port UNI to dfatool

4 files changed, 80 insertions, 14 deletions

diff --git a/UNI/baselines/cpu/Makefile b/UNI/baselines/cpu/Makefile
index a1b4766..10e4f9b 100644
--- a/UNI/baselines/cpu/Makefile
+++ b/UNI/baselines/cpu/Makefile
@@ -1,6 +1,29 @@
-all:
-	gcc -o uni -fopenmp app_baseline.c 
+.PHONY: all
+all: uni
 
-clean:
-	rm uni
+TYPE ?= int64_t
+
+uni: app_baseline.c
+	gcc -O2 -o uni -fopenmp -DT=${TYPE} app_baseline.c
+
+uni_O0: app_baseline.c
+	gcc -o uni_O0 -fopenmp app_baseline.c
+
+uni_O2: app_baseline.c
+	gcc -O2 -o uni_O2 -fopenmp app_baseline.c
 
+.PHONY: run
+run: uni
+	./uni -i 1258291200 -t 4
+
+.PHONY: run_O0
+run_O0: uni_O0
+	./uni_O0 -i 1258291200 -t 4
+
+.PHONY: run_O2
+run_O2: uni_O2
+	./uni_O2 -i 1258291200 -t 4
+
+.PHONY: clean
+clean:
+	rm -f uni uni_O0 uni_O2
diff --git a/UNI/baselines/cpu/app_baseline.c b/UNI/baselines/cpu/app_baseline.c
index 9d3184c..39f225a 100644
--- a/UNI/baselines/cpu/app_baseline.c
+++ b/UNI/baselines/cpu/app_baseline.c
@@ -10,7 +10,12 @@
 #include <omp.h>
 #include "../../support/timer.h"
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
+#ifndef T
 #define T int64_t
+#endif
 
 static int pos;
 
@@ -124,17 +129,31 @@ int main(int argc, char **argv) {
     create_test_file(file_size);
 
     Timer timer;
-    start(&timer, 0, 0);
-
-    total_count = unique_host(file_size, p.n_threads);
-
-    stop(&timer, 0);
 
-    printf("Total count = %d\t", total_count);
-
-    printf("Kernel ");
-    print(&timer, 0, 1);
-    printf("\n");
+    for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+        start(&timer, 0, 0);
+        total_count = unique_host(file_size, p.n_threads);
+        stop(&timer, 0);
+
+        unsigned int nr_threads = 0;
+#pragma omp parallel
+#pragma omp atomic
+        nr_threads++;
+
+        if (rep >= p.n_warmup) {
+            printf("[::] n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_cpu_MBps=%f\n",
+                nr_threads, XSTR(T), file_size,
+                file_size * 2 * sizeof(T) / timer.time[0]);
+            printf("[::] n_threads=%d e_type=%s n_elements=%d "
+                "| throughput_cpu_MOpps=%f\n",
+                nr_threads, XSTR(T), file_size,
+                file_size / timer.time[0]);
+            printf("[::] n_threads=%d e_type=%s n_elements=%d | ",
+                nr_threads, XSTR(T), file_size);
+            printall(&timer, 0);
+        }
+    }
 
     free(A);
     free(B);
diff --git a/UNI/baselines/cpu/run.sh b/UNI/baselines/cpu/run.sh
new file mode 100755
index 0000000..3c99ad2
--- /dev/null
+++ b/UNI/baselines/cpu/run.sh
@@ -0,0 +1,17 @@
+#!/bin/sh
+
+set -e
+
+echo "prim-benchmarks UNI CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+	for i in 19660800 1258291200 629145600 314572800 157286400 78643200 39321600; do
+		for dt in int8_t int16_t int32_t int64_t float double; do
+			if make -B TYPE=${dt}; then
+				timeout -k 1m 30m ./uni -i ${i} -w 0 -e 100 -t ${nr_threads} || true
+			fi
+		done
+	done
+done
diff --git a/UNI/support/timer.h b/UNI/support/timer.h
index b53d95f..4d597b9 100755
--- a/UNI/support/timer.h
+++ b/UNI/support/timer.h
@@ -57,3 +57,10 @@ void stop(Timer *timer, int i) {
 }

 

 void print(Timer *timer, int i, int REP) { printf("Time (ms): %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");

+}