MLIR News 71st Edition (3rd Nov 2024)

MLIR Newsletter
1

Welcome to the 71st issue of the MLIR Newsletter covering developments in MLIR, and related projects in the ecosystem. We welcome your contributions (contact: javed.absar@gmail.com). Click here to see previous editions.

Highlights, Discussions & RFCs

  • Renato initiated an [RFC] MLIR Project Charter and Restructuring – "In recent years we have discussed many changes to the MLIR infrastructure, dialects, conversions, representations, transforms and their future…We need to agree to a common charter, at both high level (what MLIR is and how it fulfils its purpose with lowest global cost) and low-level (dialect semantics and conversions, what does canonicalization means, type systems, etc). The proposed next steps are: (a) Reach a consensus in this RFC ; (b) Create a survey on dialects and gather feedback from all in the community; (c) Collect the survey’s results and publish a summary (End of Dec 2024); Create PRs with the proposed code move and ownership models after LLVM 20 branches off (End of Jan 2025) "

  • Another RFC by Renato on Op explosion in Linalg – “…move it out of OpDSL (Python, YAML), into ODS (table-gen, C++) so that we can common up implementation details (such as broadcast, transpose, quantize, loop order). …have a discussion on how to stop the op explosion and focus on a few operations that we agree is reasonable for such a core dialect?” [click here].

  • Matthias Springer put up an RFC for Merging 1:1 and 1:N Dialect Conversions – " There are currently two incompatible dialect conversion drivers is MLIR: the “default” driver and the 1:N dialect conversion driver . This RFC proposes: 1. Extending the default driver with full 1:N conversion support. 2. Deleting the 1:N driver". [click here for more information]

  • Maksim Levental has written an RFC: an incubator for language frontends (bindings) – " The TL;DR can be gleaned from the presentation I gave at US LLVM Dev meeting just recently but basically there are a lot of language communities working to target MLIR and I think we should support them with first-class bindings."

  • [Question] How to raise `llvm` to `scf`? Or lower C code to `scf`?

  • Matthias Springer has been cleaning up code base to remove deprecated dialect-based bufferization and move everything to one-shot-buferizer.

MLIR Commits Recently:

  • Matthias added "support for bufferizing external functions that have no
    body. Such functions were previously rejected by One-Shot Bufferize if
    *they returned a tensor value. [click here].

  • Jakub added a fast walk-based pattern rewrite driver: walkAndApplyPatterns. It does not iterate until a fixpoint and does not perform folding or DCE. It should have much lower overhead than the greed pattern rewrite driver for simple pattern sets. [PR]

  • Lialan added VectorEmulateNarrowType to “support loading of unaligned vectors” [click here].

  • Andrzej split GenericPadOpVectorizationPattern into two patterns – "With this change, we gain the following: 1 a clear separation between pre-processing and vectorization transformations/stages; 2. a path to support masked vectorisation for tensor.insert_slice (with a dedicated pattern for vectorization, it is much easier to specify the input vector sizes used in masking); 3. more opportunities to vectorize tensor.insert_slice." [[click here]](Split GenericPadOpVectorizationPattern into two patterns).

  • This [commit] adds extra checks/assertions to the
    ConversionPatternRewriterImpl::notifyOpReplaced to improve its
    robustness. This change is in preparation of merging the 1:1 and 1:N dialect
    conversion drivers.

  • The 1:N type converter derived from the 1:1 type converter and extends it with 1:N target materializations. This [commit] merges the two type converters and stores 1:N target materializations in the 1:1 type converter. This is in preparation of merging the 1:1 and 1:N dialect conversion infrastructures.

  • [Commit] adds a canonicalization pattern for scf.forall that replaces constant induction variables with a constant index. There is a similar
    canonicalization that completely removes constant induction variables from the loop, but that pattern does not apply on foralls with mappings, so this one is necessary for those cases.

  • Shahid introduced transpose semantic to ‘linalg.matmul’ ops. [click here]. The main goal of this patch is to extend the semantic of ‘linalg.matmul’ named op to include per operand transpose semantic while also laying out a way to move ops definition from OpDSL to tablegen. Hence, it is implemented in tablegen.

Related Projects

Useful Links

2 Likes