3 files changed, 39 insertions, 11 deletions
diff --git a/TRNS/baselines/cpu/Makefile b/TRNS/baselines/cpu/Makefile
index 236f7bb..2f28738 100644
--- a/TRNS/baselines/cpu/Makefile
+++ b/TRNS/baselines/cpu/Makefile
@@ -32,16 +32,30 @@
 #  THE SOFTWARE.
 # 
 
-NUMA ?= 0
-NUMA_MEMCPY ?= 0
-FLAGS =
+benchmark ?= 1
+debug ?= 0
+native ?= 1
+nop_sync ?= 0
+numa ?= 0
+numa_memcpy ?= 0
 
-ifeq (${NUMA}, 1)
-	FLAGS += -lnuma
+CFLAGS =
+LDFLAGS =
+
+ifeq (${debug}, 1)
+	CFLAGS += -g
+endif
+
+ifeq (${native}, 1)
+	CFLAGS += -march=native
+endif
+
+ifeq (${numa}, 1)
+	LDFLAGS += -lnuma
 endif
 
 CXX=g++
-CXX_FLAGS=-std=c++11 -Wall -Wextra -pedantic -DNUMA=${NUMA} -DNUMA_MEMCPY=${NUMA_MEMCPY}
+CXX_FLAGS=-std=c++11 -Wall -Wextra -pedantic -DNUMA=${numa} -DNUMA_MEMCPY=${numa_memcpy} -DNOP_SYNC=${nop_sync} -DWITH_BENCHMARK=${benchmark}
 
 LIB=-L/usr/lib/ -lm -pthread
 
@@ -52,7 +66,7 @@ EXE=trns
 all: trns
 
 trns: ${SRC}
-	$(CXX) -O2 $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE) $(FLAGS)
+	$(CXX) -O3 $(CXX_FLAGS) ${CFLAGS} $(SRC) $(LIB) -o $(EXE) ${LDFLAGS}
 
 trns_O0: ${SRC}
 	$(CXX) $(CXX_FLAGS) $(SRC) $(LIB) -o $(EXE)_O0
diff --git a/TRNS/baselines/cpu/main.cpp b/TRNS/baselines/cpu/main.cpp
index c8cccaf..b4cd149 100644
--- a/TRNS/baselines/cpu/main.cpp
+++ b/TRNS/baselines/cpu/main.cpp
@@ -36,9 +36,18 @@
 #include "support/setup.h"
 #include "kernel.h"
 #include "support/common.h"
-#include "support/timer.h"
 #include "support/verify.h"
 
+#if WITH_BENCHMARK
+#include "support/timer.h"
+#else
+#include <string>
+struct Timer {
+    inline void start(std::string name) {(void)name;}
+    inline void stop(std::string name) {(void)name;}
+};
+#endif
+
 #include <unistd.h>
 #include <thread>
 #include <string.h>
@@ -362,6 +371,7 @@ int main(int argc, char **argv) {
             timer.stop("free");
 #endif
 
+#if WITH_BENCHMARK
         if (rep >= p.n_warmup) {
 #if NUMA_MEMCPY
             printf("[::] TRNS-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%d"
@@ -396,10 +406,8 @@ int main(int argc, char **argv) {
                 timer.get("Step 1") + timer.get("Step 2") + timer.get("Step 3"));
 #endif // NUMA_MEMCPY
         }
+#endif // WITH_BENCHMARK
     }
-    //timer.print("Step 1", p.n_reps);
-    //timer.print("Step 2", p.n_reps);
-    //timer.print("Step 3", p.n_reps);
 
     // Verify answer
     //verify(h_local, h_in_backup, M_ * m, N_ * n, 1);
diff --git a/TRNS/baselines/cpu/run-perf.sh b/TRNS/baselines/cpu/run-perf.sh
new file mode 100755
index 0000000..f16a3b1
--- /dev/null
+++ b/TRNS/baselines/cpu/run-perf.sh
@@ -0,0 +1,6 @@
+#!/bin/zsh
+
+make -B numa=1
+
+perf stat record -o t1.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 1 -a 4 -c 4
+perf stat record -o t4.perf -e ${(j:,:):-$(grep -v '^#' ../../../perf-events.txt | cut -d ' ' -f 1)} ./trns -w 0 -r 20 -p 2048 -o 2048 -m 16 -n 8 -t 4 -a 4 -c 4