ttir-decomposition-for-ttmetal

To decompose TTIR composite ops for the TTMetal backend, you add and register an OpRewritePattern-based decomposition. This lowers high-level ops like rms_norm, sdpa, layer_norm, and softmax into primitive TTIR ops such as matmul, add, multiply, and reduce.

High-level TTIR op lowering into primitive TTIR ops is needed when fused TTIR support is not available or desirable for the TTMetal backend. This decomposition enables the compiler to break down composite operations into executable primitives.

To register a new TTIR decomposition pattern, you apply it through the TTIRDecomposeComposites pipeline with correct benefit ordering. You must also specify the files to modify and ensure proper TTMetal pipeline scheduling.

Yes, TTMetal op decomposition maintains compatibility with the TTNN backend when the decomposition pattern is not triggered. You must update tests and docs to validate the new decomposition path and ensure this compatibility is preserved.

The pattern rewriter can lower high-level TTIR ops including rms_norm, sdpa, layer_norm, and softmax. These composite ops are decomposed into primitive TTIR ops like matmul, add, multiply, and reduce for the TTMetal backend.

Adding a new TTIR composite op decomposition requires modifying files to define the pattern, register the pipeline, and add tests. This ensures the pattern is properly integrated into the TTMetal backend lowering pipeline.