port BS CPU to dfatool

4 files changed, 63 insertions, 6 deletions

diff --git a/BS/baselines/cpu/Makefile b/BS/baselines/cpu/Makefile
index e907fc8..1b0ceb5 100644
--- a/BS/baselines/cpu/Makefile
+++ b/BS/baselines/cpu/Makefile
@@ -1,4 +1,25 @@
-all:
-	gcc bs_omp.c -o bs_omp -fopenmp
-run:
+.PHONY: all
+all: bs_omp
+
+bs_omp: bs_omp.c
+	gcc -O2 bs_omp.c -o bs_omp -fopenmp
+
+bs_omp_O0: bs_omp.c
+	gcc bs_omp.c -o bs_omp_O0 -fopenmp
+
+bs_omp_O2: bs_omp.c
+	gcc -O2 bs_omp.c -o bs_omp_O2 -fopenmp
+
+.PHONY: run run_O0 run_O2
+run: bs_omp
 	./bs_omp 262144 16777216
+
+run_O0: bs_omp_O0
+	./bs_omp_O0 262144 16777216
+
+run_O2: bs_omp_O2
+	./bs_omp_O2 262144 16777216
+
+.PHONY: clean
+clean:
+	rm -f bs_omp bs_omp_O0 bs_omp_O2
diff --git a/BS/baselines/cpu/bs_omp.c b/BS/baselines/cpu/bs_omp.c
index 3ad83ed..b8b3023 100644
--- a/BS/baselines/cpu/bs_omp.c
+++ b/BS/baselines/cpu/bs_omp.c
@@ -93,12 +93,24 @@ uint64_t binarySearch(DTYPE * input, uint64_t input_size, DTYPE* querys, unsigne
     result_host = binarySearch(input, input_size - 1, querys, n_querys);   
     stop(&timer, 0);
 
+    unsigned int nr_threads = 0;
+#pragma omp parallel
+#pragma omp atomic
+    nr_threads++;
 
     int status = (result_host);
     if (status) {
-        printf("[OK] Execution time: ");
-	print(&timer, 0, 1);
-	printf("ms.\n");
+        printf("[::] n_threads=%d e_type=%s n_elements=%d "
+            "| throughput_cpu_omp_MBps=%f\n",
+            nr_threads, "uint64_t", input_size,
+            n_querys * sizeof(DTYPE) / timer.time[0]);
+        printf("[::] n_threads=%d e_type=%s n_elements=%d "
+            "| throughput_cpu_omp_MOpps=%f\n",
+            nr_threads, "uint64_t", input_size,
+            n_querys / timer.time[0]);
+        printf("[::] n_threads=%d e_type=%s n_elements=%d |",
+            nr_threads, "uint64_t", input_size);
+        printall(&timer, 0);
     } else {
         printf("[ERROR]\n");
     }
diff --git a/BS/baselines/cpu/run.sh b/BS/baselines/cpu/run.sh
new file mode 100755
index 0000000..80c596c
--- /dev/null
+++ b/BS/baselines/cpu/run.sh
@@ -0,0 +1,17 @@
+#!/bin/sh
+
+set -e
+
+echo "prim-benchmarks BS CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+make
+
+for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+	for vs in 262144 524288 1048576 2097152; do
+		for i in `seq 1 20`; do
+			OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./bs_omp ${vs} 16777216 || true
+		done
+	done
+done
diff --git a/BS/baselines/cpu/timer.h b/BS/baselines/cpu/timer.h
index 969ef97..08027af 100755
--- a/BS/baselines/cpu/timer.h
+++ b/BS/baselines/cpu/timer.h
@@ -57,3 +57,10 @@ void stop(Timer *timer, int i) {
 }

 

 void print(Timer *timer, int i, int REP) { printf("%f\t", timer->time[i] / (1000 * REP)); }

+

+void printall(Timer *timer, int maxt) {

+    for (int i = 0; i <= maxt; i++) {

+        printf(" timer%d_us=%f", i, timer->time[i]);

+    }

+    printf("\n");

+}