openvino icon indicating copy to clipboard operation
openvino copied to clipboard

Published 20 hours ago • verified publisher icon openvinotoolkit

[CPU] Add test case for precision fp16

Open tiger100256-hu opened this issue 1 year ago • 11 comments

Details:

  • https://github.com/openvinotoolkit/openvino/pull/16500#issuecomment-1602581987
  • add test case for conv dconv fullconnect matmul mvn pad pooling subgraph softmax

Tickets:

  • CVS-110112

tiger100256-hu avatar Sep 11 '23 01:09 tiger100256-hu

This PR will be closed in a week because of 2 weeks of no activity.

github-actions[bot] avatar Oct 27 '23 00:10 github-actions[bot]

@usstq help to continue to review thanks

tiger100256-hu avatar Dec 05 '23 02:12 tiger100256-hu

@usstq continue to review, thanks

tiger100256-hu avatar Jan 15 '24 08:01 tiger100256-hu

@dmitry-gorokhov Could you please help review? Thanks!

yuxu42 avatar Jan 23 '24 05:01 yuxu42

@maxnick @EgorDuplensky Guys, please review the PR

dmitry-gorokhov avatar Jan 23 '24 07:01 dmitry-gorokhov

@tiger100256-hu, general comment. You introduced fp16 precision in almost every subgraph test. But, not all of them even need to be run in fp16 as the testing code path is not even impacted by the precision. Also, in the significant part of them the real primitive precision is not checked (because it's not necessary in many cases), thus running them in fp16 with nodes that don't support this precision ends up in running fp32 code path once again "silently". So, as the first step of the fp16 tests introduction, I suggest adding the fp16 precision only to those subgraph tests that have bf16 instances. Also, for the SL tests a rule of thumb is adding fp16 where bf16 instances exist, plus the operations, which already support fp16 in the plugin. Please perform this step first so that we can proceed with the review without being distracted by fp16 test instances that need to be removed.

maxnick avatar Jan 26 '24 10:01 maxnick

@maxnick thanks, I will update it, add fp16 where bf16 instances exist

tiger100256-hu avatar Jan 29 '24 06:01 tiger100256-hu

@maxnick help to continue to review thanks

tiger100256-hu avatar Feb 05 '24 01:02 tiger100256-hu

@iefode could you please review again?

yuxu42 avatar Feb 07 '24 01:02 yuxu42

@usstq @maxnick @iefode help to review

tiger100256-hu avatar Apr 16 '24 11:04 tiger100256-hu

Hi @maxnick Could you please review again? Thanks!

yuxu42 avatar May 11 '24 08:05 yuxu42

@iefode , do you have any additional comments or objections?

maxnick avatar Jun 26 '24 08:06 maxnick

@iefode , do you have any additional comments or objections?

@maxnick Honestly, no. Just FYI: to hardcode some specific like type, hw device and etc in shared lib is not got good practice

iefode avatar Jun 27 '24 08:06 iefode