Egor Duplenskii

@maxnick Comments applied, could you please take a look

> @EgorDuplensky rebased on the latest master and also add the following changes to the latest commit [12b9a9f](https://github.com/openvinotoolkit/openvino/commit/12b9a9f2fa743b484cc10aec370176a8b1b8bc9c): > > * Added SHL [tests](https://github.com/openvinotoolkit/openvino/blob/866bbbeee8c0850bd0022a28d79a974ea471714d/src/plugins/intel_cpu/tests/functional/custom/single_layer_tests/instances/riscv64/shl/matmul.cpp) for FC > * Disabled SHL...

I have also applied the patch locally. Working perfectly, great improvement for the example-region!

@dmitry-gorokhov @RICKIE777 I have applied the change to the original commit in scope of oneDNN 3.0 migration. This one can be closed. Thank you.

PR: - https://github.com/gagbo/consult-lsp/pull/49

just FYI, I am looking at this and also trying to come up with a way to somehow simplify the idea in its core

@liubo-intel What about changing the approach a bit. I was just thinking that this optimization is very similar (probably almost the same) as those ones: "MoveEltwiseUpThroughDataMov, MoveEltwiseUpThroughDataMovPerChannel". The idea is...

@maxnick Could you please take a look?

Just trying to clarify the point regarding insufficient shape size, specifically a channel size 12. If a kernel's internal loop is performed only over the channel dimension, then, for fp16...

@dmitry-gorokhov Ready for further steps