rfc: support implicit concat

[dzarukin]

This RFC covers oneDNN vision on implicit concat support on user side, provides an example how it may be implemented and potential challenges to implement it.

Link to rendered document.

[densamoilov]

@dzarukin, have you shared the PoC to estimate the performance gain we can get?

[dzarukin]

@dzarukin, have you shared the PoC to estimate the performance gain we can get?

Discussed offline. I'll post if/once I get any interesting data.

[dzarukin]

Promote this RFC for the purpose of documenting the approach since this will likely be the way to go - use concat primitive to get descriptors and re-create original primitive descriptors to make an implicit concat approach. What's not covered in the library:

• CPU backend (entirely), including reorder.
• Validation. Benchdnn doesn't work with submemory at all. This is likely a discussion for further RFC once this direction gets priority back again.