[Mlir-commits] [mlir] [MLIR] Parallel loop fusion extended to interchanged loops. (PR #191245)

[Ivan Butygin]

================

----------------
Hardcode84 wrote:

If I understood the code correctly, current version tries to fuse specifically `scf.parallel %a, %b` + `scf.parallel %b, %a`. More generic version will be `scf.parallel %a, %b, %c` + `scf.parallel  %c, %b, %a`, where the second loop IVs are the arbitrary permutation of the first one. Now, thinking about it, we should probably still limit it to some fixed number of permutation attempts to avoid explosion on the loops with lot of IVs.

https://github.com/llvm/llvm-project/pull/191245