properly set barrier when using iterrative aggregation

qdp_project/src/Benchmark.cpp

@@ -254,7 +254,12 @@ void aggr_j(size_t gid, size_t tid) {
 
     THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid * gid][0][TIME_STAMP_BEGIN] = std::chrono::steady_clock::now();
 
-    BARRIERS_[gid]->arrive_and_wait();
+    if constexpr (AGGRJ_ITERATIVE) {
+       if (tid == 0) BARRIERS_[gid]->arrive_and_wait();
+    }
+    else {
+        BARRIERS_[gid]->arrive_and_wait();
+    }
 
     THREAD_TIMING_[AGGRJ_TIMING_INDEX][tid * gid][0][TIME_STAMP_WAIT] = std::chrono::steady_clock::now();
 
@@ -341,7 +346,12 @@ int main() {
         std::vector<std::thread>  agg_pool;
 
         for(uint32_t gid = 0; gid < GROUP_COUNT; ++gid) {
-            BARRIERS_.emplace_back(new std::barrier<NopStruct>(TC_COMBINED));
+            if constexpr (AGGRJ_ITERATIVE) {
+                BARRIERS_.emplace_back(new std::barrier<NopStruct>(TC_SCANA + TC_SCANB + 1));
+            }
+            else {
+                BARRIERS_.emplace_back(new std::barrier<NopStruct>(TC_COMBINED));
+            }
 
             for(uint32_t tid = 0; tid < TC_SCANA; ++tid) {
                 filter_pool.emplace_back(scan_a, gid, tid);