diff --git a/MEMO b/MEMO
index 6fa8a95c5c13f13dc8d0a566c4b5aa4f89d1d65b..19e54706ddaf37dd845abd595dcee519df7c3280 100644
--- a/MEMO
+++ b/MEMO
@@ -39,8 +39,6 @@ Things to consider
 To-do
 ^^^^^
 
-- Automatically generate testing code vs. sequential.
-
 - Fix all tests
 
 - Deal with equality constraints.
@@ -83,6 +81,8 @@ Future ideas
 Dealt with
 ^^^^^^^^^^
 
+- Automatically generate testing code vs. sequential.
+
 - If isl can prove that all operands are positive, may use '/' instead of
   'floor_div'.
 
diff --git a/loopy/__init__.py b/loopy/__init__.py
index 451a175835d31d4fd241ce6f3713a63edeecaa98..ba05ddff496701fbd367609a5d08d55510d7c859 100644
--- a/loopy/__init__.py
+++ b/loopy/__init__.py
@@ -381,7 +381,7 @@ def join_dimensions(kernel, inames, new_iname=None, tag=AutoFitLocalIndexTag()):
 
 # {{{ dimension tag
 
-def tag_dimensions(kernel, iname_to_tag):
+def tag_dimensions(kernel, iname_to_tag, force=False):
     from loopy.kernel import parse_tag
 
     iname_to_tag = dict((iname, parse_tag(tag))
@@ -406,7 +406,7 @@ def tag_dimensions(kernel, iname_to_tag):
             raise ValueError("cannot tag '%s' as parallel--"
                     "iname requires sequential execution" % iname)
 
-        if old_tag is not None and (old_tag != new_tag):
+        if (not force) and old_tag is not None and (old_tag != new_tag):
             raise RuntimeError("'%s' is already tagged '%s'--cannot retag"
                     % (iname, old_tag))
 
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 06807c2e101ab946cd9852c756e0dfeb847f07b6..a08c2fb3e75bb09dfc309d70e2d3aa41d01f14f0 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -99,8 +99,10 @@ def realize_reduction(kernel):
         for iname in ilp_inames:
             bounds = kernel.get_iname_bounds(iname)
 
+            from loopy.symbolic import pw_aff_to_expr
             ilp_iname_lengths.append(
-                static_max_of_pw_aff(bounds.size, constants_only=True))
+                    int(pw_aff_to_expr(
+                        static_max_of_pw_aff(bounds.size, constants_only=True))))
 
         # }}}
 
diff --git a/loopy/symbolic.py b/loopy/symbolic.py
index 83fd8bde3df74a8a81ecedc130d2c80450db09b4..a5bc8ea1200f14f41d9df9cb187577126b322de9 100644
--- a/loopy/symbolic.py
+++ b/loopy/symbolic.py
@@ -258,7 +258,7 @@ class ArrayAccessFinder(CombineMapper):
 # {{{ C code mapper
 
 class LoopyCCodeMapper(CCodeMapper):
-    def __init__(self, kernel, cse_name_list=[], var_subst_map={}):
+    def __init__(self, kernel, cse_name_list=[], var_subst_map={}, with_annotation=True):
         def constant_mapper(c):
             if isinstance(c, float):
                 # FIXME: type-variable
@@ -270,6 +270,7 @@ class LoopyCCodeMapper(CCodeMapper):
                 cse_name_list=cse_name_list)
         self.kernel = kernel
 
+        self.with_annotation = with_annotation
         self.var_subst_map = var_subst_map.copy()
 
     def copy(self, var_subst_map=None, cse_name_list=None):
@@ -292,8 +293,12 @@ class LoopyCCodeMapper(CCodeMapper):
 
     def map_variable(self, expr, prec):
         if expr.name in self.var_subst_map:
-            return " /* %s */ %s" % (
-                    expr.name, self.rec(self.var_subst_map[expr.name], prec))
+            if self.with_annotation:
+                return " /* %s */ %s" % (
+                        expr.name,
+                        self.rec(self.var_subst_map[expr.name], prec))
+            else:
+                return str(self.rec(self.var_subst_map[expr.name], prec))
         else:
             return CCodeMapper.map_variable(self, expr, prec)
 
diff --git a/test/test_linalg.py b/test/test_linalg.py
index 6a968648cbf9fa807bd298ab515a03cff953ecb9..6fd15542d277fb79c309f77f8f1629e2a240f6d9 100644
--- a/test/test_linalg.py
+++ b/test/test_linalg.py
@@ -59,7 +59,7 @@ def check_error(refsol, sol):
     if rel_err > 1e-5 or np.isinf(rel_err) or np.isnan(rel_err):
         if 1:
             import matplotlib.pyplot as pt
-            pt.imshow(refsol-sol)
+            pt.imshow(refsol-sol, interpolation="nearest")
             pt.colorbar()
             pt.show()
         elif 0:
@@ -623,7 +623,7 @@ def test_image_matrix_mul_ilp(ctx_factory):
     queue = cl.CommandQueue(ctx,
             properties=cl.command_queue_properties.PROFILING_ENABLE)
 
-    n = 2*get_suitable_size(ctx)
+    n = 32
 
     knl = lp.make_kernel(ctx.devices[0],
             "{[i,j,k]: 0<=i,j,k<%d}" % n,
@@ -639,15 +639,18 @@ def test_image_matrix_mul_ilp(ctx_factory):
 
     ilp = 4
     knl = lp.split_dimension(knl, "i", 2, outer_tag="g.0", inner_tag="l.1")
-    j_inner_split = 16
+    j_inner_split = 2
     knl = lp.split_dimension(knl, "j", ilp*j_inner_split, outer_tag="g.1")
     knl = lp.split_dimension(knl, "j_inner", j_inner_split, outer_tag="ilp", inner_tag="l.0")
     knl = lp.split_dimension(knl, "k", 2)
     # conflict-free
-    knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
+    #knl = lp.add_prefetch(knl, 'a', ["i_inner", "k_inner"])
     knl = lp.add_prefetch(knl, 'b', ["j_inner_outer", "j_inner_inner", "k_inner"],
             ["b_j_io", "b_j_ii", "b_k_i"])
-    knl = lp.join_dimensions(knl, ["b_j_io", "b_j_ii"])
+    if 1:
+        knl = lp.join_dimensions(knl, ["b_j_io", "b_j_ii"])
+    else:
+        knl = lp.tag_dimensions(knl, {"b_j_io": "unr"}, force=True)
 
     #print lp.preprocess_kernel(knl)
     #1/0