Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td | ||
---|---|---|
366 | @jpienaar, is ReifyRankedShapedTypeOpInterface enough for the type inference we need for Linalg Ops ? | |
367 | Can you also add DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>, here and make the result SSA name be called "transposed" ? This will improve readability of IR. | |
372 | Please add doc that unlike memref.transpose which is a metadata operation only that produces a transposed "view", this op actually moves data. | |
379 | nit: I think I would put the permutation attr before the ins | |
387 | why not call output ? | |
396 | I would make both builders not take a resultType and use the shape reification mechanism. | |
406 | why do we also need this? | |
407 | why not just return getInput() ? | |
408 | why not just getInit ? |
mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td | ||
---|---|---|
379 | I'm not sure about that. I would prefer to have ins and out before attributes for clarity and consistency with ReduceOp, but it's just my mental model. |
mlir/include/mlir/Dialect/Linalg/IR/LinalgStructuredOps.td | ||
---|---|---|
387 | I know that there is a bigger discussion if it should init or output in DPS, but here I used init to be consistent with Reduce and Map. I can rename init into output in other ops in a separate patch if you prefer. |
@jpienaar, is ReifyRankedShapedTypeOpInterface enough for the type inference we need for Linalg Ops ?