Maksim Kutakov

> @dmitry-gorokhov @maxnick we will have difficulty when CPU_DENORMALS_OPTIMIZATION wasn't set and internal flag `denormalsOptMode` will be `DenormalsOptMode::DO_Keep` and we have no way to tell whether DAZ is set other...

> The code LGTM. We need to run Bench to prove the performance gain. I'll try to do it soon. Did you have a chance to check the performance?

@ceciliapeng2011, do we have enough data on performance validation to merge the PR?

> @maxnick I find that it's difficult to get graph's config inside constructor of Input node(for checking CPU_DENORMALS_OPTIMIZATION flag), actually it's hard to add any new input parameter into constructor...

@EgorDuplensky, could you please review?

@ceciliapeng2011 , could you please review?

Waiting for the internal validation.

>Deconv only has one ref entry(CPU_INSTANCE(ref_deconvolution_fwd_t) for JIT,gemm and c++ reference implement. So one layout can't use RTinfo to filter implement. Update test cases with 1x1 related or SSE related....

>Change stride to 1 unit because of brgemm limitation Probably should be set only for brgemm cases.

> subgraph concat deconv test can't hit ref implement because the gemm and ref share one entry implement. The same as above, simply remove the implementation type from config instead...