[Mlir-commits] [mlir] [MLIR][Linalg] Add aggregate ops decomposition pass and softmax decom… (PR #97582)

[Petr Kurapov]

kurapov-peter wrote:

> This seems to be a problem to @MaheshRavishankar and I want to understand it better. My guess is that there are pattern matchers that won't work with the generic version of `fill` (and why we want named ops in the first place).

I would like to understand the next steps here. @MaheshRavishankar, could you please elaborate ^^^?

Regarding the `fill` issue, I think customizable decomposition would be a reasonable solution - helps preserve the downstream usage and doesn't hold back the upstream.

Regarding `broadcast`, I could work on setting the semantics for implicit casting. One thing that is unclear to me though is whether having implicit cast semantics for named ops is beneficial. Wasn't the whole point of named ops to have a very explicit IR that is easy to analyze? In that regard, the absence of implicit casts is actually a good thing (I also don't see how it is ambiguous, could you please clarify?). Is there any real problem with broadcasts except for not being succinct? Wouldn't implicit casting just add unnecessary burden for analyses and transforms to handle various cases of arguments?



https://github.com/llvm/llvm-project/pull/97582