From ab850a64229bcd3f9794652940197f208161b833 Mon Sep 17 00:00:00 2001
From: Phillip Cloud <417981+cpcloud@users.noreply.github.com>
Date: Tue, 6 Jan 2026 08:12:52 -0500
Subject: [PATCH] feat: improve structured dtype array support in
 `StridedMemoryView` (#1425)

* test: add test for structured dtype properties

* feat: improve support for structured dtype arrays held by `StridedMemoryView`

* chore: simplify using an lru cache instead of manual caching

(cherry picked from commit 36742ddc5b91ed388f55d5ec7ade8d245835fb35)
---
 cuda_core/cuda/core/_memoryview.pyx | 32 +++++++++--------------------
 cuda_core/tests/test_utils.py       | 17 +++++++++++++++
 2 files changed, 27 insertions(+), 22 deletions(-)

diff --git a/cuda_core/cuda/core/_memoryview.pyx b/cuda_core/cuda/core/_memoryview.pyx
index c12cbbaa8a..1dbebe7789 100644
--- a/cuda_core/cuda/core/_memoryview.pyx
+++ b/cuda_core/cuda/core/_memoryview.pyx
@@ -365,8 +365,7 @@ cdef class StridedMemoryView:
             if self.dl_tensor != NULL:
                 self._dtype = dtype_dlpack_to_numpy(&self.dl_tensor.dtype)
             elif self.metadata is not None:
-                # TODO: this only works for built-in numeric types
-                self._dtype = _typestr2dtype[self.metadata["typestr"]]
+                self._dtype = _typestr2dtype(self.metadata["typestr"])
         return self._dtype
 
 
@@ -486,25 +485,14 @@ cdef StridedMemoryView view_as_dlpack(obj, stream_ptr, view=None):
     return buf
 
 
-_builtin_numeric_dtypes = [
-    numpy.dtype("uint8"),
-    numpy.dtype("uint16"),
-    numpy.dtype("uint32"),
-    numpy.dtype("uint64"),
-    numpy.dtype("int8"),
-    numpy.dtype("int16"),
-    numpy.dtype("int32"),
-    numpy.dtype("int64"),
-    numpy.dtype("float16"),
-    numpy.dtype("float32"),
-    numpy.dtype("float64"),
-    numpy.dtype("complex64"),
-    numpy.dtype("complex128"),
-    numpy.dtype("bool"),
-]
-# Doing it once to avoid repeated overhead
-_typestr2dtype = {dtype.str: dtype for dtype in _builtin_numeric_dtypes}
-_typestr2itemsize = {dtype.str: dtype.itemsize for dtype in _builtin_numeric_dtypes}
+@functools.lru_cache
+def _typestr2dtype(str typestr):
+    return numpy.dtype(typestr)
+
+
+@functools.lru_cache
+def _typestr2itemsize(str typestr):
+    return _typestr2dtype(typestr).itemsize
 
 
 cdef object dtype_dlpack_to_numpy(DLDataType* dtype):
@@ -664,7 +652,7 @@ cdef _StridedLayout layout_from_cai(object metadata):
     cdef _StridedLayout layout = _StridedLayout.__new__(_StridedLayout)
     cdef object shape = metadata["shape"]
     cdef object strides = metadata.get("strides")
-    cdef int itemsize = _typestr2itemsize[metadata["typestr"]]
+    cdef int itemsize = _typestr2itemsize(metadata["typestr"])
     layout.init_from_tuple(shape, strides, itemsize, True)
     return layout
 
diff --git a/cuda_core/tests/test_utils.py b/cuda_core/tests/test_utils.py
index 06ee3520e2..eb883cd3f3 100644
--- a/cuda_core/tests/test_utils.py
+++ b/cuda_core/tests/test_utils.py
@@ -446,3 +446,20 @@ def test_from_buffer_with_non_power_of_two_itemsize():
     buffer = dev.memory_resource.allocate(required_size)
     view = StridedMemoryView.from_buffer(buffer, shape=shape, strides=layout.strides, dtype=dtype, is_readonly=True)
     assert view.dtype == dtype
+
+
+def test_struct_array():
+    cp = pytest.importorskip("cupy")
+
+    x = np.array([(1.0, 2), (2.0, 3)], dtype=[("array1", np.float64), ("array2", np.int64)])
+
+    y = cp.empty(2, dtype=x.dtype)
+    y.set(x)
+
+    smv = StridedMemoryView.from_cuda_array_interface(y, stream_ptr=0)
+    assert smv.size * smv.dtype.itemsize == x.nbytes
+    assert smv.size == x.size
+    assert smv.shape == x.shape
+    # full dtype information doesn't seem to be preserved due to use of type strings,
+    # which are lossy, e.g., dtype([("a", "int")]).str == "V8"
+    assert smv.dtype == np.dtype(f"V{x.itemsize}")