dialects (tosa): initialize dialect and add Clamp and Rescale operation

[jorendumoulin]

No description provided.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.01%. Comparing base (80e2129) to head (10aa2b7).
Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3158   +/-   ##
=======================================
  Coverage   90.00%   90.01%           
=======================================
  Files         427      428    +1     
  Lines       53860    53890   +30     
  Branches     8344     8346    +2     
=======================================
+ Hits        48477    48507   +30     
  Misses       4036     4036           
  Partials     1347     1347           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[superlopuh]

nit: I assume there will be more ops

Suggested change

	TOSA = Dialect("tosa", [ClampOp, RescaleOp], [])
	TOSA = Dialect("tosa", [
	ClampOp,
	RescaleOp,
	],
	[]
	)

[superlopuh]

What's the plan for TOSA? Are you using it in a project?

[jorendumoulin]

We're currently running some MLPerf Tiny benchmarks on SNAX, which are provided in tflite flatbuffers, and tensorflow has a conversion to TOSA. Most of mlir's builtin tosa-to-linalg lowerings are perfectly fine, but the layer-to-layer quantization rescaling comes out looking really ugly. As we have dedicated hardware for this rescaling, I'd like to lower this straight from these tosa ops.