#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/simple_hardware.h"

#include "tensorflow/compiler/mlir/lite/experimental/tac/common/targets.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/common/utils.h"
#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"
#include "tensorflow/compiler/mlir/lite/utils/arithmetic_count_util.h"

namespace mlir {
namespace TFL {
namespace tac {

#include "tensorflow/compiler/mlir/lite/experimental/tac/common/targets.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/common/utils.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/target_hardware.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/transforms/device_transform_gpu.h"
#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"

#include "tensorflow/compiler/mlir/lite/experimental/tac/common/cost.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/common/targets.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/common/utils.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/target_hardware.h"
#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"

namespace mlir {
namespace TFL {
namespace tac {
namespace {

syntax = "proto3";

package third_party.tensorflow.compiler.mlir.lite.experimental.tac;

// A list of filters for TAC users to run ops/functions on ML hardwares. The
// intuition is that, for ops/functions that can be run on ML hardware (e.g.
// EdgeTPU) and TFLite CPU, TAC users give a hint that they're more performant
// to run on TFLite CPU. These filters give the TAC users freedom to specify the
// parts that they want to use other hardware to accelerate.

#include "tensorflow/compiler/mlir/lite/experimental/tac/common/targets.h"
#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/nnapi_hardware.h"
#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h"

namespace mlir {
namespace TFL {
namespace tac {
namespace {

struct DeviceTransformNNAPIPass

#include "tensorflow/compiler/mlir/lite/experimental/tac/hardwares/target_hardware.h"
#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"

namespace mlir {
namespace TFL {
namespace tac {
// Gpu hardware class which handles GPU capabilities in TFLite.
// This is used by TAC to get op supported/ op cost estimates on GPU.
class GpuHardware : public TargetHardware {
 public:

  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

#include "mlir/Pass/PassManager.h"  // from @llvm-project
#include "mlir/Support/LLVM.h"  // from @llvm-project
#include "tensorflow/compiler/mlir/lite/common/tfl_pass_config.h"
#include "tensorflow/lite/toco/toco_flags.pb.h"

namespace tensorflow {

// Add the TF to TFLite passes, specified in the pass_config, into a
// pass_manager. The session object will be provided when the TF MLIR is

        "//tensorflow/lite/toco:toco_tooling",
        "//tensorflow/lite/toco:tooling_util",
        "//tensorflow/lite/toco:types_proto_cc",
        "//tensorflow/lite/toco/logging:conversion_log_util",
        "//tensorflow/lite/toco/logging:toco_conversion_log_proto_cc",
        "//third_party/python_runtime:headers",  # build_cleaner: keep; DNR: b/35864863
        "@com_google_absl//absl/container:flat_hash_set",

#include <string>

#include "llvm/ADT/StringRef.h"
#include "mlir/IR/Operation.h"  // from @llvm-project

namespace mlir {
namespace TFL {
namespace tac {

// Interface name here is the "hook" between the CallOp and FuncOps.
// Take the following example:
//
// call @func_1_CPU {tac.interface_name = "func_1"}
//
// "func_1" is the interface name where "func_1_cpu" is the real implementation