Skip to content

README.md

Latest commit

 

History

History
1254 lines (1076 loc) · 67.9 KB

README.md

File metadata and controls

1254 lines (1076 loc) · 67.9 KB
Raw

Menu:

QCNode Sample Application

This QCNode sample is an application to demonstrate how to use the QCNode components.

1. QCNode Sample Application command line arguments

Below is a sample that how to use command line arguments to tell the QCNode sample application to create the component with the specified arguments.

export QC_LOG_LEVEL=INFO
./bin/qcrun ./bin/QCNodeSampleApp \
  -n component0_name -t component0_type \
    -k component0_attr1_name -v component0_attr1_value \
    -k component0_attr2_name -v component0_attr2_value \
    ... \
  -n componentX_name -t componentX_type \
    -k componentX_attr1_name -v componentX_attr1_value \
    -k componentX_attr2_name -v componentX_attr2_value \

Note: the "-n componentX_name -t componentX_type" must be in the begin for each component, and all the repeated "-k componentX_attr?_name -v componentX_attr?_value" are going to specify the attributes for this component.

parameter required type comments
-n true string The unique component name
-t true string The component type name, options from [DataReader, Camera, Remap, Qnn, C2D, PostProcCenternet, TinyViz, VideoEncoder, VideoDecoder, Recorder, PlrPre, PlrPost, DataOnline, CL2DFlex, GL2DFlex, SharedRing, FpsAdapter, OpticalFlow, OpticalFlowViz, FrameSync, DepthFromStereo, DepthFromStereoViz, Radar, C2C, Temporal, ComputeLidarCoord]
-k true string The unique component attribute name
-v true string The attribute value for the previous attribute name
-c false string Specify the client name for HeteroScheduler
-d false - Direct the QCNode log to stdout
-T false int Specify the time in seconds that the QCNodeSampleApp runs, if not specified or value 0, it means that the QCNodeSampleApp will run forever until stop signal(Ctrl + C).
-V false - Prints the QCNode application version information to the standard output.

2. QCNode Samples

2.1 QCNode DataReader Sample

attribute required type default comments
number false int 1 The number of simulated sensors
typeX false string "image" The buffer type for the simulated sensor X, options from [image, tensor]
formatX false string "nv12" The image format for the simulated camera X, options from [nv12, uyvy, rgb, bgr, p010, nv12_ubwc]
widthX false int 1920 The image width for the simulated camera X
heightX false int 1024 The image height for the simulated camera X
data_pathX true string - The data path for the simulated sensor X that contain the image files
tensor_typeX false string float32 The tensor type, options from [int8, int16, int32, int64, uint8, uint16, uint32, uint64, float16, float32, float64, sfixed_point8, sfixed_point16, sfixed_point32, ufixed_point8, ufixed_point16, ufixed_point32 ]
dimsX true string - The tensor dimensions, in format "N,H,W,C", "N,S,C", "N,C", or "N" depends on the number of dimensions.
offset false int 0 Specifies the starting index for the simulated sensor data.
fps false int 30 The frame rate per second
pool_size false int 4 the image memory pool size
cache false bool true use cached memory or not for the image memory
topic true string - the output topic name

Note: "X" is value from 0 to number-1, thus the attribute with suffix "X" is repeated for different simulated camera.

The command line template example for image type data reader that simulate a camera:

  -n CAM0 -t DataReader -k number -v 1 \
    -k format0 -v uyvy -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k pool_size -v 4 \
    -k topic -v /sensor/camera/CAM0/raw \

Refer DataReader Utils for how to generate a data reader inputs from video(*.mp4).

The command line template example for tensor type data reader that simulate a lidar:

  -n LIDAR0 -t DataReader -k number -v 2 \
    -k type0 -v tensor -k tensor_type0 -v float32 -k dims0 -v "300000,4"  \
    -k data_path0 -v /data/LIDAR0 \
    -k type1 -v image -k format1 -v nv12 -k width1 -v 1920 -k height1 -v 1024 \
    -k data_path1 -v /data/LIDAR0 \
    -k pool_size -v 4 -k fps -v 10 \
    -k topic -v /sensor/lidar/LIDAR0/raw \

Please note that for lidar pipeline, TinyViz was used to visualize the pointcould, and to save computing resource to dynamic generate images from pointcloud, pre-generated images was used, that's why another "data_path1".

Refer DataReader Utils for how to generate a data reader inputs point cloud files.

2.2 QCNode Camera Sample

attribute required type default comments
number false int 1 The number of camera streams
client_id false int 0 client id, used for multi client usecase, set to 0 by default for single client usecase
is_primary false bool false Flag to indicate if the session is primary or not when configured the clientId
recovery false bool false Flag to enable the self-recovery for the session
input_id true int - The camera input id
src_id false int 0 Input source identifier
input_mode false int 0 The input mode id
width true int - The image width for the stream 0
height true int - The image height for the stream 0
format false string "nv12" The camera frame format for the stream 0, options from [nv12, uyvy, p010, nv12_ubwc, tp10_ubwc, rgb, bgr]
stream_id false int 0 The camera stream id for the stream 0
pool_size false int 4 The image memory pool size for the stream 0
submit_request_pattern false int 0 The submit request pattern for the stream 0
request_mode false bool false The camera request mode
frame_drop_patten false int 0 The frame drop patten defined by qcarcam
frame_drop_period false int 0 The frame drop period defined by qcarcam
isp_use_case false int 3 The ISP use case
op_mode false int 2 The input operation mode, 1: Inline ISP, 2: Injection to ISP.
multi_stream_frame_ready false bool false Flag to set multiple streams frame ready event in one callback.
ignore_error false bool false Ignore the error of Camera Init&Start
immediate_release false bool false Perform an immediate camera frame release in the camera frame ready callback if true. Be cautious, as this approach does not provide life cycle management for the camera buffer, and data consistency is not guaranteed.
topic true string - The output topic name
widthX false int - The image width for the stream X
heightX false int - The image height for the stream X
formatX false string "nv12" The camera frame format for the stream X, options from [nv12, uyvy, p010, nv12_ubwc, tp10_ubwc, rgb, bgr]
stream_idX false int ${X} The camera stream id for the stream X
pool_sizeX false int 4 The image memory pool size for the stream X
submit_request_patternX false int 0 The submit request pattern for the stream X
topicX false string - The output topic name for the stream X

Note: "X" is value from 1 to number-1, thus the attribute with suffix "X" is repeated for different streams.

