Fix for gh-1330

When source array must be broadcast, non-zero strides corresponding
to unit size dimensions must be zeroed out, otherwise if such dimension is
broadcasted, constructor would expected a bigger buffer than is really
necessary.

Author: oleksandr-pavlyk

dpctl/tensor/_copy_utils.py

@@ -279,7 +279,10 @@ def _copy_from_usm_ndarray_to_usm_ndarray(dst, src):
         common_shape = common_shape[ones_count:]
 
     if src.ndim < len(common_shape):
-        new_src_strides = (0,) * (len(common_shape) - src.ndim) + src.strides
+        pad_count = len(common_shape) - src.ndim
+        new_src_strides = (0,) * pad_count + tuple(
+            s if d > 1 else 0 for s, d in zip(src.strides, src.shape)
+        )
         src_same_shape = dpt.usm_ndarray(
             common_shape, dtype=src.dtype, buffer=src, strides=new_src_strides
         )

dpctl/tests/test_usm_ndarray_ctor.py

@@ -1012,6 +1012,15 @@ def test_setitem_same_dtype(dtype, src_usm_type, dst_usm_type):
     Zusm_empty[Ellipsis] = Zusm_3d[0, 0, 0:0]
 
 
+def test_setitem_boradcasting():
+    get_queue_or_skip()
+    dst = dpt.ones((2, 3, 4), dtype="u4")
+    src = dpt.zeros((3, 1), dtype=dst.dtype)
+    dst[...] = src
+    expected = np.zeros(dst.shape, dtype=dst.dtype)
+    assert np.array_equal(dpt.asnumpy(dst), expected)
+
+
 @pytest.mark.parametrize(
     "dtype",
     _all_dtypes,