wait per iteration for caching mode

qdp_project/src/benchmark/pipelines/MAX_scan_filter_pipe.h

@@ -232,7 +232,7 @@ private:
 
 public:
     void scan_b(size_t gid, size_t gcnt, size_t tid) {
-        size_t tcnt = thread_count_fc;
+        size_t tcnt = thread_count_ag;
         assert(chunk_size_w % tcnt == 0);
         assert(chunk_size_w % 16   == 0);
         assert(chunk_size_w % tcnt * 16 == 0);
@@ -261,8 +261,11 @@ public:
             trt->stop_timer(1, tid * gcnt + gid);
             
             bt->timed_wait(*(*sync_barrier)[barrier_idx], 1, tid * gcnt + gid);
+
+            if constexpr (caching) (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
         }
-        (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
+
+        if constexpr (!caching) (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
     }
 
     void scan_a(size_t gid, size_t gcnt, size_t tid) {
@@ -291,8 +294,11 @@ public:
             pvc->stop("scan_a", tid * gcnt + gid);
             trt->stop_timer(0, tid * gcnt + gid);
             bt->timed_wait(*(*sync_barrier)[barrier_idx], 0, tid * gcnt + gid);
+
+            if constexpr (caching) (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
         }
-        (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
+
+        if constexpr (!caching) (*(*sync_barrier)[barrier_idx]).arrive_and_drop();
     }
 
     void aggr_j(size_t gid, size_t gcnt, size_t tid) {