[Mlir-commits] [mlir] [mlir] [linalg] fix failure on specializing matmul with permuted loops (PR #184294)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Thu Mar 5 00:11:32 PST 2026
ziereis wrote:
I did add some more tests for permuted loops + transposed indexing maps. They are not exhaustive but i thought actually testing every possible combination thereof would be a too much? but i could also add more if you think we have to test every combination.
I also added 3 tests with broadcasted indexing maps which currently don't get specialized. I thought about it a bit and now i share you concern that if we would want to add support for specializing (broadcasted+transposed+permuted loops) maps as well, the current implementation as is would probably need a more significant refactor, as this would get quite complex.
I guess the logic for verifying the indexing_maps should exist somewhere in the matmul op verifier, maybe it would be possible to expose this somehow to be used here as well?
https://github.com/llvm/llvm-project/pull/184294
More information about the Mlir-commits
mailing list