[OpenMP] [OMPT] [amdgpu] Collect timestamps for D2D transfer.

Now that a synchronous D2D version has been added, enable OMPT
timestamps collection for this event.

Fixes ROCm/aomp#851

Change-Id: I70c0aeda17f67c4250cdd188539a23f4432e9ab2

Author: dhruvachak

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -3136,7 +3136,9 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
     AMDGPUDeviceTy &DstDevice = static_cast<AMDGPUDeviceTy &>(DstGenericDevice);
 
     // For large transfers use synchronous behavior.
-    if (Size >= OMPX_MaxAsyncCopyBytes) {
+    // If OMPT is enabled or synchronous behavior is explicitly requested:
+    if (ompt::CallbacksInitialized || OMPX_ForceSyncRegions ||
+        Size >= OMPX_MaxAsyncCopyBytes) {
       if (AsyncInfoWrapper.hasQueue())
         if (auto Err = synchronize(AsyncInfoWrapper))
           return Err;
@@ -3153,6 +3155,8 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
       if (auto Err = Signal.wait(getStreamBusyWaitMicroseconds()))
         return Err;
 
+      OMPT_IF_TRACING_ENABLED(recordCopyTimingInNs(Signal.get()););
+
       return Signal.deinit();
     }