SWDEV-482086 - Fix hipGraphInstantiate leak

* In a scenario where kernel is launched with hipExtLaunchKernelGGL and stop event is used, hipGraphInstantiate leaks. Since stop event is used, profiling is enabled and Timestamp (ReferencedCountedObject) is created, but it doesn't get released.
* The idea behind this solution is that profiling should be disabled when command is captured, hence the timestamp should not be created. Because information about capturing isn't available when kernel command is created, packet capturing state is used to determine whether to create a timestamp or not.

Change-Id: Ia23adac4592ded4fb5e236acf99e12e729f63692

hipamd/src/hip_graph_internal.hpp

@@ -250,7 +250,7 @@ struct GraphNode : public hipGraphNodeDOTAttribute {
       node->RemoveEdge(this);
     }
     for (auto packet : gpuPackets_) {
-      delete packet;
+      delete[] packet;
     }
     amd::ScopedLock lock(nodeSetLock_);
     nodeSet_.erase(this);
@@ -275,7 +275,7 @@ struct GraphNode : public hipGraphNodeDOTAttribute {
     hipError_t status = CreateCommand(capture_stream);
     gpuPackets_.clear();
     for (auto& command : commands_) {
-      command->setCapturingState(true, &gpuPackets_, kernArgMgr, &capturedKernelName_);
+      command->setPktCapturingState(true, &gpuPackets_, kernArgMgr, &capturedKernelName_);
       // Enqueue command to capture GPU Packet. The packet is not submitted to the device.
       // The packet is stored in gpuPacket_ and submitted during graph launch.
       command->submit(*(command->queue())->vdev());

rocclr/device/rocm/rocvirtual.cpp

@@ -1545,7 +1545,9 @@ address VirtualGPU::allocKernelArguments(size_t size, size_t alignment) {
 * and then calls start() to get the current host timestamp.
 */
 void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
-  if (command.profilingInfo().enabled_) {
+  // Disable profiling when command is being captured to prevent memory leak from created timestamp_
+  // which won't get freed, since the command is not being executed until graph launch
+  if (!command.getPktCapturingState() && command.profilingInfo().enabled_) {
     if (timestamp_ != nullptr) {
       LogWarning("Trying to create a second timestamp in VirtualGPU. \
                   This could have unintended consequences.");
@@ -1581,7 +1583,7 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
       }
     }
   }
-  if (command.getCapturingState()) {
+  if (command.getPktCapturingState()) {
     currCmd_ = &command;
   }
 }
@@ -1592,7 +1594,7 @@ void VirtualGPU::profilingBegin(amd::Command& command, bool sdmaProfiling) {
 * current host timestamp if no signal is available.
 */
 void VirtualGPU::profilingEnd(amd::Command& command) {
-  if (command.profilingInfo().enabled_) {
+  if (!command.getPktCapturingState() && command.profilingInfo().enabled_) {
     if (timestamp_->HwProfiling() == false) {
       timestamp_->end();
     }
@@ -3111,7 +3113,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
 
   amd::Memory* const* memories =
       reinterpret_cast<amd::Memory* const*>(parameters + kernelParams.memoryObjOffset());
-  bool isGraphCapture = currCmd_ != nullptr && currCmd_->getCapturingState();
+  bool isGraphCapture = currCmd_ != nullptr && currCmd_->getPktCapturingState();
   for (int j = 0; j < iteration; j++) {
     // Reset global size for dimension dim if split is needed
     if (dim != -1) {
@@ -3424,9 +3426,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes,
     if (isGraphCapture) {
       // Dispatch the packet
       if (!dispatchAqlPacket(&dispatchPacket, aqlHeaderWithOrder,
-                        (sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS),
-                        GPU_FLUSH_ON_EXECUTION, currCmd_->getCapturingState(),
-                        currCmd_->getAqlPacket())) {
+                             (sizes.dimensions() << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS),
+                             GPU_FLUSH_ON_EXECUTION, currCmd_->getPktCapturingState(),
+                             currCmd_->getAqlPacket())) {
         return false;
       }
     } else {

rocclr/platform/command.hpp

@@ -268,7 +268,7 @@ class Command : public Event {
   std::vector<void*> data_;
   const Event* waitingEvent_;  //!< Waiting event associated with the marker
 
-  bool capturing_ = false;           //!< Flag to enable/disable graph gpu packet capture
+  bool packetCapturing_ = false;           //!< Flag to enable/disable graph gpu packet capture
   std::vector<uint8_t*>* gpuPackets_;  //!< GPU packets captured when graph capturing is enabled
   GraphKernelArgManager* graphKernArgMgr_ = nullptr;  //!< KernelMgr for graph
   address kernArgOffset_ = nullptr;  //!< KernelArg buffer to used when graph capturing is enabled
@@ -316,13 +316,13 @@ class Command : public Event {
       command_pool_ = nullptr;
     }
   }
-  bool getCapturingState() const { return capturing_; }
+  bool getPktCapturingState() const { return packetCapturing_; }
 
   //! Sets AQL capture state, aql packet to capture and where to copy kernArgs
-  void setCapturingState(bool state, std::vector<uint8_t*>* packet,
+  void setPktCapturingState(bool state, std::vector<uint8_t*>* packet,
                          amd::GraphKernelArgManager* graphKernArgMgr,
                          std::string* capturedKernelName) {
-    capturing_ = state;
+    packetCapturing_ = state;
     gpuPackets_ = packet;
     graphKernArgMgr_ = graphKernArgMgr;
     capturedKernelName_ = capturedKernelName;