// Definition of AOTI runtime interface functions

#include <torch/csrc/inductor/aoti_runtime/interface.h>
#include <torch/csrc/inductor/aoti_runtime/model_container.h>

#include <iostream>
#include <vector>

#define CONVERT_EXCEPTION_TO_ERROR_CODE(...)      \
  try {                                           \
    __VA_ARGS__                                   \
  } catch (const std::exception& e) {             \
    std::cerr << "Error: " << e.what() << '\n';   \
    return AOTI_RUNTIME_FAILURE;                  \
  } catch (...) {                                 \
    std::cerr << "Unknown exception occurred.\n"; \
    return AOTI_RUNTIME_FAILURE;                  \
  }                                               \
  return AOTI_RUNTIME_SUCCESS;

#define AOTI_VECTOR_SIZE_CHECK(actual_size, expected_size, name)  \
  do {                                                            \
    AOTI_RUNTIME_CHECK(                                           \
        actual_size == expected_size,                             \
        "expected " + std::string(name) + " vector size to be " + \
            std::to_string(expected_size) + ", but got " +        \
            std::to_string(actual_size));                         \
  } while (0)

// AOTInductor uses at::addmm_out, which doesn't supports
// arguments that requires gradient. For this reason, we
// enforce no_grad context for run APIs.
//
// A RAII, thread local (!) guard that enables or disables grad mode upon
// construction, and sets it back to the original value upon destruction.
struct AOTINoGradGuard {
  AOTINoGradGuard() {
    aoti_torch_grad_mode_set_enabled(false);
  }
  AOTINoGradGuard(const AOTINoGradGuard&) = delete;
  AOTINoGradGuard(AOTINoGradGuard&&) noexcept = delete;
  ~AOTINoGradGuard() {
    aoti_torch_grad_mode_set_enabled(prev_mode);
  }
  AOTINoGradGuard& operator=(const AOTINoGradGuard&) = delete;
  AOTINoGradGuard& operator=(AOTINoGradGuard&&) noexcept = delete;
  bool prev_mode{aoti_torch_grad_mode_is_enabled()};
};

extern "C" {

AOTIRuntimeError AOTInductorModelContainerCreate(
    AOTInductorModelContainerHandle* container_handle,
    size_t num_models,
    bool is_cpu,
    const char* cubin_dir) {
      return AOTInductorModelContainerCreateWithDevice(
        container_handle,
        num_models,
        is_cpu ? "cpu" : "cuda",
        cubin_dir);
}

AOTIRuntimeError AOTInductorModelContainerCreateWithDevice(
    AOTInductorModelContainerHandle* container_handle,
    size_t num_models,
    const char* device_str,
    const char* cubin_dir) {

  if (num_models == 0) {
    std::cerr << "Error: num_models must be positive, but got 0\n";
    return AOTI_RUNTIME_FAILURE;
  }
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    std::optional<std::string> cubin_dir_opt;
    if (cubin_dir != nullptr) {
      cubin_dir_opt.emplace(cubin_dir);
    }
    auto* container = new torch::aot_inductor::AOTInductorModelContainer(
        num_models, std::string(device_str), cubin_dir_opt);
    *container_handle =
        reinterpret_cast<AOTInductorModelContainerHandle>(container);
  })
}


AOTIRuntimeError AOTInductorModelContainerDelete(
    AOTInductorModelContainerHandle container_handle) {
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    auto* container =
        reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
            container_handle);
    delete container;
  });
}

AOTIRuntimeError AOTInductorModelContainerRun(
    AOTInductorModelContainerHandle container_handle,
    AtenTensorHandle* input_handles, // array of input AtenTensorHandle; handles
                                     // are stolen; the array itself is borrowed
    size_t num_inputs,
    AtenTensorHandle*
        output_handles, // array for writing output AtenTensorHandle; handles
                        // will be stolen by the caller; the array itself is
                        // borrowed
    size_t num_outputs,
    AOTInductorStreamHandle stream_handle,
    AOTIProxyExecutorHandle proxy_executor_handle) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  AOTI_VECTOR_SIZE_CHECK(num_inputs, container->num_inputs(), "inputs");
  AOTI_VECTOR_SIZE_CHECK(num_outputs, container->num_outputs(), "outputs");

  auto stream =
      reinterpret_cast<torch::aot_inductor::DeviceStreamType>(stream_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    AOTINoGradGuard guard;
    container->run(
        input_handles, output_handles, stream, proxy_executor_handle);
  })
}

