You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Running --torch-mlir-opt --torch-backend-to-linalg-on-tensors-backend-pipeline on some (rather large IR) causes a crash during the pass --resolve-shaped-type-result-dims (from memref).
If adding the pass resolve-shaped-type-result-dims to torch-mlir-opt.cpp, you can reproduce the error with torch-mlir-opt --resolve-shaped-type-result-dims applied to the above IR.
torch-mlir-opt --torch-backend-to-linalg-on-tensors-backend-pipeline applied to the above IR also reproduces the crash.
However, mlir-opt --resolve-shaped-type-result-dims does NOT reproduce the crash, but instead generates the valid IR:
Why are there memref passes used in a pass pipeline, but not available by themselves? This makes debugging an error during this pass more difficult. In particular, because of the position in the pipeline, there are still torch_c objects floating around everywhere in the ir, and the mlir-opt pass will not work properly since torch dialects aren't loaded there.
Is there a reason the affine dialect is not included in torch-mlir? I believe this is part of why the mlir-opt version works, but the torch-mlir-opt version does not.
How should I fix this issue? Is this something to fix in the structure of torch-mlir, or something to upstream to llvm-project despite not being a reproducible issue there?
Debug Dumps
When running ``mlir-opt --resolve-shaped-type-result-dims --debug```, the successful application details are here:
The Issue
Running
--torch-mlir-opt --torch-backend-to-linalg-on-tensors-backend-pipeline
on some (rather large IR) causes a crash during the pass --resolve-shaped-type-result-dims (from memref).Here is some IR that can reproduce the issue:
resolve-shaped-type-result-dims
to torch-mlir-opt.cpp, you can reproduce the error withtorch-mlir-opt --resolve-shaped-type-result-dims
applied to the above IR.torch-mlir-opt --torch-backend-to-linalg-on-tensors-backend-pipeline
applied to the above IR also reproduces the crash.mlir-opt --resolve-shaped-type-result-dims
does NOT reproduce the crash, but instead generates the valid IR:Questions:
Debug Dumps
When running ``mlir-opt --resolve-shaped-type-result-dims --debug```, the successful application details are here:
The original crash backtrace:
The text was updated successfully, but these errors were encountered: