[cherry-pick]Reduce performance influence by record event in python (#42142)

* fix kenrel name apperance (#42071)

* Reduce performance influence by record event in python (#42040)

* optimize performance

* fix

* improve coverage

* fix

* fix

python/paddle/fluid/dataloader/dataloader_iter.py

@@ -31,6 +31,7 @@ import queue
 
 import paddle
 import paddle.profiler as profiler
+from paddle.profiler.utils import in_profiler_mode
 from .. import core, layers
 from ..framework import _non_static_mode, in_dygraph_mode, _in_legacy_dygraph
 from ..multiprocess_utils import _set_SIGCHLD_handler, MP_STATUS_CHECK_INTERVAL, CleanupFuncRegistrar
@@ -252,6 +253,7 @@ class _DataLoaderIterSingleProcess(_DataLoaderIterBase):
         self._exit_thread_expectedly()
 
     def __next__(self):
+        if in_profiler_mode():
             trace_event = profiler.RecordEvent(
                 name="_DataLoaderIterSingleProcess",
                 event_type=profiler.TracerEventType.Dataloader)
@@ -294,6 +296,7 @@ class _DataLoaderIterSingleProcess(_DataLoaderIterBase):
             self._try_shutdown_all()
             six.reraise(*sys.exc_info())
         finally:
+            if in_profiler_mode():
                 trace_event.end()
 
     def _shutdown_thread(self):
@@ -708,6 +711,7 @@ class _DataLoaderIterMultiProcess(_DataLoaderIterBase):
         self._try_shutdown_all(1)
 
     def __next__(self):
+        if in_profiler_mode():
             trace_event = profiler.RecordEvent(
                 name="_DataLoaderIterMultiProcess",
                 event_type=profiler.TracerEventType.Dataloader)
@@ -765,6 +769,7 @@ class _DataLoaderIterMultiProcess(_DataLoaderIterBase):
                 self._try_shutdown_all()
             six.reraise(*sys.exc_info())
         finally:
+            if in_profiler_mode():
                 trace_event.end()
 
     # python2 compatibility

python/paddle/fluid/dygraph/layers.py

@@ -26,6 +26,7 @@ import inspect
 
 import paddle
 import paddle.profiler as profiler
+from paddle.profiler.utils import in_profiler_mode
 
 from . import parallel_helper
 from .. import unique_name
@@ -906,9 +907,12 @@ class Layer(object):
 
             self._built = True
 
+        if in_profiler_mode():
             with profiler.RecordEvent(self.full_name(),
                                       profiler.TracerEventType.Forward):
                 outputs = self.forward(*inputs, **kwargs)
+        else:
+            outputs = self.forward(*inputs, **kwargs)
 
         for forward_post_hook in self._forward_post_hooks.values():
             hook_result = forward_post_hook(self, inputs, outputs)
@@ -919,7 +923,7 @@ class Layer(object):
 
     def __call__(self, *inputs, **kwargs):
         if (not in_declarative_mode()) and (not self._forward_pre_hooks) \
-            and (not self._forward_post_hooks) and (not self._built) and in_dygraph_mode():
+            and (not self._forward_post_hooks) and (not self._built) and in_dygraph_mode() and (not in_profiler_mode()):
             self._build_once(*inputs, **kwargs)
             return self.forward(*inputs, **kwargs)
         else:

python/paddle/fluid/dygraph/varbase_patch_methods.py

@@ -30,6 +30,7 @@ from .parallel import scale_loss
 from paddle.fluid.data_feeder import convert_dtype, _PADDLE_DTYPE_2_NUMPY_DTYPE
 import paddle.utils.deprecated as deprecated
 import paddle.profiler as profiler
+from paddle.profiler.utils import in_profiler_mode
 from paddle import _C_ops
 
 _grad_scalar = None
@@ -247,6 +248,7 @@ def monkey_patch_varbase():
 
