// is just an estimation.
constexpr float kQuantizedInferenceEfficiency = 0.3;

inline float InferenceTypeEfficiency(InferenceType inference_type) {
  if (inference_type == QUANTIZED_INT8 || inference_type == QUANTIZED_UINT8) {
    return kQuantizedInferenceEfficiency;
  }
  return 1.0;
}

// CPU hardware class which handles CPU capabilities in TFLite.

    case toco::IODataType::QUANTIZED_UINT8:
      return tflite::TensorType_UINT8;
    case toco::IODataType::UINT8:
      return tflite::TensorType_UINT8;
    case toco::IODataType::QUANTIZED_INT8:
      return tflite::TensorType_INT8;
    case toco::IODataType::INT8:
      return tflite::TensorType_INT8;
    default:
      return tflite::TensorType_FLOAT32;
  }
}

    case toco::IODataType::FLOAT16:
      return DT_HALF;
    case toco::IODataType::FLOAT64:
      return DT_DOUBLE;
    case toco::IODataType::QUANTIZED_UINT8:
      return DT_QUINT8;
    case toco::IODataType::QUANTIZED_INT8:
      return DT_QINT8;
    case toco::IODataType::QUANTIZED_INT16:
      return DT_QINT16;
    case toco::IODataType::INT8:
      return DT_INT8;
    case toco::IODataType::INT16:
      return DT_INT16;

  ::tflite::optimize::BufferType quantized_type;
  switch (quant_specs.inference_type) {
    case DT_QINT8:
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_INT8;
      break;
    case DT_HALF:
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_FLOAT16;
      break;
    default:
      return absl::InvalidArgumentError("Quantized type not supported");