Merge remote-tracking branch 'upstream/main' into dching/add-compute-sanitizer-to-ci

Author: carterbox

.bandit

@@ -0,0 +1,2 @@
+[bandit]
+skips = B101,B311

.github/workflows/test-wheel-windows.yml

@@ -55,9 +55,9 @@ jobs:
 
           if ('${{ inputs.local-ctk }}' -eq '1') {
             if ($TEST_CUDA_MAJOR -eq '12') {
-              $MINI_CTK_DEPS = '["nvcc", "nvrtc", "nvjitlink"]'
+              $MINI_CTK_DEPS = '["nvcc", "nvrtc", "nvjitlink", "thrust"]'
             } else {
-              $MINI_CTK_DEPS = '["nvcc", "nvrtc"]'
+              $MINI_CTK_DEPS = '["nvcc", "nvrtc", "thrust"]'
             }
           }
 

.pre-commit-config.yaml

@@ -20,6 +20,9 @@ repos:
     rev: 8ff25e07e487f143571cc305e56dd0253c60bc7b  #v1.8.3
     hooks:
       - id: bandit
+        args:
+          - --ini
+          - .bandit
 
 default_language_version:
       python: python3

cuda_bindings/tests/test_cuda.py

@@ -652,7 +652,8 @@ def test_get_error_name_and_string():
 
 @pytest.mark.skipif(not callableBinary("nvidia-smi"), reason="Binary existance needed")
 def test_device_get_name():
-    import subprocess
+    # TODO: Refactor this test once we have nvml bindings to avoid the use of subprocess
+    import subprocess  # nosec B404
 
     (err,) = cuda.cuInit(0)
     assert err == cuda.CUresult.CUDA_SUCCESS
@@ -661,12 +662,12 @@ def test_device_get_name():
     err, ctx = cuda.cuCtxCreate(0, device)
     assert err == cuda.CUresult.CUDA_SUCCESS
 
-    p = subprocess.run(
-        ["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"], stdout=subprocess.PIPE, stderr=subprocess.PIPE
-    )
+    p = subprocess.check_output(
+        ["nvidia-smi", "--query-gpu=name", "--format=csv,noheader"], shell=False, stderr=subprocess.PIPE
+    )  # nosec B603, B607
 
     delimiter = b"\r\n" if platform.system() == "Windows" else b"\n"
-    expect = p.stdout.split(delimiter)
+    expect = p.split(delimiter)
     size = 64
     _, got = cuda.cuDeviceGetName(size, device)
     got = got.split(b"\x00")[0]

cuda_core/cuda/core/experimental/_event.py

@@ -8,7 +8,15 @@
 from dataclasses import dataclass
 from typing import TYPE_CHECKING, Optional
 
-from cuda.core.experimental._utils.cuda_utils import CUDAError, check_or_create_options, driver, handle_return
+from cuda.core.experimental._utils.cuda_utils import (
+    CUDAError,
+    check_or_create_options,
+    driver,
+    handle_return,
+)
+from cuda.core.experimental._utils.cuda_utils import (
+    _check_driver_error as raise_if_driver_error,
+)
 
 if TYPE_CHECKING:
     import cuda.bindings
@@ -117,13 +125,31 @@ def __rsub__(self, other):
 
     def __sub__(self, other):
         # return self - other (in milliseconds)
+        err, timing = driver.cuEventElapsedTime(other.handle, self.handle)
         try:
-            timing = handle_return(driver.cuEventElapsedTime(other.handle, self.handle))
+            raise_if_driver_error(err)
+            return timing
         except CUDAError as e:
-            raise RuntimeError(
-                "Timing capability must be enabled in order to subtract two Events; timing is disabled by default."
-            ) from e
-        return timing
+            if err == driver.CUresult.CUDA_ERROR_INVALID_HANDLE:
+                if self.is_timing_disabled or other.is_timing_disabled:
+                    explanation = (
+                        "Both Events must be created with timing enabled in order to subtract them; "
+                        "use EventOptions(enable_timing=True) when creating both events."
+                    )
+                else:
+                    explanation = (
+                        "Both Events must be recorded before they can be subtracted; "
+                        "use Stream.record() to record both events to a stream."
+                    )
+            elif err == driver.CUresult.CUDA_ERROR_NOT_READY:
+                explanation = (
+                    "One or both events have not completed; "
+                    "use Event.sync(), Stream.sync(), or Device.sync() to wait for the events to complete "
+                    "before subtracting them."
+                )
+            else:
+                raise e
+            raise RuntimeError(explanation) from e
 
     @property
     def is_timing_disabled(self) -> bool:
@@ -164,5 +190,11 @@ def is_done(self) -> bool:
 
     @property
     def handle(self) -> cuda.bindings.driver.CUevent:
-        """Return the underlying CUevent object."""
+        """Return the underlying CUevent object.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(Event.handle)``.
+        """
         return self._mnff.handle

cuda_core/cuda/core/experimental/_linker.py

@@ -503,6 +503,11 @@ def handle(self) -> LinkerHandleT:
         .. note::
 
            The type of the returned object depends on the backend.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(Linker.handle)``.
         """
         return self._mnff.handle
 

cuda_core/cuda/core/experimental/_memory.py

@@ -6,7 +6,7 @@
 
 import abc
 import weakref
-from typing import Optional, Tuple, TypeVar
+from typing import Optional, Tuple, TypeVar, Union
 
 from cuda.core.experimental._dlpack import DLDeviceType, make_py_capsule
 from cuda.core.experimental._stream import default_stream
@@ -18,6 +18,9 @@
 # TODO: define a memory property mixin class and make Buffer and
 # MemoryResource both inherit from it
 
+DevicePointerT = Union[driver.CUdeviceptr, int, None]
+"""A type union of `Cudeviceptr`, `int` and `None` for hinting Buffer.handle."""
+
 
 class Buffer:
     """Represent a handle to allocated memory.
@@ -81,8 +84,14 @@ def close(self, stream=None):
         self._mnff.close(stream)
 
     @property
-    def handle(self):
-        """Return the buffer handle object."""
+    def handle(self) -> DevicePointerT:
+        """Return the buffer handle object.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(Buffer.handle)``.
+        """
         return self._mnff.ptr
 
     @property

cuda_core/cuda/core/experimental/_module.py

@@ -354,5 +354,11 @@ def code(self) -> CodeTypeT:
     @property
     @precondition(_lazy_load_module)
     def handle(self):
-        """Return the underlying handle object."""
+        """Return the underlying handle object.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(ObjectCode.handle)``.
+        """
         return self._handle

cuda_core/cuda/core/experimental/_program.py

@@ -524,5 +524,10 @@ def handle(self) -> ProgramHandleT:
         .. note::
 
            The type of the returned object depends on the backend.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(Program.handle)``.
         """
         return self._mnff.handle

cuda_core/cuda/core/experimental/_stream.py

@@ -189,7 +189,13 @@ def __cuda_stream__(self) -> Tuple[int, int]:
 
     @property
     def handle(self) -> cuda.bindings.driver.CUstream:
-        """Return the underlying ``CUstream`` object."""
+        """Return the underlying ``CUstream`` object.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int(Stream.handle)``.
+        """
         return self._mnff.handle
 
     @property