VA: port to dfatool

4 files changed, 97 insertions, 21 deletions

diff --git a/VA/baselines/cpu/Makefile b/VA/baselines/cpu/Makefile
index f320d87..3171e19 100644
--- a/VA/baselines/cpu/Makefile
+++ b/VA/baselines/cpu/Makefile
@@ -1,5 +1,29 @@
-all:
-	gcc -o va -fopenmp app_baseline.c 
+.PHONY: all
+all: va
 
+TYPE ?= int32_t
+
+va: app_baseline.c
+	gcc -O2 -o va -fopenmp -DT=${TYPE} app_baseline.c
+
+va_O0: app_baseline.c
+	gcc -o va_O0 -fopenmp app_baseline.c
+
+va_O2: app_baseline.c
+	gcc -O2 -o va_O2 -fopenmp app_baseline.c
+
+.PHONY: run
+run: va
+	./va -t 4
+
+.PHONY: run_O0
+run_O0: va_O0
+	./va_O0 -t 4
+
+.PHONY: run_O2
+run_O2: va_O2
+	./va_O2 -t 4
+
+.PHONY: clean
 clean:
-	rm va
+	rm -f va va_O0 va_O2
diff --git a/VA/baselines/cpu/app_baseline.c b/VA/baselines/cpu/app_baseline.c
index ecd8efa..b4396d2 100644
--- a/VA/baselines/cpu/app_baseline.c
+++ b/VA/baselines/cpu/app_baseline.c
@@ -15,10 +15,17 @@
 #include <omp.h>
 #include "../../support/timer.h"
 
-static int32_t *A;
-static int32_t *B;
-static int32_t *C;
-static int32_t *C2; 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
+#ifndef T
+#define T int32_t
+#endif
+
+static T *A;
+static T *B;
+static T *C;
+static T *C2; 
 
 /**
 * @brief creates a "test file" by filling a buffer of 64MB with pseudo-random values
@@ -27,14 +34,13 @@ static int32_t *C2;
 */
 void  *create_test_file(unsigned int nr_elements) {
     srand(0);
-    printf("nr_elements\t%u\t", nr_elements);
-    A = (uint32_t*) malloc(nr_elements * sizeof(uint32_t));
-    B = (uint32_t*) malloc(nr_elements * sizeof(uint32_t));
-    C = (uint32_t*) malloc(nr_elements * sizeof(uint32_t));
+    A = (T*) malloc(nr_elements * sizeof(T));
+    B = (T*) malloc(nr_elements * sizeof(T));
+    C = (T*) malloc(nr_elements * sizeof(T));
     
     for (int i = 0; i < nr_elements; i++) {
-        A[i] = (int) (rand());
-        B[i] = (int) (rand());
+        A[i] = (T) (rand());
+        B[i] = (T) (rand());
     }
 
 }
@@ -115,14 +121,31 @@ int main(int argc, char **argv) {
     create_test_file(file_size);
 
     Timer timer;
-    start(&timer, 0, 0);
-
-    vector_addition_host(file_size, p.n_threads);
-	
-    stop(&timer, 0);
-    printf("Kernel ");
-    print(&timer, 0, 1);
-    printf("\n");
+
+    for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+        start(&timer, 0, 0);
+        vector_addition_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 * 3 * 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/VA/baselines/cpu/run.sh b/VA/baselines/cpu/run.sh
new file mode 100755
index 0000000..dd4acad
--- /dev/null
+++ b/VA/baselines/cpu/run.sh
@@ -0,0 +1,22 @@
+#!/bin/sh
+
+set -e
+
+echo "prim-benchmarks VA CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# default input size: 16777216
+# default threads: 4
+# default type: int32_t
+
+for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+	for i in 16777216 8388608 4194304 33554432 67108864; do
+		for dt in int8_t int16_t int32_t int64_t float double; do
+			if make -B TYPE=${dt}; then
+				# -w 1 to make sure that target array (C) is allocated
+				timeout -k 1m 30m ./va -i ${i} -w 1 -e 100 -t ${nr_threads} || true
+			fi
+		done
+	done
+done
diff --git a/VA/support/timer.h b/VA/support/timer.h
index eedc385..5c00213 100755
--- a/VA/support/timer.h
+++ b/VA/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");

+}