Refine statistic table (#41524)

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

@@ -185,20 +185,22 @@ class TestProfilerStatistic(unittest.TestCase):
                 profiler.TracerEventType.Communication), 5)
         self.assertEqual(len(event_summary.items), 2)
         self.assertEqual(len(event_summary.userdefined_items), 1)
-        self.assertEqual(len(event_summary.model_perspective_items), 3)
+        self.assertEqual(len(event_summary.model_perspective_items), 4)
         self.assertEqual(len(event_summary.memory_manipulation_items), 1)
         self.assertEqual(event_summary.items['conv2d'].cpu_time, 15)
-        self.assertEqual(event_summary.items['conv2d'].gpu_time, 25)
+        self.assertEqual(event_summary.items['conv2d'].general_gpu_time, 25)
         self.assertEqual(
             event_summary.model_perspective_items['Forward'].cpu_time, 100)
         self.assertEqual(
-            event_summary.model_perspective_items['Forward'].gpu_time, 135)
+            event_summary.model_perspective_items['Forward'].general_gpu_time,
+            135)
         self.assertEqual(
-            event_summary.model_perspective_items['Backward'].gpu_time, 0)
+            event_summary.model_perspective_items['Backward'].general_gpu_time,
+            0)
         self.assertEqual(
             event_summary.memory_manipulation_items['AsyncMemcpy'].cpu_time, 15)
-        self.assertEqual(
-            event_summary.memory_manipulation_items['AsyncMemcpy'].gpu_time, 60)
+        self.assertEqual(event_summary.memory_manipulation_items['AsyncMemcpy']
+                         .general_gpu_time, 60)
         print(
             profiler.profiler_statistic._build_table(
                 statistic_data,
@@ -226,31 +228,31 @@ class TestProfilerStatistic(unittest.TestCase):
         userdefined_node = HostPythonNode('Communication Time',
                                           profiler.TracerEventType.UserDefined,
                                           100, 110, 1000, 1001)
-        reduce_all_launchkernel0 = HostPythonNode(
+        allreduce_launchkernel0 = HostPythonNode(
             'cudalaunchkernel', profiler.TracerEventType.CudaRuntime, 102, 104,
             1000, 1001)
 
-        nccl_reduce_all_kernel0 = DevicePythonNode(
-            'nccl_reduce_all_kernel', profiler.TracerEventType.Kernel, 105, 120,
+        nccl_allreduce_kernel0 = DevicePythonNode(
+            'nccl_allreduce_kernel', profiler.TracerEventType.Kernel, 105, 120,
             0, 0, 2)
 
         communication_node = HostPythonNode(
             'Communication', profiler.TracerEventType.Communication, 105, 110,
             1000, 1001)
 
-        reduce_all_op1 = HostPythonNode('reduce_all_op1',
-                                        profiler.TracerEventType.Operator, 105,
-                                        108, 1000, 1001)
-        reduce_all_op1_infershape = HostPythonNode(
-            'reduce_all_op1::infershape',
-            profiler.TracerEventType.OperatorInner, 105, 106, 1000, 1001)
+        allreduce_op1 = HostPythonNode('allreduce_op1',
+                                       profiler.TracerEventType.Operator, 105,
+                                       108, 1000, 1001)
+        allreduce_op1_infershape = HostPythonNode(
+            'allreduce_op1::infershape', profiler.TracerEventType.OperatorInner,
+            105, 106, 1000, 1001)
 
-        reduce_all_launchkernel1 = HostPythonNode(
+        allreduce_launchkernel1 = HostPythonNode(
             'cudalaunchkernel', profiler.TracerEventType.CudaRuntime, 106, 107,
             1000, 1001)
 
-        nccl_reduce_all_kernel1 = DevicePythonNode(
-            'nccl_reduce_all_kernel', profiler.TracerEventType.Kernel, 130, 150,
+        nccl_allreduce_kernel1 = DevicePythonNode(
+            'nccl_allreduce_kernel', profiler.TracerEventType.Kernel, 130, 150,
             0, 0, 2)
 
         backward_node = HostPythonNode('Gradient Backward',
@@ -305,19 +307,19 @@ class TestProfilerStatistic(unittest.TestCase):
             'sync_batch_norm_memcpy', profiler.TracerEventType.Memcpy, 150, 200,
             0, 0, 1)
 
-        reduce_all_node2 = HostPythonNode('reduce_all',
-                                          profiler.TracerEventType.Operator,
-                                          230, 250, 1000, 1001)
+        allreduce_node2 = HostPythonNode('allreduce',
+                                         profiler.TracerEventType.Operator, 230,
+                                         250, 1000, 1001)
 
