more debugging

Author: majosm

.github/workflows/ci.yml

@@ -85,8 +85,20 @@ jobs:
                 curl -L -O https://tiker.net/ci-support-v0
                 . ./ci-support-v0
                 build_py_project_in_conda_env
-                export PYTEST_FLAGS=-sv
-                export CISUPPORT_PYTEST_NRUNNERS=1
+
+                # Diagnose the Intel CPU OpenCL heap corruption: run serially
+                # in-process (no xdist workers, via CISUPPORT_PARALLEL_PYTEST=no)
+                # so a glibc abort happens at the culprit test with a full
+                # faulthandler traceback, instead of detonating later in another
+                # worker's GC. ARRAYCONTEXT_INTEL_DIAG makes the conftest force a
+                # gc pass at each test teardown so the abort lands on the test
+                # that caused it. A fresh cache dir keeps a crash from leaving a
+                # poisoned sqlite cache that cascades into spurious failures.
+                export CISUPPORT_PARALLEL_PYTEST=no
+                export ARRAYCONTEXT_INTEL_DIAG=1
+                export PYTHONFAULTHANDLER=1
+                export XDG_CACHE_HOME="$(mktemp -d)"
+                export PYTEST_FLAGS="-sv -p no:cacheprovider"
                 export NO_COLOR=1
                 test_py_project
 

test/conftest.py

@@ -0,0 +1,21 @@
+from __future__ import annotations
+
+import gc
+import os
+
+import pytest
+
+
+@pytest.hookimpl(hookwrapper=True)
+def pytest_runtest_teardown(item, nextitem):
+    # Diagnostic only (enabled via ARRAYCONTEXT_INTEL_DIAG): force a garbage
+    # collection after each test's fixtures have been torn down. The Intel CPU
+    # OpenCL runtime corrupts the host heap during some kernel executions; the
+    # abort only fires when the smashed block is next freed, which otherwise
+    # happens at an arbitrary later GC (often in an unrelated test's startup).
+    # Collecting here frees this test's OpenCL objects immediately, so the abort
+    # detonates at the teardown of the test that caused it, pinning the culprit.
+    yield
+
+    if os.environ.get("ARRAYCONTEXT_INTEL_DIAG") == "1":
+        gc.collect()