Qualcomm AI Engine Direct - add more profile event (#10227)

### Summary
- add more profile event for important metrics. e.g. RPC execution time
- change test cases accordingly


![image](https://github.com/user-attachments/assets/03e5a423-90c7-4e63-970b-6793c1e93915)

### Test plan
```python backends/qualcomm/tests/test_qnn_delegate.py -k TestQNNQuantizedUtils.test_qnn_backend_profile_op -s $DEVICE_ID -m SM8650 -b build-android/```

Author: haowhsu-quic

backends/qualcomm/runtime/backends/QnnProfiler.cpp

@@ -84,6 +84,22 @@ Qnn_ErrorHandle_t QnnProfile::ProfileData(
         "ProfileData failed to get events: %d", QNN_GET_ERROR_CODE(error));
     return error;
   }
+
+  auto get_unit = [](QnnProfile_EventUnit_t unit) {
+    switch (unit) {
+      case QNN_PROFILE_EVENTUNIT_MICROSEC:
+        return " (us)";
+      case QNN_PROFILE_EVENTUNIT_BYTES:
+        return " (bytes)";
+      case QNN_PROFILE_EVENTUNIT_COUNT:
+        return " (count)";
+      case QNN_PROFILE_EVENTUNIT_BACKEND:
+      // cycle unit is default appeared
+      case QNN_PROFILE_EVENTUNIT_CYCLES:
+      default:
+        return "";
+    }
+  };
   QnnProfile_EventData_t event_data;
   for (std::uint32_t i = 0; i < num_events; ++i) {
     error =
@@ -96,6 +112,16 @@ Qnn_ErrorHandle_t QnnProfile::ProfileData(
           QNN_GET_ERROR_CODE(error));
       return error;
     }
+    // add events for other important metrics, e.g. RPC execution time
+    std::string identifier =
+        std::string(event_data.identifier) + get_unit(event_data.unit);
+    executorch::runtime::event_tracer_log_profiling_delegate(
+        event_tracer,
+        identifier.c_str(),
+        /*delegate_debug_id=*/
+        static_cast<executorch::runtime::DebugHandle>(-1),
+        0,
+        event_data.value);
     // Check an event's sub events only if it relates to graph execution time
     // (and its sub events are the individual op executions):
     if (backend_->IsProfileEventTypeParentOfNodeTime(event_data.type)) {
@@ -109,6 +135,7 @@ Qnn_ErrorHandle_t QnnProfile::ProfileData(
             QNN_GET_ERROR_CODE(error));
         return error;
       }
+
       QnnProfile_EventData_t sub_event_data;
       for (std::uint32_t j = 0; j < num_sub_events; ++j) {
         error = qnn_interface.qnn_profile_get_event_data(

backends/qualcomm/tests/test_qnn_delegate.py

@@ -2505,7 +2505,7 @@ def test_qnn_backend_profile_op(self):
             module,
             sample_input,
             expected_partitions=1,
-            expected_profile_events=24,
+            expected_profile_events=34,
         )
 
     def test_qnn_backend_shared_buffer(self):
@@ -3120,7 +3120,7 @@ def test_qnn_backend_profile_op(self):
             module,
             sample_input,
             expected_partitions=1,
-            expected_profile_events=25,
+            expected_profile_events=35,
         )
 
     def test_qnn_backend_shared_buffer(self):

backends/qualcomm/tests/utils.py

@@ -30,6 +30,7 @@
     to_edge_transform_and_lower_to_qnn,
 )
 from executorch.devtools import generate_etrecord, Inspector
+from executorch.devtools.inspector._inspector_utils import TimeScale
 from executorch.examples.qualcomm.utils import (
     generate_inputs,
     make_output_dir,
@@ -290,7 +291,12 @@ def post_process():
                     outputs.append(output)
 
             def validate_profile():
-                inspector = Inspector(etdump_path=etdump_path, etrecord=etrecord_path)
+                inspector = Inspector(
+                    etdump_path=etdump_path,
+                    etrecord=etrecord_path,
+                    source_time_scale=TimeScale.CYCLES,
+                    target_time_scale=TimeScale.CYCLES,
+                )
                 self.assertTrue(
                     len(inspector.to_dataframe().index) == expected_profile_events
                 )