Create a canonicalization pass in TT-MLIR compiler #1264
Assignees
Labels
TTIR Dialect
Issues related to TTIR dialect
Comments
azecevicTT
added a commit
that referenced
this issue
Jan 15, 2025
Canonicalization can be roughly divided into three forms: 1. Traits (properties) of some ops that allows them to be folded (Involution, Idempotence) 2. Per op foldings 4. Per op canonicalization (pattern rewriters) The first one allows us to decleratively add folding for a large class of ops. While traits like `Involution` already exists in `MLIR` infrastructure it doesn't account for DPS, so I added a new one that takes care of it. The second one allows us in practice to define some simple graph rewritings where we replace the producing value of op with some existing value (or constant). The third one gives us the most freedom, where we can do arbitrary graph rewritings, we usually use it when we have to create a new op during rewriting. I added a `canonicalize` pass both before and after `ttir-to-ttir-decomposition-pass` in `ttir-to-ttnn-backend-pipeline`, as there are many graph rewritings during that pass, so we might benefit form canonicalization both before and after. I plan to cover one big part of canonicalization in the future, and that's constant folding. I will also write a short document on adding a canonicalization for new and existing ops. While this PR covers a lot of patterns, it's definitely not an exhaustive list. This MLIR [doc](https://mlir.llvm.org/docs/Canonicalization/) is already a great source of information, I just believe we might also benefit from the additional context of TTIR dialect. This PR should cover a big part of #1264, as I said the missing part is constant folding.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
To improve the efficiency and reliability of our MLIR compiler, we need to introduce a canonicalizer pass. This pass should apply a range of canonicalization patterns that simplify and normalize the intermediate representation (IR). Canonicalization makes IR easier to optimize and analyze by reducing redundant patterns and ensuring consistency in how operations are represented. Currently, our compiler lacks a centralized approach for defining and managing these patterns, making it difficult to ensure consistent transformations across components.
The text was updated successfully, but these errors were encountered: