-
Notifications
You must be signed in to change notification settings - Fork 511
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Query] The refback-munge-calling-conventions pass employed in the backend from linalg to llvm, causes runtime error #3682
Comments
Hi @Abhishek-TyRnT, the support for this was added here #3615 by @mgehre-amd. Please refer to this PR and ask the author if you have any issues. |
Is the question about |
@mgehre-amd , the question was originally about |
Otherwise the
I think the Conv_Transpose1d lowering needs to be fixed, but I have not looked into the details of it. |
If the issue is with |
I was hoping to fix it myself and send as a contribution, |
I was running the Conv1d_transpose op test case, lowering it to the llvm IR from linalg-from-tensors Dialect. I notice that this test is marked as crashing test case in x_fail test cases
torch-mlir/projects/pt1/e2e_testing/xfail_sets.py
Lines 50 to 54 in 3180704
I tried to debug this issue.
Upon investigating I noticed that there is an following Runtime Error from llvm
You can see that the memref is trying to cast an unranked tensor to ranked tensor, which is not allowed in the
generate-runtime- verification
pass , called heretorch-mlir/projects/pt1/python/torch_mlir_e2e_test/linalg_on_tensors_backends/refbackend.py
Lines 184 to 187 in 3180704
The above casting operation is called after
refback-munge-calling-conventions
which gets rid of the ranked memrefs. My question is why is this pass necessary, this pass will clearly fail runtime verification since there is no way one could know at runtime what input we are getting and what is it's rank. Am I missing here something ? Please let me knowThe text was updated successfully, but these errors were encountered: