tensorflow/core/framework/dataset.h

/* Copyright 2017 The TensorFlow 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.
==============================================================================*/
#ifndef TENSORFLOW_CORE_FRAMEWORK_DATASET_H_
#define TENSORFLOW_CORE_FRAMEWORK_DATASET_H_

#include <deque>
#include <memory>
#include <unordered_map>

#include "absl/memory/memory.h"
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/attr_value_util.h"
#include "tensorflow/core/framework/cancellation.h"
#include "tensorflow/core/framework/dataset_stateful_op_whitelist.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/model.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/thread_factory.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/framework/variant_encode_decode.h"
#include "tensorflow/core/framework/variant_tensor_data.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/core/threadpool.h"
#include "tensorflow/core/lib/core/threadpool_interface.h"
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/strcat.h"
#include "tensorflow/core/platform/cpu_info.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/tracing.h"

// Polymorphic datasets should support all primitive TensorFlow
// types. Use this macro to expand `m(T)` once for each primitive type
// `T`, e.g. to build a `switch` statement.
#define TF_CALL_DATASET_TYPES(m) TF_CALL_ALL_TYPES(m) TF_CALL_QUANTIZED_TYPES(m)

namespace tensorflow {

// Forward declarations to avoid introducing a dependency on headers in
// "tensorflow/core/graph/...".
class GraphDefBuilder;
class Node;

namespace data {

constexpr int kInfiniteCardinality = -1;
constexpr int kUnknownCardinality = -2;

class DatasetBase;
class SerializationContext;

// Interface for reading values from a key-value store.
// Used for restoring iterator state.
class IteratorStateReader {
 public:
  virtual Status ReadScalar(StringPiece key, int64* val) = 0;
#ifdef USE_TSTRING
  // TODO(dero): Temp guard to prevent duplicate declaration during tstring
  // migration.
  virtual Status ReadScalar(StringPiece key, string* val) = 0;
#endif
  virtual Status ReadScalar(StringPiece key, tstring* val) = 0;
  virtual Status ReadTensor(StringPiece key, Tensor* val) = 0;
  virtual bool Contains(StringPiece key) = 0;

  virtual ~IteratorStateReader() {}
};

// Interface for writing values to a key-value store.
// Used for saving iterator state.
class IteratorStateWriter {
 public:
  virtual Status WriteScalar(StringPiece key, const int64 val) = 0;
#ifdef USE_TSTRING
  // TODO(dero): Temp guard to prevent duplicate declaration during tstring
  // migration.
  virtual Status WriteScalar(StringPiece key, const string& val) = 0;
#endif
  virtual Status WriteScalar(StringPiece key, const tstring& val) = 0;
  virtual Status WriteTensor(StringPiece key, const Tensor& val) = 0;

  virtual ~IteratorStateWriter() {}
};

// Wrapper around GraphDefBuilder. Used to serialize Dataset graph.
class GraphDefBuilderWrapper {
 public:
  explicit GraphDefBuilderWrapper(GraphDefBuilder* b) : b_(b) {}

  // Adds a Const node with scalar value to the Graph.
  // `*output` contains a pointer to the output `Node`. It is guaranteed to be
  // non-null if the method returns with an OK status.
  // The returned Node pointer is owned by the backing Graph of GraphDefBuilder.
  template <typename T>
  Status AddScalar(const T& val, Node** output) {
    Tensor val_t = Tensor(DataTypeToEnum<T>::v(), TensorShape({}));
    val_t.scalar<T>()() = val;
    AddTensorInternal(val_t, output);
    if (*output == nullptr) {
      return errors::Internal("AddScalar: Failed to build Const op.");
    }
    return Status::OK();
  }

  // Adds a Const node with vector value to the Graph.
  // `*output` contains a pointer to the output `Node`. It is guaranteed to be
  // non-null if the method returns with an OK status.
  // The returned Node pointer is owned by the backing Graph of GraphDefBuilder.
  // TODO(shivaniagrawal): Consider changing to gtl::ArraySlice?
  template <typename T>
  Status AddVector(const std::vector<T>& val, Node** output) {
    Tensor val_t = Tensor(DataTypeToEnum<T>::v(),
                          TensorShape({static_cast<int64>(val.size())}));
    for (size_t i = 0; i < val.size(); i++) {
      val_t.flat<T>()(i) = val[i];
    }
    AddTensorInternal(val_t, output);
    if (*output == nullptr) {
      return errors::Internal("AddVector: Failed to build Const op.");
    }
    return Status::OK();
  }

#ifdef USE_TSTRING
  // TODO(dero): Temp guard to prevent duplicate declaration during tstring
  // migration.
  Status AddVector(const std::vector<string>& val, Node** output) {
    Tensor val_t = Tensor(DataTypeToEnum<tstring>::v(),
                          TensorShape({static_cast<int64>(val.size())}));
    for (size_t i = 0; i < val.size(); i++) {
      val_t.flat<tstring>()(i) = val[i];
    }
    AddTensorInternal(val_t, output);
    if (*output == nullptr) {
      return errors::Internal("AddVector: Failed to build Const op.");
    }
    return Status::OK();
  }
#endif  // USE_TSTRING