-        reduce_all_node2_infershape = HostPythonNode(
-            'reduce_all_node2::infershape',
+        allreduce_node2_infershape = HostPythonNode(
+            'allreduce_node2::infershape',
             profiler.TracerEventType.OperatorInner, 231, 232, 1000, 1001)
-        reduce_all_launchkernel2 = HostPythonNode(
+        allreduce_launchkernel2 = HostPythonNode(
             'cudalaunchkernel', profiler.TracerEventType.CudaRuntime, 235, 240,
             1000, 1001)
 
-        nccl_reduce_all_kernel2 = DevicePythonNode(
-            'nccl_reduce_all_kernel', profiler.TracerEventType.Kernel, 250, 280,
+        nccl_allreduce_kernel2 = DevicePythonNode(
+            'nccl_allreduce_kernel', profiler.TracerEventType.Kernel, 250, 280,
             0, 0, 2)
 
         root_node.children_node.append(profilerstep_node)
@@ -329,12 +331,12 @@ class TestProfilerStatistic(unittest.TestCase):
         yolonet_node.children_node.extend(
             [sync_batch_norm_node, userdefined_node])
         userdefined_node.children_node.append(communication_node)
-        userdefined_node.runtime_node.append(reduce_all_launchkernel0)
-        reduce_all_launchkernel0.device_node.append(nccl_reduce_all_kernel0)
-        communication_node.children_node.append(reduce_all_op1)
-        reduce_all_op1.children_node.append(reduce_all_op1_infershape)
-        reduce_all_op1.runtime_node.append(reduce_all_launchkernel1)
-        reduce_all_launchkernel1.device_node.append(nccl_reduce_all_kernel1)
+        userdefined_node.runtime_node.append(allreduce_launchkernel0)
+        allreduce_launchkernel0.device_node.append(nccl_allreduce_kernel0)
+        communication_node.children_node.append(allreduce_op1)
+        allreduce_op1.children_node.append(allreduce_op1_infershape)
+        allreduce_op1.runtime_node.append(allreduce_launchkernel1)
+        allreduce_launchkernel1.device_node.append(nccl_allreduce_kernel1)
         conv2d_node.children_node.extend(
             [conv2d_infer_shape, conv2d_compute, conv2d_MemCpy])
         conv2d_compute.runtime_node.append(conv2d_launchkernel)
@@ -350,10 +352,10 @@ class TestProfilerStatistic(unittest.TestCase):
         sync_batch_norm_MemCpy.runtime_node.append(sync_batch_norm_cudaMemCpy)
         sync_batch_norm_launchkernel.device_node.append(sync_batch_norm_kernel)
         sync_batch_norm_cudaMemCpy.device_node.append(sync_batch_norm_memcpy)
-        optimization_node.children_node.append(reduce_all_node2)
-        reduce_all_node2.children_node.append(reduce_all_node2_infershape)
-        reduce_all_node2.runtime_node.append(reduce_all_launchkernel2)
-        reduce_all_launchkernel2.device_node.append(nccl_reduce_all_kernel2)
+        optimization_node.children_node.append(allreduce_node2)
+        allreduce_node2.children_node.append(allreduce_node2_infershape)
+        allreduce_node2.runtime_node.append(allreduce_launchkernel2)
+        allreduce_launchkernel2.device_node.append(nccl_allreduce_kernel2)
         thread_tree = {'thread1001': root_node}
         extra_info = {
             'Process Cpu Utilization': '1.02',
@@ -415,20 +417,22 @@ class TestProfilerStatistic(unittest.TestCase):
                 distributed_summary.overlap_range), 85)
         self.assertEqual(len(event_summary.items), 4)
         self.assertEqual(len(event_summary.userdefined_items), 1)
-        self.assertEqual(len(event_summary.model_perspective_items), 3)
+        self.assertEqual(len(event_summary.model_perspective_items), 4)
         self.assertEqual(len(event_summary.memory_manipulation_items), 1)
         self.assertEqual(event_summary.items['conv2d'].cpu_time, 15)
-        self.assertEqual(event_summary.items['conv2d'].gpu_time, 25)
+        self.assertEqual(event_summary.items['conv2d'].general_gpu_time, 25)
         self.assertEqual(
             event_summary.model_perspective_items['Forward'].cpu_time, 100)
         self.assertEqual(
-            event_summary.model_perspective_items['Forward'].gpu_time, 315)
+            event_summary.model_perspective_items['Forward'].general_gpu_time,
+            315)
         self.assertEqual(
-            event_summary.model_perspective_items['Backward'].gpu_time, 0)
+            event_summary.model_perspective_items['Backward'].general_gpu_time,
+            0)
         self.assertEqual(
             event_summary.memory_manipulation_items['AsyncMemcpy'].cpu_time, 15)
-        self.assertEqual(
-            event_summary.memory_manipulation_items['AsyncMemcpy'].gpu_time, 60)
+        self.assertEqual(event_summary.memory_manipulation_items['AsyncMemcpy']
+                         .general_gpu_time, 60)
         print(
             profiler.profiler_statistic._build_table(
                 statistic_data,