From d2a63a144453cce3de04c377b9ce2ae1a9c1967a Mon Sep 17 00:00:00 2001
From: Vivek Kale <11766050+vlkale@users.noreply.github.com>
Date: Thu, 3 Aug 2023 11:13:50 -0700
Subject: [PATCH 1/2] Using kval hashmap (draft) for nestedID

---
 common/kokkos-sampler/kp_sampler_skip.cpp | 24 +++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)

diff --git a/common/kokkos-sampler/kp_sampler_skip.cpp b/common/kokkos-sampler/kp_sampler_skip.cpp
index 773753f8b..87bfb4cf8 100644
--- a/common/kokkos-sampler/kp_sampler_skip.cpp
+++ b/common/kokkos-sampler/kp_sampler_skip.cpp
@@ -5,10 +5,11 @@
 #include <dlfcn.h>
 #include "../../profiling/all/kp_core.hpp"
 #include "kp_config.hpp"
+#include <atomic>
 
 namespace KokkosTools {
 namespace Sampler {
-static uint64_t uniqID           = 0;
+static atomic<uint64_t> uniqID   = 0;
 static uint64_t kernelSampleSkip = 101;
 static int tool_verbosity        = 0;
 static int tool_globFence        = 0;
@@ -160,8 +161,10 @@ void kokkosp_begin_parallel_for(const char* name, const uint32_t devID,
              (unsigned long long)(*kID));
     }
 
-    if (NULL != beginForCallee) {
-      (*beginForCallee)(name, devID, kID);
+    if (NULL != beginForCallee) { 
+       uint64_t nestedID; 
+      (*beginForCallee)(name, devID, nestedID);
+      // map.insert(kID, nestedID);  
     }
   }
 }
@@ -174,7 +177,10 @@ void kokkosp_end_parallel_for(const uint64_t kID) {
     }
 
     if (NULL != endForCallee) {
-      (*endForCallee)(kID);
+        (*endForCallee)(kID);
+      // (*endForCallee)(map.find(kID));
+      
+      // map.clear(nestedID);
     }
   }
 }
@@ -182,7 +188,6 @@ void kokkosp_end_parallel_for(const uint64_t kID) {
 void kokkosp_begin_parallel_scan(const char* name, const uint32_t devID,
                                  uint64_t* kID) {
   *kID = uniqID++;
-
   if (((*kID) % kernelSampleSkip) == 0) {
     if (tool_verbosity > 0) {
       printf("KokkosP: sample %llu calling child-begin function...\n",
@@ -190,7 +195,9 @@ void kokkosp_begin_parallel_scan(const char* name, const uint32_t devID,
     }
 
     if (NULL != beginScanCallee) {
+      uint64_t nestedID = 0;
       (*beginScanCallee)(name, devID, kID);
+         // map.insert(kID, nestedID);  
     }
   }
 }
@@ -204,6 +211,8 @@ void kokkosp_end_parallel_scan(const uint64_t kID) {
 
     if (NULL != endScanCallee) {
       (*endScanCallee)(kID);
+      // (*endForCallee)(map.find(kID)); 
+      // map.clear(nestedID);
     }
   }
 }
@@ -219,7 +228,8 @@ void kokkosp_begin_parallel_reduce(const char* name, const uint32_t devID,
     }
 
     if (NULL != beginReduceCallee) {
-      (*beginReduceCallee)(name, devID, kID);
+      (*beginReduceCallee)(name, devID, kID); 
+        
     }
   }
 }
@@ -233,6 +243,8 @@ void kokkosp_end_parallel_reduce(const uint64_t kID) {
 
     if (NULL != endReduceCallee) {
       (*endReduceCallee)(kID);
+     // (*endForCallee)(map.find(kID)); 
+      // map.clear(nestedID);
     }
   }
 }

From 3becbe2d34236444fa3e3b0e44a089380a99ab56 Mon Sep 17 00:00:00 2001
From: Vivek Kale <11766050+vlkale@users.noreply.github.com>
Date: Sat, 5 Aug 2023 08:17:27 -0700
Subject: [PATCH 2/2] take out erroneous changes - uniqID

---
 common/kokkos-sampler/kp_sampler_skip.cpp | 13 +------------
 1 file changed, 1 insertion(+), 12 deletions(-)

diff --git a/common/kokkos-sampler/kp_sampler_skip.cpp b/common/kokkos-sampler/kp_sampler_skip.cpp
index 87bfb4cf8..c611db0c9 100644
--- a/common/kokkos-sampler/kp_sampler_skip.cpp
+++ b/common/kokkos-sampler/kp_sampler_skip.cpp
@@ -5,11 +5,10 @@
 #include <dlfcn.h>
 #include "../../profiling/all/kp_core.hpp"
 #include "kp_config.hpp"
-#include <atomic>
 
 namespace KokkosTools {
 namespace Sampler {
-static atomic<uint64_t> uniqID   = 0;
+static uint64_t uniqID   = 0;
 static uint64_t kernelSampleSkip = 101;
 static int tool_verbosity        = 0;
 static int tool_globFence        = 0;
@@ -153,7 +152,6 @@ void kokkosp_finalize_library() {
 
 void kokkosp_begin_parallel_for(const char* name, const uint32_t devID,
                                 uint64_t* kID) {
-  *kID = uniqID++;
 
   if (((*kID) % kernelSampleSkip) == 0) {
     if (tool_verbosity > 0) {
@@ -178,16 +176,12 @@ void kokkosp_end_parallel_for(const uint64_t kID) {
 
     if (NULL != endForCallee) {
         (*endForCallee)(kID);
-      // (*endForCallee)(map.find(kID));
-      
-      // map.clear(nestedID);
     }
   }
 }
 
 void kokkosp_begin_parallel_scan(const char* name, const uint32_t devID,
                                  uint64_t* kID) {
-  *kID = uniqID++;
   if (((*kID) % kernelSampleSkip) == 0) {
     if (tool_verbosity > 0) {
       printf("KokkosP: sample %llu calling child-begin function...\n",
@@ -197,7 +191,6 @@ void kokkosp_begin_parallel_scan(const char* name, const uint32_t devID,
     if (NULL != beginScanCallee) {
       uint64_t nestedID = 0;
       (*beginScanCallee)(name, devID, kID);
-         // map.insert(kID, nestedID);  
     }
   }
 }
@@ -211,8 +204,6 @@ void kokkosp_end_parallel_scan(const uint64_t kID) {
 
     if (NULL != endScanCallee) {
       (*endScanCallee)(kID);
-      // (*endForCallee)(map.find(kID)); 
-      // map.clear(nestedID);
     }
   }
 }
@@ -243,8 +234,6 @@ void kokkosp_end_parallel_reduce(const uint64_t kID) {
 
     if (NULL != endReduceCallee) {
       (*endReduceCallee)(kID);
-     // (*endForCallee)(map.find(kID)); 
-      // map.clear(nestedID);
     }
   }
 }