  // Adds a `Const` node for the given tensor value to the graph.
  //
  // `*output` contains a pointer to the output `Node`. It is guaranteed to be
  // non-null if the method returns with an OK status. The returned `Node`
  // pointer is owned by the backing graph of `GraphDefBuilder`.
  Status AddTensor(const Tensor& val, Node** output) {
    AddTensorInternal(val, output);
    if (*output == nullptr) {
      return errors::Internal("AddTensor: Failed to build Const op.");
    }
    return Status::OK();
  }

  // Adds a `Placeholder` node for the given tensor value to the graph.
  //
  // `*output` contains a pointer to the output `Node`. It is guaranteed to be
  // non-null if the method returns with an OK status. The returned `Node`
  // pointer is owned by the backing graph of `GraphDefBuilder`.
  Status AddPlaceholder(const Tensor& val, Node** output) {
    AddPlaceholderInternal(val, output);
    if (*output == nullptr) {
      return errors::Internal(
          "AddPlaceholder: Failed to build Placeholder op.");
    }
    return Status::OK();
  }

  Status AddDataset(const DatasetBase* dataset,
                    const std::vector<Node*>& inputs, Node** output) {
    return AddDataset(dataset, inputs, {}, output);
  }

  // Adds a node corresponding to the `DatasetType` to the Graph.
  // Return value of `DatasetType::type_string()` is used as the op type for the
  // node.
  // Values for the output_types and output_shapes node attributes are also
  // written if those attributes are defined in the OpDef.
  // `*output` contains a pointer to the output `Node`. It is guaranteed to be
  // non-null if the method returns with an OK status.
  // The returned Node pointer is owned by the backing Graph of GraphDefBuilder.
  Status AddDataset(const DatasetBase* dataset,
                    const std::vector<Node*>& inputs,
                    const std::vector<std::pair<StringPiece, AttrValue>>& attrs,
                    Node** output) {
    std::vector<std::pair<size_t, Node*>> enumerated_inputs(inputs.size());
    for (size_t i = 0; i < inputs.size(); i++) {
      enumerated_inputs[i] = std::make_pair(i, inputs[i]);
    }
    return AddDataset(dataset, enumerated_inputs, {}, attrs, output);
  }

  Status AddDataset(
      const DatasetBase* dataset,
      const std::vector<std::pair<size_t, Node*>>& inputs,
      const std::vector<std::pair<size_t, gtl::ArraySlice<Node*>>>& list_inputs,
      const std::vector<std::pair<StringPiece, AttrValue>>& attrs,
      Node** output);

  // Adds a user-defined function with name `function_name` to the graph and
  // recursively adds all functions it references. If a function with a matching
  // name has already been added, returns with OK status. If a user-defined with
  // name `function_name` is not found in the context's function library,
  // returns an InvalidArgumentError. If the function with name `function_name`
  // or any of its dependent functions are stateful, and the context does not
  // explicitly permit stateful functions, returns an InvalidArgument error.
  Status AddFunction(SerializationContext* ctx, const string& function_name,
                     const FunctionLibraryDefinition& lib_def);

  template <typename T>
  void BuildAttrValue(const T& value, AttrValue* attr) {
    SetAttrValue(value, attr);
  }

 private:
  void AddPlaceholderInternal(const Tensor& val, Node** output);
  void AddTensorInternal(const Tensor& val, Node** output);
  bool HasAttr(const string& op_type_name, const string& attr_name) const;

  bool HasAttr(const OpDef* op_def, const string& attr_name) const {
    for (auto attr : op_def->attr()) {
      if (attr.name() == attr_name) {
        return true;
      }
    }
    return false;
  }

  Status AddAttrFunctions(SerializationContext* ctx,
                          const AttrValue& attr_value,
                          const FunctionLibraryDefinition& lib_def) {
    if (attr_value.has_func()) {
      TF_RETURN_IF_ERROR(AddFunction(ctx, attr_value.func().name(), lib_def));
    } else if (attr_value.has_list()) {
      for (const NameAttrList& name_attr_list : attr_value.list().func()) {
        TF_RETURN_IF_ERROR(AddFunction(ctx, name_attr_list.name(), lib_def));
      }
    }
    return Status::OK();
  }

  GraphDefBuilder* b_;
};

class StatsAggregator;
class FunctionHandleCache;

// A utility class for running a function and ensuring that there is always a
// `tensorflow::data` symbol on the stack.
class Runner {
 public:
  virtual ~Runner() {}

  // Runs the given function.
  virtual void Run(const std::function<void()>& f) = 0;

