Bump Cython to >=3.3,<3.4

cuda_bindings/pixi.toml

@@ -28,7 +28,7 @@ cuda-python-test-helpers = { path = "../cuda_python_test_helpers", editable = tr
 [feature.docs.dependencies]
 cuda-bindings = "13.2.*"
 python = "3.12.*"
-cython = ">=3.2.5,<3.3"
+cython = ">=3.3,<3.4"
 enum_tools = "*"
 make = "*"
 myst-nb = "*"
@@ -44,7 +44,7 @@ sphinx-copybutton = "*"
 sphinx-toolbox = "*"
 
 [feature.cython-tests.dependencies]
-cython = ">=3.2.5,<3.3" # for tests that exercise APIs from cython
+cython = ">=3.3,<3.4" # for tests that exercise APIs from cython
 setuptools = "*"      # for distutils
 gxx = "*"             # to compile the generated code
 # These are necessary because running the Cython tests requires compiling
@@ -118,7 +118,7 @@ python = "*"
 cuda-version = "*"
 setuptools = ">=80"
 setuptools-scm = ">=8"
-cython = ">=3.2.5,<3.3"
+cython = ">=3.3,<3.4"
 cuda-pathfinder = { path = "../cuda_pathfinder" }
 cuda-cudart-static = "*"
 cuda-nvrtc-dev = "*"

cuda_bindings/pyproject.toml

@@ -4,7 +4,7 @@
 requires = [
     "setuptools>=80.0.0",
     "setuptools_scm[simple]>=8,!=10.1",
-    "cython>=3.2.5,<3.3",
+    "cython>=3.3,<3.4",
     "cuda-pathfinder>=1.5",
 ]
 build-backend = "build_hooks"
@@ -43,7 +43,7 @@ all = [
 
 [dependency-groups]
 test = [
-    "cython>=3.2.5,<3.3",
+    "cython>=3.3,<3.4",
     "setuptools>=80.0.0",
     # TODO: remove the Python 3.15 guard once 3.15 is officially supported
     "matplotlib>=3.5.0,<=3.10.9; python_version < '3.15'",

cuda_bindings/tests/test_cuda.py

@@ -579,12 +579,12 @@ def test_eglFrame():
     assert int(val.frame.pArray[0]) == 0
     assert int(val.frame.pArray[1]) == 0
     assert int(val.frame.pArray[2]) == 0
-    val.frame.pArray = [1, 2, 3]
+    val.frame.pArray = [cuda.CUarray(x) for x in (1, 2, 3)]
     # [<CUarray 0x1>, <CUarray 0x2>, <CUarray 0x3>]
     assert int(val.frame.pArray[0]) == 1
     assert int(val.frame.pArray[1]) == 2
     assert int(val.frame.pArray[2]) == 3
-    val.frame.pArray = [cuda.CUarray(4), 2, 3]
+    val.frame.pArray = [cuda.CUarray(x) for x in (4, 2, 3)]
     # [<CUarray 0x4>, <CUarray 0x2>, <CUarray 0x3>]
     assert int(val.frame.pArray[0]) == 4
     assert int(val.frame.pArray[1]) == 2

cuda_bindings/tests/test_cudart.py

@@ -1285,12 +1285,12 @@ def test_cudart_eglFrame():
     assert int(frame.frame.pArray[0]) == 0
     assert int(frame.frame.pArray[1]) == 0
     assert int(frame.frame.pArray[2]) == 0
-    frame.frame.pArray = [1, 2, 3]
+    frame.frame.pArray = [cudart.cudaArray_t(x) for x in (1, 2, 3)]
     # [<cudaArray_t 0x1>, <cudaArray_t 0x2>, <cudaArray_t 0x3>]
     assert int(frame.frame.pArray[0]) == 1
     assert int(frame.frame.pArray[1]) == 2
     assert int(frame.frame.pArray[2]) == 3
-    frame.frame.pArray = [1, 2, cudart.cudaArray_t(4)]
+    frame.frame.pArray = [cudart.cudaArray_t(x) for x in (1, 2, 4)]
     # [<cudaArray_t 0x1>, <cudaArray_t 0x2>, <cudaArray_t 0x4>]
     assert int(frame.frame.pArray[0]) == 1
     assert int(frame.frame.pArray[1]) == 2

cuda_core/cuda/core/_device_resources.pyi

@@ -40,7 +40,7 @@ class SMResourceOptions:
     preferred_coscheduled_sm_count: int | SequenceABC[int] | None = None
     backfill: bool | SequenceABC[bool] = False
 
-@dataclass
+@dataclass(init=False)
 class WorkqueueResourceOptions:
     """Customizable :obj:`WorkqueueResource.configure` options.
 
