Improve queueLock and lastCmdLock

Reduce the size of the queueLock and lastCmdLock critical sections
to improve lock contention performance. The smaller the critical
sections are the better.

lasCmdLock is still needed to guarantee that getLastEnqueueCommand_
can retain the command before it is swapped out and released.

Change-Id: Id35d4a77c035b2da0de4c15568b153d49e958bb7

rocclr/platform/commandqueue.cpp

@@ -185,17 +185,26 @@ void HostQueue::append(Command& command) {
   }
   command.retain();
   command.setStatus(CL_QUEUED);
-  ScopedLock l(lastCmdLock_);
   queue_.enqueue(&command);
   if (!IS_HIP) {
     return;
   }
   // Set last submitted command
-  if (lastEnqueueCommand_ != nullptr) {
-    lastEnqueueCommand_->release();
+  Command* prevLastEnqueueCommand;
+  command.retain();
+  {
+    // lastCmdLock_ ensures that lastEnqueueCommand() can retain the command before it is swapped
+    // out. We want to keep this critical section as short as possible, so the command should be
+    // released outside this section.
+    ScopedLock l(lastCmdLock_);
+
+    prevLastEnqueueCommand = lastEnqueueCommand_;
+    lastEnqueueCommand_ = &command;
+  }
+
+  if (prevLastEnqueueCommand != nullptr) {
+    prevLastEnqueueCommand->release();
   }
-  lastEnqueueCommand_ = &command;
-  lastEnqueueCommand_->retain();
 }
 
 bool HostQueue::isEmpty() {
@@ -207,6 +216,8 @@ Command* HostQueue::getLastQueuedCommand(bool retain) {
   // Get last submitted command
   ScopedLock l(lastCmdLock_);
 
+  // Since the lastCmdLock_ is acquired, it is safe to read and retain the lastEnqueueCommand. It is
+  // guaranteed that the pointer will not change.
   if (retain && lastEnqueueCommand_ != nullptr) {
     lastEnqueueCommand_->retain();
   }

rocclr/platform/commandqueue.hpp

@@ -234,9 +234,6 @@ class HostQueue : public CommandQueue {
 
   //! Get last enqueued command
   Command* getLastQueuedCommand(bool retain);
-
-  //! Set last enqueued command
-  void setLastQueuedCommand(Command* lastCommand) {}
 };