scaleWeights.cu - Cuda Error in NCHWToNCQHW4 with detection sample

I’m hacking deepstream detection sample, it has play function, but video messed up if I decode some video. here is the output:

scaleWeights.cu (1025) - Cuda Error in NCHWToNCQHW4: 11
[ERROR][16:13:42] CUDA runtime error 11 at line 214 in file src/cudaImage.cu

the converted grey stream tensor of color space converter is okay by check saved file.
I made up a demo video which can duplicate the problem, please download “delogo.h264” file from attachment.
SampleTestVideo.zip (1.56 MB)

Hi,

We have checked your video with sample decPerf and nvDecInfer_detection.
Both can run correctly.

Guess that there is something wrong in your implementation.
Could you recheck your source?

Thanks

have you set -gui=1 in the run.sh file with nvDecInfer_detection for visible playback? the function is run okay, just the visible is awful context. I’m sorry for the missing information.

Hi,

Although we can reproduce the issue on our side, the testing video looks strange.

Attachment is the screenshot of VLC player for your video.
Could you check if it is under your expectation?

Thanks.

the screenshot is okay as I applied a delogo filter to the video. Thanks for you reproduce.

Hi,

The video size is not aligned.
Real video size: 1024x770
Deepstream parameter: 1024x752

Could you resize your video to 1024x752 and check if this issue still occurs?

Thanks.

I resized the video to 1024x752, it is okay now for GL display. But I think the original video size is 1012x740.

Hi,

For alignment, we require the input resolution to be divided by 16.
If not, please pad your video to the valid resolution.

The 1024x770 is the auto padding resolution from my video viewer.
Sorry for the misunderstanding.

Thanks.

I heard about video width/height need 16 alignment, and 1012 can’t be divided by 16 with integer. we will no test with such video.
Thanks.