rfcs: add a proposal on MatMul reduction API

[densamoilov]

Link to rendered document

[mgouicem]

Thanks Denis.
Just a quick comment: it seems there are two problems you are trying to tackle:

• simplify internals by using matmul under inner-product,
• and simplify external interface to ditch inner-product.

I would recommend to focus on the first one and keep the matmul extension internal as a first step. We can work on gathering feedback from users for the second part in parallel. What do you think?

Regarding using extension enum and reuse bias_md, that works for me (and much less complex that pre-op/attribute).

[densamoilov]

@mgouicem, thanks for the prompt review!

To be precise, the first bullet is the primary goal, while the second represents my thought that the feature could provide value to some users, leading me to suggest making it external. I'm totally fine with making the extensions internal at first.