The command line template example:

  • Basic request mode:
  -n CAM0 -t Camera -k input_id -v 0 \
    -k width -v 1928 -k height -v 1208 \
    -k request_mode -v true \
    -k topic -v /sensor/camera/CAM0/raw \
  • Multi-stream mode with request pattern:
-n CAM0 -t Camera -k input_id -v 8 -k number -v 2 \
    -k isp_use_case -v 65 -k request_mode -v true \
    -k multi_stream_frame_ready -v true \
    -k stream_id -v 1 \
    -k submit_request_pattern -v 1 \
    -k width -v 3840 -k height -v 2160 \
    -k pool_size -v 4 \
    -k topic -v /sensor/camera/CAM0_0/raw \
    -k stream_id1 -v 5 \
    -k submit_request_pattern1 -v 0 \
    -k width1 -v 3840 -k height1 -v 2160 \
    -k pool_size1 -v 4 \
    -k topic1 -v /sensor/camera/CAM0_1/raw

2.3 QCNode C2D Sample

attribute required type default comments
output_width false int 1920 The output image width
output_height false int 1024 The output image height
output_format false string nv12 The output image format, options from [uyvy, nv12, rgb, bgr, p010 ]
batch_size false int 1 The Remap input batch size
input_widthX false int 1920 The input X image width
input_heightX false int 1024 The input X image height
input_formatX false string uyvy The input X image format, options from [uyvy, nv12, rgb, bgr, p010 ]
roi_xX false int 0 The ROI top x for input X
roi_yX false int 0 The ROI top y for input X
roi_widthX false int =output_width The ROI width for input X
roi_heightX false int =output_height The ROI height for input X
pool_size false int 4 the image memory pool size
cache false bool true use cached memory or not for the image memory
input_topic true string - the input topic name
output_topic true string - the output topic name

Note: "X" is value from 0 to batch_size-1, thus the attribute with suffix "X" is repeated for different input batch.

The command line template example:

  -n C2D0 -t C2D -k batch_size -v 1 \
    -k input_width0 -v 2048 -k input_height0 -v 1216 -k input_format0 -v nv12 \
    -k roi_x0 -v 0 -k roi_y0 -v 0 -k roi_width0 -v 1928 -k roi_height0 -v 1208 \
    -k output_width -v 2048 -k output_height -v 1216 -k output_format -v uyvy \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/uyvy \
    -k cache -v false

2.4 QCNode Remap Sample

attribute required type default comments
processor false string "htp0" The processor type, options from [htp0, htp1, cpu, gpu]
core_id false int 0 The processor core ID, options from [0,1,2,3]
rsm_priority false int 0 the RSM request priority, options [0,1,2,3], 0 is the lowest and 3 is highest priority.
output_width false int 1152 The output image width
output_height false int 800 The output image height
output_format false string rgb The output image format, options from [rgb, bgr]
batch_size false int 1 The Remap input batch size
input_widthX false int 1920 The input X image width
input_heightX false int 1024 The input X image height
input_formatX false string uyvy The input X image format, options from [uyvy, nv12, nv12_ubwc]
map_widthX false int =output_width The map width for input X
map_heightX false int =output_height The map height for input X
roi_xX false int 0 The ROI top x for input X
roi_yX false int 0 The ROI top y for input X
roi_widthX false int =output_width The ROI width for input X
roi_heightX false int =output_height The ROI height for input X
pool_size false int 4 the image memory pool size
normalize false bool true enable or disable normalization
map_table false bool false enable or disable create map from map table pointers
mapX_pathX false string "./data/test/remap/mapX.raw" map table path of width direction for input X
mapY_pathX false string "./data/test/remap/mapY.raw" map table path of height direction for input X
quant_scale false float 0.0186584480106831 The quantization scale of the quantize model input
quant_offset false int 114 The quantization offset of the quantize model input
Rsub false float 123.675 The sub normalization parameter of R channel
Rmul false float 0.0171 The mul normalization parameter of R channel
Radd false float 0.0 The add normalization parameter of R channel
Gsub false float 116.28 The sub normalization parameter of G channel
Gmul false float 0.0175 The mul normalization parameter of G channel
Gadd false float 0.0 The add normalization parameter of G channel
Bsub false float 103.53 The sub normalization parameter of B channel
Bmul false float 0.0174 The mul normalization parameter of B channel
Badd false float 0.0 The add normalization parameter of B channel
input_topic true string - the input topic name
output_topic true string - the output topic name

Note: "X" is value from 0 to batch_size-1, thus the attribute with suffix "X" is repeated for different input batch.

The command line template example:

  -n REMAP0 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 2048 -k input_height0 -v 1216 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM0/uyvy \
    -k output_topic -v /sensor/camera/CAM0/remap \

2.5 QCNode Qnn Sample

attribute required type default comments
processor false string "htp0" The processor type, options from [htp0, htp1, cpu, gpu]
core_ids false int list 0 Comma-separated list of core IDs on which the QNN model will run.
rsm_priority false int 0 the RSM request priority, options [0,1,2,3], 0 is the lowest and 3 is highest priority.
async fasle bool false If true, enable to use QNN Async Execute API, else use QNN Sync Execute API.
model_path true string - The QNN model path
load_type false string binary The QNN model load type, options: binary, library
image_convert false string default The image convert method for input image, options from [default, gray, chroma_first]. default: the default way to convert the image to 1 or more tenors. gray: convert the image to 1 tensor with the gray(luminance) part of the image. chroma_first: the same as default for NV12 or P010, but with chroma tensor first and then luma tensor.
pool_size false int 4 the image memory pool size
input_topic true string - the input topic name
output_topic true string - the output topic name
udo false string - udo lib path and interface provider name. e.g. libQnnAutoAiswOpPackage.so:AutoAiswOpPackageInterfaceProvider
model_io_info_topic false string "" if configured, this topic will be used to publish the input/output tensor informatin of the model, generally used by the QCNode DataOnline Sample for the QNN online inference.
perf_profile false string burst Specifies perf profile to set.
Options: low_balanced, balanced, default, high_performance, sustained_high_performance, burst, low_power_saver, power_saver, high_power_saver, extreme_power_saver
Default: default
weight_sharing_enabled fasle bool false If true, enable the weight sharing.
extended_udma fasle bool false If true, enable the extended udma feature.