AOTIRuntimeError AOTInductorModelContainerRunSingleThreaded(
    AOTInductorModelContainerHandle container_handle,
    AtenTensorHandle* input_handles, // array of input AtenTensorHandle; handles
                                     // are stolen; the array itself is borrowed
    size_t num_inputs,
    AtenTensorHandle*
        output_handles, // array for writing output AtenTensorHandle; handles
                        // will be stolen by the caller; the array itself is
                        // borrowed
    size_t num_outputs,
    AOTInductorStreamHandle stream_handle,
    AOTIProxyExecutorHandle proxy_executor_handle) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  AOTI_VECTOR_SIZE_CHECK(num_inputs, container->num_inputs(), "inputs");
  AOTI_VECTOR_SIZE_CHECK(num_outputs, container->num_outputs(), "outputs");

  auto stream =
      reinterpret_cast<torch::aot_inductor::DeviceStreamType>(stream_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    AOTINoGradGuard guard;
    container->run_single_threaded(
        input_handles, output_handles, stream, proxy_executor_handle);
  })
}

AOTIRuntimeError AOTInductorModelContainerGetNumConstants(
    AOTInductorModelContainerHandle container_handle,
    size_t* num_constants) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { *num_constants = container->num_constants(); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantName(
    AOTInductorModelContainerHandle container_handle,
    size_t idx,
    const char** name) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { *name = container->constant_name(idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantOriginalFQN(
    AOTInductorModelContainerHandle container_handle,
    size_t idx,
    const char** original_fqn) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { *original_fqn = container->constant_original_fqn(idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantFromFolded(
    AOTInductorModelContainerHandle container_handle,
    size_t idx,
    bool* from_folded) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({ *from_folded = container->constant_from_folded(idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantType(
    AOTInductorModelContainerHandle container_handle,
    size_t idx,
    int32_t* type) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({ *type = container->constant_type(idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantDtype(
    AOTInductorModelContainerHandle container_handle,
    size_t idx,
    int32_t* dtype) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { *dtype = container->constant_dtype(idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantDataSize(
  AOTInductorModelContainerHandle container_handle,
  size_t idx,
  size_t* data_size) {
  auto* container =
    reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
        container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { *data_size = container->constant_data_size(idx); })
}

AOTIRuntimeError AOTInductorModelContainerExtractConstantsMap(
    AOTInductorModelContainerHandle container_handle,
    AOTInductorConstantMapHandle constant_map_handle,
    bool use_inactive) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  auto constants_map = reinterpret_cast<std::unordered_map<std::string, AtenTensorHandle>*>(constant_map_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
    { const auto ret = container->extract_constants_map(use_inactive);
      for (const auto& pair: ret) {
        constants_map->emplace(pair.first, pair.second);
      }
    })
}

AOTIRuntimeError AOTInductorModelContainerUpdateUserManagedConstantBuffer(
    AOTInductorModelContainerHandle container_handle,
    AOTInductorConstantMapHandle constant_map_handle,
    bool use_inactive,
    bool validate_full_update) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  auto input_map = reinterpret_cast<std::unordered_map<std::string, AtenTensorHandle>*>(constant_map_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    container->update_constant_buffer(
        *input_map, use_inactive, validate_full_update, /* user_managed = */ true);
  })
}

AOTIRuntimeError AOTInductorModelContainerUpdateUserManagedConstantBufferPairs(
    AOTInductorModelContainerHandle container_handle,
    const AOTInductorConstantMapEntry* pairs,
    size_t num_pairs,
    bool use_inactive,
    bool validate_full_update) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(container_handle);
  // Build a local unordered_map inside
  std::unordered_map<std::string, AtenTensorHandle> input_map;
  input_map.reserve(num_pairs);
  for (size_t i = 0; i < num_pairs; ++i) {
      input_map.emplace(pairs[i].name, pairs[i].handle);
  }
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    container->update_constant_buffer(
        input_map, use_inactive, validate_full_update, /*user_managed=*/true);
  })
}

AOTIRuntimeError AOTInductorModelContainerUpdateConstantBuffer(
    AOTInductorModelContainerHandle container_handle,
    AOTInductorConstantMapHandle constant_map_handle,
    bool use_inactive,
    bool validate_full_update) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  auto input_map = reinterpret_cast<std::unordered_map<std::string, AtenTensorHandle>*>(constant_map_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    container->update_constant_buffer(
        *input_map, use_inactive, validate_full_update);
  })
}

AOTIRuntimeError AOTInductorModelContainerUpdateInactiveConstantBuffer(
    AOTInductorModelContainerHandle container_handle,
    AOTInductorConstantMapHandle constant_map_handle) {
  return AOTInductorModelContainerUpdateConstantBuffer(container_handle,
          constant_map_handle,
          /*use_inactive*/ true,
          /*validate_full_update*/ true);
}

AOTIRuntimeError AOTInductorModelContainerFreeInactiveConstantBuffer(
    AOTInductorModelContainerHandle container_handle) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    container->free_inactive_constant_buffer();
  })
}

AOTIRuntimeError AOTInductorModelContainerRunConstantFolding(
    AOTInductorModelContainerHandle container_handle,
    bool use_inactive,
    AOTInductorStreamHandle stream_handle,
    AOTIProxyExecutorHandle proxy_executor_handle) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  auto stream =
      reinterpret_cast<torch::aot_inductor::DeviceStreamType>(stream_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    AOTINoGradGuard guard;
    container->run_const_fold(use_inactive, stream, proxy_executor_handle);
  })
}

