SWDEV-301667 - Refactor graph code

- Remove Last graph node optimization and instead submit a barrier NOP
packet always. This simplifies the code.

Change-Id: Ied443173ba47a08b6df148ac7e3ead712acda11c

hipamd/src/hip_graph_internal.cpp

@@ -576,21 +576,17 @@ hipError_t GraphExec::Run(hipStream_t stream) {
     repeatLaunch_ = true;
   }
 
-  if (parallelLists_.size() == 1 &&
-      instantiateDeviceId_ == hip_stream->DeviceId()) {
+  if (parallelLists_.size() == 1 && instantiateDeviceId_ == hip_stream->DeviceId()) {
+    // Accumulate command tracks all the AQL packet batch that we submit to the HW. For now
+    // we track only kernel nodes.
     amd::AccumulateCommand* accumulate = nullptr;
-    bool isLastKernelWithoutHiddenHeap =
-        ((topoOrder_.back()->GetType() == hipGraphNodeTypeKernel) &&
-         !reinterpret_cast<hip::GraphKernelNode*>(topoOrder_.back())->HasHiddenHeap());
+
     if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
-      uint8_t* lastCapturedPacket =
-          isLastKernelWithoutHiddenHeap ? topoOrder_.back()->GetAqlPacket() : nullptr;
-      if (topoOrder_.back()->GetEnabled()) {
-        accumulate = new amd::AccumulateCommand(*hip_stream, {}, nullptr, lastCapturedPacket);
-      }
+      accumulate = new amd::AccumulateCommand(*hip_stream, {}, nullptr);
     }
-    for (int i = 0; i < topoOrder_.size() - 1; i++) {
-      if (DEBUG_CLR_GRAPH_PACKET_CAPTURE && topoOrder_[i]->GetType() == hipGraphNodeTypeKernel &&
+    for (int i = 0; i < topoOrder_.size(); i++) {
+      if (DEBUG_CLR_GRAPH_PACKET_CAPTURE &&
+          topoOrder_[i]->GetType() == hipGraphNodeTypeKernel &&
           !reinterpret_cast<hip::GraphKernelNode*>(topoOrder_[i])->HasHiddenHeap()) {
         if (topoOrder_[i]->GetEnabled()) {
           hip_stream->vdev()->dispatchAqlPacket(topoOrder_[i]->GetAqlPacket(), accumulate);
@@ -603,19 +599,6 @@ hipError_t GraphExec::Run(hipStream_t stream) {
       }
     }
 
-    // If last captured packet is kernel, optimize to detect completion of last kernel
-    // This saves on extra packet submitted to determine end of graph
-    if (DEBUG_CLR_GRAPH_PACKET_CAPTURE && isLastKernelWithoutHiddenHeap) {
-      // Add the last kernel node name to the accumulate command
-      if (topoOrder_.back()->GetEnabled()) {
-        accumulate->addKernelName(topoOrder_.back()->GetKernelName());
-      }
-    } else {
-      topoOrder_.back()->SetStream(hip_stream, this);
-      status = topoOrder_.back()->CreateCommand(topoOrder_.back()->GetQueue());
-      topoOrder_.back()->EnqueueCommands(stream);
-    }
-
     if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
       accumulate->enqueue();
       accumulate->release();

rocclr/device/rocm/rocvirtual.cpp

@@ -3477,23 +3477,11 @@ void VirtualGPU::submitAccumulate(amd::AccumulateCommand& vcmd) {
   amd::ScopedLock lock(execution());
   profilingBegin(vcmd, true);
 
-  uint8_t* aqlPacket = vcmd.getLastPacket();
-  if (aqlPacket != nullptr) {
-    dispatchBlockingWait();
-    auto packet = reinterpret_cast<hsa_kernel_dispatch_packet_t*>(aqlPacket);
-    dispatchGenericAqlPacket(packet, packet->header, packet->setup, false);
-    // We need to set fence_dirty_ flag as we would use a dispatch packet with  a completion signal
-    // to track graph finish for the last. The sync logic assumes HW event to a barrier packet that
-    // has a system scope release. This would cause isFenceDirty() check at top level to insert
-    // barrier packet wherever needed
-    fence_dirty_ = true;
+  const Settings& settings = dev().settings();
+  if (settings.barrier_value_packet_) {
+    dispatchBarrierValuePacket(kBarrierVendorPacketNopScopeHeader, true);
   } else {
-    const Settings& settings = dev().settings();
-    if (settings.barrier_value_packet_) {
-      dispatchBarrierValuePacket(kBarrierVendorPacketNopScopeHeader, true);
-    } else {
-      dispatchBarrierPacket(kNopPacketHeader, false);
-    }
+    dispatchBarrierPacket(kNopPacketHeader, false);
   }
 
   profilingEnd(vcmd);

rocclr/platform/command.hpp

@@ -1251,8 +1251,6 @@ class Marker : public Command {
 
 class AccumulateCommand : public Command {
  private:
-  uint8_t* lastPacket_;
-
   //! Kernel names and timestamps list for activity profiling
   std::vector<std::string> kernelNames_;
   std::vector<std::pair<uint64_t, uint64_t>> tsList_;
@@ -1260,12 +1258,9 @@ class AccumulateCommand : public Command {
  public:
   //! Create a new Marker
   AccumulateCommand(HostQueue& queue, const EventWaitList& eventWaitList = nullWaitList,
-         const Event* waitingEvent = nullptr, uint8_t* lastPacket = nullptr)
-      : Command(queue, CL_COMMAND_TASK, eventWaitList, 0, waitingEvent),
-        lastPacket_(lastPacket)
+         const Event* waitingEvent = nullptr)
+      : Command(queue, CL_COMMAND_TASK, eventWaitList, 0, waitingEvent)
       {}
-  //! Return last packet
-  uint8_t* getLastPacket() const { return lastPacket_; }
 
   //! Add kernel name to the list if available
   void addKernelName(const std::string& kernelName) {