The command line template example:

  -n CNT0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM0/remap \
    -k output_topic -v /sensor/camera/CAM0/qnn \

2.6 QCNode PostProcCenternet Sample

attribute required type default comments
processor false string "cpu" The processor type, options from [cpu, gpu]
roi_x false int 0 The ROI top x
roi_y false int 0 The ROI top y
width false int 1920 The ROI width
height false int 1024 The ROI height
score_threshold false float 0.6 The score threshold
nms_threshold false float 0.6 The NMS threshold
pool_size false int 4 the image memory pool size
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n POSTPROC_CNT0 -t PostProcCenternet \
    -k processor -v cpu \
    -k width -v 1928 -k height -v 1028 \
    -k input_topic -v /sensor/camera/CAM0/qnn \
    -k output_topic -v /sensor/camera/CAM0/objs \

2.7 QCNode TinyViz Sample

attribute required type default comments
winW false int 1920 The window width
winH false int 1080 The window height
cameras true string list - The cameras' name list
cam_topicX false string /sensor/camera/${cameras[X]}/raw the input camera frame topic name for camera X
obj_topicX false string /sensor/camera/${cameras[X]}/objs the input road object topic name for camera X
batch_indexX false int 0 The camera frame index in the data frames

The command line template example:

  -n VIZ -t TinyViz -k cameras -v CAM0,CAM1,CAM2,CAM3

The command line template example for Lidar pipeline:

  -n VIZ -t TinyViz -k cameras -v LIDAR0 \
    -k batch_index0 -v 1 \
    -k cam_topic0 -v /sensor/lidar/LIDAR0/raw \
    -k obj_topic0 -v /sensor/lidar/LIDAR0/objs

2.8 QCNode VideoEncoder Sample

attribute required type default comments
width true int - The image width
height true int - The image height
pool_size false int 4 The image memory pool size
bitrate false int 8000000 The encoding bitrate
fps false int 30 The frame rate per second
buffer_size false int 2*width*height The image buffer size
input_topic true string - the input topic name
output_topic true string - the output topic name
format false string "nv12" The image format, options from [nv12, nv12_ubwc]
output_format false string h265 The output image format, options from [h264, h265]
profile false string HEVC_MAIN for h265 or H264_MAIN for h264 The video encoder profile, options from [H264_BASELINE, H264_HIGH, H264_MAIN, HEVC_MAIN, HEVC_MAIN10]
gop false int 20 No of P-Pictures between 2 I-Frames
numInputBufferReq false int $pool_size Number of input buffers
numOutputBufferReq false int $pool_size Number of output buffers

The command line template example:

  -n VENC0 -t VideoEncoder -k width -v 1920 -k height -v 1024 \
    -k bitrate -v 8000000 \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/hevc \

2.9 QCNode Recorder Sample

attribute required type default comments
max false int 1000 The maximum recorded images
topic true string - the input topic name

For compressed image format, all the images are saved into 1 file with name "/tmp/${name}.raw". For the non-compressed image format, with separated images with name "/tmp/${name}_${id}_${batch_id}.raw" to save the real image content. The file with name "/tmp/${name}.meta" is used to record image information, below is an example:

$ cat /tmp/REC0.meta
0: frameId 0 timestamp 322037864334718: batch=3 resolution=1024x768 stride=2048 actual_height=768 format=2
$ ls -l /tmp/*.raw
-rw-rw-r--   2 root      root        1572864 Jan 04 17:27 /tmp/REC0_0_0.raw
-rw-rw-r--   2 root      root        1572864 Jan 04 17:27 /tmp/REC0_0_1.raw
-rw-rw-r--   2 root      root        1572864 Jan 04 17:27 /tmp/REC0_0_2.raw
-rw-rw-r--   2 root      root            107 Jan 04 17:27 /tmp/REC0.meta

The command line template example:

  -n REC0 -t Recorder -k max -v 100 -k topic -v /sensor/camera/CAM0/hevc \

2.10 QCNode PlrPre Sample

attribute required type default comments
processor false string "htp0" The processor type, options from [htp0, htp1, cpu, gpu]
core_id false int 0 The processor core ID, options from [0,1,2,3]
rsm_priority false int 0 the RSM request priority, options [0,1,2,3], 0 is the lowest and 3 is highest priority.
pillar_size_x false float 0.16 Pillar size in x direction in meters
pillar_size_y false float 0.16 Pillar size in y direction in meters
pillar_size_z false float 4.0 Pillar size in z direction in meters
min_x false float 0.0 Minimum range value in x direction
min_y false float -39.68 Minimum range value in y direction
min_z false float -3.0 Minimum range value in z direction
max_x false float 69.12 Maximum range value in x direction
max_y false float 39.68 Maximum range value in x direction
max_z false float 1 Maximum range value in x direction
max_points false int 300000 Maximum number of points in input point cloud
in_feature_dim false int 4 Number of features for each point in the input point cloud data
max_pillars false int 12000 Maximum number of point pillars that can be created
max_points_per_pillar false int 32 Maximum number of pMaximum number of points to map to each pillar
out_feature_dim false int 10 Number of features for each point in point pillars
input_mode false string "xyzr" The input pointclouds type used for OpenCL implementation, options from [xyzr, xyzrt]
pool_size false int 4 the image memory pool size
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n PLRPRE0 -t PlrPre -k processor -v htp0 \
    -k input_topic -v /sensor/lidar/LIDAR0/raw \
    -k output_topic -v /sensor/lidar/LIDAR0/plrpre \