AOTIRuntimeError AOTInductorModelContainerSwapConstantBuffer(
    AOTInductorModelContainerHandle container_handle) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    container->swap_constant_buffer();
  })
}

AOTIRuntimeError AOTInductorModelContainerGetNumInputs(
    AOTInductorModelContainerHandle container_handle,
    size_t* ret_num_inputs) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      { *ret_num_inputs = container->num_inputs(); })
}

AOTIRuntimeError AOTInductorModelContainerGetInputName(
    AOTInductorModelContainerHandle container_handle,
    size_t input_idx,
    const char** ret_input_names) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      { *ret_input_names = container->input_name(input_idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetNumOutputs(
    AOTInductorModelContainerHandle container_handle,
    size_t* ret_num_outputs) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      { *ret_num_outputs = container->num_outputs(); })
}

AOTIRuntimeError AOTInductorModelContainerGetOutputName(
    AOTInductorModelContainerHandle container_handle,
    size_t output_idx,
    const char** ret_output_names) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      { *ret_output_names = container->output_name(output_idx); })
}

AOTIRuntimeError AOTInductorModelContainerGetCallSpec(
    AOTInductorModelContainerHandle container_handle,
    const char** in_spec,
    const char** out_spec) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    *in_spec = container->get_in_spec();
    *out_spec = container->get_out_spec();
  })
}

AOTIRuntimeError AOTInductorModelCreate(
    AOTInductorModelHandle* model_handle,
    AOTInductorConstantMapHandle constant_map_handle){
    CONVERT_EXCEPTION_TO_ERROR_CODE({
      auto constant_map = std::make_shared<torch::aot_inductor::ConstantMap>();
      auto constant_array = std::make_shared<std::vector<torch::aot_inductor::ConstantHandle>>();
      auto input_map = reinterpret_cast<std::unordered_map<std::string, AtenTensorHandle>*>(constant_map_handle);

      auto model = new torch::aot_inductor::AOTInductorModel(
          constant_map,
          constant_array,
          "cpu", // device_str is hardcoded, as AOTInductorModelCreate is only use for CPU models
          ""
      );

      if (input_map) {
        for (auto const& kv : *input_map) {
          constant_map->emplace(kv.first, kv.second);
        }
      } else {
        model->load_constants();
      }

      *model_handle = reinterpret_cast<AOTInductorModelHandle>(model);
    })}

AOTIRuntimeError AOTInductorModelRun(
    AOTInductorModelHandle model_handle,
    AtenTensorHandle* input_handles,
    AtenTensorHandle* output_handles) {
  auto model =
      reinterpret_cast<torch::aot_inductor::AOTInductorModel*>(model_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    AOTINoGradGuard guard;
    model->run_impl(
        input_handles,
        output_handles,
        (torch::aot_inductor::DeviceStreamType) nullptr,
        nullptr);
  })
}

AOTIRuntimeError AOTInductorModelDelete(AOTInductorModelHandle model_handle){
    CONVERT_EXCEPTION_TO_ERROR_CODE({
      auto model = reinterpret_cast<torch::aot_inductor::AOTInductorModel*>(
          model_handle);
      delete model;
    })}

AOTIRuntimeError AOTInductorModelGetNumOutputs(
    AOTInductorModelHandle model_handle,
    size_t* ret_num_outputs) {
  CONVERT_EXCEPTION_TO_ERROR_CODE({
      auto model = reinterpret_cast<torch::aot_inductor::AOTInductorModel*>(model_handle);
      *ret_num_outputs = model->num_outputs();
  })
}

AOTIRuntimeError AOTInductorModelUpdateConstantsMap(
    AOTInductorModelHandle model_handle,
    AOTInductorConstantMapHandle constant_map_handle) {
  auto model =
      reinterpret_cast<torch::aot_inductor::AOTInductorModel*>(model_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE({
    auto constant_map = std::make_shared<torch::aot_inductor::ConstantMap>();
    auto input_map =
        reinterpret_cast<std::unordered_map<std::string, AtenTensorHandle>*>(
            constant_map_handle);

    for (auto const& kv : *input_map) {
      constant_map->emplace(kv.first, kv.second);
    }
    model->update_constants_map(std::move(constant_map));
  })
}

AOTIRuntimeError AOTInductorModelContainerGetConstantsBlobSize(
    AOTInductorModelContainerHandle container_handle,
    uint64_t* ret_size) {
  auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      { *ret_size = container->constant_blob_size(); })
}


// Load weights from a single blob in weight_blob_ptr
AOTIRuntimeError AOTInductorModelUpdateConstantsFromBlob(
    AOTInductorModelContainerHandle container_handle,
    const uint8_t* weight_blob_ptr){
    auto* container =
      reinterpret_cast<torch::aot_inductor::AOTInductorModelContainer*>(
          container_handle);
  CONVERT_EXCEPTION_TO_ERROR_CODE(
      {container->update_constants_from_blob(weight_blob_ptr); })
    }


} // extern "C"