Proposal for 6.0: Winnograd

[krzysz00]

Motivation

The Winnograd algorithm is often the fastest way to do a convolution.

When we're performing inference in whole graph mode, we have access to the filter tensor, so we can pre-compute the Winnograd constants at compile time (I think) for extra performance.

Why?

We get faster, which is useful for whole graph workloads.

Why not?

It's an entire separate code path.

It'll make tuning a bit tricky