2.11 QCNode PlrPost Sample

attribute required type default comments
processor false string "htp0" The processor type, options from [htp0, htp1, cpu]
rsm_priority false int 0 the RSM request priority, options [0,1,2,3], 0 is the lowest and 3 is highest priority.
pillar_size_x false float 0.16 Pillar size in x direction in meters
pillar_size_y false float 0.16 Pillar size in y direction in meters
min_x false float 0.0 Minimum range value in x direction
min_y false float -39.68 Minimum range value in y direction
max_x false float 69.12 Maximum range value in x direction
max_y false float 39.68 Maximum range value in x direction
max_points false int 300000 Maximum number of points in input point cloud
in_feature_dim false int 4 Number of features for each point in the input point cloud data
max_det_out false int 500 Maximum number of 3D bounding boxes expected in the output
stride false int 2 The downsample ratio
thresh_score false float 0.4 Confidence score threshold
thresh_iou false float 0.4 NMS Overlap threshold
offset_x false float 514 The min_x corresponding pixel position x of the pre-generated lidar image
offset_y false float 514 The min_y corresponding pixel position y of the pre-generated lidar image
ratio_w false float 12.903225806451614 The ration to transfrom the point cloud position x to image pixel position x
ratio_h false float 12.903225806451614 The ration to transfrom the point cloud position y to image pixel position y
debug false bool false print out the detected 3d bounding box
output_indexs false int list 3,0,1,4,2 The index of the pointpillar model outputs "heapmap", "center", "center_z", "dim" and "rot"
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n PLRPOST0 -t PlrPost -k processor -v htp0 -k debug -v false \
    -k input_lidar_topic -v /sensor/lidar/LIDAR0/raw \
    -k input_topic -v /sensor/lidar/LIDAR0/qnn \
    -k output_topic -v /sensor/lidar/LIDAR0/objs

2.12 QCNode DataOnline Sample

attribute required type default comments
port false int 6666 The TCP server socket port number
pool_size false int 4 the image memory pool size
input_topic true string - the input topic name
output_topic true string - the output topic name
model_io_info_topic false string "/data/online/${name}/model/info" This topic will be used to subscribe the input/output tensor informatin of the QNN model which will consume the online datas from host PC tool.
cache false bool true use cached memory or not for the buffer pool
timeout false int 2 socket send timeout in seconds

The command line template example:

  -n DO0 -t DataOnline -k port -v 6666 \
    -k input_topic -v /data/online/DO0/input \
    -k output_topic -v /data/online/DO0/output \
    -k model_io_info_topic -v /data/online/DO0/model/info \

2.13 QCNode CL2DFlex Sample

attribute required type default comments
batch_size false int 1 The CL2DFlex input batch size
input_widthX false int 1920 The input X image width
input_heightX false int 1024 The input X image height
input_formatX false string nv12 The input X image format, options from [uyvy, nv12, rgb, nv12_ubwc]
deviceId false int 0 The OpenCL device id
output_width false int 1920 The output image width
output_height false int 1024 The output image height
output_format false string rgb The output image format, options from [rgb, nv12]
work_modeX false string resize_nearest The input X work mode, options from [convert, resize_nearest, letterbox_nearest, letterbox_nearest_multiple, resize_nearest_multiple, convert_ubwc]
roi_xX false int 0 The ROI top x for input X
roi_yX false int 0 The ROI top y for input X
roi_widthX false int =input_widthX The ROI width for input X
roi_heightX false int =input_heightX The ROI height for input X
roi_number false int 1 The roi configurations number used for ExecuteWithROI
map_table false bool false enable or disable create map from map table pointers
mapX_pathX false string "./data/test/CL2DFlex/mapX.raw" map table path of width direction for input X
mapY_pathX false string "./data/test/CL2DFlex/mapY.raw" map table path of height direction for input X
pool_size false int 4 the image memory pool size
cache false bool true use cached memory or not for the image memory
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n CL2D -t CL2DFlex \
    -k deviceId -v 0 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v nv12 \
    -k batch_size -v 1 -k work_mode0 -v letterbox_nearest \
    -k roi_x0 -v 960 -k roi_y0 -v 512 -k roi_width0 -v 960 -k roi_height0 -v 512 \
    -k output_width -v 960 -k output_height -v 512 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/cl2d \

2.14 QCNode GL2DFlex Sample

attribute required type default comments
input_widthX false int 1920 The input X image width
input_heightX false int 1024 The input X image height
input_formatX false string nv12 The input X image format, options from [nv12, uyvy]
output_format false string rgb The output image format, options from [rgb, nv12]
roi_xX false int 0 The ROI top x for input frame X
roi_yX false int 0 The ROI top y for input frame X
roi_widthX false int =output_width The ROI width for input X
roi_heightX false int =output_height The ROI height for input X
pool_size false int 4 the image memory pool size
cache false bool true use cached memory or not for the image memory
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n GL2D0 -t GL2DFlex -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k roi_x0 -v 100 -k roi_y0 -v 100 -k roi_width0 -v 600 -k roi_height0 -v 600 \
    -k output_width -v 1024 -k output_height -v 768 -k output_format -v nv12 \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM1/raw

2.15 QCNode Shared Ring Sample

attribute required type default comments
topic true string - the topic name
type false string pub valid type: "pub" or "sub". If the type is "pub", get the message from DataBroker and forward it to the Shared Ring, else get the message from the Shared Ring and forward it to DataBroker.
queue_depth false int 2 the subscriber queue depth

The command line template example:

# for type pub
  -n CAM0_P -t SharedRing -k type -v pub \
    -k topic -v /sensor/camera/CAM0/raw \

# for type sub
  -n CAM0_S0 -t SharedRing -k type -v sub \
    -k topic -v /sensor/camera/CAM0/raw \

2.16 QCNode FpsAdapter Sample

attribute required type default comments
frame_drop_patten false int 0 The frame drop bit pattern is defined by FpsAdapter, where a value of 1 keeps the frame and a value of 0 drops the frame. The period is determined by the index of the highest bit set to 1, plus 1. For example, with 0b100, the highest bit set to 1 is at index 2, so the period is 3. This means that only 1 frame is kept out of every 3 frames, reducing the FPS from 30 to 10.
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n FpsAdpter0 -t FpsAdapter \
    -k frame_drop_patten -v 0b100 \
    -k input_topic -v /sensor/camera/CAM0/raw_30fps \
    -k output_topic -v /sensor/camera/CAM0/raw_10fps

2.17 QCNode VideoDecoder Sample

attribute required type default comments
width true int - The frame width
height true int - The frame height
input_format false string h265 The input frame format, options from [h264, h265]
output_format false string nv12 The output image format, options from [nv12, nv12_ubwc, p010]
pool_size false int 4 The image memory pool size
fps false int 30 The frame rate per second
input_topic true string - the input topic name
output_topic true string - the output topic name
numInputBufferReq false int $pool_size Number of input buffers
numOutputBufferReq false int $pool_size Number of output buffers

