[Help Needed] Running sample_path_perception from DriveWorks 2.2 on x86 Host — Engine Incompatible Issue

Autonomous Vehicles DRIVE PX 2 DRIVE - Linux
,
1

Hi NVIDIA Dev Team,

I’m working on running the sample_path_perception (PathNet) sample using DriveWorks SDK 2.2 on an Ubuntu 18.04 x86_64 host machine with a dedicated NVIDIA RTX 3070 Ti GPU.

My Current Setup

Installed via SDK Manager (or manually from .deb and .tar packages):

  • DriveWorks SDK: 2.2.3136
  • DRIVE Software SDK: 5.1
  • CUDA Toolkit: 10.2.19
  • cuDNN: 7.5.1.14
  • TensorRT: 5.1.4.2
  • CMake: 3.22.6 (manually installed, since 3.10 was too old for some build steps)
  • GCC: 7.5.0 (close to required 7.4)
  • Ubuntu: 18.04

All DriveWorks samples including sample_hello_world are compiling and running fine.

I want to use the sample_path_perception application to run inference on offline .h264 videos using PathNet (ego-lane, adjacent lanes). I have the H264 videos ready and the sample builds correctly.

Error -
[01-04-2025 13:01:38] Platform: Detected Generic x86 Platform
[01-04-2025 13:01:38] TimeSource: monotonic epoch time offset is 1743490553847085
[01-04-2025 13:01:38] Platform: number of GPU devices detected 1
[01-04-2025 13:01:38] Platform: currently selected GPU device discrete ID 0
[01-04-2025 13:01:38] SDK: Resources mounted from .././data/
[01-04-2025 13:01:38] TimeSource: monotonic epoch time offset is 1743490553847085
[01-04-2025 13:01:38] Initialize DriveWorks SDK v2.2.3136
[01-04-2025 13:01:38] Release build with GNU 7.4.0 from heads/buildbrain-branch-0-gca7b4b26e65
[01-04-2025 13:01:38] Initialize DriveWorks VisualizationSDK v2.2.3136
[01-04-2025 13:01:38] Initialize DriveWorksGL SDK v2.2.3136
[01-04-2025 13:01:38] GL-SDK: initialize OpenGL
[01-04-2025 13:01:38] SensorFactory::createSensor() → camera.virtual, debugView=false,detectionThreshold=0.5,fps=30,lookAheadDistance=1000000.0,offscreen=0,precision=fp32,profiling=1,rig=.././data/samples/pathDetection/rig.json,roi.height=800,roi.width=1920,roi.x=0,roi.y=400,temporalSmoothingFactor=0.1,useCudaGraph=0,video=.././data/samples/pathDetect
[01-04-2025 13:01:38] CameraNVCUVID: no seek table found at .././data/samples/pathDetection/video_paths.h264.seek, seeking is not available.
SimpleCamera: Camera image: 1920x1208
[PathDetectorApp]: Camera image with 1208x1920 at 30 FPS
[01-04-2025 13:01:38] Loading rig file: .././data/samples/pathDetection/rig.json
[01-04-2025 13:01:38] PathNet: Running in FP32 mode.
[01-04-2025 13:01:38] PathNet: Running from model: gv1xx-discrete/pathnet/pathnet_v1_fov60/fp32/pathnet.dnn
[01-04-2025 13:01:38] PathNet: Running from metadata: gv1xx-discrete/pathnet/pathnet_v1_fov60/fp32/pathnet_metadata.json
[01-04-2025 13:01:38] PathNet: Running from trt_json: gv1xx-discrete/pathnet/pathnet_v1_fov60/fp32/tensorrt_metadata.json
[01-04-2025 13:01:38] FileHeader: file header created with development build (patch = -1), assumingcompatibility
[01-04-2025 13:01:38] FileHeader: file header created with development build (patch = -1), assumingcompatibility
[01-04-2025 13:01:38] The engine plan file is generated on an incompatible device, expecting compute 8.6 got compute 7.0, please rebuild.
Segmentation fault (core dumped)

What I Need Help With

  1. How can I rebuild the .engine file for PathNet (DriveWorks 2.2) on x86 host with compute capability 8.6?
3

Dear @lokesh6,
The pathnet model is not publicly available in DRIVE SW 10 release. So you can’t re-generate for ampere GPU. BTW Are you using DRIVE PX2 or DRIVE AGX Xavier/Pegasus ?

4

Thanks @SivaRamaKrishnaNV . i was able to run the inference on Turing Architecture GPU.