@GlebKazantaev, @ilyachur, @mvafin, @yekruglov, @rkazants, @vladimir-paramuzov, @dmitry-gorokhov, please, see this PR (#4354)

@GlebKazantaev, @ilyachur, @mvafin, @yekruglov, @rkazants, @vladimir-paramuzov, please, see this PR (#4354)

Do we have a performance numbers or tests which proof that the performance was fixed?

Do we have a performance numbers or tests which proof that the performance was fixed?

According to ticket content (see link to a performance report in the ticket), we have the following.

Before the merge of #970, for any Interpolate-4 nodes with 4D input data, the transformation ConvertInterpolate1ToInterpolate4 set the attribute mode of the operation Interpolate-4 as linear_onnx instead of linear, in the case of 4D input tensors.

Generally speaking, this behavior is not correct for non-ONNX models, and this behavior was corrected in the PR #970.

Next, plugins have no any vectorized implementation of the mode = linear, but have vectorized implementations of the mode linear_onnx. Hence, we have a performance drop (see the ticket).

This PR corrects the reason of this performance drop.

I created tickets to implement vectorized implementation in CPU and GPU plugins.

@GlebKazantaev, @ilyachur, @mvafin, @yekruglov, @rkazants, @vladimir-paramuzov, please, see this PR (#4354)