Fixed gh-1468

Function _reduce_over_axis promotes input array to requested
result data type and carries out reduction computation in that
data type.

Author: oleksandr-pavlyk

dpctl/tensor/_reduction.py

@@ -142,21 +142,24 @@ def _reduction_over_axis(
                 "Automatically determined reduction data type does not "
                 "have direct implementation"
             )
-        tmp_dt = _default_reduction_type_fn(inp_dt, q)
         tmp = dpt.empty(
-            res_shape, dtype=tmp_dt, usm_type=res_usm_type, sycl_queue=q
+            arr2.shape, dtype=res_dt, usm_type=res_usm_type, sycl_queue=q
         )
-        ht_e_tmp, r_e = _reduction_fn(
-            src=arr2, trailing_dims_to_reduce=red_nd, dst=tmp, sycl_queue=q
+        ht_e_cpy, cpy_e = ti._copy_usm_ndarray_into_usm_ndarray(
+            src=arr2, dst=tmp, sycl_queue=q
         )
-        host_tasks_list.append(ht_e_tmp)
+        host_tasks_list.append(ht_e_cpy)
         res = dpt.empty(
             res_shape, dtype=res_dt, usm_type=res_usm_type, sycl_queue=q
         )
-        ht_e, _ = ti._copy_usm_ndarray_into_usm_ndarray(
-            src=tmp, dst=res, sycl_queue=q, depends=[r_e]
+        ht_e_red, r_e = _reduction_fn(
+            src=tmp,
+            trailing_dims_to_reduce=red_nd,
+            dst=res,
+            sycl_queue=q,
+            depends=[cpy_e],
         )
-        host_tasks_list.append(ht_e)
+        host_tasks_list.append(ht_e_red)
 
     if keepdims:
         res_shape = res_shape + (1,) * red_nd