[Codegen][GPU] Add pass to unroll to native mma widths

[qedawkins]

Left as TODO is to extend this pass to also unroll consumers/producers. To eventually connect everything we should add a pattern that unrolls iter args of scf.for loops.

Note that the unrolling pattern is already tested in its own test, so simply updating the pipeline test is fine here. Once extending this pass to do more unrolling a test can be added.

[hanhanW]

I only see the pipeline test. Where is the test for the pass? Can you add one?

[qedawkins]

I only see the pipeline test. Where is the test for the pass? Can you add one?

I deduplicated the logic so now it is testing the same code that is already tested here: https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/Dialect/GPU/TransformExtensions/test/unroll_multi_mma.mlir

And I didn't feel it was worth adding a test for pass boilerplate for now. If you want I can make a dummy test for the pass version based on this test.

[hanhanW]

I only see the pipeline test. Where is the test for the pass? Can you add one?

I deduplicated the logic so now it is testing the same code that is already tested here: https://github.com/iree-org/iree/blob/main/compiler/src/iree/compiler/Codegen/Dialect/GPU/TransformExtensions/test/unroll_multi_mma.mlir

And I didn't feel it was worth adding a test for pass boilerplate for now. If you want I can make a dummy test for the pass version based on this test.

Ah I see, somehow I missed the changes IREEGPUExtensions.cpp. And it explains that why I'm feeling that I've seen these code somewhere. My conclusion was that is on other branch (not main).

Thanks thanks, LGTM