Remove python-gil requirement and enable free-threaded Python

.github/workflows/run-tests-from-dppy-bits.yaml

@@ -28,6 +28,10 @@ jobs:
     strategy:
       matrix:
         python: ['3.10', '3.11', '3.12', '3.13', '3.14']
+        python_spec: ['']
+        include:
+          - python: '3.14'
+            python_spec: '3.14.* *_cp314'
         experimental: [false]
         runner: [ubuntu-22.04, ubuntu-24.04]
     continue-on-error: ${{ matrix.experimental }}
@@ -47,7 +51,12 @@ jobs:
 
       - name: Install dpctl
         run: |
-          conda create -n "${{ env.TEST_ENV_NAME }}" -c dppy/label/dev "${{ env.CHANNELS }}" dpctl pytest pytest-cov cython setuptools c-compiler cxx-compiler
+          PYTHON_SPEC="${{ matrix.python_spec }}"
+          if [ -z "${PYTHON_SPEC}" ]; then
+            PYTHON_SPEC="${{ matrix.python }}"
+          fi
+          export PYTHON_SPEC
+          conda create -n "${{ env.TEST_ENV_NAME }}" -c dppy/label/dev "${{ env.CHANNELS }}" dpctl pytest cython setuptools c-compiler cxx-compiler python="${PYTHON_SPEC}"
 
       - name: Smoke test
         run: |
@@ -79,6 +88,10 @@ jobs:
     strategy:
       matrix:
         python: ['3.10', '3.11', '3.12', '3.13', '3.14']
+        python_spec: ['']
+        include:
+          - python: '3.14'
+            python_spec: '3.14.* *_cp314'
         experimental: [false]
         runner: [windows-latest]
 
@@ -106,7 +119,9 @@ jobs:
 
       - name: Install dpctl
         run: |
-          conda install -n ${{ env.TEST_ENV_NAME }} -c dppy/label/dev ${{ env.CHANNELS }} dpctl pytest pytest-cov cython setuptools c-compiler cxx-compiler
+          SET "PYTHON_SPEC=${{ matrix.python_spec }}"
+          IF "%PYTHON_SPEC%"=="" SET "PYTHON_SPEC=${{ matrix.python }}"
+          conda install -n ${{ env.TEST_ENV_NAME }} -c dppy/label/dev ${{ env.CHANNELS }} dpctl pytest cython setuptools c-compiler cxx-compiler python="%PYTHON_SPEC%"
 
       # intel-opencl-rt is not being installed when running conda install dpctl, so do it manually
       - name: Install intel-opencl-rt

conda-recipe/meta.yaml

@@ -28,7 +28,6 @@ requirements:
         - {{ compiler('dpcpp') }} >={{ required_compiler_version }}
     host:
         - python
-        - python-gil     # [py>=314]
         - pip >=24.0
         - level-zero-devel >=1.16
         - pybind11 >=2.12
@@ -52,7 +51,6 @@ requirements:
         - tomli # [py<311]
     run:
         - python
-        - python-gil     # [py>=314]
         - {{ pin_compatible('intel-sycl-rt', min_pin='x.x', max_pin='x') }}
         - {{ pin_compatible('intel-cmplr-lib-rt', min_pin='x.x', max_pin='x') }}
         - numpy
@@ -65,7 +63,6 @@ test:
         - cython
         - setuptools
         - pytest
-        - pytest-cov
 
 about:
     home: https://github.com/IntelPython/dpctl.git

conda-recipe/run_test.sh

@@ -4,11 +4,4 @@ set -e
 
 ${PYTHON} -c "import dpctl; print(dpctl.__version__)"
 ${PYTHON} -m dpctl -f
-# don't use coverage for Python 3.13 due to crashes related to
-# Cython >= 3.1.0 and Python >= 3.13
-# TODO: remove if crash is triaged
-if ${PYTHON} --version 2>&1 | grep -q '^Python 3\.13'; then
-    ${PYTHON} -m pytest -q -ra --disable-warnings --pyargs dpctl -vv
-else
-    ${PYTHON} -m pytest -q -ra --disable-warnings --cov dpctl --cov-report term-missing --pyargs dpctl -vv
-fi
+${PYTHON} -m pytest -q -ra --disable-warnings --pyargs dpctl -vv

dpctl/_diagnostics.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements developer utilities.
 """
@@ -60,6 +61,10 @@ def _shutdown_logger():
 def syclinterface_diagnostics(verbosity="warning", log_dir=None):
     """Context manager that activate verbosity of DPCTLSyclInterface
     function calls.
+
+    .. warning::
+        This context manager modifies the ``DPCTL_VERBOSITY`` environment
+        variable and should only be used from a single thread.
     """
     _allowed_verbosity = ["warning", "error"]
     if verbosity not in _allowed_verbosity:

dpctl/_sycl_context.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements SyclContext Cython extension type.
 """

dpctl/_sycl_device.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements SyclDevice Cython extension type.
 """

dpctl/_sycl_device_factory.pyx

@@ -17,8 +17,10 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
-""" This module implements several device creation helper functions:
+"""
+This module implements several device creation helper functions:
 
