diff --git a/driver/sw/rtlsim/Makefile b/driver/sw/rtlsim/Makefile
index 5b95c7cb..ca3991b4 100644
--- a/driver/sw/rtlsim/Makefile
+++ b/driver/sw/rtlsim/Makefile
@@ -1,7 +1,7 @@
-#CFLAGS += -std=c++11 -O3 -Wall -Wextra -pedantic -Wfatal-errors
-CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -pedantic -Wfatal-errors
+CFLAGS += -std=c++11 -O3 -Wall -Wextra -pedantic -Wfatal-errors
+#CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -pedantic -Wfatal-errors
 
-#USE_MULTICORE=1
+USE_MULTICORE=1
 
 CFLAGS += -I../../include -I../../../../rtl/simulate -I../../../../runtime
 
@@ -15,6 +15,7 @@ ifdef USE_MULTICORE
 	CFLAGS += -DUSE_MULTICORE
 	RTL_TOP = Vortex_SOC
 else
+    VL_FLAGS += -DSINGLE_CORE_BENCH
 	RTL_TOP = Vortex
 endif
 
diff --git a/driver/tests/demo/Makefile b/driver/tests/demo/Makefile
index cc505cfc..9d3049f9 100644
--- a/driver/tests/demo/Makefile
+++ b/driver/tests/demo/Makefile
@@ -50,7 +50,7 @@ run-ase: $(PROJECT)
 
 run-rtlsim: $(PROJECT)
 	LD_LIBRARY_PATH=../../sw/rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
-
+	
 run-simx: $(PROJECT)
 	LD_LIBRARY_PATH=../../sw/simx:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
 
diff --git a/driver/tests/demo/demo b/driver/tests/demo/demo
index a67bb433..58215168 100755
Binary files a/driver/tests/demo/demo and b/driver/tests/demo/demo differ
diff --git a/driver/tests/demo/demo.cpp b/driver/tests/demo/demo.cpp
index aeaaae84..9e486d82 100644
--- a/driver/tests/demo/demo.cpp
+++ b/driver/tests/demo/demo.cpp
@@ -115,7 +115,8 @@ int main(int argc, char *argv[]) {
 
   // allocate shared memory  
   std::cout << "allocate shared memory" << std::endl;    
-  ret = vx_alloc_shared_mem(device, buf_size, &buffer);
+  uint32_t alloc_size = std::max<uint32_t>(buf_size, sizeof(kernel_arg_t));
+  ret = vx_alloc_shared_mem(device, alloc_size, &buffer);
   if (ret != 0) {
     cleanup();
     return -1;  
@@ -205,6 +206,7 @@ int main(int argc, char *argv[]) {
   if (0 == errors) {
     printf("PASSED!\n");
   } else {
+    printf("Found %d errors!\n", errors);
     printf("FAILED!\n");
   }
 
diff --git a/rtl/VX_warp_scheduler.v b/rtl/VX_warp_scheduler.v
index 22ad21da..38885e6e 100644
--- a/rtl/VX_warp_scheduler.v
+++ b/rtl/VX_warp_scheduler.v
@@ -133,6 +133,7 @@ module VX_warp_scheduler (
 			thread_masks[0]       <= 1; // Activating first thread in first warp
 			warp_stalled          <= 0;
 			didnt_split           <= 0;
+			warp_lock             <= 0;      
 			// total_barrier_stall    = 0;
 			for (curr_w_help = 1; curr_w_help < `NW; curr_w_help=curr_w_help+1) begin
 				warp_pcs[curr_w_help]        <= 0;