[Mlir-commits] [mlir] [mlir][linalg] Introduce transpose semantic to 'linalg.matmul' ops. (PR #104783)
Stella Laurenzo
llvmlistbot at llvm.org
Tue Oct 8 07:50:41 PDT 2024
================
@@ -335,7 +412,7 @@ static void printNamedStructuredOp(OpAsmPrinter &p, Operation *op,
/*elidedAttrs=*/{"operandSegmentSizes",
// See generated code in
// LinalgNamedStructuredOps.yamlgen.cpp.inc
- "linalg.memoized_indexing_maps"});
+ "linalg.memoized_indexing_maps", "indexing_maps"});
----------------
stellaraccident wrote:
I'm trying to decide whether this is a good idea or not to have as one list for every op vs pushing to the caller.
If we fast forward and introduce an op that wants its indexing maps atte printed as a default attrdict attr, then we're going to have to unwind what this was in support of.
Maybe just move elidedAttrs to an argument and pass it explicitly from each call site. Then it will be easy to find exactly what is being done. I would move the operandSegmentSizes to the call site too in order to avoid splicing the list here.
https://github.com/llvm/llvm-project/pull/104783
More information about the Mlir-commits
mailing list