#include "tensorflow/compiler/mlir/lite/ir/tfl_ops.h"
#include "tensorflow/compiler/mlir/tensorflow/ir/tf_ops.h"

namespace mlir {
namespace TFL {
namespace tac {

// TODO(renjieliu): add more patterns.

// This basically:
// Pack => (Concat -> Reshape)
struct LowerPackIntoConcatReshape : public OpRewritePattern<TFL::PackOp> {
  using OpRewritePattern<TFL::PackOp>::OpRewritePattern;

//     plugin or core implementation:
//     * Struct that should be filled by the plugin: `TP_OptimizerConfigs`,
//       `TP_Optimizer`, `TP_OptimizerRegistrationParams`
//     * Struct that should be filled by the proper: `TF_GrapplerItem`,
//       `TF_GraphProperties`, `TF_FunctionLibraryDefinition`
//   * We use `struct_size` for version checking. It should be set both by
//     core and the plugin.

// Permutation from the OIHW (== (output features, input features, height,
// width)) tensor format to HWIO. This is commonly used to transpose convolution
// weights represented as OIHW format to HWIO, which is more desirable for
// certain downstream optimization passes (e.g. XLA).
inline constexpr std::array<int64_t, 4> kOihwToHwioPermutation = {2, 3, 1, 0};

// Returns true if the value has static shape.

#include "tensorflow/compiler/mlir/quantization/tensorflow/exported_model.pb.h"
#include "tensorflow/compiler/mlir/quantization/tensorflow/quantization_options.pb.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/protobuf/meta_graph.pb.h"
#include "tensorflow/python/lib/core/pybind11_lib.h"
#include "tsl/platform/protobuf.h"  // IWYU pragma: keep

namespace pybind11::detail {
namespace internal {

// Loads a TFLite flatbuffer, convert to TOSA for backend compilation and
// produce an MLIR bytecode file as output.
// TODO(jpienaar): Refactor this when we use more implicit module passing
// between calls to avoid serialization overhead.
void ExperimentalTFLiteToTosaBytecode(
    const std::string &flatbuffer_file, const std::string &tosa_bytecode_file,
    bool use_external_constant,

  std::shared_ptr<ShapeRefiner> refiner_ = nullptr;

  // If scope_used_ is not nullptr, op_name_ should be empty and
  // GetUniqueNameForOp can only be called once on this scope. More calls to
  // GetUniqueNameForOp will cause an error status to be set on this scope.
  std::shared_ptr<bool> scope_used_ = nullptr;

  const std::vector<Operation> control_deps_;

//   * Struct prefix indicates whether struct fields should be filled by the
//     plugin or core implementation:
//     * SE_ : set/filled by core unless explicitly marked otherwise.
//     * SP_ : set/filled by plugin unless explicitly marked otherwise.
//   * We use `struct_size` for version checking. It is exempt from the `SE/SP`
//     rule above and should be set both by core and the plugin.

TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_NewTensorHandle(const TF_Tensor* t,
                                                            TF_Status* status);
// Indicates that the caller will not be using `h` any more.
TF_CAPI_EXPORT extern void TFE_DeleteTensorHandle(TFE_TensorHandle* h);
TF_CAPI_EXPORT extern TF_DataType TFE_TensorHandleDataType(TFE_TensorHandle* h);

//     grad_outputs[0]->Ref();
//     grad_outputs[1]->Ref();
//     return OkStatus();
//   }
//   ~AddGradientFunction() override {}
// };
//
// GradientFunction* AddRegisterer(const ForwardOperation& op) {
//   // More complex gradient functions can use inputs/attrs etc. from the
//   // forward `op`.
//   return new AddGradientFunction;
// }
//
// Status RegisterGradients(GradientRegistry* registry) {

/// pointers) owned by core TensorFlow and guaranteed to not be `nullptr`.
///
/// To ensure ABI and API compatibility, we have out-of-bounds data that is used
/// by both core TensorFlow and the plugin at load time. We don't include this
/// data in the structures here to prevent cases when padding/packing enabled by
/// different compiler options breaks compatibility. For more details about how