  // Returns a global singleton Runner.
  static Runner* get();
};

// A cut-down version of `OpKernelContext` for running computations in
// iterators. Note that we cannot simply use `OpKernelContext` here because we
// might run computation in an iterator whose lifetime is not nested within the
// lifetime of a single `OpKernelContext` (e.g. asynchronous prefetching).
//
// TODO(mrry): We're making some daring assumptions about the lifetime of the
// runner passed in here. A runner will be deleted when the original step ends,
// but all existing runners only close over session-lifetime (or longer-lived)
// state, so we can make a copy of the function. There's nothing in the
// definition of the API from which we took the runner to guarantee that what we
// are doing is safe. We should formalize the properties here.
class IteratorContext {
 public:
  struct Params {
    explicit Params(IteratorContext* ctx)
        : allocator_getter(ctx->allocator_getter()),
          cancellation_manager(ctx->cancellation_manager()),
          env(ctx->env()),
          flr(ctx->flr()),
          function_handle_cache(ctx->function_handle_cache()),
          resource_mgr(ctx->resource_mgr()),
          model(ctx->model()),
          runner(*(ctx->runner())),
          runner_threadpool_size(ctx->runner_threadpool_size()),
          stats_aggregator(ctx->stats_aggregator()),
          thread_factory(ctx->thread_factory()),
          thread_pool(ctx->thread_pool()) {}

    explicit Params(OpKernelContext* ctx)
        : env(ctx->env()), flr(ctx->function_library()) {
      // NOTE: need reinterpret_cast because function.h forward-declares Device.
      DeviceBase* device =
          reinterpret_cast<DeviceBase*>(ctx->function_library()->device());
      allocator_getter = [device](AllocatorAttributes attrs) {
        return device->GetAllocator(attrs);
      };
      thread::ThreadPool* thread_pool =
          ctx->device()->tensorflow_device_thread_pool();
      if (thread_pool) {
        runner_threadpool_size = thread_pool->NumThreads();
      } else {
        runner_threadpool_size = port::MaxParallelism();
      }

      // NOTE: Wrap every runner invocation in a call to Runner()->Run(), so
      // that a symbol in the tensorflow::data namespace is always on the stack
      // when executing a function inside a Dataset.
      runner = std::bind(
          [](
              // Note: `runner` is a const reference to avoid copying it.
              const std::function<void(std::function<void()>)>& ctx_runner,
              std::function<void()> fn) {
            std::function<void()> wrapped_fn = std::bind(
                [](const std::function<void()>& fn) { Runner::get()->Run(fn); },
                std::move(fn));
            ctx_runner(std::move(wrapped_fn));
          },
          *ctx->runner(), std::placeholders::_1);
    }

    // The Allocator to be used to allocate the output of an iterator.
    std::function<Allocator*(AllocatorAttributes)> allocator_getter = nullptr;

    // The CancellationManager to be used to cancel execution of ops.
    CancellationManager* cancellation_manager;

    // Interface to operating system functionality.
    Env* env = nullptr;

    // The FunctionLibraryRuntime object to be used to make function calls.
    FunctionLibraryRuntime* flr = nullptr;

    // A FunctionHandleCache that owns all the function handles. Not owned.
    FunctionHandleCache* function_handle_cache = nullptr;

    // A resource manager for storing dataset-related state, e.g. random
    // seeds or cached tensors. Not owned.
    ResourceMgr* resource_mgr = nullptr;

    // If non-null, identifies the object used for performance modeling.
    std::shared_ptr<model::Model> model = nullptr;

    // Function call support.
    std::function<void(std::function<void()>)> runner = nullptr;

    // Number of threads used for executing user-defined functions.
    int32 runner_threadpool_size = 0;

    // The `StatsAggregator` object to record statistics about the iterator.
    std::shared_ptr<StatsAggregator> stats_aggregator = nullptr;

    // A factory for creating threads to perform blocking work.
    std::shared_ptr<ThreadFactory> thread_factory = nullptr;

    // A shared thread pool to schedule computation into.
    thread::ThreadPoolInterface* thread_pool = nullptr;
  };

  explicit IteratorContext(IteratorContext* ctx) : params_(Params{ctx}) {}

  explicit IteratorContext(OpKernelContext* ctx) : params_(Params{ctx}) {}

  explicit IteratorContext(Params params) : params_(std::move(params)) {}

  Allocator* allocator(AllocatorAttributes attrs) {
    return params_.allocator_getter(attrs);
  }

  std::function<Allocator*(AllocatorAttributes)> allocator_getter() {
    return params_.allocator_getter;
  }

  CancellationManager* cancellation_manager() {
    return params_.cancellation_manager;
  }

  Env* env() const { return params_.env; }

  FunctionLibraryRuntime* flr() { return params_.flr; }

  FunctionHandleCache* function_handle_cache() {
    return params_.function_handle_cache;
  }

