Proposal for 6.0: NCHWC

[krzysz00]

What this is

Allow the C dimension of the input tensor, and, analogously, the K dimension of the output tensor, to be split into two pieces which I'll call C_o and C_i. The point of this layout is to improve vectorization, and to deal with things like N[# number of channels we're MLing on]HW[RGBA] data, for istannce.

I think this is something MIOpen might want.

The hardest part of the plan is determining how we'll communicate the layout

Why?

Once we've agreed on the API, this is just one more coordinate transformation applied at the beginning of conv-to-gemm

Why not?

Increased tuning surface, and we don't know what the demand for this feature is