         """
         if framework._non_static_mode():
+            if in_profiler_mode():
                 record_event = profiler.RecordEvent(
                     "Gradient Backward", profiler.TracerEventType.Backward)
                 record_event.begin()
@@ -288,6 +290,7 @@ def monkey_patch_varbase():
                     core.dygraph_run_backward([self], [grad_tensor],
                                               retain_graph,
                                               framework._dygraph_tracer())
+            if in_profiler_mode():
                 record_event.end()
         else:
             raise ValueError(

python/paddle/fluid/tests/unittests/test_newprofiler.py

@@ -134,6 +134,42 @@ class TestProfiler(unittest.TestCase):
         prof.export(path='./test_profiler_pb.pb', format='pb')
         prof.summary()
         result = profiler.utils.load_profiler_result('./test_profiler_pb.pb')
+        prof = None
+        dataset = RandomDataset(10 * 4)
+        simple_net = SimpleNet()
+        opt = paddle.optimizer.SGD(learning_rate=1e-3,
+                                   parameters=simple_net.parameters())
+        loader = DataLoader(
+            dataset, batch_size=4, shuffle=True, drop_last=True, num_workers=2)
+        prof = profiler.Profiler(on_trace_ready=lambda prof: None)
+        prof.start()
+        for i, (image, label) in enumerate(loader()):
+            out = simple_net(image)
+            loss = F.cross_entropy(out, label)
+            avg_loss = paddle.mean(loss)
+            avg_loss.backward()
+            opt.minimize(avg_loss)
+            simple_net.clear_gradients()
+            prof.step()
+        prof.stop()
+        prof.summary()
+        prof = None
+        dataset = RandomDataset(10 * 4)
+        simple_net = SimpleNet()
+        loader = DataLoader(dataset, batch_size=4, shuffle=True, drop_last=True)
+        opt = paddle.optimizer.Adam(
+            learning_rate=1e-3, parameters=simple_net.parameters())
+        prof = profiler.Profiler(on_trace_ready=lambda prof: None)
+        prof.start()
+        for i, (image, label) in enumerate(loader()):
+            out = simple_net(image)
+            loss = F.cross_entropy(out, label)
+            avg_loss = paddle.mean(loss)
+            avg_loss.backward()
+            opt.step()
+            simple_net.clear_gradients()
+            prof.step()
+        prof.stop()
 
 
 class TestNvprof(unittest.TestCase):

python/paddle/profiler/profiler_statistic.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 import collections
 from enum import Enum
+import re
 
 from paddle.fluid.core import TracerEventType
 
@@ -1317,10 +1318,11 @@ def _build_table(statistic_data,
         append(header_sep)
         append(row_format.format(*headers))
         append(header_sep)
+        kernel_name_pattern = re.compile('(.+?)(<.*>)(\(.*\))')
         for row_values in all_row_values:
-            indx = row_values[0].find('(')
-            if indx != -1:
-                name = row_values[0][:indx]
+            match = kernel_name_pattern.match(row_values[0])
+            if match:
+                name = match.group(1) + match.group(2)
             else:
                 name = row_values[0]
             if len(name) > name_column_width:

python/paddle/profiler/utils.py

@@ -21,6 +21,7 @@ from paddle.fluid import core
 from paddle.fluid.core import (_RecordEvent, TracerEventType)
 
 _is_profiler_used = False
+_has_optimizer_wrapped = False
 
 _AllowedEventTypeList = [
     TracerEventType.Dataloader, TracerEventType.ProfileStep,
@@ -154,20 +155,31 @@ def load_profiler_result(filename: str):
     return core.load_profiler_result(filename)
 
 
+def in_profiler_mode():
+    return _is_profiler_used == True
+
+
 def wrap_optimizers():
     def optimizer_warpper(func):
         @functools.wraps(func)
         def warpper(*args, **kwargs):
+            if in_profiler_mode():
                 with RecordEvent(
                         'Optimization Step',
                         event_type=TracerEventType.Optimization):
                     return func(*args, **kwargs)
+            else:
+                return func(*args, **kwargs)
 
         return warpper
 
+    global _has_optimizer_wrapped
+    if _has_optimizer_wrapped == True:
+        return
     import paddle.optimizer as optimizer
     for classname in optimizer.__all__:
         if classname != 'Optimizer':
             classobject = getattr(optimizer, classname)
             if getattr(classobject, 'step', None) != None:
                 classobject.step = optimizer_warpper(classobject.step)
+    _has_optimizer_wrapped = True