From 9c1560b8bb36e9b1d195af6cb9eb38b5463d83da Mon Sep 17 00:00:00 2001
From: vedithal-amd <Vignesh.Edithal@amd.com>
Date: Sat, 27 Dec 2025 09:47:42 -0500
Subject: [PATCH] [rocprofiler-compute] Fix merging logic for multi process
 (#2445)

* Fix merging logic for multi process

* Fix dispatch id reset logic in case of rocpd format

* Fix kernel id reset logic in case of csv format

* Revert correlation logic change in csv format

* Do inner join instead of left join
---
 projects/rocprofiler-compute/CHANGELOG.md     |  3 +--
 .../rocprofiler-compute/src/utils/utils.py    | 27 ++++++++++++++-----
 2 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/projects/rocprofiler-compute/CHANGELOG.md b/projects/rocprofiler-compute/CHANGELOG.md
index e55041e700..4bc6e2c438 100644
--- a/projects/rocprofiler-compute/CHANGELOG.md
+++ b/projects/rocprofiler-compute/CHANGELOG.md
@@ -45,8 +45,7 @@ Full documentation for ROCm Compute Profiler is available at [https://rocm.docs.
   * Fix the check to prevent showing table where a column is full of N/A
   * Improve detection of empty values when metric evalulation fails due to counter data missing
 
-* Fix the wrong logic in native counter csv to rocprofv3 csv conversion
-  * Use kernel_id instead of dispatch_id to correlate native counter csv and kernel trace csv
+* Fix issue where counter collection data was empty when profiling workload which spawn multiple child processes
 
 ### Removed
 
diff --git a/projects/rocprofiler-compute/src/utils/utils.py b/projects/rocprofiler-compute/src/utils/utils.py
index 956191bae6..54dc7bcfdd 100644
--- a/projects/rocprofiler-compute/src/utils/utils.py
+++ b/projects/rocprofiler-compute/src/utils/utils.py
@@ -890,9 +890,17 @@ def run_prof(
         )
         combined_df = pd.read_csv(workload_dir + f"/results_{fbase}.csv")
         # Reset Dispatch_ID based on PID, Kernel_Name, Grid_Size,
-        # Workgroup_Size, LDS_Per_Workgroup
+        # Workgroup_Size, LDS_Per_Workgroup, Start_Timestamp, End_Timestamp
         combined_df["Dispatch_ID"] = combined_df.groupby(
-            ["PID", "Kernel_Name", "Grid_Size", "Workgroup_Size", "LDS_Per_Workgroup"],
+            [
+                "PID",
+                "Kernel_Name",
+                "Grid_Size",
+                "Workgroup_Size",
+                "LDS_Per_Workgroup",
+                "Start_Timestamp",
+                "End_Timestamp",
+            ],
             sort=False,
         ).ngroup()
         # Reset Kernel_ID based on Kernel_Name, Grid_Size,
@@ -959,6 +967,13 @@ def run_prof(
         # Overwrite column to ensure unique IDs.
         combined_results["Dispatch_ID"] = range(0, len(combined_results))
 
+        # Reset Kernel_ID based on Kernel_Name, Grid_Size,
+        # Workgroup_Size, LDS_Per_Workgroup
+        combined_results["Kernel_ID"] = combined_results.groupby(
+            ["Kernel_Name", "Grid_Size", "Workgroup_Size", "LDS_Per_Workgroup"],
+            sort=False,
+        ).ngroup()
+
         combined_results.to_csv(
             workload_dir + "/out/pmc_1/results_" + fbase + ".csv", index=False
         )
@@ -1097,13 +1112,13 @@ def convert_native_counter_collection_csv(workload_dir: str) -> None:
         )[0]
         kernel_data = pd.read_csv(kernel_data_filename)
 
-        # Merge counter_data with kernel_data on kernel_id
+        # Merge counter_data with kernel_data on dispatch_id
         merged_data = pd.merge(
             counter_data,
             kernel_data,
-            left_on="kernel_id",
-            right_on="Kernel_Id",
-            how="left",
+            left_on="dispatch_id",
+            right_on="Dispatch_Id",
+            how="inner",
         )
 
         rocprofv3_counter_data = pd.DataFrame({