diff --git a/driver/sw/rtlsim/Makefile b/driver/sw/rtlsim/Makefile
index ca3991b4..c64f1b12 100644
--- a/driver/sw/rtlsim/Makefile
+++ b/driver/sw/rtlsim/Makefile
@@ -1,5 +1,5 @@
-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
 
@@ -23,10 +23,9 @@ SRCS = vortex.cpp ../vx_utils.cpp ../../../rtl/simulate/$(RTL_TOP).cpp
 
 RTL_INCLUDE = -I../../../rtl -I../../../rtl/interfaces -I../../../rtl/cache -I../../../rtl/VX_cache -I../../../rtl/shared_memory -I../../../rtl/pipe_regs -I../../../rtl/compat
 
-THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
-VL_FLAGS += --threads $(THREADS)
-
-VL_FLAGS += -Wno-UNOPTFLAT -Wno-WIDTH 
+# Enable Verilator multithreaded simulation
+#THREADS ?= $(shell python3 -c 'import multiprocessing as mp; print(max(1, mp.cpu_count() // 2))')
+#VL_FLAGS += --threads $(THREADS)
 
 VL_FLAGS += -Wno-UNDRIVEN --Wno-PINMISSING -Wno-STMTDLY -Wno-WIDTH -Wno-UNSIGNED -Wno-UNOPTFLAT -Wno-LITENDIAN
 
diff --git a/driver/tests/demo/Makefile b/driver/tests/demo/Makefile
index 9d3049f9..48effb87 100644
--- a/driver/tests/demo/Makefile
+++ b/driver/tests/demo/Makefile
@@ -43,16 +43,16 @@ $(PROJECT): $(SRCS)
 	$(CXX) $(CXXFLAGS) $^ $(LDFLAGS) -L../../sw/simx -lvortex -o $@
 
 run-fpga: $(PROJECT)
-	LD_LIBRARY_PATH=../../sw/opae:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
+	LD_LIBRARY_PATH=../../sw/opae:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
 
 run-ase: $(PROJECT)
-	LD_LIBRARY_PATH=../../sw/opae/ase:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
+	LD_LIBRARY_PATH=../../sw/opae/ase:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
 
 run-rtlsim: $(PROJECT)
-	LD_LIBRARY_PATH=../../sw/rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
+	LD_LIBRARY_PATH=../../sw/rtlsim:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
 	
 run-simx: $(PROJECT)
-	LD_LIBRARY_PATH=../../sw/simx:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin
+	LD_LIBRARY_PATH=../../sw/simx:$(LD_LIBRARY_PATH) ./$(PROJECT) -f kernel.bin -n 16
 
 .depend: $(SRCS)
 	$(CXX) $(CXXFLAGS) -MM $^ > .depend;
diff --git a/driver/tests/demo/demo b/driver/tests/demo/demo
index 58215168..72482b23 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 9e486d82..386b8ff4 100644
--- a/driver/tests/demo/demo.cpp
+++ b/driver/tests/demo/demo.cpp
@@ -178,6 +178,66 @@ int main(int argc, char *argv[]) {
     return -1;  
   }
 
+  // flush the destination buffer caches
+  std::cout << "flush the destination buffer caches" << std::endl;
+  ret = vx_flush_caches(device, kernel_arg.dst_ptr, buf_size);
+  if (ret != 0) {
+    cleanup();
+    return -1;  
+  }
+
+  // download destination buffer
+  std::cout << "download destination buffer" << std::endl;
+  ret = vx_copy_from_dev(buffer, kernel_arg.dst_ptr, buf_size, 0);
+  if (ret != 0) {
+    cleanup();
+    return -1;  
+  }
+
+  // verify result
+  std::cout << "verify result" << std::endl;  
+  {
+    auto buf_ptr = (int*)vx_host_ptr(buffer);
+    for (uint32_t i = 0; i < num_points; ++i) {
+      int ref = i * i; 
+      int cur = buf_ptr[i];
+      if (cur != ref) {
+        ++errors;
+      }
+    }
+  }
+
+  if (errors != 0) {
+    printf("Found %d errors!\n", errors);
+    printf("FAILED!\n");
+    cleanup();
+    return -1;  
+  }
+
+  // start device
+  std::cout << "start device" << std::endl;
+  ret = vx_start(device);
+  if (ret != 0) {
+    cleanup();
+    return -1;  
+  }
+
+  // wait for completion
+  std::cout << "wait for completion" << std::endl;
+  ret = vx_ready_wait(device, -1);
+  if (ret != 0) {
+    cleanup();
+    return -1;  
+  }
+
+  // flush the destination buffer caches
+  std::cout << "flush the destination buffer caches" << std::endl;
+  ret = vx_flush_caches(device, kernel_arg.dst_ptr, buf_size);
+  if (ret != 0) {
+    cleanup();
+    return -1;  
+  }
+
   // download destination buffer
   std::cout << "download destination buffer" << std::endl;
   ret = vx_copy_from_dev(buffer, kernel_arg.dst_ptr, buf_size, 0);
diff --git a/driver/tests/demo/kernel.bin b/driver/tests/demo/kernel.bin
index 80593892..2e2205bb 100755
Binary files a/driver/tests/demo/kernel.bin and b/driver/tests/demo/kernel.bin differ
diff --git a/rtl/simulate/Vortex.cpp b/rtl/simulate/Vortex.cpp
index 050b4fd1..7de2504e 100644
--- a/rtl/simulate/Vortex.cpp
+++ b/rtl/simulate/Vortex.cpp
@@ -257,7 +257,7 @@ bool Vortex::is_busy() {
 
 void Vortex::send_snoops(uint32_t mem_addr, uint32_t size) {
   // align address to LLC block boundaries
-  auto aligned_addr_start = GLOBAL_BLOCK_SIZE_BYTES * ((mem_addr + GLOBAL_BLOCK_SIZE_BYTES - 1) / GLOBAL_BLOCK_SIZE_BYTES);
+  auto aligned_addr_start = GLOBAL_BLOCK_SIZE_BYTES * (mem_addr / GLOBAL_BLOCK_SIZE_BYTES);
   auto aligned_addr_end = GLOBAL_BLOCK_SIZE_BYTES * ((mem_addr + size + GLOBAL_BLOCK_SIZE_BYTES - 1) / GLOBAL_BLOCK_SIZE_BYTES);
 
   // submit snoop requests for the needed blocks
diff --git a/rtl/simulate/Vortex_SOC.cpp b/rtl/simulate/Vortex_SOC.cpp
index d12c2d40..8bfc94be 100644
--- a/rtl/simulate/Vortex_SOC.cpp
+++ b/rtl/simulate/Vortex_SOC.cpp
@@ -195,7 +195,7 @@ bool Vortex_SOC::is_busy() {
 
 void Vortex_SOC::send_snoops(uint32_t mem_addr, uint32_t size) {
   // align address to LLC block boundaries
-  auto aligned_addr_start = GLOBAL_BLOCK_SIZE_BYTES * ((mem_addr + GLOBAL_BLOCK_SIZE_BYTES - 1) / GLOBAL_BLOCK_SIZE_BYTES);
+  auto aligned_addr_start = GLOBAL_BLOCK_SIZE_BYTES * (mem_addr / GLOBAL_BLOCK_SIZE_BYTES);
   auto aligned_addr_end = GLOBAL_BLOCK_SIZE_BYTES * ((mem_addr + size + GLOBAL_BLOCK_SIZE_BYTES - 1) / GLOBAL_BLOCK_SIZE_BYTES);
 
   // submit snoop requests for the needed blocks