The command line template example:

  -n VDEC1 -t VideoDecoder -k width -v 1920 -k height -v 1024 \
    -k input_format -v h265 \
    -k output_format -v nv12 \
    -k input_topic -v /sensor/camera/CAM0/hevc \
    -k output_topic -v /sensor/camera/CAM0_DEC/raw \

2.18 QCNode OpticalFlow Sample

attribute required type default comments
eva_mode false string dsp the eval filter mode, options from [dsp, cpu, disable]
direction false string forward the opticalflow direction, options from [forward, backward]
step_size false int 1 the step size, options from [0, 1, 2]
width true int - The input image width
height true int - The input image height
format false string nv12 The input image format, options from [nv12, nv12_ubwc]
pool_size false int 4 The image memory pool size
fps false int 30 The frame rate per second
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n OFL0 -t OpticalFlow -k eva_mode -v cpu \
    -k width -v 1920 -k height -v 1024 -k format -v nv12 \
    -k fps -v 30 -k step_size -v 0 \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/mv \

2.19 QCNode OpticalFlowViz Sample

The Sample OpticalFlowViz converts the motion vection output from the Sample OpticalFlow to RGB image.

attribute required type default comments
processor false string gpu The processor type, options from [cpu, gpu]
width true int - The input image width
height true int - The input image height
pool_size false int 4 The image memory pool size
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n OFLVIZ0 -t OpticalFlowViz \
    -k width -v 960 -k height -v 512 \
    -k input_topic -v /sensor/camera/CAM0/mv \
    -k output_topic -v /sensor/camera/CAM1/raw \

2.20 QCNode FrameSync Sample

attribute required type default comments
number false int 1 The number of input topics
mode false string window The frame sync mode, options from [window]
window false int 100 the window time in ms
perms false int list - A list of integers to permute the output frame order
input_topicX true string - the input topic name for input X
output_topic true string - the output topic name

The command line template example:

  -n FS0 -t FrameSync -k number -v 4 \
    -k input_topic0 -v /sensor/camera/CAM0/raw \
    -k input_topic1 -v /sensor/camera/CAM1/raw \
    -k input_topic2 -v /sensor/camera/CAM2/raw \
    -k input_topic3 -v /sensor/camera/CAM3/raw \
    -k output_topic -v /sensor/camera/FS0/raw \

2.21 QCNode DepthFromStereo Sample

attribute required type default comments
direction false string l2r the search direction, options from [l2r, r2l]
width true int - The input image width
height true int - The input image height
format false string nv12 The input image format, options from [nv12, nv12_ubwc, p010, tp10_ubwc]
pool_size false int 4 The image memory pool size
fps false int 30 The frame rate per second
cache false bool true use cached memory or not for the buffer pool
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n DFS0 -t DepthFromStereo \
    -k width -v 1280 -k height -v 416 -k format -v nv12 \
    -k fps -v 30 -k direction -v l2r \
    -k input_topic -v /sensor/camera/DFS0/raw \
    -k output_topic -v /sensor/camera/DFS0/dfs \

2.22 QCNode DepthFromStereoViz Sample

The Sample DepthFromStereoViz converts the disparity and confidence map output from the Sample DepthFromStereo to RGB image.

attribute required type default comments
processor false string gpu The processor type, options from [cpu, gpu]
width true int - The input image width
height true int - The input image height
pool_size false int 4 The image memory pool size
disparity_max false int 1008 The maximum value of the disparity
conf_threshold false int 0 The threshold value of the confidence
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n DFSVIZ0 -t DepthFromStereoViz -k processor -v gpu \
    -k width -v 1280 -k height -v 416 \
    -k input_topic -v /sensor/camera/DFS0/dfs \
    -k output_topic -v /sensor/camera/CAM2/raw \

2.23 QCNode VideoDemuxer Sample

attribute required type default comments
input_file true string - The input video file path
start_frame_idx false int 0 The index of beginning frame for demuxer
replay_mode true bool false The option to enable replay mode
fps false int 30 The frame rate per second
topic true string - the topic name

The command line template example:

  -n VDMUX0 -t VideoDemuxer \
    -k input_file -v /data/test/test.mp4 \
    -k replay_mode -v true \
    -k start_time -v 0 \
    -k playback_time 2000 \
    -k topic  -v /sensor/camera/CAM0/hevc

2.24 QCNode Radar Sample

attribute required type default comments
max_input_buffer_size false int 1048576 Maximum input buffer size in bytes (1MB default, max 100MB)
max_output_buffer_size false int 1048576 Maximum output buffer size in bytes (1MB default, max 100MB)
service_name false string "/dev/radar0" Radar service device path
timeout_ms false int 5000 Processing timeout in milliseconds (max 60 seconds)
enable_performance_log false bool false Enable/disable performance logging
pool_size false int 4 Number of buffers in output buffer pool (max 32)
cache false bool true Enable/disable cached memory allocation
input_topic true string - Input topic name for radar data
output_topic true string - Output topic name for processed results

The QC Radar Sample demonstrates hardware-accelerated radar data processing using the QC Radar Node component. It provides a complete radar processing pipeline that can read radar data from input topics, process it using hardware acceleration, and publish the results to output topics.

Key Features:

  • Hardware-accelerated radar processing via QC Radar Node
  • Configurable buffer sizes and processing parameters
  • Zero-copy buffer sharing between pipeline components
  • Performance monitoring and logging capabilities
  • Robust error handling and recovery mechanisms
  • Thread-safe operation with dedicated processing thread

The command line template example:

  -n RADAR0 -t Radar \
    -k max_input_buffer_size -v 2097152 \
    -k max_output_buffer_size -v 2097152 \
    -k service_name -v "/dev/radar0" \
    -k timeout_ms -v 10000 \
    -k enable_performance_log -v true \
    -k pool_size -v 8 \
    -k cache -v true \
    -k input_topic -v /sensor/radar/RADAR0/raw \
    -k output_topic -v /sensor/radar/RADAR0/processed \

Typical Usage Patterns:

