%out_c, %out_tensor = "tfd.delegate_kernel"(
        %in_c, %in1_tensor, %in2_tensor) {
        _name = "MatMul",
        attr1_name = "transpose_a", attr1_value = "bool$false",
        attr2_name = "transpose_b", attr2_value = "bool$false"
      } : (!tfrt.chain, !tfd.tf_tensor, !tfd.tf_tensor) -> (
        !tfrt.chain, !tfd.tf_tensor)
  }];

#include "tensorflow/compiler/mlir/tfrt/runtime_fallback/runtime_fallback_ops.h"

// This optimizes the following scenario:
// %tft0, %c2 = "tfd.move_dht_to_tft"(%dht0, %c1)
//     : (!dht.host_tensor, !tfrt.chain) -> (!tfd.tf_tensor, !tfrt.chain)
// %dht1, %c3 = "tfd.convert_tft_to_dht"(%tft0, %c2)
//     : (!tfd.tf_tensor, !tfrt.chain) -> (!dht.host_tensor, !tfrt.chain)
// some_op %dht1, %c3
//
// becomes
// some_op %dht0, %c1

namespace mlir {
namespace tfd {

//===----------------------------------------------------------------------===//
// TfrtDelegate Dialect
//===----------------------------------------------------------------------===//

RuntimeFallbackDialect::RuntimeFallbackDialect(MLIRContext *context)
    : Dialect(/*name=*/"tfd", context, TypeID::get<RuntimeFallbackDialect>()) {
  allowUnknownTypes();

#include "tfrt/tensor/opdefs/tensor.h"  // from @tf_runtime

namespace mlir {
namespace tfd {

// Dialect for TFRT delegate operations.
class RuntimeFallbackDialect : public Dialect {
 public:
  explicit RuntimeFallbackDialect(MLIRContext* context);
  static StringRef getDialectNamespace() { return "tfd"; }
};

}  // namespace tfd
}  // namespace mlir

#define GET_OP_CLASSES

                              diag) { LOG(ERROR) << diag.message(); },
      tfrt::CreateMallocAllocator(),
      tfrt::CreateMultiThreadedWorkQueue(16, 16));
  CHECK(corert);
  auto fallback_op_handler = tensorflow::tfd::CreateKernelFallbackOpHandler(
      corert->get(), corert->get()->GetHostContext()->GetHostDeviceRef());
  CHECK(fallback_op_handler);
  corert.get()->RegisterOpHandler("tfkernel", fallback_op_handler.get());

      resource_context_
          .GetOrCreateResource<tensorflow::tfd::EagerContextResource>(
              tensorflow::tfd::kEagerContextResourceName);
  auto expected_eager_context = eager_context_resource->GetTFEagerContext();
  auto* eager_context = expected_eager_context.get();

  // Initialize fallback kernels state with a custom intra-op thread pool.
  auto status = tensorflow::tfd::SetUpKernelFallbackCompatRequestContext(