diff --git a/sumpy/kernel_common.py b/sumpy/kernel_common.py
index d91f5949792455ea0a740c88ee34818038bdfa2c..0940e0e7737774026043404c41d0462ea6e8a58a 100644
--- a/sumpy/kernel_common.py
+++ b/sumpy/kernel_common.py
@@ -14,7 +14,7 @@ COMMON_PREAMBLE = r"""//CL//
 % if double_support:
     #pragma OPENCL EXTENSION cl_khr_fp64: enable
 % endif
-#pragma OPENCL EXTENSION cl_amd_printf: enable
+## #pragma OPENCL EXTENSION cl_amd_printf: enable
 
 <%def name="load_vector_g(tgt_name, src_name, src_base_offset)">
   % for i in range(dimensions):
diff --git a/sumpy/tools.py b/sumpy/tools.py
index f1b8ee3f769f7a7f049e06491eed650baf99bab1..b9d6c8baa304a550bccb006529b24c4b6e16f910 100644
--- a/sumpy/tools.py
+++ b/sumpy/tools.py
@@ -5,6 +5,9 @@ from __future__ import division
 
 # {{{ multi_index helpers
 
+def add_mi(mi1, mi2):
+    return tuple(mi1i+mi2i for mi1i, mi2i in zip(mi1, mi2))
+
 def mi_factorial(mi):
     from pytools import factorial
     result = 1