Fix bad casts in tools.

Also virtualize queue profiling enable.

Change-Id: I761b41269be3df7eb64a5914ee9951ed6b51bb04


[ROCm/ROCR-Runtime commit: 6455a69b03]

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/hsa_ext_amd.cpp

@@ -266,11 +266,9 @@ hsa_status_t hsa_amd_profiling_set_profiler_enabled(hsa_queue_t* queue, int enab
   IS_OPEN();
 
   core::Queue* cmd_queue = core::Queue::Convert(queue);
-
   IS_VALID(cmd_queue);
 
-  AMD_HSA_BITS_SET(cmd_queue->amd_queue_.queue_properties,
-                   AMD_QUEUE_PROPERTIES_ENABLE_PROFILING, (enable != 0));
+  cmd_queue->SetProfiling(enable);
 
   return HSA_STATUS_SUCCESS;
   CATCH;

projects/rocr-runtime/runtime/hsa-runtime/core/runtime/intercept_queue.cpp

@@ -235,6 +235,7 @@ void InterceptQueue::StoreRelaxed(hsa_signal_value_t value) {
 
   next_packet_ = i;
   Cursor.queue = nullptr;
+  atomic::Store(&amd_queue_.read_dispatch_id, next_packet_, std::memory_order_release);
 }
 
 }  // namespace core