[SYCL][CUDA] Cleanup of profiling events (#1956)

Additional CUDA events created for profiling are now
released upon destruction of the PI Event.

Signed-off-by: Ruyman Reyes <ruyman@codeplay.com>

Author: Ruyman

sycl/plugins/cuda/pi_cuda.cpp

@@ -405,6 +405,18 @@ pi_result _pi_event::wait() {
   return retErr;
 }
 
+pi_result _pi_event::release() {
+  assert(queue_ != nullptr);
+  PI_CHECK_ERROR(cuEventDestroy(evEnd_));
+
+  if (queue_->properties_ & PI_QUEUE_PROFILING_ENABLE) {
+    PI_CHECK_ERROR(cuEventDestroy(evQueued_));
+    PI_CHECK_ERROR(cuEventDestroy(evStart_));
+  }
+
+  return PI_SUCCESS;
+}
+
 // makes all future work submitted to queue wait for all work captured in event.
 pi_result enqueueEventWait(pi_queue queue, pi_event event) {
   // for native events, the cuStreamWaitEvent call is used.
@@ -2889,8 +2901,7 @@ pi_result cuda_piEventRelease(pi_event event) {
     pi_result result = PI_INVALID_EVENT;
     try {
       ScopedContext active(event->get_context());
-      auto cuEvent = event->get();
-      result = PI_CHECK_ERROR(cuEventDestroy(cuEvent));
+      result = event->release();
     } catch (...) {
       result = PI_OUT_OF_RESOURCES;
     }

sycl/plugins/cuda/pi_cuda.hpp

@@ -374,6 +374,8 @@ class _pi_event {
     return new _pi_event(type, queue->get_context(), queue);
   }
 
+  pi_result release();
+
   ~_pi_event();
 
 private: