[WIP] src: cpu: x64: jit_brdgmm_dw_conw: make brdgmm support fusing sum op

[amakarev]

Brdgmm kernel could not be selected when using the --attr-post-ops=sum flag. jit_dw was used instead.
This PR makes it possible to use brdgmm kernel with sum post-op

The changes look trivial, In fact, there was already support, but brdgmm could not be dispatched because the function(has_default_values) signature was updated a long time ago but the function call was not updated. As a result, data_type_undef was always used inside, which prevented the brdgmm kernel from being selected

[inteldimitrius]

LGTM

[dzarukin]

dst_dt argument is responsible to support s8 sum post-op while dst data type is u8, or vice versa. It doesn't have anything about enabling the sum post-op per se. undef data type passed requires sum post-op to check if it uses exactly the same data type as destination tensor, not allowing divergence from the first sentence.

It would be nice to see the following:

• There's actually some coverage for brdgmm + sum post+op in ci (or some nightly) file.
• The support flipped from jit to brgemm for a regular sum post-op case (same data type) and for a sum_dt != dst_dt (updating description with verbose before and after should be fine).

[amakarev]

dst_dt argument is responsible to support s8 sum post-op while dst data type is u8, or vice versa. It doesn't have anything about enabling the sum post-op per se. undef data type passed requires sum post-op to check if it uses exactly the same data type as destination tensor, not allowing divergence from the first sentence.

It would be nice to see the following:

• There's actually some coverage for brdgmm + sum post+op in ci (or some nightly) file.
• The support flipped from jit to brgemm for a regular sum post-op case (same data type) and for a sum_dt != dst_dt (updating description with verbose before and after should be fine).

Got it, thanks for the clarification