Firmware Configuration

When SCPU activates NCPU to run models, kmdw_inference_app_config_t (KL520/KL720) or VMF_NNM_INFERENCE_APP_CONFIG_T (KL630/KL730), which contains configurations of NCPU, is required to pass to NCPU.

kmdw_inference_app_config_t and VMF_NNM_INFERENCE_APP_CONFIG_T contain following configurable options:

• image_buf:

  • type: void *
  • The buffer address of the image.

• image_width:

  • type: uint32_t
  • The width of the image.

• image_height:

  • type: uint32_t
  • The height of the image.

• image_channel:

  • type: uint32_t
  • The channel count of the image.

• image_resize:

  • type: uint32_t
  • This is used for image resize in pre-process.
  • Please refer to kp_resize_mode_t.

• image_padding:

  • type: uint32_t
  • This is used for image padding in pre-process.
  • Please refer to kp_padding_mode_t.

• image_format:

  • type: uint32_t
  • This is used for color space conversion in pre-process.
  • Please refer to kp_image_format_t.

• image_norm:

  • type: uint32_t
  • This is used for data normalization in pre-process.
  • Please refer to kp_normalize_mode_t.

• bypass_pre_proc:

  • type: bool
  • If this is true, NCPU will NOT execute the pre-process. The size, channel, format, ... of the input image must be exactly the same as the requirement of the model.
  • If this is true, image_width, image_height, image_channel, image_resize, image_padding, image_format, and image_norm will be ignored.
  • If this is true, image_buf_size must be provided correctly.

• image_buf_size:

  • type: uint32_t
  • The size of image buffer.
  • This is only used when bypass_pre_proc is true.

• model_id:

  • type: int
  • The ID of the target model to be inferenced in NCPU.

• enable_crop:

  • type: bool
  • Whether crop a partial area of the image to be inferenced.
  • If this is true, crop_area must be set properly.

• crop_area:

  • type: kp_inf_crop_box_t
  • The cropping area of the image to be inferrenced.

• enable_raw_output:

  • type: bool
  • If this is true, NCPU does not execute the post-process. The post-process may need to be execute at the software on the host server.
  • If this is false, NCPU will execute the post-process. The customized post-process must be registered via kdpio_post_processing_register().

• enable_parallel:

  • type: bool
  • This is only available when single model is adapted and the post-process is executed in NCPU.
  • When one inference is in the post-process, the next inference will be start parallelly.
  • After one inference is fully finished, the callback function set to result_callback will be invoked.

• result_callback:

  • type [KL520/KL720]: kmdw_inference_result_callback_t
  • type [KL630/KL730]: VMF_NNM_INFERENCE_APP_RESULT_CALLBACK_T
  • The callback function for parallel mode

• inf_result_buf:

  • type: void*
  • This only works for parallel mode to carry it back to user callback function.

• ncpu_result_buf

  • type: void*
  • The buffer address where NCPU put the inference result.
  • It will be passed to result_callback under parallel mode.

• pad_value:

  • type: kp_pad_value_t*
  • The pad_value for the pre-processing in NCPU.

• user_define_data (KL520/KL720/KL630)

  • type: void*
  • The user define data for the pre-processing and post-processing in NCPU.

• pre_proc_config (KL730)

  • type: void*
  • The user define data for the pre-processing in NCPU.

• post_proc_config (KL730)

  • type: void*
  • The user define data for the post-processing in NCPU.

• pre_proc_func (KL630/KL730)

  • type: void*
  • The pointer for the customized pre-processing function.

• post_proc_func (KL630/KL730)

  • type: void*
  • The pointer for the customized post-processing function.