diff --git a/example/Performance/memBW/Makefile b/example/Performance/memBW/Makefile
index c1cb1db9782cec726f7088768b662ce8b958b6f3..329fcab0e90a66c497577d5ae2d29dbae2bebb40 100644
--- a/example/Performance/memBW/Makefile
+++ b/example/Performance/memBW/Makefile
@@ -53,10 +53,10 @@ memBW: $(OBJ)
 all: memBW
 
 run: memBW
-	mpirun --oversubscribe -np 2 ./miniBUDE
+	mpirun --oversubscribe -np 2 ./memBW
 
 .PHONY: clean all run
 
 clean:
-	rm -f *.o *~ core miniBUDE
+	rm -f *.o *~ core memBW
 
diff --git a/example/Performance/memBW/main.cu b/example/Performance/memBW/main.cu
index 07c516f7e2b08eaa5035ffae4c4d640a13cf1699..ae78ade4ceba701d5bfd4a8d13b203af1f997bb0 100644
--- a/example/Performance/memBW/main.cu
+++ b/example/Performance/memBW/main.cu
@@ -48,10 +48,10 @@ int main(int argc, char *argv[])
         timer t;
         t.start();
 
-	auto vout = out.toKernel();
-	auto vin = in.toKernel();
+//	auto v_out = out.toKernel();
+//	auto v_in = in.toKernel();
 
-        CUDA_LAUNCH(translate_fill_prop,ite,vout,vin);
+        CUDA_LAUNCH(translate_fill_prop,ite,out.toKernel(),in.toKernel());
 
         cudaDeviceSynchronize();