[Tuner][Codegen] Allow for tuning specs to override codegen pipeline without setting full lowering config

[kuhar]

This is when we want to force a specific codegen pipeline but still want to rely on compiler heuristics choosing the exact lowering config.

[qedawkins]

What kind of case do you have in mind? I would expect any case where we want to force a pipeline is just an artifact of that pipeline not being the default choice.

[kuhar]

Basically any time we don't like the default choice, e.g., sending an MMT down TileAndFuse or WarpReduction instead of VectorDistribution.

[qedawkins]

I see that as a transient issue with having too many pipelines. Over time we expect to only have 2 or 3 (a main default one and a few others for very specific cases). Think some amalgamation of TileAndFuse and VectorDistribution that handles WarpReduction also.

We could add optimization hints that indicate the kind of algorithm we want to use. For matmuls that could be an enum for WarpReduce, MFMA, and SIMT (as an example). Today that would just select between the different pipelines, and in the future a unified pipeline would just pick a config based on the hint. WDYT?

[kuhar]

I think that's reasonable, let's chat about this next week.