pytorch/torch/csrc/autograd/forward_grad.h

#pragma once

#include <ATen/core/Tensor.h>
#include <unordered_set>

namespace torch {
namespace autograd {

// [ Using ForwardGrad ]
// ForwardGrad needs to be a shared_ptr to satisfy constraints of its inner
// design. But this shared_ptr must be uniquely associated with the object that
// stores it (as of writing, either AutogradMeta or SavedVariable). This object
// is called the "owning object" in the discussions below. This owning object
// must call `ForwardGrad::clear()` when it is destroyed to ensure that the
// ForwardGrad is properly de-allocated.

struct ForwardGrad;

// This file contains two classes that are used to store forward AD gradients
// and ensure that they are scoped properly. Because forward AD runs
// concurrently with the evaluation of the function, we need a mechanism to
// separate different forward AD invocations and be able to compute the right
// gradients. We model such invocations as levels here. The particular scoping
// issue mentioned above has two main drivers:
//   - Ensure that we can conveniently use forward AD within a high level API
//   without
//     leaking the forward AD states outside.
//   - Ensure that we can keep the level that we expose to the user API simple
//   (an integer
//     that represents the nesting depth) while avoiding confusions when the
//     level index is re-used.

// The important external APIs from this file are:
//   - ForwardADLevel::get_next_idx() that can be used to enter a new level and
//   get its index
//   - ForwardADLevel::release_idx() that can be used to exit a given level.
//   - ForwardGrad() can be used to store a given forward gradient that will
//   handle the level
//     tracking automatically.

// The basic implementation strategy is as follows:
// Every tensor has a ForwardGrad, maintaining a map from levels to tangents.
// ForwardGrad is responsible for registering itself to the appropriate
// ForwardADLevel when a new tangent is added to it via ForwardGrad::set_value
// and to un-register itself from this same level if that tangent is removed via
// ForwardGrad::reset. The ForwardADLevel is created when a new level is entered
// via ForwardADLevel::get_next_idx. A reference to the new ForwardADLevel is
// stored into a global (for the whole process) vector that ensure it can be
// accessed via ForwardADLevel::get_by_idx. This reference is deleted when the
// index is released by the user when calling ForwardADLevel::release_idx. When
// it is destructed, the ForwardADLevel is responsible for clearing all the
// tangents for its level stored in all the ForwardGrad that registered with it.
//
// This process-wide level design, compared to a thread local one, allows us to
// use very simple user facing handle for the level (an int) while enabling
// cross-thread forward AD. The only required synchronization for the user is
// when entering and exiting the levels. Some discussion on alternative design
// is in https://github.com/pytorch/pytorch/pull/49097#discussion_r543716453 and
// can be refined in the future.

// Correctness of concurrency:
// Each class uses its own lock when reading or modifying internal storages.
// This allows in particular to safely remove tangents from ForwardGrad when the
// ForwardADLevel is being exited. We ensure no deadlock by ensuring that a
// methods never calls into another class's method while the local class's lock
// is held except in one single case: calling from ForwardADLevel's destructor
// into ForwardGrad::reset with update_level=false.

// The lifetime of these objects is as follows:
// The ForwardADLevel can be in three states:
//      - Initialized: where one of its reference is held by the global vector
//      and there may be more
//        references held by temporary variables in ForwardGrad's methods.
//      - About to be destructed: where "release_idx" has been called and the
//      only reason for the
//        ForwardADLevel not to be destructed right away is that some methods in
//        ForwardGrad have owning reference to it. This is done so that a
//        ForwardADLevel can never be destructed when a ForwardGrad is
//        registered with it and in the process of adding something to its
//        internal state.
//      - Being destructed: Here the ForwardADLevel is not referenced anymore
//      and can be safely reset
//        all of the ForwardGrad. Note that we can have more than one reset
//        being called here (which is ok) but we are guaranteed that there is at
//        least one.
// The ForwardGrad is simpler as there is no intermediary state and no special
// destructor for. The logic to unregister it from the different ForwardADLevel
// is done when the owning object (AutogradMeta or SavedVariable) is being
// destroyed.

