diff --git a/loopy/kernel/array.py b/loopy/kernel/array.py
index 186597c64734b54b8d08f0db43b57826d79f9567..6bf733a84f9df48fbb8433015e1e137f6dc0392c 100644
--- a/loopy/kernel/array.py
+++ b/loopy/kernel/array.py
@@ -833,6 +833,7 @@ class ArrayBase(ImmutableRecord):
                 dim_names=dim_names,
                 order=order,
                 alignment=alignment,
+                for_atomic=for_atomic,
                 **kwargs)
 
     def __eq__(self, other):
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index fc950c78e465a1d9051ee8c3cc6cb9650a6865ea..2afcd3db4331d57e1e61c48ba521ebaa296ddbb2 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -53,7 +53,7 @@ def prepare_for_caching(kernel):
     for arg in kernel.args:
         dtype = arg.dtype
         if dtype is not None and dtype is not lp.auto and dtype.target is not tgt:
-            arg = arg.copy(dtype=dtype.with_target(kernel.target))
+            arg = arg.copy(dtype=dtype.with_target(tgt), target=tgt)
 
         new_args.append(arg)
 
@@ -61,7 +61,7 @@ def prepare_for_caching(kernel):
     for name, temp in six.iteritems(kernel.temporary_variables):
         dtype = temp.dtype
         if dtype is not None and dtype is not lp.auto and dtype.target is not tgt:
-            temp = temp.copy(dtype=dtype.with_target(tgt))
+            temp = temp.copy(dtype=dtype.with_target(tgt), target=tgt)
 
         new_temporary_variables[name] = temp