  ResourceMgr* resource_mgr() { return params_.resource_mgr; }

  const std::shared_ptr<model::Model>& model() { return params_.model; }

  std::function<void(std::function<void()>)>* runner() {
    return &params_.runner;
  }

  int32 runner_threadpool_size() { return params_.runner_threadpool_size; }

  std::shared_ptr<StatsAggregator> stats_aggregator() {
    return params_.stats_aggregator;
  }

  const std::shared_ptr<ThreadFactory>& thread_factory() {
    return params_.thread_factory;
  }

  thread::ThreadPoolInterface* thread_pool() { return params_.thread_pool; }

  Params params() { return params_; }

  std::unique_ptr<thread::ThreadPool> CreateThreadPool(const string& name,
                                                       int num_threads) {
    if (params_.thread_pool) {
      // Create a `ThreadPool` instance by wrapping `params_.thread_pool` (which
      // is an instance of `thread::ThreadPoolInterface`). Notably, the
      // ownership of `params_.thread_pool` is *not* transferred onto the newly
      // created `ThreadPool` instance.
      return absl::make_unique<thread::ThreadPool>(params_.thread_pool);
    } else {
      return absl::make_unique<thread::ThreadPool>(params_.env, ThreadOptions(),
                                                   name, num_threads,
                                                   /*low_latency_hint=*/false);
    }
  }

  std::unique_ptr<Thread> StartThread(const string& name,
                                      std::function<void()> fn) {
    if (params_.thread_factory) {
      return params_.thread_factory->StartThread(name, std::move(fn));
    } else {
      return absl::WrapUnique(
          Env::Default()->StartThread({}, name, std::move(fn)));
    }
  }

 private:
  Params params_;
};

// Aggregates runtime support needed for dataset and iterator serialization.
class SerializationContext {
 public:
  struct Params {
    std::vector<std::pair<string, Tensor>>* input_list = nullptr;  // Not owned.

    // Indicates whether serialization should check if the dataset depends on
    // external state. If the check is enabled and external state is
    // encountered, then the serialization will fail.
    bool check_external_state = true;

    // Indicates whether an attempt to serialize a dataset that does not
    // implement serialization should result in an error. If set to `false`, the
    // serialized graph will replace the dataset with a placeholder returned in
    // `input_list`.
    bool fail_if_unimplemented = true;

    // Indicates whether (potentionally large) data tensors should be
    // serialized, or replaced with a placeholder returned in `input_list`. The
    // latter makes sense to do when performing data agnostic graph rewrites to
    // reduce the memory usage.
    bool serialize_data_tensors = true;
  };

  explicit SerializationContext(Params params) : params_(std::move(params)) {}

  std::vector<std::pair<string, Tensor>>* input_list() {
    return params_.input_list;
  }

  bool check_external_state() const { return params_.check_external_state; }

  bool fail_if_unimplemented() const { return params_.fail_if_unimplemented; }

  bool serialize_data_tensors() const { return params_.serialize_data_tensors; }

 private:
  Params params_;

  TF_DISALLOW_COPY_AND_ASSIGN(SerializationContext);
};

// Represents the current position in a range of outputs, where the
// range of outputs is typically represented by an `DatasetBase`,
// defined below.
class IteratorBase {
 public:
  virtual ~IteratorBase() {
    for (auto rit = cleanup_fns_.rbegin(); rit != cleanup_fns_.rend(); ++rit) {
      (*rit)();
    }
  }

  // Gets the next output from the range that this iterator is traversing.
  //
  // If at least one output remains in this iterator's range, that
  // output will be stored in `*out_tensors` and `false` will be
  // stored in `*end_of_sequence`.
  //
  // If no more outputs remain in this iterator's range, `true` will
  // be stored in `*end_of_sequence`, and the content of
  // `*out_tensors` will be undefined.
  //
  // Implementations should never return `OutOfRange` error. If at end of
  // sequence, set `*end_of_sequence = true` and return `Status::OK()`.
  // Internally raised `OutOfRange` errors that do not imply end of sequence
  // should be converted to a different error type before being propagated to
  // the caller.
  //
  // This method is thread-safe.
  //
  // TODO(mrry): Define `GetNextAsync()` or `GetNextManyAsync()`, and
  // potentially remove this method.
  virtual Status GetNext(IteratorContext* ctx, std::vector<Tensor>* out_tensors,
                         bool* end_of_sequence) = 0;

  Status GetNext(IteratorContext&& ctx, std::vector<Tensor>* out_tensors,
                 bool* end_of_sequence) {
    return GetNext(&ctx, out_tensors, end_of_sequence);
  }

  // Returns a vector of DataType values, representing the respective
  // element types of each tuple component in the outputs of this
  // iterator.
  virtual const DataTypeVector& output_dtypes() const = 0;

