Repository navigation
CI: use the Prescott OpenBLAS kernel for ASE tests - #8348
Conversation
|
That adds quite a few complexity. Would it be possible to simply choose the generic kernel for GitHub CI? Btw I'm surprised that such fault can be related to system OpenBLAS... |
This comment was marked as outdated.
This comment was marked as outdated.
|
@Growl1234 It's already been simplified to 4 lines |
|
Thanks, looks good! Maybe, we should consider updating the images base to Ubuntu 24.04, which gets rid of the OpenBLAS-0.3.20 problem described in PR summary... (@mohanchen) |
ABACUS-PR-BOT
left a comment
There was a problem hiding this comment.
Static review of a6296f6a found no actionable issues in the job-wide Prescott override; the original ASE tests and numerical assertions remain unchanged.
No builds or tests were run, and the reported runtime and CI results were not independently verified.
kirk0830
left a comment
There was a problem hiding this comment.
Hi,
Sounds nice. If this is true, it is necessary to inform the ABACUS developers and users of updating the openblas or use a proper release.
Do you have some minimal examples to reproduce the issue you found? That would be more convincing.
Thanks,
Yike
|
I guess, one can easily reproduce this with an Ubuntu 22.04 docker image setting |
The reproducer of OpenBLAS issue seems to be covered by early commit 741ca21... |
|
I cannot reproduce the runtime error with |
|
I confirmed the OpenBLAS bug through rerunning the minimal The results are below
|
|
The ASE testing failure can also be reproduced locally by using SDE to emulate Cooperlake instructor: |
|
#8357 is an alternative; it updates the docker image to Ubuntu 24.04 so we can fully get rid of this OpenBLAS bug in CI... |
The GNU CI image's OpenBLAS 0.3.20 Cooperlake kernel returns incorrect DSYMM results, causing the Gamma-only Si calculator and restart tests to abort at SCF iteration 4 with a charge integral of
-17.6846.The affected library is Ubuntu 22.04's system package
libopenblas0-openmp 0.3.20+ds-1, installed byDockerfile.gnu. In this version'sDYNAMIC_ARCHbuilds, Cooperlake inherits SkylakeX's 4-by-8 multiplication kernel but retains 16-by-2 packing parameters. OpenBLAS PR #3631, included in 0.3.21, restores consistent kernels and parameters. The controlled builds below confirm that this upstream fix resolves the observed failure.What's changed?
Set
OPENBLAS_CORETYPE=Prescottfor the entire x86-64 ASE CI job, using a conservative kernel supported by the current image. The change is four workflow lines; existing tests, numerical assertions, and ABACUS source remain unchanged. Remove the override when a corrected library in the GNU image has been verified.Linked Issue
Related: #7794.
Unit Tests and/or Case Tests for my changes
Fresh validation used the original failing GNU image (
sha256:b195875e7f3262d26eb4e07960afc0e7a193861699cd281eb08f65562880b17d) and the independently built upstreameba09c937c73executable on AMD Ryzen 9 9950X3D.With
OPENBLAS_CORETYPE=Prescottand one OpenMP/BLAS thread:TestAbacusCalculator.test_calculator_resultsandtest_restartpassed with every original assertion.Workflow YAML parsing, staged governance, and whitespace checks passed:
The original failure was isolated by replacing only DSYMM while retaining Cooperlake for other BLAS operations. A hosted Intel run also directly recorded automatic Cooperlake selection and the incorrect DSYMM result. The hosted ASE job for the fixed Prescott revision passed: 56 inline tests, 2 existing skips, and 7 integration tests.
To verify the upstream fix, OpenBLAS 0.3.20 was also rebuilt in the same image with and without only the two commits from PR #3631, including LAPACK. Both runs retained Cooperlake and used the same ABACUS executable, two MPI ranks,
genelpa, and one OpenMP/BLAS thread:-17.6846.2.22e-16), and both Si tests pass every original assertion.Governance Notes