Is VisionWorks dead for SM 8.x?

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

I upgraded my host machine to an RTX 3090 from a 1070Ti and after trying to execute the graph I get hit with “CUDA error: no kernel image is available for execution on the device”. The target device is an AGX Xavier so visionworks is fine for that, but I want to do the bulk of work on my primary machine. Is this just a case of forced migration to VPI or can you release compatible binaries for newer GPUs? (preferably not bundled with SDK manager, for example VPI can be downloaded separately).

Cheers

3

Hi,

Which application do you try to run?
The error indicates the application doesn’t build with the corresponding GPU archutecture.

https://developer.nvidia.com/cuda-gpus

Since sm_86 support is added into CUDA toolkit until v11.0.
VPI, which is built with CUDA 10.2, doesn’t have sm_86 support.

Thanks.

4

This is my own application which currently uses VisionWorks for image pre-processing before inference, so I guess I cannot do any work on my host machine?

Will VPI be built with 11.0 for its 1.0 release or is it going to stay 10.2 since Jetpack is still 10.2 (with no mention of upgrading on the roadmap).

For Visionworks does /usr/share/visionworks/sources contain all the source code that I need to rebuild for sm_86?

5

Hi,

Usually, we will align the package version from host and device to ensure compatibility.
But, please let us check this with our internal team first.

Thanks.

6

As asked before, are there plans for CUDA 11.1 releases for Visionworks or VPI anytime soon? Or do I have to go back to OpenCV for image processing…

7

To get CUDA 11.1 release for Visionworks or VPI is under planning, not that quick to be introduced in recent JetPack release. The VPI will go to stay CUDA 10.2 at coming JetPack relese.