@@ -60,7 +60,7 @@ class WorkqueueResourceOptions:
     sharing_scope: WorkqueueSharingScopeType | str | None = None
     concurrency_limit: int | None = None
 
-    def __post_init__(self): ...
+    def __init__(self, sharing_scope: WorkqueueSharingScopeType | str | None=None, concurrency_limit: int | None=None): ...
 
 class SMResource:
     """Represent an SM (streaming multiprocessor) resource partition.

cuda_core/cuda/core/_device_resources.pyx

@@ -126,7 +126,12 @@ cdef class SMResourceOptions:
     backfill: bool | SequenceABC[bool] = False
 
 
-@dataclass
+# init=False: Cython 3.3 wraps the dataclass-generated __init__ in a
+# critical_section and then warns (promoted to an error by our
+# warning_errors=True build) that the __post_init__ call inside it is
+# unprotected Python attribute access. A hand-written __init__ sidesteps
+# that; construction of a not-yet-shared object needs no lock anyway.
+@dataclass(init=False)
 cdef class WorkqueueResourceOptions:
     """Customizable :obj:`WorkqueueResource.configure` options.
 
@@ -147,13 +152,19 @@ cdef class WorkqueueResourceOptions:
     sharing_scope: WorkqueueSharingScopeType | str | None = None
     concurrency_limit: int | None = None
 
-    def __post_init__(self):
+    def __init__(
+        self,
+        sharing_scope: WorkqueueSharingScopeType | str | None = None,
+        concurrency_limit: int | None = None,
+    ):
         from cuda.core.typing import WorkqueueSharingScopeType
-        check_str_enum(self.sharing_scope, WorkqueueSharingScopeType, allow_none=True)
-        if self.concurrency_limit is not None and self.concurrency_limit < 1:
+        check_str_enum(sharing_scope, WorkqueueSharingScopeType, allow_none=True)
+        if concurrency_limit is not None and concurrency_limit < 1:
             raise ValueError(
-                f"concurrency_limit must be >= 1, got {self.concurrency_limit}"
+                f"concurrency_limit must be >= 1, got {concurrency_limit}"
             )
+        self.sharing_scope = sharing_scope
+        self.concurrency_limit = concurrency_limit
 
 
 cdef inline int _validate_split_field_length(

cuda_core/cuda/core/_memory/_buffer.pyx

@@ -404,16 +404,24 @@ cdef class Buffer:
         return _ipc.Buffer_from_ipc_descriptor(cls, mr, ipc_descriptor, stream)
 
     @property
-    @cython.critical_section
     def ipc_descriptor(self) -> IPCBufferDescriptor:
         """Descriptor for sharing this buffer with other processes."""
         Buffer_check_open(self)
         cdef object ipc_data
-        if self._ipc_data is None:
+        # The critical section only protects the cdef-level check-and-set of
+        # self._ipc_data; the Python-level IPCDataForBuffer construction and
+        # .ipc_descriptor read happen outside the lock (Cython 3.3 warns that
+        # Python attribute access is not usefully protected by critical_section).
+        with cython.critical_section(self):
+            ipc_data = self._ipc_data
+        if ipc_data is None:
             ipc_data = IPCDataForBuffer(_ipc.Buffer_get_ipc_descriptor(self), False)
-            if self._ipc_data is None:
-                self._ipc_data = ipc_data
-        return self._ipc_data.ipc_descriptor
+            with cython.critical_section(self):
+                if self._ipc_data is None:
+                    self._ipc_data = ipc_data
+                else:
+                    ipc_data = self._ipc_data
+        return ipc_data.ipc_descriptor
 
     def close(self, stream: Stream | GraphBuilder | None = None) -> None:
         """Deallocate this buffer asynchronously on the given stream.

cuda_core/cuda/core/_memoryview.pyx

@@ -560,28 +560,36 @@ cdef class StridedMemoryView:
                 self._layout = layout
         return self._layout
 
-    @cython.critical_section
     cdef inline object get_buffer(self):
         """
         Returns Buffer instance with the underlying data.
         If the SMV was created from a Buffer, it will return the same Buffer instance.
         Otherwise, it will create a new instance with owner set to the exporting object.
         """
         cdef object buffer
-        if self._buffer is None:
+        with cython.critical_section(self):
+            buffer = self._buffer
+        if buffer is None:
             if isinstance(self.exporting_obj, Buffer):
                 buffer = self.exporting_obj
             else:
                 buffer = Buffer.from_handle(self.ptr, 0, owner=self.exporting_obj)
-            if self._buffer is None:
-                self._buffer = buffer
-        return self._buffer
+            # The critical section only protects the cdef-level check-and-set
+            # of self._buffer; the Python-level Buffer construction above
+            # happens outside the lock (Cython 3.3 warns that Python attribute
+            # access is not usefully protected by critical_section).
+            with cython.critical_section(self):
+                if self._buffer is None:
+                    self._buffer = buffer
+                else:
+                    buffer = self._buffer
+        return buffer
 
