Welcome to the 54th 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
-
Tanya, “I am pleased to announce the 2023 LLVM Developers’ Meeting Program for October 11-12th in Santa Clara, CA! We have some amazing keynotes, talks, community track talks, tutorials, quick, lightning, and student talks. This may be our best program yet! Be sure to register for the early bird discount by Sept 10th”.
-
Properties adopted as storage for attribute by default.[click here].
MLIR Commits
-
Sometimes we need to add additional code after the result patterns, e.g. coyping the attributes of the source op to the result ops. These can be specified via
SupplementalPatterns
parameter. Similar to auxiliary patterns, they are not for replacing results in the source pattern. [click here]. -
This patch makes the
transform.structured.pad
op return also a handle to the copy op that it inserts. Patch was originally authored by Matthias Springer - [click here]. -
This adds support for lowering vector.broadcast ops to SME, if the source is either a scalar, 0-d vector, or 1-d vector, and the result a 2-d scalable vector that aligns with SME tiles [click here].
-
Vector canonicalizer improvement for create mask removing trialing unit dims [click here].
-
New ops in IRDL - irdl.region and irdl.regions [click here]. c.f.r. An ODS-like dialect to define dialects at runtime.
MLIR RFC Discussions
-
JIT focused (x86) Code Generation for MLIR without the
llvm
dialect -
“MLIR is unrelated to “machine learning”, so I would look into projects like IREE if you need an end-to-end flow instead of a Swiss Army knife”
Useful Links