kp_sampler_skip.cpp: put begin for callee check before fence

This improves performance in the case there is no callee for the kokkosp_begin_parallel_for. This is actually done correctly in the kokkosp_begin_parallel_scan and begin_parallel_reduce.
kokkos · May 11, 2024 · 69221e5 · 69221e5
1 parent dd39767
commit 69221e5
Showing 1 changed file with 4 additions and 3 deletions.
diff --git a/common/kokkos-sampler/kp_sampler_skip.cpp b/common/kokkos-sampler/kp_sampler_skip.cpp
@@ -200,10 +200,11 @@ void kokkosp_begin_parallel_for(const char* name, const uint32_t devID,
       std::cout << "KokkosP: sample " << *kID
                 << " calling child-begin function...\n";
     }
-    if (tool_globFence) {
-      invoke_ktools_fence(0);
-    }
+
     if (NULL != beginForCallee) {
+      if (tool_globFence) {
+        invoke_ktools_fence(0);
+      }
       uint64_t nestedkID = 0;
       (*beginForCallee)(name, devID, &nestedkID);
       if (tool_verbosity > 0) {