#include "tensorflow/core/platform/types.h"

struct MyCustomKernel {
  bool created;
  bool compute_called;
};

static bool delete_called = false;
static bool async_kernel_done = false;

static void* MyCreateFunc(TF_OpKernelConstruction* ctx) {
  struct MyCustomKernel* s = new struct MyCustomKernel;
  s->created = true;
  s->compute_called = false;

  // Exercise attribute reads.
  TF_DataType type;

  // Create return op for device computation region in the paralle_execute op
  Operation* after_op_r = builder.create<mlir::tf_device::ReturnOp>(
      new_device_cluster.getLoc(), device_results);

  builder.setInsertionPointToEnd(&parallel_execute_device_block);

  // Create return op for the new device cluster op
  builder.create<mlir::tf_device::ReturnOp>(device_cluster.getLoc(),

//    only worker 1.
// 2. Create a context B using A.
// 3. Create the variable in B.
// 4. Create another single host server def C with only worker 0.
// 5. Start the GRPC server for worker 0 using C.
// 6. Create a context D using the full cluster server def E.
// 7. Read the variable in D.
TEST(CAPI, SingleHostServerDefV1Works) {
  // Create a server def that represents a 2-process cluster and a client.

    auto begin_op =
        rewriter.create<arith::ConstantOp>(op->getLoc(), type, begin_attr);
    auto end_attr = DenseElementsAttr::get<int32_t>(type, padded_end);
    auto end_op =
        rewriter.create<arith::ConstantOp>(op->getLoc(), type, end_attr);
    auto stride_attr = DenseElementsAttr::get<int32_t>(type, padded_stride);
    auto stride_op =
        rewriter.create<arith::ConstantOp>(op->getLoc(), type, stride_attr);

      tensorflow::GetTypeFromTFTensorShape({1}, shape_dtype);
  auto expanded_start_index = rewriter->create<TF::ExpandDimsOp>(
      loc, vector_type, start_index, scalar_zero);
  auto start_position = rewriter->create<TF::ConcatOp>(
      loc, position_type, scalar_zero,
      ArrayRef<Value>({expanded_start_index, partial_start_position}));

  // Create the slice size tensor. This is done by concatenating `size` and
  // `partial_size`.

  TF_Run(session, run_options, nullptr, nullptr, 0, nullptr, nullptr, 0,
         nullptr, 0, run_metadata, s);
  EXPECT_EQ(TF_INVALID_ARGUMENT, TF_GetCode(s)) << TF_Message(s);
  EXPECT_EQ("Session was not created with a graph before Run()!",
            string(TF_Message(s)));
  TF_DeleteBuffer(run_metadata);
  TF_DeleteBuffer(run_options);

  TF_DeleteDeprecatedSession(session, s);

    // Creates a _Retval node for the src node of edge, and add it to results_,
    // if none exists yet. If a new _Retval node is created, also adds the edge
    // within the subgraph from the src to the _Retval node.
    Status RecordResult(
        const Edge* edge,
        const absl::flat_hash_map<const Node*, Node*>& node_images);

  TensorType zero_point_type = scale_type.clone(rewriter.getI32Type());
  scale = rewriter.create<TF::ConstOp>(
      loc, scale_type,
      DenseFPElementsAttr::get(scale_type,
                               {static_cast<float>(qtype.getScale())}));
  zero_point = rewriter.create<TF::ConstOp>(
      loc, zero_point_type,
      DenseIntElementsAttr::get(zero_point_type,

    // if a special Identity node in the following pattern is clustered in.
    // That is, an Identity node in the loop cond computation is used to drive
    // const nodes consumed by the loop body.  If this Identity node goes into
    // the same cluster with nodes from the loop body, extra dependency is
    // created between the loop cond and body computations and it hinders the

  // (a) read-only. These reads can be replaced by a hoisted read created
  //        before the WhileOp (similar to if and case).
  // (b) written: since the value is written in the loop (which can only in
  //        loop body, all these will become loop variables. Since all resource
  //        variables are removed from the loop variabled during
  //        canonicalizationW, we need to create new operand/result slots. The