  // Returns a vector of tensor shapes, representing the respective
  // (and possibly partially defined) shapes of each tuple component
  // in the outputs of this iterator.
  virtual const std::vector<PartialTensorShape>& output_shapes() const = 0;

  // Returns a string that identifies the sequence of iterators leading up to
  // this iterator.
  virtual const string& prefix() const = 0;

  // Performs initialization that needs to happen outside of a constructor to
  // properly propagate errors.
  virtual Status Initialize(IteratorContext* ctx) { return Status::OK(); }

  // Saves the state of this iterator.
  virtual Status Save(SerializationContext* ctx, IteratorStateWriter* writer) {
    return SaveInternal(writer);
  }

  // Restores the state of this iterator.
  virtual Status Restore(IteratorContext* ctx, IteratorStateReader* reader) {
    return RestoreInternal(ctx, reader);
  }

 protected:
  // Returns a node that models this iterator.
  virtual std::shared_ptr<model::Node> CreateNode(
      IteratorContext* ctx, model::Node::Args args) const = 0;

  // This is needed so that sub-classes of IteratorBase can call
  // `SaveInternal` on their input iterators.
  Status SaveInput(IteratorStateWriter* writer,
                   const std::unique_ptr<IteratorBase>& input) {
    return input->SaveInternal(writer);
  }

  // This is needed so that sub-classes of IteratorBase can call
  // `RestoreInternal` on their input iterators.
  Status RestoreInput(IteratorContext* ctx, IteratorStateReader* reader,
                      const std::unique_ptr<IteratorBase>& input) {
    return input->RestoreInternal(ctx, reader);
  }

  // Saves the state of this iterator.
  //
  // This method is used to store the state of the iterator in a checkpoint.
  //
  // TODO(jsimsa): Make this method pure virtual once all `IteratorBase`
  // implementations have an override.
  virtual Status SaveInternal(IteratorStateWriter* writer) {
    return errors::Unimplemented("SaveInternal");
  }

  // Restores the state of this iterator.
  //
  // This method is used to restore the state of the iterator from a checkpoint.
  //
  // TODO(jsimsa): Make this method pure virtual once all `IteratorBase`
  // implementations have an override.
  virtual Status RestoreInternal(IteratorContext* ctx,
                                 IteratorStateReader* reader) {
    return errors::Unimplemented("RestoreInternal");
  }

  // Returns the number of elements produced by this itertaor.
  int64 num_elements() const {
    if (node_) return node_->num_elements();
    return 0;
  }

 private:
  friend class DatasetBase;          // for access to `AddCleanupFunction`
  friend class DatasetBaseIterator;  // for access to `node_`

  // Registers a cleanup function to be called upon object destruction.
  //
  // Registered functions are invoked in the reserve order of registration.
  void AddCleanupFunction(std::function<void()>&& cleanup_fn) {
    cleanup_fns_.push_back(std::move(cleanup_fn));
  }

  // Associates the given performance modeling `Node` with this iterator.
  void SetNode(std::shared_ptr<model::Node> node) { node_ = node.get(); }

  std::vector<std::function<void()>> cleanup_fns_;
  model::Node* node_ = nullptr;  // Not owned.
};

// Represents runtime information needed to construct a dataset.
class DatasetContext {
 public:
  struct Params {
    string type_string;  // op type name of this dataset.
    string node_name;    // graph node name of this dataset op, uniquely
                         // identifying the dataset in the graph.
  };

  explicit DatasetContext(Params params) : params_(std::move(params)) {}

  explicit DatasetContext(OpKernelContext* ctx) {
    params_.type_string = ctx->op_kernel().type_string();
    params_.node_name = ctx->op_kernel().name();
  }

  const string& type_string() const { return params_.type_string; }
  const string& node_name() const { return params_.node_name; }

 private:
  Params params_;
};

// Returns the number of bytes allocated for the given tensor.
int64 GetAllocatedBytes(const std::vector<Tensor>& element);

// Validates and extracts a `DatasetBase` object from `tensor`.
//
// `tensor` must have been written by a call to SetVariantTensorToDataset().
//
// The retrieved pointer is a borrowed reference to the dataset, which is owned
// by the tensor. The consumer must either acquire its own reference to the
// dataset by calling `(*out_dataset)->Ref()`, or ensure that `tensor` is not
// destroyed or mutated while the retrieved pointer is in use.
Status GetDatasetFromVariantTensor(const Tensor& tensor,
                                   DatasetBase** out_dataset);

// Stores a `DatasetBase` object in `tensor`.
//
// The ownership of `dataset` is transferred to `tensor`.
Status StoreDatasetInVariantTensor(DatasetBase* dataset, Tensor* tensor);

// Represents a (potentially infinite) range of outputs, where each
// output is a tuple of tensors.
class DatasetBase : public core::RefCounted {
 public:
  // Key for storing the Dataset graph in the serialized format.
  TF_EXPORT static const char kDatasetGraphKey[];

