Verification successful for Resolute: openblas 0.3.32+ds-5ubuntu0.1, arm64 (2026-11-10).
Environment: a clean Ubuntu 26.04.1 (Resolute) ARM64 chroot made from the official Ubuntu Base image, with its Ubuntu signature and SHA-256 verified, running natively on Apple silicon in OrbStack (GCC/GFortran 15.2.0, glibc 2.43). The chroot was unprivileged: a user namespace, with dpkg run under fakeroot. Installed the original 0.3.32+ds-5 runtime/development package pairs, then upgraded all three LP64 flavors through APT to exactly 0.3.32+ds-5ubuntu0.1 from resolute-proposed. apt policy and dpkg-query confirm all six package versions. Switched both OpenBLAS alternatives explicitly for each flavor and checked the resolved library paths. Results: - Original pthread and OpenMP packages reproduce the bug: the seeded-v0 length-64 all-ones dot product returns 4000064.0, exit 1. The original serial package did not reproduce this symptom on this machine and returned 64.0, as on Stonking. - Proposed pthread, OpenMP and serial: the unmodified debian/tests/sdot-v0 script from this source package passes for each flavor, using default dispatch (neoversen1), THUNDERX2T99, NEOVERSEN1 and THUNDERX3T110. All 12 reproducer invocations pass, with no CPU-feature skips. The seeded result and all three repeated-call results are 64.0. - Proposed pthread, OpenMP and serial: the unmodified debian/tests/upstream-testsuite script passes for each flavor against the installed libraries. I also checked all 24 level 2/3 .SUMM files and the level 1 output for failed, suspect or fatal results; none were found. - Additional native sdot/ddot tests pass all 15,552 assertions across the three flavors, four dispatch settings, requested thread counts 1 and 4 (serial remains at one effective thread), 18 lengths from 0 to 32769, four positive stride pairs, and three exact dyadic input patterns. Selected core, effective thread count and library paths were checked. Unlike the Stonking run, the test program no longer calls dlclose() at exit: with OpenMP at 4 threads, unloading the library segfaulted after the checks had completed, with both the original and the proposed package. The test source, its SHA-256 and the one-line dlclose() removal are in the attachment, with the segfault log. The reproducer, upstream suite and numerical matrix above cover LP64 only; SVE-only cores were not tested. Autopkgtest regressions (numpy, petsc, slepc): local comparisons with the original library and historical runs with other triggers are described below. Excerpts of the hosted logs (run URLs, versions, triggers, failure lines) are in the attachment. - petsc 3.24.4+dfsg1-1ubuntu1 and slepc 3.24.2+dfsg1-1 (amd64, arm64, armhf, ppc64el): test-petsc-common and test-slepc pass. The complex and 64-bit tests stop with "Permission denied" and exit 126 because their wrappers execute debian/tests/test-petsc-common and debian/tests/test-slepc, which are not executable in these source packages, so no PETSc/SLEPc code runs. The same package versions also failed in Stonking from May to September 2026 with many triggers unrelated to openblas. The arm64 PETSc openmpi/scotch and SLEPc openmpi/petsc logs I inspected show the same exit-126 helper failure. I reproduced the six failures locally on arm64 with both the original and the proposed openblas. Calling the helpers with bash makes all six wrappers pass locally on arm64 with either openblas version (patches included in the attachment); the recorded BLAS alternatives point LP64 to reference BLAS and ILP64 to the OpenBLAS pthread library, so only the 64-bit variants can exercise OpenBLAS here. - numpy 1:2.3.5+ds-3ubuntu1 (amd64, arm64, s390x, and ppc64el before its latest run): the latest amd64, arm64 and s390x logs and the 2026-10-28 ppc64el log show the test run killed after 33 completed tests in _core/tests/test_shape_base.py. In Stonking, the same numpy version passed in May and June 2026 and then failed from July to August 2026 with lapack, blis, sphinx and glibc triggers; the amd64 and arm64 glibc and lapack runs I checked were killed at the same point. In a local collection, the next test is TestConcatenate::test_huge_list_error, which builds a tuple of 2^31 references (16 GiB) while its memory guard asks for about 4 GiB (source excerpt attached); these LXD jobs have an 8 GiB memory limit. Running only that test on a separate Resolute arm64 machine (not the clean chroot) in an 8 GiB memory cgroup is OOM-killed with both the installed original pthread library and the extracted proposed pthread library selected through LD_LIBRARY_PATH. The hosted logs have no kernel messages, so I cannot confirm the hosted kill was an OOM kill. - numpy ppc64el: the latest job (2026-11-03) failed before the testbed started (the LXD host could not be reached); no test ran. The same package versions passed in Resolute in April 2026, including with openblas 0.3.32+ds-5. Since then, no Resolute run of these tests with any other trigger has finished. This verification applies to Resolute only. I have not changed the verification tag. Testing and this report were prepared with Claude Opus 5.5 (Anthropic). The attached log has the APT policy, reproducer, upstream test and numerical test logs, the numerical test source, hosted autopkgtest excerpts and history, and the PETSc/SLEPc and NumPy logs. ** Attachment added: "Resolute ARM64 verification log: APT policy, reproducer, upstream suite, numerical tests and autopkgtest comparisons" https://bugs.launchpad.net/ubuntu/+source/openblas/+bug/2169719/+attachment/6017165/+files/resolute-arm64-verification.log -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2169719 Title: sdot on arm64 adds the caller's v0 register to the result (Neoverse N1, AmpereOne, Apple silicon, ThunderX2/3) To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/openblas/+bug/2169719/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