   - wrapper functions to create a SyclDevice from the standard SYCL
     device selector classes.
@@ -46,7 +48,7 @@ from ._backend cimport (  # noqa: E211
     _device_type,
 )
 
-from contextvars import ContextVar
+import threading
 
 from ._sycl_device import SyclDeviceCreationError
 from .enum_types import backend_type
@@ -286,7 +288,8 @@ cpdef int get_num_devices(
 
 
 cpdef cpp_bool has_cpu_devices():
-    """ A helper function to check if there are any SYCL CPU devices available.
+    """
+    A helper function to check if there are any SYCL CPU devices available.
 
     Returns:
         bool:
@@ -298,7 +301,8 @@ cpdef cpp_bool has_cpu_devices():
 
 
 cpdef cpp_bool has_gpu_devices():
-    """ A helper function to check if there are any SYCL GPU devices available.
+    """
+    A helper function to check if there are any SYCL GPU devices available.
 
     Returns:
         bool:
@@ -310,7 +314,8 @@ cpdef cpp_bool has_gpu_devices():
 
 
 cpdef cpp_bool has_accelerator_devices():
-    """ A helper function to check if there are any SYCL Accelerator devices
+    """
+    A helper function to check if there are any SYCL Accelerator devices
     available.
 
     Returns:
@@ -325,7 +330,8 @@ cpdef cpp_bool has_accelerator_devices():
 
 
 cpdef SyclDevice select_accelerator_device():
-    """A wrapper for ``sycl::device{sycl::accelerator_selector_v}`` constructor.
+    """
+    A wrapper for ``sycl::device{sycl::accelerator_selector_v}`` constructor.
 
     Returns:
         dpctl.SyclDevice:
@@ -347,7 +353,8 @@ cpdef SyclDevice select_accelerator_device():
 
 
 cpdef SyclDevice select_cpu_device():
-    """A wrapper for ``sycl::device{sycl::cpu_selector_v}`` constructor.
+    """
+    A wrapper for ``sycl::device{sycl::cpu_selector_v}`` constructor.
 
     Returns:
         dpctl.SyclDevice:
@@ -369,7 +376,8 @@ cpdef SyclDevice select_cpu_device():
 
 
 cpdef SyclDevice select_default_device():
-    """A wrapper for ``sycl::device{sycl::default_selector_v}`` constructor.
+    """
+    A wrapper for ``sycl::device{sycl::default_selector_v}`` constructor.
 
     Returns:
         dpctl.SyclDevice:
@@ -391,7 +399,8 @@ cpdef SyclDevice select_default_device():
 
 
 cpdef SyclDevice select_gpu_device():
-    """A wrapper for ``sycl::device{sycl::gpu_selector_v}`` constructor.
+    """
+    A wrapper for ``sycl::device{sycl::gpu_selector_v}`` constructor.
 
     Returns:
         dpctl.SyclDevice:
@@ -414,46 +423,45 @@ cpdef SyclDevice select_gpu_device():
 
 cdef class _DefaultDeviceCache:
     cdef dict __device_map__
+    cdef object _cache_lock
 
     def __cinit__(self):
         self.__device_map__ = {}
-
-    cdef get_or_create(self):
-        """Return instance of SyclDevice and indicator if cache
-        has been modified"""
-        key = 0
-        if key in self.__device_map__:
-            return self.__device_map__[key], False
-        dev = select_default_device()
-        self.__device_map__[key] = dev
-        return dev, True
-
-    cdef _update_map(self, dev_map):
+        self._cache_lock = threading.Lock()
+
+    def get_or_create(self):
+        """Return cached default SyclDevice, creating it if needed."""
+        with self._cache_lock:
+            key = 0
+            if key in self.__device_map__:
+                return self.__device_map__[key]
+            dev = select_default_device()
+            self.__device_map__[key] = dev
+            return dev
+
+    def _update_map(self, dev_map):
         self.__device_map__.update(dev_map)
 
     def __copy__(self):
         cdef _DefaultDeviceCache _copy = _DefaultDeviceCache.__new__(
             _DefaultDeviceCache)
-        _copy._update_map(self.__device_map__)
+        # lock must be held to avoid race conditions on map state
+        with self._cache_lock:
+            _copy._update_map(self.__device_map__.copy())
         return _copy
 