  1. DataReader → Radar → Recorder Pipeline:
./bin/qcrun ./bin/QCNodeSampleApp \
  -n RADAR_DATA -t DataReader -k number -v 1 \
    -k type0 -v tensor -k tensor_type0 -v uint8 -k dims0 -v "2097152,1" \
    -k data_path0 -v /data/radar_input \
    -k topic -v /sensor/radar/RADAR0/raw \
  -n RADAR0 -t Radar \
    -k max_input_buffer_size -v 2097152 \
    -k max_output_buffer_size -v 2097152 \
    -k service_name -v "/dev/radar0" \
    -k timeout_ms -v 10000 \
    -k enable_performance_log -v true \
    -k input_topic -v /sensor/radar/RADAR0/raw \
    -k output_topic -v /sensor/radar/RADAR0/processed \
  -n RADAR_REC -t Recorder -k max -v 100 \
    -k topic -v /sensor/radar/RADAR0/processed

This pipeline demonstrates:

  • DataReader: Reads raw radar data files (0.raw, 1.raw, 2.raw, etc.) from /data/radar_input/ directory as QC_BUFFER_TYPE_TENSOR buffers
  • Radar: Processes the raw radar data using hardware acceleration via the Radar Node component
  • Recorder: Writes the processed results back to disk as individual files (/tmp/RADAR_REC_0_0.raw, /tmp/RADAR_REC_1_0.raw, etc.) with metadata in /tmp/RADAR_REC.meta
  1. Radar Processing with Performance Monitoring:
./bin/qcrun ./bin/QCNodeSampleApp \
  -n RADAR_DATA -t DataReader -k number -v 1 \
    -k type0 -v tensor -k tensor_type0 -v uint8 -k dims0 -v "1048576,1" \
    -k data_path0 -v /data/radar_input \
    -k fps -v 10 \
    -k topic -v /sensor/radar/RADAR0/raw \
  -n RADAR0 -t Radar \
    -k max_input_buffer_size -v 1048576 \
    -k max_output_buffer_size -v 1048576 \
    -k service_name -v "/dev/radar0" \
    -k timeout_ms -v 5000 \
    -k enable_performance_log -v true \
    -k pool_size -v 8 \
    -k input_topic -v /sensor/radar/RADAR0/raw \
    -k output_topic -v /sensor/radar/RADAR0/processed

This pipeline demonstrates radar processing with detailed performance monitoring and logging enabled.

Data Preparation: To use the radar pipeline, prepare your radar data files in the following format:

  • Create a directory (e.g., /data/radar_input/)
  • Place raw radar data files named sequentially: 0.raw, 1.raw, 2.raw, etc.
  • Each file should contain raw radar data up to the specified buffer size
  • The DataReader will read these files sequentially and create QC_BUFFER_TYPE_TENSOR buffers
  • The Radar component will process these tensors using hardware acceleration
  • The Recorder will save the processed results back to disk with metadata

2.25 QCNode C2C Sample

attribute required type default comments
topic true string - the topic name
type false string pub valid type: "pub" or "sub". If the type is "pub", get the message from DataBroker and forward it to the PCIe C2C, else get the message from the PCIe C2C and forward it to DataBroker.
queue_depth false int 2 the subscriber queue depth
width true int - The image width
height true int - The image height
format false string nv12 The image format, options from [nv12, nv12_ubwc, uyvy, p010, rgb, bgr]
pool_size false int 4 The image memory pool size
channel false int 0 The PCIe C2C channel ID, start from 0
buffers_name true string - The input buffers name, only valid if type is "pub".

The command line template example:

# for type pub
  -n C2C_CAM0 -t C2C -k channel -v 0 -k type -v pub \
    -k buffers_name -v CAM0.0 \
    -k width -v 1920 -k height -v 1080 -k format -v nv12 \
    -k topic -v /sensor/camera/CAM0/raw \

# for type sub
  -n C2C_CAM0 -t C2C -k channel -v 0 -k type -v sub \
    -k width -v 1920 -k height -v 1080 -k format -v nv12 \
    -k topic -v /sensor/camera/CAM0/raw \

2.26 QCNode Temporal Sample

This is a specially designed sample for temporal-type models such as VAD, which has a temporal input derived from one of the previous model outputs.

graph LR
    subgraph Inputs
        IPP["Image PreProc"]
        SN["Misc"]
    end

    Temp["Temporary"]
    FS["FrameSync"]
    QNN["QNN"]

    IPP -- image --> FS
    SN -- canbus --> FS
    SN -- lidar2img --> FS
    Temp -- prev_bev --> FS
    Temp -- use_prev_flag --> FS

    FS --> QNN

    QNN -- bev_embed --> Temp

    PP["PostProc"]

    QNN -- all_cls_scores --> PP
    QNN -- all_bbox_preds --> PP
    QNN -- all_traj_preds --> PP
    QNN -- "..." --> PP
Loading
attribute required type default comments
number false int 1 The number of temporal tensors
temporal_tensor_type false string ufixed_point8 The default temporal tensor type, options from [float32,ufixed_point8, ufixed_point16]
temporal_tensor_dims true string - The default temporal tensor dimensions, in format "N,H,W,C", "N,S,C", "N,C", or "N" depends on the number of dimensions.
temporal_quant_scale false float 1.0f The default quantization scale of the quantize temporal tensor
temporal_quant_offset false int 0 The default quantization offset of the quantize temporal tensor
temporal_index false int 0 The default temporal tensor index of the outputs of the temporal model.
temporal_tensor_typeX false string =temporal_tensor_type The temporal tensor type for tensor X
temporal_tensor_dimsX false string =temporal_tensor_dims The temporal tensor dimensions for tensor X
temporal_quant_scaleX false float =temporal_quant_scale The quantization scale for tensor X
temporal_quant_offsetX false int =temporal_quant_offset The quantization offset for tensor X
temporal_indexX false int =temporal_index The temporal tensor index for tensor X
use_flag_tensor_type false string - The use flag tensor type, options from [float32,ufixed_point8, ufixed_point16]
use_flag_tensor_dims false string "1" The use flag tensor dimensions, in format "N,H,W,C", "N,S,C", "N,C", or "N" depends on the number of dimensions.
use_flag_quant_scale false float 1.0f The quantization scale of the quantize use flag tensor
use_flag_quant_offset false int 0 The quantization offset of the quantize use flag tensor
window false int 200 The receive window time in ms
input_topic true string - the input topic name
output_topic true string - the output topic name

Note: "X" is value from 0 to number-1.

