TST use aligned data with unstable openblas configuration (scikit-learn#21702)

jjerphan · adrinjalali · ogrisel · samronsin · commit 9defec717a29 · 2021-11-30T17:28:12.000+01:00
* TST XFAIL test when unstable openblas configuration * Adapt versions parsing Co-authored-By: Adrin Jalali <adrin.jalali@gmail.com> * DOC Prefer ifskip and reference scikit-learn#21361 Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org> * Fix Julien's scheduler incorrect dispatch * Simplify Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org> * Do not remove blank line * fixup! Do not remove blank line * Retrigger CI * Prudently assume Prescott might be the architecture if it is unknown Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org> Co-authored-by: Adrin Jalali <adrin.jalali@gmail.com> Co-authored-by: Olivier Grisel <olivier.grisel@ensta.org>
diff --git a/sklearn/utils/estimator_checks.py b/sklearn/utils/estimator_checks.py
@@ -53,6 +53,7 @@
 from ..model_selection import ShuffleSplit
 from ..model_selection._validation import _safe_split
 from ..metrics.pairwise import rbf_kernel, linear_kernel, pairwise_distances
+from ..utils.fixes import threadpool_info
 from ..utils.validation import check_is_fitted
 
 from . import shuffle
@@ -2080,7 +2081,19 @@ def check_classifiers_train(
         X_b -= X_b.min()
 
     if readonly_memmap:
-        X_m, y_m, X_b, y_b = create_memmap_backed_data([X_m, y_m, X_b, y_b])
+        # OpenBLAS is known to segfault with unaligned data on the Prescott architecture
+        # See: https://github.com/scipy/scipy/issues/14886
+        has_prescott_openblas = any(
+            True
+            for info in threadpool_info()
+            if info["internal_api"] == "openblas"
+            # Prudently assume Prescott might be the architecture if it is unknown.
+            and info.get("architecture", "prescott").lower() == "prescott"
+        )
+
+        X_m, y_m, X_b, y_b = create_memmap_backed_data(
+            data=[X_m, y_m, X_b, y_b], aligned=has_prescott_openblas
+        )
 
     problems = [(X_b, y_b)]
     tags = _safe_tags(classifier_orig)
