Use a numpy array rather than ctypes object

python/cuda_parallel/cuda/parallel/experimental/_cccl.py

@@ -13,7 +13,7 @@
 
 from ._utils.cai import get_dtype, is_contiguous
 from .iterators._iterators import IteratorBase
-from .typing import DeviceArrayLike, GpuStruct
+from .typing import DeviceArrayLike
 
 
 # MUST match `cccl_type_enum` in c/include/cccl/c/types.h
@@ -215,10 +215,7 @@ def to_cccl_iter(array_or_iterator) -> Iterator:
     return _device_array_to_cccl_iter(array_or_iterator)
 
 
-def to_cccl_value(array_or_struct: np.ndarray | GpuStruct) -> Value:
-    info = _numpy_type_to_info(array_or_struct.dtype)
-    if isinstance(array_or_struct, np.ndarray):
-        data = ctypes.cast(array_or_struct.ctypes.data, ctypes.c_void_p)
-    else:
-        data = ctypes.cast(ctypes.pointer(array_or_struct._data), ctypes.c_void_p)
+def to_cccl_value(array: np.ndarray) -> Value:
+    info = _numpy_type_to_info(array.dtype)
+    data = ctypes.cast(array.ctypes.data, ctypes.c_void_p)
     return Value(info, data)

python/cuda_parallel/cuda/parallel/experimental/algorithms/reduce.py

@@ -74,7 +74,9 @@ def __init__(
             d_in_cccl,
             d_out_cccl,
             self.op_wrapper.handle(),
-            cccl.to_cccl_value(h_init),
+            cccl.to_cccl_value(
+                h_init if isinstance(h_init, np.ndarray) else h_init._data
+            ),
             cc_major,
             cc_minor,
             ctypes.c_char_p(cub_path),
@@ -121,7 +123,9 @@ def __call__(
             d_out_cccl,
             ctypes.c_ulonglong(num_items),
             self.op_wrapper.handle(),
-            cccl.to_cccl_value(h_init),
+            cccl.to_cccl_value(
+                h_init if isinstance(h_init, np.ndarray) else h_init._data
+            ),
             None,
         )
         if error != enums.CUDA_SUCCESS:

python/cuda_parallel/cuda/parallel/experimental/struct.py

@@ -55,13 +55,15 @@ def gpu_struct(this: type) -> Type[GpuStruct]:
     # create CuPy/NumPy arrays of this type.
     setattr(this, "dtype", np.dtype(list(anns.items())))
 
-    # Define __post_init__ to create a ctypes object from the fields,
-    # and keep a reference to it in the `._data` attribute. This
-    # ctypes object is what is ultimately passed to the underlying C
-    # library and we need to make sure we keep it alive.
+    # Define __post_init__ to create a numpy struct from the fields,
+    # and keep a reference to it in the `._data` attribute. The data
+    # underlying this array is what is ultimately passed to the C
+    # library, and we need to keep a reference to it for the lifetime
+    # of the object.
     def __post_init__(self):
-        ctypes_typ = np.ctypeslib.as_ctypes_type(this.dtype)
-        self._data = ctypes_typ(*(getattr(self, name) for name in this.dtype.names))
+        self._data = np.array(
+            [tuple(getattr(self, name) for name in anns)], dtype=self.dtype
+        )
 
     setattr(this, "__post_init__", __post_init__)
 

python/cuda_parallel/tests/test_reduce_api.py

@@ -181,9 +181,13 @@ def square_op(a):
 
 
 def test_reduce_struct_type():
+    # example-begin reduce-struct
+    import cupy as cp
+    import numpy as np
+
+    from cuda.parallel.experimental import algorithms
     from cuda.parallel.experimental.struct import gpu_struct
 
-    # example-begin reduce-struct
     @gpu_struct
     class Pixel:
         r: np.int32
@@ -209,4 +213,3 @@ def max_g_value(x, y):
 
     np.testing.assert_equal(expected["g"], d_out.get()["g"])
     # example-end reduce-struct
-