  // Key for storing the output node of the Dataset graph in the serialized
  // format.
  TF_EXPORT static const char kDatasetGraphOutputNodeKey[];

  explicit DatasetBase(DatasetContext&& ctx)
      : type_string_(ctx.type_string()), node_name_(ctx.node_name()) {}

  // Op type name of this dataset.
  const string& type_string() const { return type_string_; }

  // Graph node name of this dataset op, uniquely identifying the dataset in
  // the graph.
  const string& node_name() const { return node_name_; }

  // Returns a new iterator for iterating over the range of elements in
  // this dataset.
  //
  // This method may be called multiple times on the same instance,
  // and the resulting iterators will have distinct state. Each
  // iterator will traverse all elements in this dataset from the
  // start.
  //
  // The prefix identifies the sequence of iterators leading up to the newly
  // created iterator.
  Status MakeIterator(IteratorContext* ctx, const string& output_prefix,
                      std::unique_ptr<IteratorBase>* iterator) const {
    *iterator = MakeIteratorInternal(output_prefix);
    if (const auto& model = ctx->model()) {
      const string& prefix = (*iterator)->prefix();
      (*iterator)->SetNode(model->AddNode(MakeNodeFactory(ctx, iterator->get()),
                                          prefix, output_prefix));
      (*iterator)->AddCleanupFunction(
          [model, prefix]() { model->RemoveNode(prefix); });
    }
    return (*iterator)->Initialize(ctx);
  }

  Status MakeIterator(IteratorContext&& ctx, const string& output_prefix,
                      std::unique_ptr<IteratorBase>* iterator) const {
    return MakeIterator(&ctx, output_prefix, iterator);
  }

  // Returns a vector of DataType values, representing the respective
  // element types of each tuple component in the outputs of this
  // dataset.
  virtual const DataTypeVector& output_dtypes() const = 0;

  // Returns a vector of tensor shapes, representing the respective
  // (and possibly partially defined) shapes of each tuple component
  // in the outputs of this dataset.
  virtual const std::vector<PartialTensorShape>& output_shapes() const = 0;

  // Returns the number of bytes allocated for tensors of this dataset.
  virtual int64 AllocatedBytes() const { return 0; }

  // Returns the cardinality of this dataset.
  virtual int64 Cardinality() const { return kUnknownCardinality; }

  // A human-readable debug string for this dataset.
  virtual string DebugString() const = 0;

  // If the dataset is stateful it will not be possible to save its graph or
  // checkpoint the state of its iterators.
  //
  // TODO(jsimsa): Remove this method once all `DatasetBase` implementations are
  // migrated over to `CheckExternalState`.
  virtual bool IsStateful() const { return false; }

  // Indicates whether the dataset depends on any external state. If so, the
  // method returns `errors::FailedPrecondition` with a message that identifies
  // the external state. Otherwise, the method returns `Status::OK()`.
  //
  // TODO(jsimsa): Make this method pure virtual once all `DatasetBase`
  // implementations have an override.
  virtual Status CheckExternalState() const {
    if (IsStateful()) {
      return errors::FailedPrecondition("Dataset cannot be serialized.");
    }
    return Status::OK();
  }

 protected:
  friend Status AsGraphDef(
      OpKernelContext* ctx, const DatasetBase* dataset,
      SerializationContext&& serialization_ctx,
      GraphDef* graph_def);  // For access to graph related members.
  friend class CapturedFunction;

  class DatasetGraphDefBuilder : public GraphDefBuilderWrapper {
   public:
    explicit DatasetGraphDefBuilder(GraphDefBuilder* b)
        : GraphDefBuilderWrapper(b) {}
    Status AddInputDataset(SerializationContext* ctx,
                           const DatasetBase* dataset, Node** output);
  };

  // Serializes the dataset into a `GraphDef`, which has two uses:
  //
  // 1) To perform static input pipeline optimizations, tf.data serializes the
  // dataset graph, applies graph rewrites, and then deserializes the graph.
  // If a subclass of `DatasetBase` does not implement this method, then it will
  // be excluded from static optimizations (and so will any upstream datasets).
  //
  // 2) To save the dataset so that it can restore at a later point (possibly in
  // different environment). If a subclass of `DatasetBase` does not implement
  // this method, then this migration will not be possible.
  virtual Status AsGraphDefInternal(SerializationContext* ctx,
                                    DatasetGraphDefBuilder* b,
                                    Node** node) const = 0;

  virtual std::unique_ptr<IteratorBase> MakeIteratorInternal(
      const string& prefix) const = 0;

 private:
  // Returns a factory for nodes that represent the given iterator.
  static model::Node::Factory MakeNodeFactory(IteratorContext* ctx,
                                              IteratorBase* iterator) {
    return [ctx, iterator](model::Node::Args args) {
      return iterator->CreateNode(ctx, std::move(args));
    };
  }

  const string type_string_;
  const string node_name_;
};