// Other considered design:
// To avoid having the ForwardGrad::clear, we considered storing weak_ptr inside
// the ForwardADLevel. While this would work, it would mean that the set inside
// the ForwardADLevel would only grow unless we do an expensive linear scan to
// remove all the dangling weak pointers. Hence this approach was not used.

// Data structures in this file are optimized for this maximum number of levels.
// The number of levels corresponds to the degree of the gradient being
// computed using forward AD and we don't expect more than second order
// gradients to be common.
#define EXPECTED_MAX_LEVEL 2

struct TORCH_API ForwardADLevel {
  ForwardADLevel(uint64_t idx) : idx_(idx) {}
  ~ForwardADLevel();

  static uint64_t get_next_idx();
  static void release_idx(uint64_t idx);
  static std::shared_ptr<ForwardADLevel> get_by_idx(uint64_t idx);
  static std::shared_ptr<ForwardADLevel> try_get_by_idx(uint64_t idx);

  void erase(const std::shared_ptr<ForwardGrad>& grad) {
    std::lock_guard<std::mutex> lock(mutex_);
    grads_.erase(grad);
  }

  void insert(const std::shared_ptr<ForwardGrad>& grad) {
    std::lock_guard<std::mutex> lock(mutex_);
    grads_.insert(grad);
  }

 private:
  std::unordered_set<std::shared_ptr<ForwardGrad>> grads_;
  std::mutex mutex_;
  uint64_t idx_;
};

struct TORCH_API ForwardGrad : std::enable_shared_from_this<ForwardGrad> {
  ForwardGrad() = default;

  // This function must only be called when AutogradMeta or SavedVariable is
  // being destructed as it ensures that:
  //   - The only (potential) other references to this ForwardGrad are the
  //     different level it is registered to
  //   - No other thread will try to call `set_value` or `value` ever from now
  //   on
  //   - Any of the ForwardADLevel that this ForwardGrad is registered with
  //   might
  //     call `reset` at any point during this function
  void clear() {
    c10::SmallVector<uint64_t, EXPECTED_MAX_LEVEL> levels_idx;

    {
      std::lock_guard<std::mutex> lock(mutex_);
      for (auto& c : content_) {
        levels_idx.push_back(c.first);
      }
    }

    for (auto l_idx : levels_idx) {
      // Use "try" version here as another thread might have deleted this
      // level before we got here
      // This is an owning reference as we want to keep the level alive
      // until we successfully unregister ourselves
      auto level = ForwardADLevel::try_get_by_idx(l_idx);
      if (level) {
        level->erase(shared_from_this());
      }
    }
  }

  void set_value(const at::Tensor& value, uint64_t level) {
    // Owning reference to ensure the forward_level is not destroyed
    // while we are updating our internal state
    auto forward_level = ForwardADLevel::get_by_idx(level);
    forward_level->insert(shared_from_this());

    std::lock_guard<std::mutex> lock(mutex_);
    content_.insert({level, value});
  }

  // This function removes the tangent for a given level from this ForwardGrad
  // Use the update_level flag to disable notifying the level about this reset
  // This flag is most notably used by the ForwardADLevel destructor.
  void reset(uint64_t level, bool update_level = true) {
    if (update_level) {
      ForwardADLevel::get_by_idx(level)->erase(shared_from_this());
    }

    std::unique_lock<std::mutex> lock(mutex_);
    const auto& it = content_.find(level);
    TORCH_INTERNAL_ASSERT(
        it != content_.end(), "Resetting a non-existent level.");
    // Keep the Tensor alive until we have released the lock
    // This is needed as we can be in a case where this function is called by
    // ForwardADLevel destructor
    auto t = (*it).second;
    content_.erase(level);
    lock.unlock();
  }

  const at::Tensor& value(uint64_t level) const;

  bool contains(uint64_t level) {
    std::lock_guard<std::mutex> lock(mutex_);
    return content_.count(level) > 0;
  }

  bool empty() const {
    return content_.empty();
  }

  static const at::Tensor& undef_grad();

 private:
  // TODO(albanD): replace this with a SmallVector
  std::unordered_map<uint64_t, at::Tensor> content_;
  mutable std::mutex mutex_;
};

} // namespace autograd
} // namespace torch