[mlir][linalg] Implement patterns for reducing rank of named linalg contraction ops #95710
Conversation
srcarroll
changed the title
srcarroll
changed the title
mlir/test/lib/Dialect/Linalg/TestLinalgRankReduceContractionOps.cpp
Outdated
Show resolved
Hide resolved
srcarroll
requested review from
MaheshRavishankar and
nicolasvasilache
as code owners
…ontraction ops (llvm#95710) This patch introduces pattern rewrites for reducing the rank of named linalg contraction ops with unit spatial dim(s) to other named contraction ops. For example `linalg.batch_matmul` with batch size 1 -> `linalg.matmul` and `linalg.matmul` with unit LHS spatial dim -> `linalg.vecmat`, etc. These patterns don't support reducing the rank along reduction dimension as those don't convert to other named contraction ops.
|
This is a neat feature! Any objection into making it a standalone pass? |
Thanks. No objection at all. If I remember correctly the pattern rewrites are public so they should be able to be reused by any pass. In fact someone requested a feature to have options to choose what ops are targeted for reduction (or choose how low to take the rank). I was going to but haven't had time so if you want to address that too that would be great. |
correction. only the populate patterns function is public. so we can either make the pattern rewrite public to make it easier to build passes with individual patterns, or just add the options to the populate function. i dont think i have a particular preference |
This patch introduces pattern rewrites for reducing the rank of named linalg contraction ops with unit spatial dim(s) to other named contraction ops. For example
linalg.batch_matmulwith batch size 1 ->linalg.matmulandlinalg.matmulwith unit LHS spatial dim ->linalg.vecmat, etc. These patterns don't support reducing the rank along reduction dimension as those don't convert to other named contraction ops.