Fix #1043: Fix memory leak in StridedMemoryView

Author: mdboom

cuda_bindings/benchmarks/test_cupy.py

@@ -12,6 +12,7 @@
 except ImportError:
     skip_tests = True
 
+from cuda.core.experimental.utils import StridedMemoryView
 from kernels import kernel_string
 
 
@@ -197,3 +198,11 @@ def test_launch_latency_small_kernel_16_args(benchmark):
     with stream:
         benchmark(launch, kernel, args)
         stream.synchronize()
+
+
+# Ensure that memory views dellocate their reference to dlpack/cupy tensors
+@pytest.mark.skipif(skip_tests, reason="cupy is not installed")
+def test_strided_memory_view_leak(benchmark):
+    for idx in range(1000):
+        arr = cupy.zeros((1024, 1024, 1024), dtype=cupy.uint8)
+        view = StridedMemoryView(arr, stream_ptr=-1)

cuda_core/cuda/core/experimental/_memoryview.pyx

@@ -105,6 +105,23 @@ cdef class StridedMemoryView:
         else:
             pass
 
+    def __dealloc__(self):
+        if self.dl_tensor == NULL:
+            return
+
+        if cpython.PyCapsule_IsValid(
+                self.metadata, DLPACK_VERSIONED_TENSOR_USED_NAME):
+            data = cpython.PyCapsule_GetPointer(
+                self.metadata, DLPACK_VERSIONED_TENSOR_USED_NAME)
+            dlm_tensor_ver = <DLManagedTensorVersioned*>data
+            dlm_tensor_ver.deleter(dlm_tensor_ver)
+        elif cpython.PyCapsule_IsValid(
+                self.metadata, DLPACK_TENSOR_USED_NAME):
+            data = cpython.PyCapsule_GetPointer(
+                self.metadata, DLPACK_TENSOR_USED_NAME)
+            dlm_tensor = <DLManagedTensor*>data
+            dlm_tensor.deleter(dlm_tensor)
+
     @property
     def shape(self) -> tuple[int]:
         if self._shape is None and self.exporting_obj is not None: