3 files changed, 45 insertions, 4 deletions
diff --git a/TS/baselines/cpu/Makefile b/TS/baselines/cpu/Makefile
index 80729b0..69121ed 100644
--- a/TS/baselines/cpu/Makefile
+++ b/TS/baselines/cpu/Makefile
@@ -1,4 +1,22 @@
-all:
-	g++ streamp_openmp.cpp tools.cpp -o streamp_openmp -std=c++11 -fopenmp
-run:
-	./streamp_openmp SampleInput/randomlist5M.txt 256
+
+all: streamp_openmp
+
+streamp_openmp: streamp_openmp.cpp tools.cpp
+	g++ -O2 streamp_openmp.cpp tools.cpp -o streamp_openmp -std=c++11 -fopenmp
+
+streamp_openmp_O0: streamp_openmp.cpp tools.cpp
+	g++ streamp_openmp.cpp tools.cpp -o streamp_openmp_O0 -std=c++11 -fopenmp
+
+streamp_openmp_O2: streamp_openmp.cpp tools.cpp
+	g++ -O2 streamp_openmp.cpp tools.cpp -o streamp_openmp_O2 -std=c++11 -fopenmp
+
+run: streamp_openmp
+	./streamp_openmp inputs/randomlist33M.txt 256
+
+run_O0: streamp_openmp_O0
+	./streamp_openmp_O0 inputs/randomlist33M.txt 256
+
+run_O2: streamp_openmp_O2
+	./streamp_openmp_O2 inputs/randomlist33M.txt 256
+
+.PHONY: all run run_O0 run_O2 clean
diff --git a/TS/baselines/cpu/run.sh b/TS/baselines/cpu/run.sh
new file mode 100755
index 0000000..7d8fe37
--- /dev/null
+++ b/TS/baselines/cpu/run.sh
@@ -0,0 +1,16 @@
+#!/bin/sh
+
+set -e
+
+echo "prim-benchmarks TS CPU (dfatool edition)"
+echo "Started at $(date)"
+echo "Revision $(git describe --always)"
+
+# input size depends on file -> strong scaling only
+
+make
+for i in $(seq 1 10); do
+	for nr_threads in 1 2 4 6 8 12 16 20 24 32; do
+		OMP_NUM_THREADS=${nr_threads} timeout --foreground -k 1m 30m ./streamp_openmp inputs/randomlist33M.txt 256 || true
+	done
+done
diff --git a/TS/baselines/cpu/streamp_openmp.cpp b/TS/baselines/cpu/streamp_openmp.cpp
index 1925e52..94f110f 100644
--- a/TS/baselines/cpu/streamp_openmp.cpp
+++ b/TS/baselines/cpu/streamp_openmp.cpp
@@ -39,6 +39,9 @@ The second column of the output file is the matrix profile index.
 #include <chrono>
 #include <omp.h>
 
+#define XSTR(x) STR(x)
+#define STR(x) #x
+
 #include "mprofile.h"
 
 bool interrupt = false;
@@ -335,6 +338,7 @@ int main(int argc, char* argv[])
   tend = std::chrono::high_resolution_clock::now();
   time_elapsed = tend - tstart;
   std::cout << "[OK] Preprocess Time:         " << std::setprecision(std::numeric_limits<double>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;
+  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_preproc_MBps=%f throughput_preproc_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));
 
   //Initialize Matrix Profile and Matrix Profile Index
   std::cout << "[>>] Initializing Profile..." << std::endl;
@@ -351,6 +355,7 @@ int main(int argc, char* argv[])
   tend = std::chrono::high_resolution_clock::now();
   time_elapsed = tend - tstart;
   std::cout << "[OK] Initialize Profile Time: " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;
+  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_init_MBps=%f throughput_init_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));
 
   // Random shuffle the diagonals
   idx.clear();
@@ -369,6 +374,7 @@ int main(int argc, char* argv[])
   tend = std::chrono::high_resolution_clock::now();
   time_elapsed = tend - tstart;
   std::cout << "[OK] STREAMP Time:            " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;
+  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_streamp_MBps=%f throughput_streamp_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));
 
   // Save profile to file
   //std::cout << "[>>] Saving Profile..." << std::endl;
@@ -383,6 +389,7 @@ int main(int argc, char* argv[])
   // Calculate total time
   time_elapsed = tend - tprogstart;
   std::cout << "[OK] Total Time:              " << std::setprecision(std::numeric_limits<DTYPE>::digits10 + 2) << time_elapsed.count() << " seconds." << std::endl;
+  printf("[::] n_threads=%d e_type=%s n_elements=%d | throughput_total_MBps=%f throughput_total_MOpps=%f\n", numThreads, XSTR(DTYPE), timeSeriesLength, timeSeriesLength * sizeof(DTYPE) / (time_elapsed.count() * 1e6), timeSeriesLength / (time_elapsed.count() * 1e6));
   std::cout << std::endl;
 
   delete profile;