SWDEV-457755 - Add TS only for kernel packets in the Accumulate command

Change-Id: I1b2f01c5763761808f49802fa117abc6306a22aa

rocclr/device/rocm/rocdevice.hpp

@@ -82,6 +82,7 @@ public:
   Timestamp*    ts_;      //!< Timestamp object associated with the signal
   HwQueueEngine engine_;  //!< Engine used with this signal
   amd::Monitor  lock_;    //!< Signal lock for update
+  bool isPacketDispatch_; //!< True if the packet associated with the signal is dispatch
 
   typedef union {
     struct {
@@ -98,6 +99,7 @@ public:
     : ts_(nullptr)
     , engine_(HwQueueEngine::Compute)
     , lock_("Signal Ops Lock", true)
+    , isPacketDispatch_(false)
     {
       signal_.handle = 0;
       flags_.done_ = true;

rocclr/device/rocm/rocvirtual.cpp

@@ -158,7 +158,7 @@ void Timestamp::checkGpuTime() {
         start = std::min(time.start, start);
         end = std::max(time.end, end);
 
-        if (command().type() == CL_COMMAND_TASK) {
+        if ((command().type() == CL_COMMAND_TASK) && (it->isPacketDispatch_ == true)) {
           static_cast<amd::AccumulateCommand&>(command()).addTimestamps(time.start, time.end);
         }
 
@@ -868,12 +868,17 @@ bool VirtualGPU::dispatchGenericAqlPacket(
     // Get active signal for current dispatch if profiling is necessary
     packet->completion_signal = Barriers().ActiveSignal(kInitSignalValueOne, timestamp_);
 
-    // If profiling is enabled, store the correlation ID in the dispatch packet. The profiler can
-    // retrieve this correlation ID to attribute waves to specific dispatch locations.
-    if (std::is_same<decltype(packet), hsa_kernel_dispatch_packet_t*>::value &&
-        activity_prof::IsEnabled(OP_ID_DISPATCH)) {
-      auto dispatchPacket = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(packet);
-      dispatchPacket->reserved2 = timestamp_->command().profilingInfo().correlation_id_;
+    if (std::is_same<decltype(packet), hsa_kernel_dispatch_packet_t*>::value) {
+      // If profiling is enabled, store the correlation ID in the dispatch packet. The profiler can
+      // retrieve this correlation ID to attribute waves to specific dispatch locations.
+      if (activity_prof::IsEnabled(OP_ID_DISPATCH)) {
+        auto dispatchPacket = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(packet);
+        dispatchPacket->reserved2 = timestamp_->command().profilingInfo().correlation_id_;
+      }
+
+      ProfilingSignal* current_signal = Barriers().GetLastSignal();
+      current_signal->isPacketDispatch_ = true;
+
     }
   }
 

rocclr/platform/activity.cpp

@@ -103,12 +103,16 @@ void ReportActivity(const amd::Command& command) {
 
   if (command.type() == CL_COMMAND_TASK) {
     auto timestamps = static_cast<const amd::AccumulateCommand&>(command).getTimestamps();
+    std::vector<std::string> kernel_names =
+        static_cast<const amd::AccumulateCommand&>(command).getKernelNames();
     for (uint32_t i = 0; i < timestamps.size(); i++) {
       auto it = timestamps[i];
       record.begin_ns = it.first;
       record.end_ns = it.second;
-      record.kernel_name =
-        static_cast<const amd::AccumulateCommand&>(command).getKernelNames()[i].c_str();
+      if (kernel_names[i].empty()) {
+        LogError("kernel name cannot be empty");
+      }
+      record.kernel_name = kernel_names[i].c_str();
       function(ACTIVITY_DOMAIN_HIP_OPS, operation_id, &record);
     }
   } else {