port SpMV to dfatool; add benchmark scripts

4 files changed, 78 insertions, 6 deletions

diff --git a/SpMV/baselines/cpu/Makefile b/SpMV/baselines/cpu/Makefile
index 9c63605..64b20db 100644
--- a/SpMV/baselines/cpu/Makefile
+++ b/SpMV/baselines/cpu/Makefile
@@ -1,7 +1,24 @@
-all:
-		gcc -o spmv -fopenmp app.c 
+all: spmv
 
-clean:
-		rm spmv
+spmv: app.c
+	gcc -O2 -o spmv -fopenmp app.c
+
+spmv_O0: app.c
+	gcc -o spmv_O0 -fopenmp app.c
+
+spmv_O2: app.c
+	gcc -O2 -o spmv_O2 -fopenmp app.c
+
+run: spmv
+	OMP_NUM_THREADS=4 ./spmv -f ../../data/bcsstk30.mtx -v 0
 
+run_O0: spmv_O0
+	OMP_NUM_THREADS=4 ./spmv_O0 -f ../../data/bcsstk30.mtx -v 0
+
+run_O2: spmv_O2
+	OMP_NUM_THREADS=4 ./spmv_O2 -f ../../data/bcsstk30.mtx -v 0
+
+clean:
+	rm -f spmv spmv_O0 spmv_O2
 
+.PHONY: all run run_O0 run_O2 clean
diff --git a/SpMV/baselines/cpu/app.c b/SpMV/baselines/cpu/app.c
index 46db2f0..8d360ee 100644
--- a/SpMV/baselines/cpu/app.c
+++ b/SpMV/baselines/cpu/app.c
@@ -29,7 +29,7 @@ int main(int argc, char** argv) {
 
     // Calculating result on CPU
     PRINT_INFO(p.verbosity >= 1, "Calculating result on CPU");
-    omp_set_num_threads(4);
+    //omp_set_num_threads(4);
     Timer timer;
     startTimer(&timer);
     #pragma omp parallel for
@@ -43,7 +43,22 @@ int main(int argc, char** argv) {
         outVector[rowIdx] = sum;
     }
     stopTimer(&timer);
-    if(p.verbosity == 0) PRINT("%f", getElapsedTime(timer)*1e3);
+
+
+    unsigned int nr_threads = 0;
+#pragma omp parallel
+#pragma omp atomic
+    nr_threads++;
+
+
+    // coomatrix / csrmatrix use uint32_t indexes and float values
+    printf("[::] SpMV CPU | n_threads=%u e_type=float n_elements=%u |"
+        " throughput_MBps=%f throughput_MOpps=%f timer0_us=%f\n",
+        nr_threads, csrMatrix.numNonzeros,
+        csrMatrix.numNonzeros * sizeof(float) / (getElapsedTime(timer)*1e6),
+        csrMatrix.numNonzeros / (getElapsedTime(timer)*1e6),
+        getElapsedTime(timer)*1e6);
+    //if(p.verbosity == 0) PRINT("%f", getElapsedTime(timer)*1e3);
     PRINT_INFO(p.verbosity >= 1, "    Elapsed time: %f ms", getElapsedTime(timer)*1e3);
 
     // Deallocate data structures
diff --git a/SpMV/baselines/cpu/run-opti.sh b/SpMV/baselines/cpu/run-opti.sh
new file mode 100755
index 0000000..62a3e8b
--- /dev/null
+++ b/SpMV/baselines/cpu/run-opti.sh
@@ -0,0 +1,15 @@
+#!/bin/sh
+
+HOST="$(hostname)"
+
+echo $HOST
+
+make clean
+
+for i in $(seq 1 50); do
+	make run_O0 | sed 's/CPU/CPU O0/'
+done | tee "${HOST}-O0.txt"
+
+for i in $(seq 1 50); do
+	make run_O2 | sed 's/CPU/CPU O2/'
+done | tee "${HOST}-O2.txt"
diff --git a/SpMV/baselines/cpu/run.sh b/SpMV/baselines/cpu/run.sh
new file mode 100755
index 0000000..a993cc0
--- /dev/null
+++ b/SpMV/baselines/cpu/run.sh
@@ -0,0 +1,25 @@
+#!/bin/sh
+
+set -e
+
+HOST="$(hostname)"
+
+echo $HOST
+
+(
+
+echo "prim-benchmarks SpMV CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# default threads: 4
+
+# input size depends on file -> strong scaling only
+
+make -B
+for i in $(seq 1 50); do
+	for nr_threads in 88 64 44 1 2 4 6 8 12 16 20 24 32; do
+		OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./spmv -f ../../data/bcsstk30.mtx -v 0 || true
+	done
+done
+) | tee "${HOST}-explore.txt"