 
-_global_default_device_cache = ContextVar(
-    "global_default_device_cache",
-    default=_DefaultDeviceCache()
-)
+# all threads share the same cached default
+_global_default_device_cache = _DefaultDeviceCache()
 
 
 cpdef SyclDevice _cached_default_device():
-    """Returns a cached device selected by default selector.
+    """
+    Returns a cached device selected by default selector.
 
     Returns:
         dpctl.SyclDevice:
             A cached default-selected SYCL device.
 
     """
-    cdef _DefaultDeviceCache _cache = _global_default_device_cache.get()
-    d_, changed_ = _cache.get_or_create()
-    if changed_:
-        _global_default_device_cache.set(_cache)
-    return d_
+    return _global_default_device_cache.get_or_create()

dpctl/_sycl_event.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements SyclEvent Cython extension type.
 """

dpctl/_sycl_platform.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements SyclPlatform Cython extension type.
 """

dpctl/_sycl_queue.pyx

@@ -17,6 +17,7 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 """ Implements SyclQueue Cython extension type.
 """

dpctl/_sycl_queue_manager.pyx

@@ -17,9 +17,10 @@
 # distutils: language = c++
 # cython: language_level=3
 # cython: linetrace=True
+# cython: freethreading_compatible = True
 
 import logging
-from contextvars import ContextVar
+import threading
 from ._sycl_context cimport SyclContext
 from ._sycl_device cimport SyclDevice
 
@@ -33,13 +34,14 @@ _logger = logging.getLogger(__name__)
 
 cdef class _DeviceDefaultQueueCache:
     cdef dict __device_queue_map__
+    cdef object _cache_lock
 
     def __cinit__(self):
         self.__device_queue_map__ = {}
+        self._cache_lock = threading.Lock()
 
     def get_or_create(self, key):
-        """Return instance of SyclQueue and indicator if cache
-        has been modified"""
+        """Return cached SyclQueue for given key, creating it if needed."""
         if (
             isinstance(key, tuple)
             and len(key) == 2
@@ -56,32 +58,34 @@ cdef class _DeviceDefaultQueueCache:
             ctx_dev = q.sycl_context, q.sycl_device
         else:
             raise TypeError
-        if ctx_dev in self.__device_queue_map__:
-            return self.__device_queue_map__[ctx_dev], False
-        if q is None:
-            q = SyclQueue(*ctx_dev)
-        self.__device_queue_map__[ctx_dev] = q
-        return q, True
-
-    cdef _update_map(self, dev_queue_map):
+        with self._cache_lock:
+            if ctx_dev in self.__device_queue_map__:
+                return self.__device_queue_map__[ctx_dev]
+            if q is None:
+                q = SyclQueue(*ctx_dev)
+            self.__device_queue_map__[ctx_dev] = q
+            return q
+
+    def _update_map(self, dev_queue_map):
         self.__device_queue_map__.update(dev_queue_map)
 
     def __copy__(self):
         cdef _DeviceDefaultQueueCache _copy = _DeviceDefaultQueueCache.__new__(
             _DeviceDefaultQueueCache
         )
-        _copy._update_map(self.__device_queue_map__)
+        # lock must be held to avoid race conditions on map state
+        with self._cache_lock:
+            _copy._update_map(self.__device_queue_map__.copy())
         return _copy
 
 
-_global_device_queue_cache = ContextVar(
-    "global_device_queue_cache",
-    default=_DeviceDefaultQueueCache()
-)
+# all threads share the same cached default
+_global_device_queue_cache = _DeviceDefaultQueueCache()
 
 
 cpdef object get_device_cached_queue(object key):
-    """Returns a cached queue associated with given device.
+    """
+    Returns a cached queue associated with given device.
 
     Args:
         key : Either a 2-tuple consisting of a :class:`dpctl.SyclContext` and
@@ -96,8 +100,4 @@ cpdef object get_device_cached_queue(object key):
         TypeError: If the input key is not one of the accepted types.
 
     """
-    _cache = _global_device_queue_cache.get()
-    q_, changed_ = _cache.get_or_create(key)
-    if changed_:
-        _global_device_queue_cache.set(_cache)
-    return q_
+    return _global_device_queue_cache.get_or_create(key)