add python profiler package

paddle/fluid/pybind/CMakeLists.txt

@@ -2,7 +2,7 @@ set(PYBIND_DEPS init pybind python proto_desc memory executor fleet_wrapper box_
  feed_fetch_method pass generate_pass pass_builder parallel_executor profiler layer tracer engine scope_pool
  analysis_predictor imperative_profiler imperative_flag save_load_util dlpack_tensor device_context
  gloo_wrapper infer_io_utils heter_wrapper generator op_version_registry ps_gpu_wrapper custom_operator
- cost_model cuda_graph_with_memory_pool fleet_executor global_utils phi_utils tcp_store)
+ cost_model cuda_graph_with_memory_pool fleet_executor global_utils phi_utils tcp_store new_profiler)
 
 if (WITH_PSCORE)
  set(PYBIND_DEPS ${PYBIND_DEPS} ps_service)

paddle/fluid/pybind/pybind.cc

@@ -77,6 +77,9 @@ limitations under the License. */
 #include "paddle/fluid/platform/monitor.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
+#include "paddle/fluid/platform/profiler/event_python.h"
+#include "paddle/fluid/platform/profiler/event_tracing.h"
+#include "paddle/fluid/platform/profiler/profiler.h"
 #include "paddle/fluid/pybind/cuda_streams_py.h"
 #include "paddle/fluid/pybind/distributed_py.h"
 #include "paddle/fluid/pybind/eager.h"
@@ -2913,6 +2916,85 @@ All parameter, weight, gradient are variables in Paddle.
  });
 