The command line template example:

  -n TEMP0 -t Temporal \
    -k temporal_tensor_type -v ufixed_point8 -k temporal_tensor_dims -v "1,100,100,256" \
    -k temporal_quant_scale -v 0.0230286 -k temporal_quant_offset -v -131 \
    -k temporal_index -v 0 \
    -k use_flag_tensor_type -v ufixed_point8 -k use_flag_tensor_dims -v "1" \
    -k use_flag_quant_scale -v 0.003921576 -k use_flag_quant_offset -v 0 \
    -k input_topic -v /sensor/camera/VAD0/qnn \
    -k output_topic -v /sensor/camera/VAD0/temp \

2.27 QCNode ResMon Sample

The Sample ResMon gets the real-time system loading, including CPU utilization, GPU utilization, NSP0/NSP1 utilization and DDR bandwidth. Users can configure the sampling interval and total sampling duration, and optionally enable loop mode. When loop mode is enabled, the total sampling duration is ignored, and ResMon will continue running indefinitely.

attribute required type default comments
metrics true string list - The metric names for ResMon. Options: cpu_util, gpu_util, nsp0_util, nsp1_util, ddr_bw
loop_mode true bool true The option to enable loop mode
sample_interval true int 5000 The ResMon sampling iterval in ms
total_time false int 0 The ResMon total sampling time in ms

loop mode:

  -n RESMON -t ResMon \
    -k metrics -v cpu_util,gpu_util,nsp0_util,nsp1_util,ddr_bw \
    -k loop_mode -v true \
    -k sample_interval -v 1000

non-loop mode:

  -n RESMON -t ResMon \
    -k metrics -v cpu_util,gpu_util,nsp0_util,nsp1_util,ddr_bw \
    -k loop_mode -v false \
    -k sample_interval -v 1000 \
    -k total_time -v 10000

2.28 QCNode Genie Sample

attribute required type default comments
config true string - The Genie Dialog JSON configuration file.
embedding_table true string - Token-to-Embedding lookup table provided as a file.
lut_data_type true string float32 Token-to-Embedding lookup table data type, options from [float32,ufixed_point8,ufixed_point16,sfixed_point8,sfixed_point16]
lut_scale false float 1.0f The quantization scale of the Token-to-Embedding lookup table
lut_offset false int 0 The quantization offset of the Token-to-Embedding lookup table
input_data_type true string float32 Input data type, options from [float32,ufixed_point8,ufixed_point16,sfixed_point8,sfixed_point16]
input_scale false float 1.0f The quantization scale of the input
input_offset false int 0 The quantization offset of the input
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

  -n GENIE0 -t Genie -k config -v qwen2vl.json \
    -k embedding_table -v embedding_weights_152064x3584_ssd.bin \
    -k input_topic -v /sensor/genie/embeding/raw \
    -k output_topic -v /sensor/genie/decoder/text \

2.29 QCNode ComputeLidarCoord Sample

The Sample ComputeLidarCoord is used for lidar raw data preprocessing, it computes points cloud coordinates from distance and azimuth angle, also do correction according to firetime matrix and angle correction matrix. The input topic should contain 3 float tensors with following dimensions:

  • raw data which contains distance, intensity, azimuth, motor information: [cols, blocks*2+2]
  • firetime correction matrix: [cols]
  • azimuth correction matrix: [cols]

The command line template example:

attribute required type default comments
cols false int 1000 The lidar raw data colums number in vertical direction
blocks false int 100 The lidar raw data blocks number in horizontal direction
pool_size false int 4 The image memory pool size
cache false bool true use cached memory or not for the image memory
input_topic true string - the input topic name
output_topic true string - the output topic name

The command line template example:

-n LIDAR0 -t DataReader \
  -k number -v 3 \
  -k type0 -v tensor -k dims0 -v "1000,202" \
  -k type1 -v tensor -k dims1 -v "1000" \
  -k type2 -v tensor -k dims2 -v "1000" \
  -k topic -v /sensor/lidar/LIDAR0/input \
-n LIDARPRE -t ComputeLidarCoord \
  -k cols -v 1000 -k blocks -v 100 \
  -k input_topic -v /sensor/lidar/LIDAR0/input \
  -k output_topic -v /sensor/lidar/LIDAR1/output \

3. Typical QCNode Sample Application pipelines

3.1 4 DataReader based QNN perception pipelines

export QC_LOG_LEVEL=INFO
./bin/qcrun ./bin/QCNodeSampleApp -n CAM0 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM0/raw \
  -n REMAP0 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/remap \
  -n CNT0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM0/remap \
    -k output_topic -v /sensor/camera/CAM0/qnn \
  -n POSTPROC_CNT0 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM0/qnn \
    -k output_topic -v /sensor/camera/CAM0/objs \
  -n CAM1 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM1/raw \
  -n REMAP1 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM1/raw \
    -k output_topic -v /sensor/camera/CAM1/remap \
  -n CNT1 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM1/remap \
    -k output_topic -v /sensor/camera/CAM1/qnn \
  -n POSTPROC_CNT1 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM1/qnn \
    -k output_topic -v /sensor/camera/CAM1/objs \
  -n CAM2 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM2/raw \
  -n REMAP2 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM2/raw \
    -k output_topic -v /sensor/camera/CAM2/remap \
  -n CNT2 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM2/remap \
    -k output_topic -v /sensor/camera/CAM2/qnn \
  -n POSTPROC_CNT2 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM2/qnn \
    -k output_topic -v /sensor/camera/CAM2/objs \
  -n CAM3 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM3/raw \
  -n REMAP3 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM3/raw \
    -k output_topic -v /sensor/camera/CAM3/remap \
  -n CNT3 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM3/remap \
    -k output_topic -v /sensor/camera/CAM3/qnn \
  -n POSTPROC_CNT3 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM3/qnn \
    -k output_topic -v /sensor/camera/CAM3/objs \
  -n VIZ -t TinyViz -k cameras -v CAM0,CAM1,CAM2,CAM3

