SWDEV-480209 - Make internal callbacks non-blocking

Change-Id: Ic918d08f341abfd9a7c167d09f9c723cdc43157f


[ROCm/clr commit: 683a942364]

projects/clr/hipamd/src/hip_graph.cpp

@@ -1476,6 +1476,8 @@ hipError_t hipGraphExecDestroy(hipGraphExec_t pGraphExec) {
   }
   hip::GraphExec* ge = reinterpret_cast<hip::GraphExec*>(pGraphExec);
   ge->release();
+  amd::ScopedLock lock(GraphExec::graphExecSetLock_);
+  GraphExec::graphExecSet_.erase(ge);
   HIP_RETURN(hipSuccess);
 }
 

projects/clr/hipamd/src/hip_graph_internal.cpp

@@ -732,21 +732,11 @@ hipError_t GraphExec::Run(hipStream_t graph_launch_stream) {
   // we may not need to flush any caches.
   CallbackCommand->setEventScope(amd::Device::kCacheStateIgnore);
   amd::Event& event = CallbackCommand->event();
-  if (!event.setCallback(CL_COMPLETE, GraphExec::DecrementRefCount, this)) {
+  constexpr bool kBlocking = false;
+  if (!event.setCallback(CL_COMPLETE, GraphExec::DecrementRefCount, this, kBlocking)) {
     return hipErrorInvalidHandle;
   }
   CallbackCommand->enqueue();
-  // Add the new barrier to stall the stream, until the callback is done
-  amd::Command::EventWaitList eventWaitList;
-  eventWaitList.push_back(CallbackCommand);
-  amd::Command* block_command = new amd::Marker(*launch_stream, kMarkerDisableFlush, eventWaitList);
-  // we may not need to flush any caches.
-  block_command->setEventScope(amd::Device::kCacheStateIgnore);
-  if (block_command == nullptr) {
-    return hipErrorInvalidValue;
-  }
-  block_command->enqueue();
-  block_command->release();
   CallbackCommand->release();
   return status;
 }

projects/clr/hipamd/src/hip_graph_internal.hpp

@@ -750,12 +750,10 @@ struct GraphExec : public amd::ReferenceCountedObject, public Graph {
   ~GraphExec() {
     for (auto stream : parallel_streams_) {
       if (stream != nullptr) {
-        stream->finish();
-        hip::Stream::Destroy(stream);
+        constexpr bool kForceDestroy = true;
+        hip::Stream::Destroy(stream, kForceDestroy);
       }
     }
-    amd::ScopedLock lock(graphExecSetLock_);
-    graphExecSet_.erase(this);
     if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
       if (kernArgManager_ != nullptr) {
         kernArgManager_->release();

projects/clr/hipamd/src/hip_internal.hpp

@@ -372,7 +372,7 @@ public:
     /// Check whether any blocking stream running
     static bool StreamCaptureBlocking();
 
-    static void Destroy(hip::Stream* stream);
+    static void Destroy(hip::Stream* stream, bool forceDestroy = false);
 
     virtual bool terminate();
 

projects/clr/hipamd/src/hip_stream.cpp

@@ -72,8 +72,9 @@ bool Stream::Create() {
 }
 
 // ================================================================================================
-void Stream::Destroy(hip::Stream* stream) {
+void Stream::Destroy(hip::Stream* stream, bool forceDestroy) {
   stream->device_->RemoveStream(stream);
+  stream->SetForceDestroy(forceDestroy);
   stream->release();
   stream = nullptr;
 }

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

@@ -239,7 +239,7 @@ bool HsaAmdSignalHandler(hsa_signal_value_t value, void* arg) {
   gpu->updateCommandsState(ts->command().GetBatchHead());
 
   // Reset API callback signal. It will release AQL queue and start commands processing
-  if (callback_signal.handle != 0) {
+  if (callback_signal.handle != 0 && ts->GetBlocking()) {
     hsa_signal_subtract_relaxed(callback_signal, 1);
   }
 
@@ -515,10 +515,13 @@ hsa_signal_t VirtualGPU::HwQueueTracker::ActiveSignal(
         // If API callback is enabled, then use a blocking signal for AQL queue.
         // HSA signal will be acquired in SW and released after HSA signal callback
         if (ts->command().Callback() != nullptr) {
-          ts->SetCallbackSignal(prof_signal->signal_);
+          bool blocking = ts->command().Callback()->blocking_;
+          ts->SetCallbackSignal(prof_signal->signal_, blocking);
           // Blocks AQL queue from further processing
-          hsa_signal_add_relaxed(prof_signal->signal_, 1);
-          init_value += 1;
+          if (blocking) {
+            hsa_signal_add_relaxed(prof_signal->signal_, 1);
+            init_value += 1;
+          }
         }
         gpu_.QueuedAsyncHandlers()++;
         hsa_status_t result = hsa_amd_signal_async_handler(prof_signal->signal_,

