dims are dynamic, the model is in paddle/fluid/inference/tests/api/analyzer_detect_tester.cc

config. SetMkldnnCacheCapacity is used to balance the memory and speed. You can see the following data from the detect model. (#18372)

speed(ms)	memory	CacheCapacity
26.1437	1.5G	CacheCapacity=10
24.0129	3.5G	CacheCapacity=50

[jczaja]

@juncaipeng Just to post status of work on that one. There are at least three sources of memory leaks. We understand two of them. Fix to first one is #33174.

[jczaja]

@juncaipeng Two PRs with fixes to this issue were merged. Please test.

[lidanqing-intel]

@OliverLPH Hi I talked with Juncai, please test this issue, since two PRs are merged and should fix this.