// Represents an iterator that is associated with a particular dataset.
class DatasetBaseIterator : public IteratorBase {
 public:
  struct BaseParams {
    // Owns one reference on the shared dataset object.
    const DatasetBase* dataset;

    // Identifies the sequence of iterators leading up to this iterator.
    const string prefix;
  };

  explicit DatasetBaseIterator(const BaseParams& params) : params_(params) {
    params_.dataset->Ref();
  }

  ~DatasetBaseIterator() override { params_.dataset->Unref(); }

  const DataTypeVector& output_dtypes() const override {
    return params_.dataset->output_dtypes();
  }

  const std::vector<PartialTensorShape>& output_shapes() const override {
    return params_.dataset->output_shapes();
  }

  // The sequence of iterators leading up to this iterator.
  const string& prefix() const override { return params_.prefix; }

  // Returns a name to be used for the TraceMe event.
  //
  // NOTE: TraceMe support passing key value pairs of "arguments" using the
  // following format "name#arg_1=value_,...,arg_n=value_n".
  virtual string BuildTraceMeName() { return params_.prefix; }

  Status GetNext(IteratorContext* ctx, std::vector<Tensor>* out_tensors,
                 bool* end_of_sequence) final;

  Status Save(SerializationContext* ctx, IteratorStateWriter* writer) final {
    TF_RETURN_IF_ERROR(params_.dataset->CheckExternalState());
    return IteratorBase::Save(ctx, writer);
  }

 protected:
  // Internal implementation of GetNext that is wrapped in tracing logic.
  virtual Status GetNextInternal(IteratorContext* ctx,
                                 std::vector<Tensor>* out_tensors,
                                 bool* end_of_sequence) = 0;

  string full_name(const string& name) const {
    return strings::StrCat(params_.prefix, ":", name);
  }

  // By default we model iterators using an unknown node, which acts as
  // pass-through with respect to performance modeling.
  std::shared_ptr<model::Node> CreateNode(
      IteratorContext* ctx, model::Node::Args args) const override {
    return model::MakeUnknownNode(std::move(args));
  }

  // When modeling is enabled, this method disables autotuning for the given
  // iterator (and the transitive closure of its inputs).
  void DisableAutotune(IteratorContext* ctx, IteratorBase* iterator) {
    if (iterator->node_) {
      iterator->node_->set_autotune(false);
    }
  }

  // When modeling is enabled, this method enables autotuning for the given
  // iterator (and the transitive closure of its inputs).
  void EnableAutotune(IteratorContext* ctx, IteratorBase* iterator) {
    if (iterator->node_) {
      iterator->node_->set_autotune(true);
    }
  }

  // When modeling is enabled, this method records the fact that this iterator
  // has dequeued an element from an internal buffer.
  void RecordBufferDequeue(IteratorContext* ctx,
                           const std::vector<Tensor>& element) {
    if (collect_resource_usage(ctx)) {
      node_->add_buffered_bytes(-GetAllocatedBytes(element));
    }
  }

  // When modeling is enabled, this method records the fact that this iterator
  // has enqueued an element in an internal buffer.
  void RecordBufferEnqueue(IteratorContext* ctx,
                           const std::vector<Tensor>& element) {
    if (collect_resource_usage(ctx)) {
      node_->add_buffered_bytes(GetAllocatedBytes(element));
    }
  }

  // When modeling is enabled, this method records the fact that this iterator
  // has produced an element.
  void RecordElement(IteratorContext* ctx) {
    if (node_) {
      node_->record_element();
    }
  }

  // When modeling is enabled, this method records the fact that a thread of
  // this iterator has started work.
  void RecordStart(IteratorContext* ctx, bool stop_output = false) {
    if (collect_resource_usage(ctx)) {
      int64 now_nanos = Env::Default()->NowNanos();
      if (stop_output && node_->output()) {
        node_->output()->record_stop(now_nanos);
      }
      node_->record_start(now_nanos);
    }
  }

  // When modeling is enabled, this method records the fact that a thread of
  // this iterator has stopped work.
  void RecordStop(IteratorContext* ctx, bool start_output = false) {
    if (collect_resource_usage(ctx)) {
      int64 now_nanos = Env::Default()->NowNanos();
      node_->record_stop(now_nanos);
      if (start_output && node_->output()) {
        node_->output()->record_start(now_nanos);
      }
    }
  }

 private:
  inline bool collect_resource_usage(IteratorContext* ctx) {
    auto model = ctx->model();
    return model && model->collect_resource_usage() && node_;
  }

  BaseParams params_;
};

// Represents an iterator that is associated with a particular dataset
// with a particular type.
template <class DatasetType>
class DatasetIterator : public DatasetBaseIterator {
 public:
  struct Params {
    // Borrowed pointer to the dataset.
    const DatasetType* dataset;