  m.def("size_of_dtype", framework::SizeOfType);
+ py::class_<paddle::platform::ProfilerResult>(m, "_ProfilerResult")
+ .def(py::init<>())
+ .def("get_data", &paddle::platform::ProfilerResult::GetData,
+ py::return_value_policy::automatic_reference)
+ .def("save", &paddle::platform::ProfilerResult::Save)
+ .def("get_extra_info", &paddle::platform::ProfilerResult::GetExtraInfo);
+
+ py::class_<paddle::platform::DevicePythonNode>(m, "DevicePythonNode")
+ .def(py::init<>())
+ .def_readwrite("name", &paddle::platform::DevicePythonNode::name)
+ .def_readwrite("type", &paddle::platform::DevicePythonNode::type)
+ .def_readwrite("start_ns", &paddle::platform::DevicePythonNode::start_ns)
+ .def_readwrite("end_ns", &paddle::platform::DevicePythonNode::end_ns)
+ .def_readwrite("device_id",
+ &paddle::platform::DevicePythonNode::device_id)
+ .def_readwrite("context_id",
+ &paddle::platform::DevicePythonNode::context_id)
+ .def_readwrite("stream_id",
+ &paddle::platform::DevicePythonNode::stream_id);
+
+ py::class_<paddle::platform::HostPythonNode>(m, "HostPythonNode")
+ .def(py::init<>())
+ .def_readwrite("name", &paddle::platform::HostPythonNode::name)
+ .def_readwrite("type", &paddle::platform::HostPythonNode::type)
+ .def_readwrite("start_ns", &paddle::platform::HostPythonNode::start_ns)
+ .def_readwrite("end_ns", &paddle::platform::HostPythonNode::end_ns)
+ .def_readwrite("process_id",
+ &paddle::platform::HostPythonNode::process_id)
+ .def_readwrite("thread_id", &paddle::platform::HostPythonNode::thread_id)
+ .def_readwrite("children_node",
+ &paddle::platform::HostPythonNode::children_node_ptrs)
+ .def_readwrite("runtime_node",
+ &paddle::platform::HostPythonNode::runtime_node_ptrs)
+ .def_readwrite("device_node",
+ &paddle::platform::HostPythonNode::device_node_ptrs);
+
+ py::class_<paddle::platform::Profiler>(m, "_Profiler")
+ .def("Create", &paddle::platform::Profiler::Create)
+ .def("Prepare",
+ [](paddle::platform::Profiler *profiler) {
+ platform::EnableHostEventRecorder();
+ profiler->Prepare();
+ })
+ .def("Start", &paddle::platform::Profiler::Start)
+ .def("Stop", &paddle::platform::Profiler::Stop,
+ py::return_value_policy::automatic_reference);
+
+ py::class_<paddle::platform::ProfilerOptions>(m, "ProfilerOptions")
+ .def(py::init<>())
+ .def_readwrite("trace_level",
+ &paddle::platform::ProfilerOptions::trace_level)
+ .def_readwrite("trace_switch",
+ &paddle::platform::ProfilerOptions::trace_switch);
+
+ py::class_<platform::RecordEvent>(m, "_RecordEvent")
+ .def(py::init([](std::string name, platform::TracerEventType type) {
+ return std::unique_ptr<platform::RecordEvent>(new platform::RecordEvent(
+ name, type, 1, paddle::platform::EventRole::kOrdinary));
+ }))
+ .def("end", [](platform::RecordEvent *event) { event->End(); });
+
+ py::enum_<paddle::platform::TracerEventType>(m, "TracerEventType")
+ .value("Operator", paddle::platform::TracerEventType::Operator)
+ .value("Dataloader", paddle::platform::TracerEventType::Dataloader)
+ .value("ProfileStep", paddle::platform::TracerEventType::ProfileStep)
+ .value("CudaRuntime", paddle::platform::TracerEventType::CudaRuntime)
+ .value("Kernel", paddle::platform::TracerEventType::Kernel)
+ .value("Memcpy", paddle::platform::TracerEventType::Memcpy)
+ .value("Memset", paddle::platform::TracerEventType::Memset)
+ .value("UserDefined", paddle::platform::TracerEventType::UserDefined)
+ .value("OperatorInner", paddle::platform::TracerEventType::OperatorInner)
+ .value("Forward", paddle::platform::TracerEventType::Forward)
+ .value("Backward", paddle::platform::TracerEventType::Backward)
+ .value("Optimization", paddle::platform::TracerEventType::Optimization)
+ .value("Communication", paddle::platform::TracerEventType::Communication)
+ .value("PythonOp", paddle::platform::TracerEventType::PythonOp)
+ .value("PythonUserDefined",
+ paddle::platform::TracerEventType::PythonUserDefined);
+ m.def("LoadProfilerResult", &paddle::platform::LoadProfilerResult);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
  m.def("set_cublas_switch", platform::SetAllowTF32Cublas);

python/paddle/fluid/core.py

@@ -283,6 +283,7 @@ def to_list(s):
  from .core_avx import _set_cached_executor_build_strategy
  from .core_avx import _device_synchronize
  from .core_avx import _get_current_stream
+ from .core_avx import _Profiler, _ProfilerResult, _RecordEvent
  from .core_avx import _set_current_stream
  if sys.platform != 'win32':
  from .core_avx import _set_process_pids
@@ -344,6 +345,7 @@ def to_list(s):
  from .core_noavx import _device_synchronize
  from .core_noavx import _get_current_stream
  from .core_noavx import _set_current_stream
+ from .core_avx import _Profiler, _ProfilerResult, _RecordEvent
  if sys.platform != 'win32':
  from .core_noavx import _set_process_pids
  from .core_noavx import _erase_process_pids

python/paddle/profiler/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .profiler import ProfilerState, ProfilerTarget
+from .profiler import make_scheduler, export_chrome_tracing, export_protobuf
+from .profiler import Profiler
+from .profiler import TracerEventType
+from .utils import RecordEvent
+from .profiler_statistic import SortedKeys
+
+__all__ = [
+ 'ProfilerState', 'ProfilerTarget', 'TracerEventType', 'make_scheduler',
+ 'export_chrome_tracing', 'export_protobuf', 'Profiler', 'RecordEvent',
+ 'SortedKeys'
+]