diff --git a/driver/sw/rtlsim/Makefile b/driver/sw/rtlsim/Makefile
index 1dc5e3f8..020694a2 100644
--- a/driver/sw/rtlsim/Makefile
+++ b/driver/sw/rtlsim/Makefile
@@ -1,5 +1,6 @@
-#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 -O2 -Wall -Wextra -pedantic -Wfatal-errors
+# CFLAGS += -std=c++11 -g -O0 -Wall -Wextra -pedantic -Wfatal-errors
 
 USE_MULTICORE=1
 
@@ -27,7 +28,7 @@ RTL_INCLUDE = -I../../../rtl -I../../../rtl/interfaces -I../../../rtl/cache -I..
 #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
+VL_FLAGS += -Wno-UNDRIVEN --Wno-PINMISSING -Wno-STMTDLY -Wno-WIDTH -Wno-UNSIGNED -Wno-UNOPTFLAT -Wno-LITENDIAN -Wno-BLKLOOPINIT
 
 # Debugigng
 #VL_FLAGS += --trace -DVL_DEBUG=1 
@@ -46,4 +47,4 @@ $(PROJECT): $(SRCS) build_config
 	make -j -C obj_dir -f V$(RTL_TOP).mk
 
 clean:
-	rm -rf $(PROJECT) obj_dir
\ No newline at end of file
+	rm -rf $(PROJECT) obj_dir
diff --git a/rtl/simulate/simulator.cpp b/rtl/simulate/simulator.cpp
index 4d7e4d8a..58070712 100644
--- a/rtl/simulate/simulator.cpp
+++ b/rtl/simulate/simulator.cpp
@@ -351,6 +351,8 @@ void Simulator::send_snoops(uint32_t mem_addr, uint32_t size) {
 }
 
 void Simulator::flush_caches(uint32_t mem_addr, uint32_t size) {
+  printf("[sim] total cycles: %lld\n", this->total_cycles_);
+
   // send snoops for L1 flush
   this->send_snoops(mem_addr, size);
   this->wait(PIPELINE_FLUSH_LATENCY);
@@ -383,4 +385,4 @@ bool Simulator::run() {
 #endif
 
   return (status == 1);
-}
\ No newline at end of file
+}
diff --git a/run_tests.sh b/run_tests.sh
new file mode 100755
index 00000000..9d249146
--- /dev/null
+++ b/run_tests.sh
@@ -0,0 +1,51 @@
+#!/bin/bash
+set -e
+
+mkdir -p test_outputs
+
+output_dir="$(pwd)/test_outputs"
+
+(cd rtl ; python3 gen_synth_configs.py ; ls -l configs)
+
+config_location=rtl/configs
+
+declare -a test_names=("sgemm" "saxpy" "bfs" "guassian" "vecadd" "nearn" "sfilter")
+
+for test_name in ${test_names[@]}; do
+ 	if [ ! -d "benchmarks/new_opencl/$test_name" ]; then
+		echo "Unknown benchmark $test_name"
+		exit 1
+	fi
+done
+
+
+for filename in "$config_location"/*.sh; do
+
+name=${filename##*/}
+base=${name%.*}
+
+. "$filename"
+
+make -C rtl build_config
+make -C runtime build_config
+make -C driver/sw/rtlsim
+
+for test_name in ${test_names[@]}; do
+
+(
+
+echo "Running $base-$test_name..."
+
+cd "benchmarks/new_opencl/$test_name"
+make clean
+make
+make run-rtlsim 2>&1 | tee "$output_dir/$base-$test_name.log"
+) &
+
+done # test_name
+
+wait
+
+done # config
+
+