fix perf

Author: allnes

scripts/generate_perf_results.sh

@@ -1,3 +1,3 @@
 mkdir build/perf_stat_dir
-source scripts/run_perf_collector.sh &> build/perf_stat_dir/perf_log.txt
+source scripts/run_perf_collector.sh | tee build/perf_stat_dir/perf_log.txt
 python3 scripts/create_perf_table.py --input build/perf_stat_dir/perf_log.txt --output build/perf_stat_dir

tasks/all/example/perf_tests/perf_all.cpp

@@ -6,7 +6,7 @@
 #include "core/perf/include/perf.hpp"
 
 TEST(nesterov_a_test_task_all, test_pipeline_run) {
-  constexpr int kCount = 900;
+  constexpr int kCount = 400;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);
@@ -51,7 +51,7 @@ TEST(nesterov_a_test_task_all, test_pipeline_run) {
 }
 
 TEST(nesterov_a_test_task_all, test_task_run) {
-  constexpr int kCount = 900;
+  constexpr int kCount = 400;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);

tasks/all/example/src/ops_all.cpp

@@ -43,14 +43,7 @@ bool nesterov_a_test_task_all::TestTaskALL::RunImpl() {
 #pragma omp critical
       { MatMul(input_, rc_size_, output_); }
     }
-  } else if (world_.rank() == 1) {
-    const int num_threads = ppc::util::GetPPCNumThreads();
-    std::vector<std::thread> threads(num_threads);
-    for (int i = 0; i < num_threads; i++) {
-      threads[i] = std::thread(MatMul, std::cref(input_), rc_size_, std::ref(output_));
-      threads[i].join();
-    }
-  } else if (world_.rank() == 2) {
+  } else {
     oneapi::tbb::task_arena arena(1);
     arena.execute([&] {
       tbb::task_group tg;
@@ -59,9 +52,15 @@ bool nesterov_a_test_task_all::TestTaskALL::RunImpl() {
       }
       tg.wait();
     });
-  } else {
-    MatMul(input_, rc_size_, output_);
   }
+
+  const int num_threads = ppc::util::GetPPCNumThreads();
+  std::vector<std::thread> threads(num_threads);
+  for (int i = 0; i < num_threads; i++) {
+    threads[i] = std::thread(MatMul, std::cref(input_), rc_size_, std::ref(output_));
+    threads[i].join();
+  }
+
   world_.barrier();
   return true;
 }

tasks/omp/example/perf_tests/main.cpp

@@ -6,7 +6,7 @@
 #include "omp/example/include/ops_omp.hpp"
 
 TEST(nesterov_a_test_task_omp, test_pipeline_run) {
-  constexpr int kCount = 500;
+  constexpr int kCount = 300;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);
@@ -47,7 +47,7 @@ TEST(nesterov_a_test_task_omp, test_pipeline_run) {
 }
 
 TEST(nesterov_a_test_task_omp, test_task_run) {
-  constexpr int kCount = 500;
+  constexpr int kCount = 300;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);

tasks/stl/example/perf_tests/main.cpp

@@ -6,7 +6,7 @@
 #include "stl/example/include/ops_stl.hpp"
 
 TEST(nesterov_a_test_task_stl, test_pipeline_run) {
-  constexpr int kCount = 1000;
+  constexpr int kCount = 700;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);
@@ -47,7 +47,7 @@ TEST(nesterov_a_test_task_stl, test_pipeline_run) {
 }
 
 TEST(nesterov_a_test_task_stl, test_task_run) {
-  constexpr int kCount = 1000;
+  constexpr int kCount = 700;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);

tasks/tbb/example/perf_tests/main.cpp

@@ -6,7 +6,7 @@
 #include "tbb/example/include/ops_tbb.hpp"
 
 TEST(nesterov_a_test_task_tbb, test_pipeline_run) {
-  constexpr int kCount = 500;
+  constexpr int kCount = 700;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);
@@ -47,7 +47,7 @@ TEST(nesterov_a_test_task_tbb, test_pipeline_run) {
 }
 
 TEST(nesterov_a_test_task_tbb, test_task_run) {
-  constexpr int kCount = 500;
+  constexpr int kCount = 700;
 
   // Create data
   std::vector<int> in(kCount * kCount, 0);