-    @cython.critical_section
     cdef inline object get_dtype(self):
         cdef object dtype
-        if self._dtype is None:
-            dtype = None
+        with cython.critical_section(self):
+            dtype = self._dtype
+        if dtype is None:
             if self.dl_tensor != NULL:
                 dtype = dtype_dlpack_to_numpy(&self.dl_tensor.dtype)
             elif isinstance(self.metadata, int):
@@ -590,9 +598,16 @@ cdef class StridedMemoryView:
                     self.metadata)
             elif self.metadata is not None:
                 dtype = _typestr2dtype(self.metadata["typestr"])
-            if self._dtype is None:
-                self._dtype = dtype
-        return self._dtype
+            # The critical section only protects the cdef-level check-and-set
+            # of self._dtype; the Python-level dtype resolution above happens
+            # outside the lock (Cython 3.3 warns that Python attribute access
+            # is not usefully protected by critical_section).
+            with cython.critical_section(self):
+                if self._dtype is None:
+                    self._dtype = dtype
+                else:
+                    dtype = self._dtype
+        return dtype
 
 
 cdef void _smv_pycapsule_deleter(object capsule) noexcept:

cuda_core/pixi.toml

@@ -56,7 +56,7 @@ CUDA_HOME = "$CONDA_PREFIX/targets/sbsa-linux"
 CUDA_HOME = "$CONDA_PREFIX/Library"
 
 [feature.cython-tests.dependencies]
-cython = ">=3.2.5,<3.3" # for tests that exercise APIs from cython
+cython = ">=3.3,<3.4" # for tests that exercise APIs from cython
 setuptools = "*"      # for distutils
 gxx = "*"             # to compile the generated code
 # These are necessary because running the Cython tests requires compiling
@@ -97,7 +97,7 @@ cuda-version = "12.*"
 
 [feature.docs.dependencies]
 cuda-core = { path = "." }
-cython = ">=3.2.5,<3.3"
+cython = ">=3.3,<3.4"
 myst-parser = "*"
 numpy = "*"
 numpydoc = "*"
@@ -214,7 +214,7 @@ python = "*"
 cuda-version = "*"
 setuptools = ">=80"
 setuptools-scm = ">=8"
-cython = ">=3.2.5,<3.3"
+cython = ">=3.3,<3.4"
 cuda-nvrtc-dev = "*"
 cuda-bindings = "*"
 dlpack = "*"

cuda_core/pyproject.toml

@@ -6,7 +6,7 @@
 requires = [
     "setuptools>=80",
     "setuptools-scm[simple]>=8,!=10.1",
-    "Cython>=3.2.5,<3.3",
+    "Cython>=3.3,<3.4",
     "cuda-pathfinder>=1.5"
 ]
 build-backend = "build_hooks"
@@ -59,7 +59,7 @@ cu13 = ["cuda-bindings[all]==13.*", "cuda-toolkit==13.*"]
 
 [dependency-groups]
 test = [
-    "cython>=3.2.5,<3.3",
+    "cython>=3.3,<3.4",
     "setuptools>=80",
     "pytest==9.1.0",
     "pytest-benchmark==5.2.3",

cuda_pathfinder/pixi.toml

@@ -20,7 +20,7 @@ pytest-run-parallel = ">=0.4.1"
 # Keep this dependency set aligned with cuda_python/docs/environment-docs.yml.
 [feature.docs.dependencies]
 python = "3.12.*"
-cython = ">=3.2.5,<3.3"
+cython = ">=3.3,<3.4"
 enum_tools = "*"
 make = "*"
 myst-nb = "*"

cuda_python/docs/environment-docs.yml

@@ -7,7 +7,7 @@ channels:
 dependencies:
   # ATTENTION: This dependency list is duplicated in
   #            toolshed/setup-docs-env.sh. Please KEEP THEM IN SYNC!
-  - cython >=3.2.5,<3.3
+  - cython >=3.3,<3.4
   - myst-parser
   - numpy
   - numpydoc

toolshed/setup-docs-env.sh

@@ -39,7 +39,7 @@ echo "Creating environment '${ENV_NAME}'…"
 #            cuda_python/docs/environment-docs.yml. Please KEEP THEM IN SYNC!
 conda create -y -n "${ENV_NAME}" \
     "python=${PYVER}" \
-    "cython>=3.2.5,<3.3" \
+    "cython>=3.3,<3.4" \
     myst-parser \
     numpy \
     numpydoc \