Add a DynamoSyncInputExecuteTime counter (pytorch#6813)

test/dynamo/test_dynamo.py

@@ -102,6 +102,8 @@ def test_simple_model(self):
                      torch_xla._XLAC._get_xla_tensor_debug_info(xla_xy))
     self.assertNotIn('XLAData: None',
                      torch_xla._XLAC._get_xla_tensor_debug_info(xla_y3))
+    # Dynamo has to sync the input since they are intermedate IR(xla_xy and xla_y3)
+    self.assertEqual(met.counter_value('DynamoSyncInputExecuteTime'), 1)
 
   # Tests that the dynamo bridge automatically moves tensors to XLA device,
   # then back to the original device.

test/test_metrics.py

@@ -192,6 +192,18 @@ def test_execute_time_metric(self):
     # of `ExecuteComputation`, but the actual async time.
     self.assertGreater(execute_time_ns, .5 * wall_time_ns)
 
+  def test_pybind_increment_counter(self):
+    met.clear_all()
+    xla_device = xm.xla_device()
+    t1 = torch.tensor(2077, device=xla_device)
+    self.assertEqual(met.counter_value('CreateXlaTensor'), 1)
+    torch_xla._XLAC._xla_increment_counter('CreateXlaTensor', 3)
+    self.assertEqual(met.counter_value('CreateXlaTensor'), 4)
+
+    # try increment a counter that does not exist
+    torch_xla._XLAC._xla_increment_counter('FakeCounter', 2)
+    self.assertEqual(met.counter_value('FakeCounter'), 2)
+
 
 if __name__ == '__main__':
   test = unittest.main()

torch_xla/core/dynamo_bridge.py

@@ -464,6 +464,7 @@ def optimized_mod(*args: tuple):
                 [a for a in args if isinstance(a, torch.Tensor)])) if x
     ]
     if len(input_tensors_to_sync) > 0:
+      torch_xla._XLAC._xla_increment_counter('DynamoSyncInputExecuteTime', 1)
       torch_xla._XLAC._xla_sync_multi(
           input_tensors_to_sync, devices=[], wait=True, sync_xla_data=True)
 

torch_xla/csrc/init_python_bindings.cpp

@@ -1769,6 +1769,12 @@ void InitXlaModuleBindings(py::module m) {
     return xla_data != nullptr ? py::cast<int64_t>(xla_data->Value())
                                : py::none();
   });
+  // TORCH_LAZY_COUNTER
+  m.def("_xla_increment_counter",
+        [](const std::string& name, uint64_t inc_val) {
+          torch::lazy::Counter* counter = new ::torch::lazy::Counter(name);
+          counter->AddValue(inc_val);
+        });
   m.def("_xla_metric_names", []() {
     auto metric_names = torch::lazy::GetMetricNames();
     auto xla_metric_names = runtime::metrics::GetMetricNames();

torch_xla/debug/metrics.py

@@ -72,7 +72,7 @@ def short_metrics_report(counter_names: list = None, metric_names: list = None):
     metric_names (list): The list of metric names whose data needs to be printed.
   """
   if not counter_names:
-    counter_names = ['CachedCompile', 'MarkStep']
+    counter_names = ['CachedCompile', 'MarkStep', 'DynamoSyncInputExecuteTime']
   if not metric_names:
     metric_names = [
         'CompileTime', 'ExecuteTime', 'ExecuteReplicatedTime',