diff --git a/loopy/compiled.py b/loopy/compiled.py
index c8b733d3d5e460618807999ee6c64f6f36309d5e..c8337f8b705b226a35fb026f896d60a53f0b548a 100644
--- a/loopy/compiled.py
+++ b/loopy/compiled.py
@@ -475,7 +475,7 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
         print_ref_code=False, print_code=True, warmup_rounds=2,
         edit_code=False, dump_binary=False, codegen_kwargs={},
         options=[],
-        fills_entire_output=True, check_result=None):
+        fills_entire_output=True, do_check=True, check_result=None):
     """Compare results of `ref_knl` to the kernels generated by the generator
     `kernel_gen`.
 
@@ -505,7 +505,7 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
         fill_value_ref = -17
         fill_value = fill_value_ref
 
-    # {{{ set up CL context for reference run
+    # {{{ find candidate devices for reference run
 
     all_devs = []
     cpu_devs = []
@@ -548,6 +548,17 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
             ref_sched_kernel = knl
             break
 
+        try:
+            ref_args, arg_descriptors = \
+                    make_ref_args(ref_sched_kernel, ref_queue, parameters,
+                            fill_value=fill_value_ref)
+        except cl.RuntimeError, e:
+            if e.code == cl.status_code.IMAGE_FORMAT_NOT_SUPPORTED:
+                continue
+
+        if not do_check:
+            break
+
         ref_compiled = CompiledKernel(ref_ctx, ref_sched_kernel,
                 options=options,
                 codegen_kwargs=codegen_kwargs)
@@ -558,13 +569,6 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
             print get_highlighted_code(ref_compiled.code)
             print 75*"-"
 
-        try:
-            ref_args, arg_descriptors = \
-                    make_ref_args(ref_sched_kernel, ref_queue, parameters,
-                            fill_value=fill_value_ref)
-        except cl.RuntimeError, e:
-            if e.code == cl.status_code.IMAGE_FORMAT_NOT_SUPPORTED:
-                continue
 
         ref_queue.finish()
         ref_start = time()
@@ -586,6 +590,8 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
 
     # {{{ compile and run parallel code
 
+    need_check = do_check
+
     queue = cl.CommandQueue(ctx,
             properties=cl.command_queue_properties.PROFILING_ENABLE)
 
@@ -610,12 +616,10 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
             print compiled.cl_program.binaries[0]
             print 75*"-"
 
-        do_check = True
-
         for i in range(warmup_rounds):
             evt, _ = compiled(queue, **args)
 
-            if do_check:
+            if need_check:
                 for arg_desc in arg_descriptors:
                     if arg_desc is None:
                         continue
@@ -637,7 +641,7 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
 
                     error_is_small, error = check_result(test_ary, ref_ary)
                     assert error_is_small, error
-                    do_check = False
+                    need_check = False
 
         events = []
         queue.finish()
@@ -687,11 +691,12 @@ def auto_test_vs_ref(ref_knl, ctx, kernel_gen, op_count=[], op_label=[], paramet
         print "elapsed: %g s event, %s s other-event %g s wall (%d rounds)%s" % (
                 elapsed, elapsed_evt_2, elapsed_wall, timing_rounds, rates)
 
-        ref_rates = ""
-        for cnt, lbl in zip(op_count, op_label):
-            ref_rates += " %g %s/s" % (cnt/ref_elapsed, lbl)
-        print "ref: elapsed: %g s event, %g s wall%s" % (
-                ref_elapsed, ref_elapsed_wall, ref_rates)
+        if do_check:
+            ref_rates = ""
+            for cnt, lbl in zip(op_count, op_label):
+                ref_rates += " %g %s/s" % (cnt/ref_elapsed, lbl)
+            print "ref: elapsed: %g s event, %g s wall%s" % (
+                    ref_elapsed, ref_elapsed_wall, ref_rates)
 
     # }}}