abstract: TensorFlow is a machine learning system that operates at large scale and in heterogeneous environments. TensorFlow uses dataflow graphs to represent computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility...

// clang-format on

namespace tensorflow {
namespace tfcompile {
namespace {

void zero_buffers(XlaCompiledCpuFunction* computation) {
  for (int i = 0; i < computation->num_args(); ++i) {
    memset(computation->arg_data(i), 0, computation->arg_size(i));
  }
}

// Trivial test that runs the generated function to ensure it doesn't crash.
TEST(TEST_NAME, NoCrash) {

        private Callable<T> computation;
        private Try<T> result;

        public Deferred(Callable<T> computation) {
            this.computation = computation;
        }

        @Override
        public T get() {
            return result().get();
        }

        private Try<T> result() {
            if (result == null) {

  let description = [{
This operation holds a replicated output from a `tpu.replicate()` computation subgraph.
Each replicated output has the same shape and type alongside the input.

For example:
```
%computation = "tf.Computation"()
%replicated_output:2 = "tf.TPUReplicatedOutput"(%computation)
```

  %1 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<i32>>>) -> tensor<i32>
  // CHECK:      [[COMPUTATION:%.+]] = "tf_device.cluster_func"([[INPUT]])
  %2 = "tf_device.cluster_func"(%1) {func = @computation, use_spmd_for_xla_partitioning = true} : (tensor<i32>) -> tensor<i32>
  // CHECK:      [[OUTPUT:%.+]]:2 = "tf.TPUPartitionedOutputV2"([[COMPUTATION]])
  // CHECK-SAME: _XlaSharding = ""
  // CHECK-SAME: partition_dims = []

// clang-format on

namespace tensorflow {
namespace tfcompile {
namespace {

void zero_buffers(XlaCompiledCpuFunction* computation) {
  for (int i = 0; i < computation->num_args(); ++i) {
    memset(computation->arg_data(i), 0, computation->arg_size(i));
  }
}

// Trivial test that runs the generated function to ensure it doesn't crash.
TEST(TEST_NAME, NoCrash) {

  return true;
}

namespace {

// Compiles the XLA computation into executable code.
Status CompileXla(xla::CompileOnlyClient* client,
                  const xla::XlaComputation& computation,
                  const xla::cpu::CpuAotCompilationOptions& aot_opts,
                  CompileResult* compile_result) {
  // Retrieves arg and result layouts from the computation.
  // TODO(toddw): Should we let the user choose the major/minor ordering?

  xla::Tuple(&builder, tuple_values);

  TF_ASSERT_OK_AND_ASSIGN(XlaComputation computation, builder.Build());
  EXPECT_EQ(computation.proto().computations_size(), 2);

  TF_ASSERT_OK(CreateMlirModule());
  TF_ASSERT_OK_AND_ASSIGN(TupleOp root_tuple,
                          ImportXlaComputationIntoModule(computation));
  EXPECT_TRUE(root_tuple);

  int num_func_ops = 0;

/**
 * Starts and exposes the given suspending [computation] as a [Future] value.
 *
 * The [computation] executes synchronously until its first suspension point.
 */
internal
fun <T> future(context: CoroutineContext = EmptyCoroutineContext, computation: suspend () -> T): Future<T> =
    FutureContinuation<T>(context).also { k ->
        computation.startCoroutine(completion = k)
    }


private

      /*lower_to_xla_hlo=*/true, computation, metadata, device_type,
      shape_determination_fns, use_tuple_args, compilation_result,
      custom_legalization_passes, arg_shapes, arg_core_mapping,
      per_core_arg_shapes);

  if (mlir_bridge_status.ok()) {
    VLOG(1) << "Successfully compiled MLIR computation to XLA HLO using MLIR "
               "tf2xla bridge";