regression in

[mattip]

I run weekly builds of HEAD building the scipy-openblas wheels. Starting 5 weeks ago between commit 624e9d1 and 3184b7f there is a test failure on aarch64, I think it is in cblas_ssyrk, there are a number of failures in the CI run. I was waiting to see if the failure dissapeared but it seems to persist from that commit forward.

[martin-frbg]

strange, there weren't that many PRs in that timeframe that could have affected ARMV8

[martin-frbg]

Also not reproducible in my local arm64 builds (and CI is quiet too). What environment does your build test run in, is this some kind of emulation ?

[mattip]

The aarch64, like all the linux builds, run inside a manylinux2104 docker image on x86_64. So the build adds the ability to use docker with qemu:

sudo apt-get -qq install -y qemu qemu-user-static
sudo docker run --rm --privileged multiarch/qemu-user-static --reset -p yes --credential yes

[martin-frbg]

Did anything change in your manylinux/docker/qemu framework lately ? This is almost certainly a fluke in your setup - I cannot reproduce it on (various) actual arm64 hardware, there have been no changes directly affecting SYRK drivers or kernels in many months, and I cannot think of any indirect changes that would manifest as SYRK breakage without general carnage in GEMM/TRMM

[martin-frbg]

Closing for now as I cannot reproduce the issue on any Arm64 setup - if anything, this could be related to your docker environment or the specific libc builds used therein - I notice even your CI jobs pass with one particular flavor of musl.

[mattip]

OK, thanks for checking. Maybe someday github aarch64 native runners will be available.