[CPU][ARM] Dynamic shapes support in ARM transformations #17548
Conversation
|
@alvoron Do you have updates there? |
Tests are added, could you please review? |
|
I am just wondering why do we need separate subgraph tests there? It seems like we can add ARM instances for Convoltuion, GroupConvolution and Reduce CPU single layer tests to cover dynamic shapes for these operations and additionally correntess of decomposition in the whole pipeline. |
|
@EgorDuplensky Do you have any additional comments? |
| concat_inputs.push_back(conv); | ||
| } | ||
| auto concat = std::make_shared<ov::opset8::Concat>(concat_inputs, 1); | ||
| function = std::make_shared<ngraph::Function>(ngraph::NodeVector{ concat }, ngraph::ParameterVector{ param }); |
There was a problem hiding this comment.
I guess it should be "function_ref", otherwise we do not actually test anything here (and some other test instances as well).
TBH, I don't really like the idea of performing the test check inside TearDown, since from the test instance perspective it is not clear, what is being tested here.
I think we should explicitly compare the functions, i.e. how it is done in convert_to_interaction.cpp tests.
| ASSERT_TRUE(res.first) << res.second; | ||
| } | ||
|
|
||
| TEST(TransformationTests, CheckConvertReduceProdTransformationIsNotAppliedForDynaimcShapes) { |
There was a problem hiding this comment.
It seems all this duplication can be avoided using
https://github.com/google/googletest/blob/main/docs/advanced.md#typed-tests
Which makes sense, since the transformation itself and the functions you are using to create the models in scope of this test are also type templates.
There was a problem hiding this comment.
Applied typed tests approach
|
@dmitry-gorokhov could we merge it? |
Ticket: CVS-101331
Propagation of #17517