Report summary stats after L0 collector is destructed (#44)

* Report summary stats after L0 collector is destructed

* Combine the two checks and fix a typo

Author: zma2

tools/unitrace/src/levelzero/ze_collector.h

@@ -4001,7 +4001,7 @@ class ZeCollector {
   }
 
   void AggregateDeviceTimeStats() const {
-    // do not acquire global_device_time_stats_mutex_. caller dos it.
+    // do not acquire global_device_time_stats_mutex_. caller does it.
     for (auto it = global_device_time_stats_->begin(); it != global_device_time_stats_->end(); it++) {
       std::string kname;
       if (it->first.tile_ >= 0) {

tools/unitrace/src/tracer.h

@@ -203,6 +203,13 @@ class UniTracer {
   ~UniTracer() {
     total_execution_time_ = correlator_.GetTimestamp();
 
+    if (ze_collector_ != nullptr) {
+      ze_collector_->DisableTracing();
+      delete ze_collector_;
+    }
+
+    // report after ze_collector_ is destructed
+    // local stats are sweeped in deconstructor
     Report();
 
     ClExtCollector::Destroy();
@@ -213,14 +220,6 @@ class UniTracer {
       cl_gpu_collector_->DisableTracing();
     }
 
-    if (ze_collector_ != nullptr) {
-      ze_collector_->DisableTracing();
-    }
-
-    if (ze_collector_ != nullptr) {
-      delete ze_collector_;
-    }
-
     if (itt_collector != nullptr){
       // Print CCL summary before deleting the object
       // If CCL summary is not enbled summary string will be empty