BUG: openblas 0.3.9 shipped with the scipy 1.7.1 wheel can cause segfaults when it fails to detect the correct CPU architecture (Prescott instead o Haswell)

Describe your issue.

Recent Intel CPUs detected as SkylakeX by recent versions of OpenBLAS are wrongfully detected as Prescott by OpenBLAS 0.3.9 shipped with scipy 1.7.1.

This in turns can cause segfaults when calling the ddot_k_PRESCOTT routine on a readonly memory buffer allocated by numpy.memmap with mode="r".

EDIT: the more probable cause is that the data is not memory aligned in this test case.

See: https://github.com/scikit-learn/scikit-learn/issues/21361 for details an reproducing code using the scikit-learn test suite.

Error message

(gdb) bt
#0  0x00007fffdbe31c31 in ddot_k_PRESCOTT () from /home/aperez/dev/sandbox/test/scikit-learn/venv/lib/python3.8/site-packages/scipy/spatial/../../scipy.libs/libopenblasp-r0-085ca80a.3.9.so
#1  0x00007fffc893b39a in ?? () from /home/aperez/dev/sandbox/test/scikit-learn/venv/lib/python3.8/site-packages/scipy/linalg/cython_blas.cpython-38-x86_64-linux-gnu.so
#2  0x00007fffc89226e0 in ?? () from /home/aperez/dev/sandbox/test/scikit-learn/venv/lib/python3.8/site-packages/scipy/linalg/cython_blas.cpython-38-x86_64-linux-gnu.so
#3  0x00007fffbe547aca in __pyx_fuse_1__pyx_f_7sklearn_5utils_12_cython_blas__dot (__pyx_v_n=<optimized out>, __pyx_v_x=<optimized out>, __pyx_v_incx=<optimized out>, __pyx_v_y=<optimized out>, 
    __pyx_v_incy=<optimized out>) at sklearn/utils/_cython_blas.c:2861
#4  0x00007fffbe258327 in svm::Kernel::Kernel (this=0x7fffffff6030, l=<optimized out>, x_=<optimized out>, param=..., blas_functions=0x7fffffff64e0) at sklearn/svm/src/libsvm/svm.cpp:394
#5  0x00007fffbe25b394 in svm::SVC_Q::SVC_Q (blas_functions=0x7fffffff64e0, y_=0x555556ff0fd0 '\001' <repeats 100 times>, '\377' <repeats 100 times>..., param=..., prob=..., this=0x7fffffff6030)
    at sklearn/svm/src/libsvm/svm.cpp:1682
#6  svm::solve_nu_svc (blas_functions=0x7fffffff64e0, si=0x7fffffff6000, alpha=0x555557450160, param=0x7fffffff69a0, prob=0x7fffffff6360) at sklearn/svm/src/libsvm/svm.cpp:1682
#7  svm::svm_train_one (prob=0x7fffffff6360, param=0x7fffffff69a0, Cp=<optimized out>, Cn=<optimized out>, status=0x7fffffff64d4, blas_functions=0x7fffffff64e0)
    at sklearn/svm/src/libsvm/svm.cpp:1856
#8  0x00007fffbe264cde in svm_train (prob=0x7fffffff6340, prob@entry=0x7fffffff6500, param=param@entry=0x7fffffff69a0, status=status@entry=0x7fffffff64d4, 
    blas_functions=blas_functions@entry=0x7fffffff64e0) at sklearn/svm/src/libsvm/svm.cpp:2504
#9  0x00007fffbe23e921 in __pyx_pf_7sklearn_3svm_7_libsvm_fit (__pyx_v_X=__pyx_v_X@entry=0x7fffbb041db0, __pyx_v_Y=__pyx_v_Y@entry=0x7fffbb041c90, __pyx_v_svm_type=__pyx_v_svm_type@entry=1, 
    __pyx_v_kernel=__pyx_v_kernel@entry=0x7fffc3052730, __pyx_v_degree=__pyx_v_degree@entry=3, __pyx_v_gamma=__pyx_v_gamma@entry=0.53188777537536391, __pyx_v_coef0=__pyx_v_coef0@entry=0, 
    __pyx_v_tol=__pyx_v_tol@entry=0.001, __pyx_v_C=__pyx_v_C@entry=0, __pyx_v_nu=0.5, __pyx_v_epsilon=0, __pyx_v_class_weight=__pyx_v_class_weight@entry=0x7fffbb041f30, 
    __pyx_v_sample_weight=0x7fffbb041e10, __pyx_v_shrinking=1, __pyx_v_probability=0, __pyx_v_cache_size=200, __pyx_v_max_iter=-1, __pyx_v_random_seed=209652396, __pyx_self=<optimized out>)
[...]

SciPy/NumPy/Python version information

1.7.1 1.21.2 3.8.2

[ogrisel]

The architecture detection bug was already fixed upstream.

The fact that the prescott kernel can segfault is probably related to the fact that the scikit-learn test is using non-memory aligned buffer. The fact that it’s readonly is probably unrelated.

Given that prescott is a really old architecture and that very few people are likely to call OpenBLAS routines on non-memory aligned data, I think it’s not worth wasting time trying to this issue. But still let me report it, just in case OpenBLAS dev are interested in it. Anyways I think we can close the scipy issue.

[ogrisel]

Note that https://github.com/scipy/scipy/blob/master/tools/openblas_support.py has already been updated to use OpenBLAS 0.3.17 and the nightly wheels already ship it:

https://anaconda.org/scipy-wheels-nightly/scipy/files