[Mlir-commits] [mlir] [mlir][linalg] Support subtracting accumulation in partial reduction … (PR #214033)

[Federico Bruzzone]

FedericoBruzzone wrote:

> You’ve raised some really good questions here, but I’m afraid I don’t know the answers. I don’t have enough historical context on the design of Linalg or the tiling infrastructure, and I also haven’t used Linalg with these overflow flags.

Thank you, and no worries at all @banach-space! I was asking because it was a little bit opaque to me to recover this info from the source code :D

> I’m not aware of tiling being blocked by these flags, but that’s worth checking and documenting, e.g. with tests.

Since subtractions can introduce subtle complexity (and eventually _unsound_ code), we should definitely check the current status of tiling in the presence of these flags (and eventually add tests for them). I believe that consistency in these scenarios is super important. @pstarkcdpr can you please take a look at this? (This is not to say that we shouldn't bail out on subtraction) :D

> From @banach-space: I’d recommend involving more people by raising these questions on Discourse.
> From @pstarkcdpr: Unfortunately I can't start a thread on Discord since Discord is disallowed due to company security policies. If there's another avenue for discussion, I can summarize my findings there.

@pstarkcdpr I believe that @banach-space wasn't referring to Discord, but rather to the [Discourse](https://discourse.llvm.org) :D
It's a happy place where folks propose RFCs and discussions related to LLVM and all subprojects etc., you should definitely check it out 🫶

Given that, I have the bandwidth to open a discussion related to this topic. Essentially, I'd like to clarify the first point in my [previous comment](https://github.com/llvm/llvm-project/pull/214033#issuecomment-5257929708), alongside your example @pstarkcdpr. But I think that we should know whether the current tiling mechanism is currently blocked by these flags. 
In the end, I’ll be able to share a HackMD file with you, and we can refine the initial message together. Let me know.

— Fede




https://github.com/llvm/llvm-project/pull/214033