[Mlir-commits] [mlir] [mlir][linalg] Introduce transpose semantic to 'linalg.matmul' ops. (PR #104783)

[Nicolas Vasilache]

https://github.com/nicolasvasilache commented:

Upon reading the super long back and forth discussion, I was afraid this would be much more intrusive on opdsl and existing behavior (e.g. prematurely deprecating `linalg.matmul_transpose_a` or other ops).

After reviewing, this looks clean and separable but I do see the potential footgun: users need to assert everywhere that the indexing map is the default one otherwise miscompiles will propagate and will be hard to track.

Is there an internal / API mechanism we can think of to avoid all silent failures propagating to all uses?





https://github.com/llvm/llvm-project/pull/104783