Check failed: error == CUDNN_STATUS_SUCCESS (4 vs. 0) CUDNN_STATUS_INTERNAL_ERROR

[Calvincs77]

After successfully installing CUDA, cuDNN and openpose via the provided instructions I get the error in the title.

I tried to run openpose from the demo, but once the window is opened it freezes until I get the error

CUDNNN_STATUS_SUCCESS (4 vs. 0) CUDNNN_STATUS_INTERNAL_ERROR

Graphics card = RTX 3050 ti laptop GPU 4 GB Windows 10 CUDA 11.6 CUDNNN 8.4

[dradamPK]

I had a similar problem with RTX3080 but on Ubuntu. To run the demo I needed to disable CUDNN while building Openpose. I don't know how it looks like in Windows, but what I've done in Ubuntu is just delete Openpose build folder and make a new one. Then in cmake-gui find and uncheck USE_CUDNN:

then configure, generate and make.

Let me know if disabling CUDNN helps you too.

[Calvincs77]

Ok I'm going to try it, before doing it please let me know, what cuda, cudnn and VisualS version are you using ?

Thanks

[dradamPK]

Now I'm using:

• Cudnn 8.3.2 (the newest one at the time of my installation)
• Cuda 11.5 (I didn't install 11.6 because there was no Cudnn dedicated to it)
• Nvidia driver 510.60
• VS Code 1.66
• Ubuntu 20

I observed this issue through different Cuda and Cudnn configurations. I'm not sure why the newest Ampere architecture makes problems. While testing the same software but older GPU (GTX 1060 6GB) Openpose with Cudnn work fine then.

[Alexleisandao]

After successfully installing CUDA, cuDNN and openpose via the provided instructions I get the error in the title.

I tried to run openpose from the demo, but once the window is opened it freezes until I get the error

CUDNNN_STATUS_SUCCESS (4 vs. 0) CUDNNN_STATUS_INTERNAL_ERROR

Graphics card = RTX 3050 ti laptop GPU 4 GB Windows 10 CUDA 11.6 CUDNNN 8.4

My configuration is RTX 3050, driver 470, cuda11.4, cudnn-11.4-linux-x64-v8.2.2.26 Tgz, Ubuntu 18.04, has the same problems as you.

Later, when I run the demo I tried to disable CUDNN while building Openpose, an error of video memory overflow will be reported. However, the error of video memory overflow can be solved by setting parameters.

After cudnn is canceled, the running settings are as follows: sudo/ build/examples/openpose/openpose. bin --video examples/media/video. avi --net_ resolution 176x256 --model_ pose COCO

[angelcarro]

I also have a laptop with a RTX 3050 and I always get a similar error. I am using the 525 driver but I also tested the previous version. If I disable the CUDA acceleration, the performance is incredibly slow. Maybe, I forgot to enable AVX2 in this test. My solution was to compile enabling OpenCL. At least, I obtain 4.4 FPS for a net_resolution of 656x368, or 15.2 FPS for 336x192.

[Alexleisandao]

您的邮件我已经收到，谢谢！