// EncapsulateSubgraphsPass.
class MarkForCompilationPass : public GraphOptimizationPass {
 public:
  MarkForCompilationPass() = default;

  Status Run(const GraphOptimizationPassOptions& options) override;

 private:
  Status RunForTest(const GraphOptimizationPassOptions& options,
                    bool disable_deadness_analysis,
                    bool deterministic_cluster_names);

  friend class MarkForCompilationPassTestHelper;

// A helper object to create GraphOptimizationPassOptions.
struct GraphOptimizationPassWrapper {
  explicit GraphOptimizationPassWrapper()
      : library(OpRegistry::Global(), FunctionDefLibrary()) {
    session_options.env = Env::Default();
  }

  // Create GraphOptimizationPassOptions with a graph passed in constructor and
  // sensible options.
  GraphOptimizationPassOptions CreateGraphOptimizationPassOptions(

// XLA compilation with `_input_hostmem`.
class ForceXlaConstantsOnHostPass : public GraphOptimizationPass {
 public:
  ForceXlaConstantsOnHostPass() = default;

  Status Run(const GraphOptimizationPassOptions& options) override;
};

}  // namespace tensorflow

// scoping information that MarkForCompilationPass will need to respect for
// clustering decision.
class ClusterScopingPass : public GraphOptimizationPass {
 public:
  Status Run(const GraphOptimizationPassOptions& options) override;
};

}  // namespace tensorflow

bool HasResourceInputOrOutput(const Node& node);

// Determines the global jit level based on GraphOptimizationPassOptions,
// --tf_xla_auto_jit and whether the graph is a single GPU graph.
OptimizerOptions::GlobalJitLevel GetGlobalJitLevelForGraph(
    const GraphOptimizationPassOptions& options);

// Returns true if `g` is a single-GPU graph.  A single-GPU graph uses exactly
// one GPU (and any number of CPUs).

  // make this more direct, but probably not worth it solely for this test.
  std::vector<std::unique_ptr<Device>> devices;
  TF_RETURN_IF_ERROR(DeviceFactory::AddDevices(session_options, "", &devices));

  GraphOptimizationPassOptions opt_options;
  opt_options.graph = graph;
  opt_options.session_options = &session_options;
  opt_options.flib_def = flib_def;

  if (options.enable_cluster_scoping) {

#include "tensorflow/compiler/jit/xla_activity_listener.h"
#include "tensorflow/compiler/jit/xla_cluster_util.h"

namespace tensorflow {
Status ReportClusteringInfoPass::Run(
    const GraphOptimizationPassOptions& options) {
  XlaAutoClusteringActivity activity;
  *activity.mutable_summary() = GetXlaAutoClusteringSummary(**options.graph);
  activity.set_global_jit_level(GetGlobalJitLevelForGraph(options));

// way; instead it computes a summary of the XLA clusters in the graph and
// broadcasts it via xla_activity_listener.
class ReportClusteringInfoPass : public GraphOptimizationPass {
 public:
  Status Run(const GraphOptimizationPassOptions& options) override;
};
}  // namespace tensorflow

//
//  - Reducing device-to-host copies.
//  - Reducing the number of XLA recompilations.
class PartiallyDeclusterPass : public GraphOptimizationPass {
 public:
  Status Run(const GraphOptimizationPassOptions& options) override;
};

}  // namespace tensorflow

  // compilation is enabled.
  explicit BuildXlaOpsPass(
      std::optional<bool> enable_lazy_compilation = std::nullopt)
      : enable_lazy_compilation_(enable_lazy_compilation) {}

  Status Run(const GraphOptimizationPassOptions& options) override;

 private:
  std::optional<bool> enable_lazy_compilation_;
};

}  // namespace tensorflow