Bringup reduction silicon tests that were previously unsupported

[jnie-TT]

We're using reduction and eltwise maximum ops in tt-xla FE but didn't have silicon tests for them, so we should bring them up.

[AleksKnezevic]

@mrakitaTT, fyi

[mrakitaTT]

@AleksKnezevic These tests will pass because keep_dim=True. The issue we have in #805 is that StableHLO reduce always kills the reducing dim, so I've used keep_dim=False in conversion, not knowing that Metal doesn't support that configuration yet. I am now trying some solution with ReduceOp with keep_dim=True and then ReshapeOp afterwards to manually kill the reducing dim, because otherwise the rest of the graph would be impacted.

@jnie-TT Thanks for adding more test cases! I've added test/ttmlir/Dialect/TTNN/simple_sum.mlir, test/ttmlir/Dialect/TTNN/simple_maximum.mlir and test/ttmlir/Dialect/TTNN/simple_max.mlir to cover these. It is weird that test/ttmlir/Dialect/TTNN/simple_max.mlir is not failing in CI since I've used keep_dim=False in that test O.o You can remove that one too, I'll add specific test for StableHLO conversion when I fix #805, but it would be good to know why it wasn't run in CI.

P.S. I've finally went through setting up a silicon machine because I'll need it for XLA, so going forward I'll manually test everything and not rely on CI.

[jnie-TT]

@AleksKnezevic These tests will pass because keep_dim=True. The issue we have in #805 is that StableHLO reduce always kills the reducing dim, so I've used keep_dim=False in conversion, not knowing that Metal doesn't support that configuration yet. I am now trying some solution with ReduceOp with keep_dim=True and then ReshapeOp afterwards to manually kill the reducing dim, because otherwise the rest of the graph would be impacted.

@jnie-TT Thanks for adding more test cases! I've added test/ttmlir/Dialect/TTNN/simple_sum.mlir, test/ttmlir/Dialect/TTNN/simple_maximum.mlir and test/ttmlir/Dialect/TTNN/simple_max.mlir to cover these. It is weird that test/ttmlir/Dialect/TTNN/simple_max.mlir is not failing in CI since I've used keep_dim=False in that test O.o You can remove that one too, I'll add specific test for StableHLO conversion when I fix #805, but it would be good to know why it wasn't run in CI.

P.S. I've finally went through setting up a silicon machine because I'll need it for XLA, so going forward I'll manually test everything and not rely on CI.

@mrakitaTT please note that if you want a test to run on silicon, you need to add it under the Silicon folder. Adding it under the dialect folder will only test the compiler's ability to compile a specific mlir module, but will not generate a flatbuffer for execution on Silicon.

[mrakitaTT]

@mrakitaTT please note that if you want a test to run on silicon, you need to add it under the Silicon folder. Adding it under the dialect folder will only test the compiler's ability to compile a specific mlir module, but will not generate a flatbuffer for execution on Silicon.

@jnie-TT Ohh, got it, thanks Jackson!

[jnie-TT]

I'll merge this in now since we want any future CI to be running these.