3 files changed, 61 insertions, 9 deletions
diff --git a/MLP/baselines/cpu/Makefile b/MLP/baselines/cpu/Makefile
index 581897e..e2e6780 100644
--- a/MLP/baselines/cpu/Makefile
+++ b/MLP/baselines/cpu/Makefile
@@ -1,4 +1,21 @@
-all:
-	gcc mlp_openmp.c -o mlp_openmp -fopenmp -std=c99
-run:
+all: mlp_openmp
+
+mlp_openmp: mlp_openmp.c
+	gcc -O2 mlp_openmp.c -o mlp_openmp -fopenmp -std=c99
+
+mlp_openmp_O0: mlp_openmp.c
+	gcc mlp_openmp.c -o mlp_openmp_O0 -fopenmp -std=c99
+
+mlp_openmp_O2: mlp_openmp.c
+	gcc -O2 mlp_openmp.c -o mlp_openmp_O2 -fopenmp -std=c99
+
+run: mlp_openmp
 	./mlp_openmp
+
+run_O0: mlp_openmp_O0
+	./mlp_openmp_O0
+
+run_O2: mlp_openmp_O2
+	./mlp_openmp_O2
+
+.PHONY: all run run_O0 run_O2
diff --git a/MLP/baselines/cpu/mlp_openmp.c b/MLP/baselines/cpu/mlp_openmp.c
index ef478c1..8f95e7c 100644
--- a/MLP/baselines/cpu/mlp_openmp.c
+++ b/MLP/baselines/cpu/mlp_openmp.c
@@ -14,6 +14,9 @@
 #include "../../support/timer.h"
 #include "../../support/common.h"
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 T** A;
 T* B;
 T* C;
@@ -136,12 +139,31 @@ void usage() {
     B = malloc(m_size*sizeof(unsigned int));
     C = malloc(m_size*sizeof(unsigned int));
 
-    // Create an input file with arbitrary data.
-    init_data(A, B, m_size, n_size);
-
-    start(&timer, 0, 1);
-    mlp_host(C, A, B, n_size, m_size);
-    stop(&timer, 0);
+    for (int i = 0; i < 100; i++) {
+        // Create an input file with arbitrary data.
+        init_data(A, B, m_size, n_size);
+
+        start(&timer, 0, 0);
+        mlp_host(C, A, B, n_size, m_size);
+        stop(&timer, 0);
+
+        unsigned int nr_threads = 0;
+#pragma omp parallel
+#pragma omp atomic
+        nr_threads++;
+
+        printf("[::] n_threads=%d e_type=%s n_elements=%lu "
+            "| throughput_cpu_omp_MBps=%f\n",
+            nr_threads, XSTR(T), n_size * m_size,
+            n_size * m_size * sizeof(T) / timer.time[0]);
+        printf("[::] n_threads=%d e_type=%s n_elements=%lu "
+            "| throughput_cpu_omp_MOpps=%f\n",
+            nr_threads, XSTR(T), n_size * m_size,
+            n_size * m_size / timer.time[0]);
+        printf("[::] n_threads=%d e_type=%s n_elements=%lu |",
+            nr_threads, XSTR(T), n_size * m_size);
+        printall(&timer, 0);
+    }
 
     uint32_t sum = mlp_host_sum(n_size, m_size);
    
diff --git a/MLP/baselines/cpu/run.sh b/MLP/baselines/cpu/run.sh
new file mode 100755
index 0000000..d0dcb25
--- /dev/null
+++ b/MLP/baselines/cpu/run.sh
@@ -0,0 +1,13 @@
+#!/bin/sh
+
+set -e
+
+echo "prim-benchmarks MLP CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+make
+
+for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+	OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./mlp_openmp || true
+done