3.2 1 DataReader and 1 Camera AR231 based QNN perception pipelines

export QC_LOG_LEVEL=INFO
./bin/qcrun ./bin/QCNodeSampleApp -n CAM0 -t Camera -k input_id -v 0 \
    -k width -v 1928 -k height -v 1208 \
    -k topic -v /sensor/camera/CAM0/raw \
  -n C2D0 -t C2D -k batch_size -v 1 \
    -k input_width0 -v 1928 -k input_height0 -v 1208 -k input_format0 -v nv12 \
    -k roi_x0 -v 0 -k roi_y0 -v 0 -k roi_width0 -v 1928 -k roi_height0 -v 1208 \
    -k output_width -v 1928 -k output_height -v 1208 -k output_format -v uyvy \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/uyvy \
  -n REMAP0 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1928 -k input_height0 -v 1208 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM0/uyvy \
    -k output_topic -v /sensor/camera/CAM0/remap \
  -n CNT0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM0/remap \
    -k output_topic -v /sensor/camera/CAM0/qnn \
  -n POSTPROC_CNT0 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1928 -k height -v 1208 \
    -k input_topic -v /sensor/camera/CAM0/qnn \
    -k output_topic -v /sensor/camera/CAM0/objs \
  -n CAM1 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM1/raw \
  -n REMAP1 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM1/raw \
    -k output_topic -v /sensor/camera/CAM1/remap \
  -n CNT1 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM1/remap \
    -k output_topic -v /sensor/camera/CAM1/qnn \
  -n POSTPROC_CNT1 -t PostProcCenternet \
    -k processor -v gpu \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM1/qnn \
    -k output_topic -v /sensor/camera/CAM1/objs \
  -n VIZ -t TinyViz -k cameras -v CAM0,CAM1

3.3 1 DataReader based Pointpillar QNN perception pipelines

./bin/qcrun ./bin/QCNodeSampleApp \
  -n LIDAR0 -t DataReader -k number -v 2 \
    -k type0 -v tensor -k tensor_type0 -v float32 -k dims0 -v "300000,4"  \
    -k data_path0 -v /data/LIDAR0 \
    -k type1 -v image -k format1 -v nv12 -k width1 -v 1920 -k height1 -v 1024 \
    -k data_path1 -v /data/LIDAR0 \
    -k pool_size -v 4 -k fps -v 10 \
    -k topic -v /sensor/lidar/LIDAR0/raw \
  -n PLRPRE0 -t PlrPre -k processor -v htp0 \
    -k input_topic -v /sensor/lidar/LIDAR0/raw \
    -k output_topic -v /sensor/lidar/LIDAR0/plrpre \
  -n PLR0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/pointpillar/program.bin \
    -k input_topic -v /sensor/lidar/LIDAR0/plrpre \
    -k output_topic -v /sensor/lidar/LIDAR0/qnn \
  -n PLRPOST0 -t PlrPost -k processor -v htp0 -k debug -v false \
    -k input_lidar_topic -v /sensor/lidar/LIDAR0/raw \
    -k input_topic -v /sensor/lidar/LIDAR0/qnn \
    -k output_topic -v /sensor/lidar/LIDAR0/objs \
  -n VIZ -t TinyViz -k cameras -v LIDAR0 \
    -k batch_index0 -v 1 \
    -k cam_topic0 -v /sensor/lidar/LIDAR0/raw \
    -k obj_topic0 -v /sensor/lidar/LIDAR0/objs -d

3.4 1 QNN model data online inference pipeline

./bin/qcrun ./bin/QCNodeSampleApp \
  -n DO0 -t DataOnline -k port -v 6666 \
    -k input_topic -v /data/online/DO0/input \
    -k output_topic -v /data/online/DO0/output \
    -k model_io_info_topic -v /data/online/DO0/model/info \
  -n CNT0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin  \
    -k input_topic -v /data/online/DO0/output \
    -k output_topic -v /data/online/DO0/input \
    -k model_io_info_topic -v /data/online/DO0/model/info -d

3.5 1 QNN model data online shared memory between process pipeline

./bin/qcrun ./bin/QCNodeSampleApp \
  -n CAM0 -t DataReader -k width0 -v 1920 -k height0 -v 1024 \
    -k data_path0 -v /data/4K_street_1000_500_1920_1024_uyvy \
    -k format0 -v uyvy -k topic -v /sensor/camera/CAM0/raw \
  -n CAM0_P -t SharedRing -k type -v pub \
    -k topic -v /sensor/camera/CAM0/raw -d &

./bin/qcrun ./bin/QCNodeSampleApp \
  -n CAM0_S0 -t SharedRing -k type -v sub \
    -k topic -v /sensor/camera/CAM0/raw \
  -n REMAP0 -t Remap -k batch_size -v 1 \
    -k input_width0 -v 1920 -k input_height0 -v 1024 -k input_format0 -v uyvy \
    -k output_width -v 1152 -k output_height -v 800 -k output_format -v rgb \
    -k input_topic -v /sensor/camera/CAM0/raw \
    -k output_topic -v /sensor/camera/CAM0/remap \
  -n REMAP0_P -t SharedRing -k type -v pub \
    -k topic -v /sensor/camera/CAM0/remap -d &

./bin/qcrun ./bin/QCNodeSampleApp \
  -n REMAP0_S0 -t SharedRing -k type -v sub \
    -k topic -v /sensor/camera/CAM0/remap \
  -n CNT0 -t Qnn -k processor -v htp0 \
    -k model_path -v data/centernet/program.bin \
    -k input_topic -v /sensor/camera/CAM0/remap \
    -k output_topic -v /sensor/camera/CAM0/qnn \
  -n CNT0_P -t SharedRing -k type -v pub \
    -k topic -v /sensor/camera/CAM0/qnn -d &

./bin/qcrun ./bin/QCNodeSampleApp \
  -n CAM0_S1 -t SharedRing -k type -v sub \
    -k topic -v /sensor/camera/CAM0/raw \
  -n CNT0_S0 -t SharedRing -k type -v sub \
    -k topic -v /sensor/camera/CAM0/qnn \
  -n POSTPROC_CNT0 -t PostProcCenternet \
    -k width -v 1920 -k height -v 1024 \
    -k input_topic -v /sensor/camera/CAM0/qnn \
    -k output_topic -v /sensor/camera/CAM0/objs \
  -n VIZ -t TinyViz -k cameras -v CAM0 -k winH -v 1050 -d