[CPU][MatMul][FC] enable 3d wei matmul for fc

[azhai219]

Details:

• 3-D matmul => FC => dnnl::matmul.
  • In this case, only 3-D with const weight will be converted to FC. Otherwise, it fallback to matmul node.
  • This change is mainly for gpt-oss model with 3-D matmul.
• Only x64 arch is affected. Both arm and risc-v remain previous behaviour.

Tickets:

• 171918

[maxnick]

@EgorDuplensky , could you please review this PR?

[EgorDuplensky]

@azhai219 Does the PR also fixes usage of dnnl::matmul as FullyConnected node executor for compressed weights?
If yes, could you please double check by running 'matmul_weights_decompression.cpp' test scope using 'sde -gnr' for example.

[EgorDuplensky]

Could you please clarify this change?
Does it mean we never want to reshape down to rank 2 anymore (the logic below this if)?
If yes, can we simplify the code by adding a lambda:

     auto weightsFormat = [](const size_t rank, const bool weightsNonTransposed) {
         switch (rank) {
             case 2:
                 return weightsNonTransposed ? dnnl::memory::format_tag::ab : dnnl::memory::format_tag::ba;
             case 3:
                 return weightsNonTransposed ? dnnl::memory::format_tag::abc : dnnl::memory::format_tag::bac;
             default:
                 OPENVINO_THROW("DnnlMatmulExecutor: unsupported weights rank: ", rank);
         }
     };

[azhai219]

@EgorDuplensky In 3d case, when weightsNonTransposed is false, the tag should be acb not bac.

[EgorDuplensky]

sure, acb
So, can we implement it this way?

[EgorDuplensky]

Could you please use a named variable for this combination of checks. The name should help to explain why we disable tensor parallel in this case (like is3DwithMultipleInputChannels or maybe there is a better name if we are talking about transformers).
Also better to use shape.getDims().back() in this case.

[azhai219]

Could you please use a named variable for this combination of checks. The name should help to explain why we disable tensor parallel in this case (like is3DwithMultipleInputChannels or maybe there is a better name if we are talking about transformers). Also better to use shape.getDims().back() in this case.

OK. fix it.

[EgorDuplensky]

what about the variable?
const bool is3DwithMultipleInputChannels = rank == 3 && shape.getDims()[2] > 1;
if (is3DwithMultipleInputChannels) ...

[yuxu42]

Hi @EgorDuplensky @praasz could you please review again? thanks!

[praasz]

@azhai219
It is OK for me, wait for CPU team member approval.