// CHECK: tac.cost = 1.000000e+03
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  // CHECK: tac.cost = 1.000000e+03
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>

    %1 = call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/lite/toco/model_flags.pb.h"
#include "tensorflow/lite/toco/toco_flags.pb.h"
#include "tensorflow/lite/toco/types.pb.h"
#include "tsl/platform/errors.h"
#include "tsl/platform/protobuf.h"  // IWYU pragma: keep

namespace tensorflow {
namespace {

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    func.return %0 : tensor<100xf32>
  }
  func.func @func_2_GPU_FLOAT(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.cost = 8.040000e+01 : f32, tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} {

// CHECK:           %[[VAL_3:.*]] = tfl.add %[[VAL_0]], %[[VAL_1]] {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
// CHECK:           %[[VAL_4:.*]] = tfl.mul %[[VAL_3]], %[[VAL_2]] {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>

// CHECK:           %[[VAL_5:.*]] = tfl.add %[[VAL_0]], %[[VAL_1]] {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
// CHECK:           %[[VAL_6:.*]] = tfl.mul %[[VAL_5]], %[[VAL_2]] {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>

#include "mlir/Pass/PassManager.h"  // from @llvm-project
#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/target_hardware.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/tac_importer_exporter.h"

namespace mlir {
namespace TFL {
namespace tac {

// Main class for using Target Aware Conversion (TAC).
// To run TAC:
// 1) users should create object form this class, with desired options
//    (TacModule::Options).

  pass_manager->addPass(mlir::TFL::tac::CreateTargetAnnotationPass(this));
  pass_manager->addPass(mlir::TFL::tac::CreateRaiseTargetSubgraphsPass());
  pass_manager->addPass(mlir::TFL::tac::CreateFoldConstantsToSubgraphPass(
      /*fold_all_constants=*/false));
  pass_manager->addPass(
      mlir::TFL::tac::CreateAlternativeSubgraphPass(device_specs));
  if (options_.legalize_to_tflite_ops) {

// RUN: tac-opt-all-backends -tfl-compute-cost %s -split-input-file -verify-diagnostics | FileCheck %s

// CHECK: tac.cost = 7.864320e+05
func.func @func_0_CPU(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<256x32x32x3xf32>, tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32>

#include "mlir/Pass/Pass.h"  // from @llvm-project
#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/target_hardware.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/tac_module.h"

namespace mlir {
namespace TFL {
namespace tac {
// An OperationPass<> with access to the TAC module instance that the
// pass is running part of.