TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);
  TF_DeleteStatus(s);
}

// Exercises tensorflow's C API.

    });
  }
}

void XlaDevice::Compute(OpKernel* op_kernel, OpKernelContext* context) {
  VLOG(2) << "XlaDevice::Compute " << op_kernel->name() << ":"
          << op_kernel->type_string();
  ShowXlaDeviceDeprecationWarning(jit_device_name_.type_string());
  op_kernel->Compute(context);
}

void XlaDevice::ComputeAsync(AsyncOpKernel* op_kernel, OpKernelContext* context,

      std::optional<absl::string_view> tf_program = std::nullopt,
      double error_tolerance = 0.1) {
    // Expected result is calculated by evaluating using TF kernels. In some
    // cases, TF kernel behaves differently from lowered graph (e.g. Hybrid
    // ops). So we optionally use a different graph to calculate the expected
    // result.
    TF_ASSERT_OK_AND_ASSIGN(
        auto expected,

    .Attr("T: type")
    .Attr("enabled: bool")
    .Attr("func_name: string")
    .Attr("node_name: string")
    .SetIsStateful();

class DumpTensorOp : public OpKernel {
 public:
  explicit DumpTensorOp(OpKernelConstruction* ctx) : OpKernel(ctx) {
    string log_dir_path;
    string file_name;
    string func_name;
    string node_name;
    OP_REQUIRES_OK(ctx, ctx->GetAttr("log_dir_path", &log_dir_path));

  def testWithKnownKernel(self):

    def biasd_dense_elu(x, y, z):
      dot = gen_composite_ops.my_biased_dense(x, y, z)
      return nn_ops.elu(dot)  # with known kernel, should not expand.

    t1 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t2 = constant_op.constant([[1.0, 2.0], [3.0, 4.0]])
    t3 = constant_op.constant([[-10.0, -10.0], [-10.0, -10.0]])

  }

  for (auto hashtable : hashtables) {
    auto key_dtype = hashtable.getKeyDtype();
    auto value_dtype = hashtable.getValueDtype();

    // Only allow string -> int64 and int64 -> string mappings due to kernel
    // capability.
    if (!((mlir::isa<TF::StringType>(key_dtype) &&
           mlir::isa<IntegerType>(value_dtype) &&
           mlir::cast<IntegerType>(value_dtype).getWidth() == 64) ||

    return status;
  }
  devices->push_back(std::move(device));
  return absl::OkStatus();
}

REGISTER_LOCAL_DEVICE_FACTORY(DEVICE_XLA_CPU, XlaCpuDeviceFactory);

// Kernel registrations

constexpr std::array<DataType, 18> kAllXlaCpuTypes = {
    {DT_UINT8, DT_QUINT8, DT_UINT16, DT_INT8, DT_QINT8, DT_INT16, DT_INT32,
     DT_QINT32, DT_INT64, DT_HALF, DT_FLOAT, DT_DOUBLE, DT_COMPLEX64,

  }];

  let arguments = (ins TFRT_ChainType);
  let results = (outs TFRT_ChainType);
}

def DelegateKernelOp : RuntimeFallbackDialect_Op<"delegate_kernel"> {
  let summary = "delegate kernel operation";
  let description = [{
    The "tfd.delegate_kernel" operation takes an input chain, and arbitrary
    number of input arguments, and runs a specified TF op via TFE C API. It

      c->set_output(3, c->MakeShape({num_bins_attr->i()}));

      return absl::OkStatus();
    });

class CustomAggregatorOp : public OpKernel {
 public:
  explicit CustomAggregatorOp(OpKernelConstruction* context)
      : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("id", &id_));

    int calibration_method_value;
    int num_bins;
    float min_percentile;

     TFRT_CostFunctionInterface, TFRT_FixedCost<1>]> {
  let summary = "Copy the CPU fallback tensor if it is small";

  let description = [{
    This kernel performs deep copy on the input tensor if it is small, to avoid
    atomic contention on its refcount.

    Note that this kernel always create a new AsyncValue for each result to
    avoid atomic contention on AsyncValue's refcount.
  }];

  let arguments = (ins