BUG: dnrm2 basic failures on a subset of platforms (observed in SciPy)

[tylerjereddy]

We recently experimented with bumping SciPy OpenBLAS version used in wheels from a pretty old version (v0.3.20-571-g3dec11c6) to 0.3.26 stable release in scipy/scipy#20215, in the hopes of fixing the reproducer in scipy/scipy#16930, which can be summarized as the following test, which we should perhaps adjust to be a tractable C-level test in this project:

def test_gh_169309():
    x = np.repeat(10, 9)
    actual = scipy.linalg.blas.dnrm2(x, 5, 3, -1)
    expected = math.sqrt(500)
    assert_allclose(actual, expected)

Bumping to 0.3.26 appeared to fix the issue on x86_64 Linux, but not on some other platforms, including MacOS ARM and Linux aarch64. In both cases, looks like actual ends up as 0..

I could probably help draft in the regression test if you tell me where to put it, but I'm unlikely to have the expertise to provide the fix.

[martin-frbg]

The only recent change to the (x86_64) NRM2 kernels happens to have been to add support for negative x increments, which reflected a relatively recent change in Reference-BLAS' definition of the NRM2 API. Looks like an "old-style" early exit has been overlooked in one or more of the aarch64 kernels. Can you tell me which other platforms besides Apple M are affected ?

[tylerjereddy]

I think I only have evidence for failure on two architectures:

1. MacOS M series (ARM)
2. Linux ARM/aarch64

[martin-frbg]

Is that Linux aarch64 across the range of processors ? And 32-bit ARM as well, or just aarch64 ? Offhand i see no similar early exit in the assembly kernel for the Cortex cpu models

[tylerjereddy]

Is that Linux aarch64 across the range of processors ?

I think we only test with one native 64-bit ARM hardware for Linux aarch64, and it would be whatever the Cirrus CI service uses by default for ARM Linux. Looking at the log, we do call some OpenBLAS info stuff in case that helps:

python /project/tools/openblas_support.py --check_version
OpenBLAS get_config returned b'OpenBLAS 0.3.26 DYNAMIC_ARCH NO_AFFINITY neoversen1 MAX_THREADS=64'
b'OpenBLAS 0.3.26'

python -c 'import openblas_support; print(openblas_support.get_plat())'
PLATFORM=linux-aarch64

And 32-bit ARM as well, or just aarch64 ?

Only the 64-bit ARM hardware is tested/distributed binaries by SciPy.

Cortex cpu models

Not sure I even have access to that, don't see it at https://portal.cfarm.net/machines/list/ for example.

[martin-frbg]

ok, Neoverse N1, Apple M and ThunderX2 share the same nrm2 kernels for which I've already posted the PR. Cortex (typical phone cpus) should be unaffected I think

[ev-br]

#4554 proposes a small addition to the test suite with a regression test.