Avoid lock for last queued command

Use atomics for last queued command update

Change-Id: I759e9d78ea72f23c0d45dbede6250b231e122276

rocclr/platform/commandqueue.cpp

@@ -192,25 +192,27 @@ bool HostQueue::isEmpty() {
 
 void HostQueue::setLastQueuedCommand(Command* lastCommand) {
   // Set last submitted command
-  ScopedLock l(lastCmdLock_);
-  if (lastEnqueueCommand_ != nullptr) {
-    lastEnqueueCommand_->release();
+  Command* lastEnqueueCommand =
+    lastEnqueueCommand_.exchange(lastCommand, std::memory_order_acq_rel);
+  if (lastEnqueueCommand != nullptr) {
+    lastEnqueueCommand->release();
   }
-  lastEnqueueCommand_ = lastCommand;
   if (lastCommand != nullptr) {
-    lastEnqueueCommand_->retain();
+    lastCommand->retain();
   }
 }
 
 Command* HostQueue::getLastQueuedCommand(bool retain) {
   // Get last submitted command
-  ScopedLock l(lastCmdLock_);
-
-  if (retain && lastEnqueueCommand_ != nullptr) {
-    lastEnqueueCommand_->retain();
+  Command* lastEnqueueCommand = lastEnqueueCommand_.load(std::memory_order_acquire);
+  if (lastEnqueueCommand == nullptr) {
+    return nullptr;
   }
 
-  return lastEnqueueCommand_;
+  if (retain) {
+    lastEnqueueCommand->retain();
+  }
+  return lastEnqueueCommand;
 }
 
 DeviceQueue::~DeviceQueue() {

rocclr/platform/commandqueue.hpp

@@ -126,7 +126,6 @@ class CommandQueue : public RuntimeObject {
         rtCUs_(rtCUs),
         priority_(priority),
         queueLock_("CommandQueue::queueLock"),
-        lastCmdLock_("LastQueuedCommand"),
         device_(device),
         context_(context),
         cuMask_(cuMask){}
@@ -135,7 +134,6 @@ class CommandQueue : public RuntimeObject {
   uint rtCUs_;                        //!< The number of used RT compute units
   Priority priority_;                 //!< Queue priority
   Monitor queueLock_;                 //!< Lock protecting the queue
-  Monitor lastCmdLock_;               //!< Lock protecting the last queued command
   Device& device_;                    //!< The device
   SharedReference<Context> context_;  //!< The context of this command queue
   const std::vector<uint32_t>& cuMask_;  //!< The CU mask
@@ -187,7 +185,7 @@ class HostQueue : public CommandQueue {
  private:
   ConcurrentLinkedQueue<Command*> queue_;  //!< The queue.
 
-  Command* lastEnqueueCommand_;  //!< The last submitted command
+  std::atomic<Command*> lastEnqueueCommand_;  //!< The last submitted command
 
   //! Await commands and execute them as they become ready.
   void loop(device::VirtualDevice* virtualDevice);