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

@@ -112,8 +112,9 @@ class Timestamp : public amd::ReferenceCountedObject {
   std::vector<ProfilingSignal*> signals_; //!< The list of all signals, associated with the TS
   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
+  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
+  bool blocking_ = true;          //!< If TRUE callback is blocking
 
   Timestamp(const Timestamp&) = delete;
   Timestamp& operator=(const Timestamp&) = delete;
@@ -177,12 +178,15 @@ class Timestamp : public amd::ReferenceCountedObject {
   VirtualGPU* gpu() const { return gpu_; }
 
   //! Updates the callback signal
-  void SetCallbackSignal(hsa_signal_t callback_signal) {
+  void SetCallbackSignal(hsa_signal_t callback_signal, bool blocking = true) {
     callback_signal_ = callback_signal;
+    blocking_ = blocking;
   }
-
   //! Returns the callback signal
   hsa_signal_t GetCallbackSignal() const { return callback_signal_; }
+
+  //! Return if callback is blocking/non-blocking
+  bool GetBlocking() { return blocking_; }
 };
 
 class VirtualGPU : public device::VirtualDevice {

projects/clr/rocclr/platform/command.cpp

@@ -190,10 +190,11 @@ bool Event::resetStatus(int32_t status) {
 }
 
 // ================================================================================================
-bool Event::setCallback(int32_t status, Event::CallBackFunction callback, void* data) {
+bool Event::setCallback(int32_t status, Event::CallBackFunction callback, void* data,
+                        bool blocking) {
   assert(status >= CL_COMPLETE && status <= CL_QUEUED && "invalid status");
 
-  CallBackEntry* entry = new CallBackEntry(status, callback, data);
+  CallBackEntry* entry = new CallBackEntry(status, callback, data, blocking);
   if (entry == NULL) {
     return false;
   }

projects/clr/rocclr/platform/command.hpp

@@ -72,10 +72,10 @@ class Event : public RuntimeObject {
 
     std::atomic<CallBackFunction> callback_;  //!< callback function pointer.
     void* data_;                              //!< user data passed to the callback function.
-    int32_t status_;                           //!< execution status triggering the callback.
-
-    CallBackEntry(int32_t status, CallBackFunction callback, void* data)
-        : callback_(callback), data_(data), status_(status) {}
+    int32_t status_;                          //!< execution status triggering the callback.
+    bool blocking_;                           //!< TRUE if callback is blocking
+    CallBackEntry(int32_t status, CallBackFunction callback, void* data, bool blocking)
+        : callback_(callback), data_(data), status_(status), blocking_(blocking) {}
   };
 
  public:
@@ -173,7 +173,7 @@ class Event : public RuntimeObject {
   int32_t status() const { return status_.load(std::memory_order_relaxed); }
 
   //! Insert the given \a callback into the callback stack.
-  bool setCallback(int32_t status, CallBackFunction callback, void* data);
+  bool setCallback(int32_t status, CallBackFunction callback, void* data, bool blocking = true);
 
   /*! \brief Set the event status.
    *

projects/clr/rocclr/platform/commandqueue.cpp

@@ -57,8 +57,9 @@ HostQueue::HostQueue(Context& context, Device& device, cl_command_queue_properti
 }
 
 bool HostQueue::terminate() {
+  // incase of force destroy skip checking on the last command
   if (AMD_DIRECT_DISPATCH) {
-    if (vdev() != nullptr) {
+    if (!forceDestroy_ && vdev() != nullptr) {
       // If the queue still has the last command, then wait and release it
       // We must be in protected way to get last command when calling
       // awaitCompletion() where lastCommand will be released and possibly

projects/clr/rocclr/platform/commandqueue.hpp

@@ -294,6 +294,9 @@ class HostQueue : public CommandQueue {
   //! Get queue status
   bool GetQueueStatus() { return isActive_; }
 
+  //! Set the force destory to terminate queue without checking last command
+  void SetForceDestroy(bool forceDestroy) { forceDestroy_ = forceDestroy; }
+
   uint64_t getQueueID() {
     return thread_.vdev()->getQueueID();
   }
@@ -305,6 +308,7 @@ private:
 
   //! True if this command queue is active
   bool isActive_;
+  bool forceDestroy_ = false;  //!< Destroy the queue in the current state
 };
 
 class DeviceQueue : public CommandQueue {