pytorch/torch/csrc/autograd/record_function_ops.cpp

#include <ATen/ThreadLocalState.h>
#include <ATen/cpp_custom_type_hack.h>
#include <ATen/record_function.h>
#include <torch/csrc/autograd/record_function_ops.h>

#include <torch/csrc/jit/runtime/operator.h>
#include <torch/library.h>

namespace caffe2 {
// Required for cpp_custom_type_hack to work
// NOLINTNEXTLINE(bugprone-exception-escape)
CAFFE_KNOWN_TYPE(at::RecordFunction);
} // namespace caffe2

namespace torch {
namespace autograd {
namespace profiler {

// Creates a new profiling scope using RecordFunction and invokes its starting
// callbacks.
static void record_function_enter(
    const std::string& name,
    const c10::optional<std::string>& args,
    at::RecordFunction& rec) {
  if (rec.isActive()) {
    if (rec.needsInputs() && args.has_value()) {
      rec.before(
          name, c10::ArrayRef<const c10::IValue>{c10::IValue{args.value()}});
    } else {
      rec.before(name);
    }
  }
}

// Legacy signature using cpp_custom_type_hack
static at::Tensor record_function_enter_legacy(
    const std::string& name,
    const c10::optional<std::string>& args) {
  auto rec = std::make_unique<at::RecordFunction>(at::RecordScope::USER_SCOPE);
  record_function_enter(name, args, *rec);
  return at::cpp_custom_type_hack::create(std::move(rec), at::TensorOptions());
}

// New signature using custom_class
c10::intrusive_ptr<PythonRecordFunction> record_function_enter_new(
    const std::string& name,
    const c10::optional<std::string>& args) {
  auto rec =
      c10::make_intrusive<PythonRecordFunction>(at::RecordScope::USER_SCOPE);
  record_function_enter(name, args, rec->record);
  return rec;
}

static at::RecordFunction& getRecordFunctionFromTensor(
    const at::Tensor& handle) {
  auto& rec = at::cpp_custom_type_hack::cast<at::RecordFunction>(handle);
  return rec;
}

// Ends the profiling scope created with record_function_enter.
static void record_function_exit(at::RecordFunction& rec) {
  rec.end();
}

// Legacy signature using cpp_custom_type_hack
static void record_function_exit_legacy(const at::Tensor& handle) {
  // We don't actually need to do anything with handle just need to persist the
  // lifetime until now.
  auto& rec = getRecordFunctionFromTensor(handle);
  record_function_exit(rec);
}

// New signature using custom_class
static void record_function_exit_new(
    const c10::intrusive_ptr<PythonRecordFunction>& record) {
  record_function_exit(record->record);
}

template <typename Func>
c10::intrusive_ptr<c10::ivalue::Future> _call_end_callbacks_on_fut(
    Func get_record,
    const c10::intrusive_ptr<c10::ivalue::Future>& fut) {
  // Profiling callback that ends the associated record_function
  // and returns the value of the passed in future.
  std::function<c10::IValue(c10::ivalue::Future&)> futureProfilingFunc =
      [get_record = std::move(get_record)](c10::ivalue::Future& fut) {
        auto& rec = get_record();
        rec.end();
        // Note: this future is returned to the user to ensure that a call to
        // wait() ensures that profiling callbacks have ran. To ensure that this
        // is transparent, we must make this future propagate the value of the
        // RPC future. Use value() here instead of constValue() to ensure we
        // propagate errors.
        return fut.value();
      };
  // Define a future that completes after the profiling callbacks are run.
  auto profiledFut = fut->then(
      at::wrapPropagateTLSState(std::move(futureProfilingFunc)),
      fut->elementType());
  return profiledFut;
}

// Legacy signature using cpp_custom_type_hack
static c10::intrusive_ptr<c10::ivalue::Future> _call_end_callbacks_on_fut_legacy(
    const at::Tensor& handle,
    const c10::intrusive_ptr<c10::ivalue::Future>& fut) {
  return _call_end_callbacks_on_fut(
      [handle]() -> at::RecordFunction& {
        TORCH_INTERNAL_ASSERT(
            handle.defined(),
            "Undefined RecordFunction handle. This can happen if the handle is "
            "not correctly persisted and is destroyed before the future is "
            "realized.");

        return getRecordFunctionFromTensor(handle);
      },
      fut);
}

// New signature using custom_class
c10::intrusive_ptr<c10::ivalue::Future> _call_end_callbacks_on_fut_new(
    const c10::intrusive_ptr<PythonRecordFunction>& record,
    const c10::intrusive_ptr<c10::ivalue::Future>& fut) {
  return _call_end_callbacks_on_fut(
      [record]() -> at::RecordFunction& { return record->record; }, fut);
}

// Internal only, do not use directly, use Python's record_function()
TORCH_LIBRARY_FRAGMENT(profiler, m) {
  m.class_<PythonRecordFunction>("_RecordFunction");

  m.def(
      "_record_function_enter(str name, str? args=None) -> Tensor",
      &record_function_enter_legacy);
  m.def(
      "_record_function_enter_new(str name, str? args=None) -> "
      "__torch__.torch.classes.profiler._RecordFunction",
      &record_function_enter_new);
  m.def("_record_function_exit", &record_function_exit_legacy);
  m.def("_record_function_exit._RecordFunction", &record_function_exit_new);

  torch::jit::registerOperator(torch::jit::Operator(
      "profiler::_call_end_callbacks_on_jit_fut(Tensor x, Future(t) y) -> Future(t)",
      [](jit::Stack& stack) {
        // Pop inputs, which should be a future and a tensor
        auto fut = jit::pop(stack).toFuture();
        auto tensor = jit::pop(stack).toTensor();
        auto profiledFut = _call_end_callbacks_on_fut_legacy(tensor, fut);
        // return future that completes when profiling callbacks have run.
        jit::push(stack, std::move(profiledFut));
      },
      c10::AliasAnalysisKind::FROM_SCHEMA));
  torch::jit::registerOperator(torch::jit::Operator(
      "profiler::_call_end_callbacks_on_jit_fut._RecordFunction("
      "__torch__.torch.classes.profiler._RecordFunction x, Future(t) y) -> Future(t)",
      [](c10::Stack& stack) {
        // Pop inputs, which should be a future and a PythonRecordFunction
        auto fut = torch::jit::pop(stack).toFuture();
        auto tensor =
            torch::jit::pop(stack).toCustomClass<PythonRecordFunction>();
        auto profiledFut = _call_end_callbacks_on_fut_new(tensor, fut);
        // return future that completes when profiling callbacks have run.
        torch::jit::push(stack, std::move(profiledFut));
      },
      c10::AliasAnalysisKind::FROM_SCHEMA));
}

} // namespace profiler
} // namespace autograd
} // namespace torch