SWDEV-393362 - Move setData() from profilingEnd() to profilingBegin() to eliminate a possible data race

- Introduce a state variable to indicate if HwProfiling is enabled to
eliminate a possible data race of vector<> signals_.

Change-Id: Id504cc76d7fa9f7e6455587dd232b60ccbbb735b


[ROCm/clr commit: afa28cdf44]

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -1386,8 +1386,8 @@ address VirtualGPU::allocKernelArguments(size_t size, size_t alignment) {
 
 // ================================================================================================
 /* profilingBegin, when profiling is enabled, creates a timestamp to save in
-* virtualgpu's timestamp_, and calls start() to get the current host
-* timestamp.
+* virtualgpu's timestamp_, saves the pointer timestamp_ to the command's data
+* and then calls start() to get the current host timestamp.
 */
 void VirtualGPU::profilingBegin(amd::Command& command, bool drmProfiling) {
   if (command.profilingInfo().enabled_) {
@@ -1398,6 +1398,7 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool drmProfiling) {
     }
     // Without barrier profiling will wait for each individual signal
     timestamp_ = new Timestamp(this, command);
+    command.setData(timestamp_);
     timestamp_->start();
   }
 
@@ -1423,15 +1424,13 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool drmProfiling) {
 // ================================================================================================
 /* profilingEnd, when profiling is enabled, checks to see if a signal was
 * created for whatever command we are running and calls end() to get the
-* current host timestamp if no signal is available. It then saves the pointer
-* timestamp_ to the command's data.
+* current host timestamp if no signal is available.
 */
 void VirtualGPU::profilingEnd(amd::Command& command) {
   if (command.profilingInfo().enabled_) {
-    if (!timestamp_->HwProfiling()) {
+    if (timestamp_->HwProfiling() == false) {
       timestamp_->end();
     }
-    command.setData(timestamp_);
     timestamp_ = nullptr;
   }
   if (AMD_DIRECT_DISPATCH) {

projects/clr/rocclr/device/rocm/rocvirtual.hpp

@@ -111,6 +111,7 @@ class Timestamp : public amd::ReferenceCountedObject {
   hsa_signal_t callback_signal_;  //!< Signal associated with a callback for possible later update
   amd::Monitor  lock_;            //!< Serialize timestamp update
   bool        accum_ena_ = false; //!< If TRUE then the accumulation of execution times has started
+  bool        hasHwProfiling_ = false; //!< If TRUE then HwProfiling is enabled for the command
 
   Timestamp(const Timestamp&) = delete;
   Timestamp& operator=(const Timestamp&) = delete;
@@ -133,11 +134,14 @@ class Timestamp : public amd::ReferenceCountedObject {
     *end = end_;
   }
 
-  void AddProfilingSignal(ProfilingSignal* signal) { signals_.push_back(signal); }
+  void AddProfilingSignal(ProfilingSignal* signal) {
+    signals_.push_back(signal);
+    hasHwProfiling_ = true;
+  }
 
   const std::vector<ProfilingSignal*>& Signals() const { return signals_; }
 
-  const bool HwProfiling() const { return !signals_.empty(); }
+  const bool HwProfiling() const { return hasHwProfiling_; }
 
   //! Finds execution ticks on GPU
   void checkGpuTime();