-
Notifications
You must be signed in to change notification settings - Fork 49
Add LIT test for gfx1201 backend bug #2018
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
Conversation
// (see the PR here https://github.com/ROCm/rocMLIR/pull/1990) | ||
|
||
// XFAIL: * | ||
// RUN: rocmlir-driver -c %s | mlir-runner -O2 --shared-libs=%linalg_test_lib_dir/libmlir_rocm_runtime%shlibext,%conv_validation_wrapper_library_dir/libconv-validation-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext,%linalg_test_lib_dir/libmlir_float16_utils%shlibext --entry-point-result=void | FileCheck %s |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nit: you don't need to do the full pipeline with -c, I guess this needs only "--host-pipeline=runner -kernel-pipeline=binary "?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I grabbed the IR fairly early into the gpu pipeline in order to try and minimize the size of the LIT test but still have the bad if
statement that is introduced without the gfx12 workaround in in GridwiseGemmToBlockwise. Because of this, there are still passes that it needs in the GPU pipeline.
mlir/test/fusion/pr-e2e/failure/attention-pipelining-failure.mlir
Outdated
Show resolved
Hide resolved
mlir/test/fusion/pr-e2e/failure/attention-pipelining-failure.mlir
Outdated
Show resolved
Hide resolved
This is passing the nightly CI here: https://ml-ci-internal.amd.com/job/MLIR/job/mlir/job/PR-2018/6/pipeline-overview/. The reason why this job isn't finishing is because there are no gfx942 nodes on the CI. Triggering an empty CI job to get this in since we have already proved that the non-gfx12 path works from this passing on other archs. |
Motivation
During the investigation for https://github.com/ROCm/rocMLIR-internal/issues/2000 it was discovered that there is a bug in the scheduler in the backend. This PR creates a LIT test to check for this failure, and to serve as a reminder that once this starts to pass we can remove the gfx1201 workarounds that were introduced in #1990.
Technical Details
This test was generated by removing the gfx1200 workaround in GridwiseGemmToBlockwise (introduced here: #1990) and taking the IR directly after this pass runs. When the backend bug is fixed, this test will start to fail with
Unexpectedly Passed
.Test Plan
Test Result
Submission Checklist