    // Identifies the sequence of iterators leading up to this iterator.
    const string prefix;
  };

  explicit DatasetIterator(const Params& params)
      : DatasetBaseIterator({params.dataset, params.prefix}),
        typed_dataset_(params.dataset) {}

  // The dataset from which this iterator was created.
  const DatasetType* dataset() const { return typed_dataset_; }

 protected:
  virtual Status GetNextInternal(IteratorContext* ctx,
                                 std::vector<Tensor>* out_tensors,
                                 bool* end_of_sequence) = 0;

 private:
  const DatasetType* const typed_dataset_;  // Not owned.
};

template <typename T>
Status ParseScalarArgument(OpKernelContext* ctx,
                           const StringPiece& argument_name, T* output) {
  const Tensor* argument_t;
  TF_RETURN_IF_ERROR(ctx->input(argument_name, &argument_t));
  if (!TensorShapeUtils::IsScalar(argument_t->shape())) {
    return errors::InvalidArgument(argument_name, " must be a scalar");
  }
  *output = argument_t->scalar<T>()();
  return Status::OK();
}

template <typename T>
Status ParseVectorArgument(OpKernelContext* ctx,
                           const StringPiece& argument_name,
                           std::vector<T>* output) {
  const Tensor* argument_t;
  TF_RETURN_IF_ERROR(ctx->input(argument_name, &argument_t));
  if (!TensorShapeUtils::IsVector(argument_t->shape())) {
    return errors::InvalidArgument(argument_name, " must be a vector");
  }
  int size = argument_t->vec<T>().size();
  output->reserve(size);
  for (int i = 0; i < size; ++i) {
    output->push_back(argument_t->vec<T>()(i));
  }
  return Status::OK();
}

// Encapsulates the work required to plug a DatasetBase into the core TensorFlow
// graph execution engine.
class DatasetOpKernel : public OpKernel {
 public:
  DatasetOpKernel(OpKernelConstruction* ctx) : OpKernel(ctx) {}
  void Compute(OpKernelContext* ctx) final;

 protected:
  // Subclasses should implement this method. It will be called during Compute
  // execution.
  virtual void MakeDataset(OpKernelContext* ctx, DatasetBase** output) = 0;
};

// Encapsulates the work required to plug unary Datasets into the core
// TensorFlow graph execution engine.
class UnaryDatasetOpKernel : public DatasetOpKernel {
 public:
  UnaryDatasetOpKernel(OpKernelConstruction* ctx) : DatasetOpKernel(ctx) {}

 protected:
  void MakeDataset(OpKernelContext* ctx, DatasetBase** output) final;
  virtual void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
                           DatasetBase** output) = 0;
};

// Encapsulates the work required to plug binary Datasets into the core
// TensorFlow graph execution engine.
class BinaryDatasetOpKernel : public DatasetOpKernel {
 public:
  BinaryDatasetOpKernel(OpKernelConstruction* ctx) : DatasetOpKernel(ctx) {}

 protected:
  void MakeDataset(OpKernelContext* ctx, DatasetBase** output) final;
  virtual void MakeDataset(OpKernelContext* ctx, DatasetBase* input,
                           DatasetBase* another_input,
                           DatasetBase** output) = 0;
};

// A simple background worker that executes closures asynchronously and without
// blocking.
//
// A `BackgroundWorker` is used to offload blocking work from an `AsyncOpKernel`
// to avoid blocking an executor thread that may be required by the blocking
// work.
//
// NOTE(mrry): We do not use a regular `tensorflow::thread::ThreadPool` for this
// purpose because its current implementation (in Eigen) uses a finite-length
// queue and will block the caller when full. This can lead to deadlock under
// heavy load. Since the number of concurrent work items in each user of a
// `BackgroundWorker` is at most one per op invocation, the dynamic allocation
// overhead is tolerable.
class BackgroundWorker {
 public:
  BackgroundWorker(Env* env, const string& name);

  ~BackgroundWorker();

  void Schedule(std::function<void()> work_item);

 private:
  void WorkerLoop();

  std::unique_ptr<Thread> thread_;
  mutex mu_;
  condition_variable cond_var_;
  bool cancelled_ GUARDED_BY(mu_) = false;
  std::deque<std::function<void()>> work_queue_ GUARDED_BY(mu_);
};

}  // namespace data

// TODO(b/114112161): Remove these aliases when all users have moved over to the
// `tensorflow::data` namespace.
using data::DatasetBase;
using data::DatasetContext;
using data::DatasetIterator;
using data::DatasetOpKernel;
using data::IteratorBase;
using data::IteratorContext;
using data::IteratorStateReader;
using data::IteratorStateWriter;
using data::SerializationContext;
using data::UnaryDatasetOpKernel;

}  // namespace tensorflow

#endif  // TENSORFLOW_CORE_FRAMEWORK_DATASET_H_