Revert "[Profiler] Move legacy profiler out of torch/csrc/autograd (#85512)"

This reverts commit 157a3d2a7c.

Reverted https://github.com/pytorch/pytorch/pull/85512 on behalf of https://github.com/DanilBaibak due to Due to files were deleted, the internal build failed. Please re-submit via codev.

build_variables.bzl

@@ -131,6 +131,7 @@ libtorch_sources_common = sorted(core_sources_common + torch_unpickler_common)
 
 # The profilers are not needed in the lite interpreter build.
 libtorch_profiler_sources = [
+    "torch/csrc/autograd/profiler_legacy.cpp",
     "torch/csrc/autograd/profiler_kineto.cpp",
     "torch/csrc/profiler/collection.cpp",
     "torch/csrc/profiler/kineto_shim.cpp",
@@ -140,7 +141,6 @@ libtorch_profiler_sources = [
     "torch/csrc/profiler/standalone/execution_graph_observer.cpp",
     "torch/csrc/profiler/standalone/itt_observer.cpp",
     "torch/csrc/profiler/standalone/nvtx_observer.cpp",
-    "torch/csrc/profiler/standalone/profiler_legacy.cpp",
     "torch/csrc/profiler/stubs/base.cpp",
     "torch/csrc/monitor/counters.cpp",
     "torch/csrc/monitor/events.cpp",

setup.py

@@ -1115,7 +1115,6 @@ def main():
         'include/torch/csrc/onnx/*.h',
         'include/torch/csrc/profiler/*.h',
         'include/torch/csrc/profiler/orchestration/*.h',
-        'include/torch/csrc/profiler/standalone/*.h',
         'include/torch/csrc/profiler/stubs/*.h',
         'include/torch/csrc/utils/*.h',
         'include/torch/csrc/tensor/*.h',

torch/csrc/autograd/profiler.h

@@ -1,20 +1,4 @@
 #pragma once
 
 #include <torch/csrc/autograd/profiler_kineto.h>
-#include <torch/csrc/profiler/orchestration/observer.h>
-#include <torch/csrc/profiler/standalone/profiler_legacy.h>
-
-// There are some components which use these symbols. Until we migrate them
-// we have to mirror them in the old autograd namespace.
-namespace torch {
-namespace autograd {
-namespace profiler {
-using namespace ::torch::profiler_legacy;
-using ::torch::profiler::impl::ActivityType;
-using ::torch::profiler::impl::getProfilerConfig;
-using ::torch::profiler::impl::ProfilerConfig;
-using ::torch::profiler::impl::profilerEnabled;
-using ::torch::profiler::impl::ProfilerState;
-} // namespace profiler
-} // namespace autograd
-} // namespace torch
+#include <torch/csrc/autograd/profiler_legacy.h>

torch/csrc/autograd/profiler_kineto.cpp

@@ -58,12 +58,9 @@ inline int64_t getTimeUs() {
 }
 
 using torch::profiler::impl::ActiveProfilerType;
-using torch::profiler::impl::ActivityType;
 using torch::profiler::impl::dtypesToStr;
 using torch::profiler::impl::EventType;
 using torch::profiler::impl::ExtraFields;
-using torch::profiler::impl::ProfilerConfig;
-using torch::profiler::impl::ProfilerState;
 using torch::profiler::impl::ProfilerStateBase;
 using torch::profiler::impl::PyExtraFieldsBase;
 using torch::profiler::impl::Result;

torch/csrc/autograd/profiler_legacy.cpp

@@ -1,4 +1,4 @@
-#include <torch/csrc/profiler/standalone/profiler_legacy.h>
+#include <torch/csrc/autograd/profiler_legacy.h>
 
 #include <torch/csrc/autograd/function.h>
 #include <torch/csrc/jit/frontend/tracer.h>
@@ -24,7 +24,8 @@
 #include <iostream>
 
 namespace torch {
-namespace profiler_legacy {
+namespace autograd {
+namespace profiler {
 
 // We decompose the profiler logic into the following components:
 //
@@ -678,5 +679,6 @@ void RecordProfile::processEvents(const std::vector<LegacyEvent*>& events) {
   writeProfilerEventsToStream(out_, events);
 }
 
-} // namespace profiler_legacy
+} // namespace profiler
+} // namespace autograd
 } // namespace torch

torch/csrc/autograd/profiler_legacy.h

@@ -16,11 +16,11 @@
 #include <torch/csrc/profiler/util.h>
 
 namespace torch {
-// namespace autograd {
-// struct Node;
-// } // namespace autograd
+namespace autograd {
 
-namespace profiler_legacy {
+struct Node;
+
+namespace profiler {
 
 enum class C10_API_ENUM EventKind : uint16_t {
   Mark,
@@ -412,5 +412,6 @@ struct TORCH_API TLSLegacyProfilerGuard {
   const c10::optional<ProfilerDisableOptions> profilerDisableOptions_;
 };
 
-} // namespace profiler_legacy
+} // namespace profiler
+} // namespace autograd
 } // namespace torch

torch/csrc/jit/mobile/profiler_edge.h

@@ -1,5 +1,5 @@
 #pragma once
-#include <torch/csrc/autograd/profiler.h>
+#include <torch/csrc/autograd/profiler_kineto.h>
 #include <torch/csrc/jit/mobile/module.h>
 
 namespace torch {

torch/csrc/profiler/api.h

@@ -1,3 +1,17 @@
 #pragma once
 
 #include <torch/csrc/profiler/orchestration/observer.h>
+
+// There are some components which use these symbols. Until we migrate them
+// we have to mirror them in the old autograd namespace.
+namespace torch {
+namespace autograd {
+namespace profiler {
+using torch::profiler::impl::ActivityType;
+using torch::profiler::impl::getProfilerConfig;
+using torch::profiler::impl::ProfilerConfig;
+using torch::profiler::impl::profilerEnabled;
+using torch::profiler::impl::ProfilerState;
+} // namespace profiler
+} // namespace autograd
+} // namespace torch

torch/csrc/profiler/kineto_shim.cpp

@@ -153,7 +153,7 @@ class ExperimentalConfigWrapper {
     // do not trace CPU or GPU events.
     bool cupti_range_profiler = config_.profiler_metrics.size() > 0;
     if (cupti_range_profiler &&
-        activities.count(torch::profiler::impl::ActivityType::CPU)) {
+        activities.count(torch::autograd::profiler::ActivityType::CPU)) {
       LOG(WARNING)
           << "Cannot run range profiler with CPU activities, please only"
           << " use CUDA activity type";
@@ -211,10 +211,10 @@ void prepareTrace(
   }
 
   std::set<libkineto::ActivityType> k_activities;
-  if (activities.count(torch::profiler::impl::ActivityType::CPU)) {
+  if (activities.count(torch::autograd::profiler::ActivityType::CPU)) {
     k_activities.insert(cpuTypes.begin(), cpuTypes.end());
   }
-  if (activities.count(torch::profiler::impl::ActivityType::CUDA)) {
+  if (activities.count(torch::autograd::profiler::ActivityType::CUDA)) {
     k_activities.insert(cudaTypes.begin(), cudaTypes.end());
   }
 

torch/csrc/profiler/kineto_shim.h

@@ -113,7 +113,7 @@ struct ActivityTraceWrapper {
 #endif
 };
 
-using ActivitySet = std::set<torch::profiler::impl::ActivityType>;
+using ActivitySet = std::set<torch::autograd::profiler::ActivityType>;
 void prepareTrace(
     const bool cpuOnly,
     const ActivitySet& activities,