Repository navigation
fix: exclude OpenBLAS 0.3.34 from the development environment - #7
Merged
Merged
Conversation
The modal and stack_modal tests stopped with an access violation inside openblas.dll on some hosted CI runners, and on every run on an AMD Zen 4 workstation. The fault is in OpenBLAS 0.3.34, not in CableDyn: a plain DGEMM with an inner dimension above 128 reproduces it with no CableDyn code loaded, and 0.3.33 runs the same calls cleanly. OpenBLAS 0.3.34 enlarges the GEMM blocking on AMD Zen 4/5 CPUs that expose AVX-512, and its Zen and Haswell DGEMM kernels then overrun their stack buffer. It is reached in builds without the AVX-512 kernels, such as the conda-forge Windows package, and with a forced OPENBLAS_CORETYPE (OpenMathLib/OpenBLAS#6013, #6021, fixed upstream after 0.3.34). The dense DSYGV reference of test_modal reached it; the solver's banded LAPACK calls never pass DGEMM an inner dimension above 128, and the release executables use reference LAPACK, so results are unchanged. The CI failures were intermittent because only some runners expose AVX-512. Exclude 0.3.34 in environment.yml, which CI also uses, so that the next OpenBLAS release is admitted, and record the cause in the changelog and the troubleshooting table of DEVELOPMENT.md. Signed-off-by: Jae Hoon Seo <jaehoon.seo@inha.ac.kr>
Owner
Author
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Codex Review: Didn't find any major issues. 🚀 Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Resolves the 0.1.1 known issue: the access violation of the
modalandstack_modaltests inopenblas.dll.The fault is in OpenBLAS 0.3.34, not in CableDyn. On an AMD Zen 4 workstation (Threadripper PRO 7965WX, Windows, conda-forge
libopenblas 0.3.34 pthreads_h952cde7_2) the tests fail on every run. A standalone C program that calls only OpenBLAS reproduces it:OPENBLAS_NUM_THREADS=1)dgemmn = 200, k = 128dgemmn = 200, k = 160dgemmn = k = 1000dsygvn = 297 (size of the test's taut-string reference)The fault offsets map to
dgemm_incopy_ZEN/dgemm_oncopy_ZEN. The cause is OpenMathLib/OpenBLAS#6013 and #6021, fixed upstream after 0.3.34:OPENBLAS_CORETYPE.On CI the failure was intermittent because only some hosted runners expose AVX-512.
Scope of the fault:
DGBTRF/DGBTRS,DPBTRF/DPBTRS,DSBGVX,DGBSVX) plus a 4 × 4DSYEV. None of them passes DGEMM an inner dimension above 128: blockedDGBTRFuses the LAPACK block size, at most 64.DSYGVreference oftest_modalreached the bug.Changes:
environment.ymlexcludesopenblas/libopenblas0.3.34. It does not pin an exact version, so the next OpenBLAS release is admitted. CI builds its environment from this file. A dry-run solve (with the glibc/macOS virtual packages set for the foreign platforms) resolves to 0.3.33 on win-64, linux-64 and osx-arm64.CHANGELOG.md: aFixedentry under[Unreleased]that gives the cause.DEVELOPMENT.md: a troubleshooting row. Afterconda env update -f environment.yml, configure a fresh build directory. Otherwise the existing build keeps its stagedopenblas.dll, whose copy is newer than the package file.Tests
modalandstack_modaltests are the regression tests: they failed with 0.3.34 and pass with 0.3.33, on the host-selected Zen kernel.ctest --test-dir build --output-on-failurewithOMP_NUM_THREADS=1,OPENBLAS_NUM_THREADS=1and--parallel 16gives 245/245, slow tier included.Checklist
pre-commit run --all-filespassesctest --test-dir build --output-on-failurepassespython -m pytest tests python/testspasses (if Python or validation records changed) — not applicable, no Python or validation records changed[Unreleased](heading added if missing) for user-visible changesdoc/updated for changed inputs, options, or outputs — not applicable, no inputs, options or outputs changedfeat:,fix:,docs:, ...)git commit -s) under theDeveloper Certificate of Origin