TensorRT Windows 10: (nvinfer.dll) Access violation

I have been able to use the trt-yolo-app example to create a classifier that is running on our hardware (which is powered by a TX2).

The code I followed was from this repo:

https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/tree/master/yolo

Unfortunately I haven’t had much joy running this same code for a Windows target. I followed the dependency requirements for the dGPU platform however my application fails with a memory access violation when trying to deserialise the TRT plan file.

I guess my question is, is there any reason (theoretically) why the trt-yolo-app sample wouldn’t work for Windows? I know the intended platform of these samples is Linux but can’t see why there would be an issue using roughly the same code for Windows.

It feels like I’m close as I can generate the plan file but just struggle to read it in to create the engine – I also tried generating the engine and then used that object for inferencing rather than trying to read a plan file and deserialise but ran into the same memory access violation:

Exception thrown at 0x00007FF9B34795E6 (nvinfer.dll) in TensorRT_Yolo.exe: 0xC0000005: Access violation reading location 0x000001EA7EDDBD51. occurred

I have uploaded a version of the code that builds under windows using Visual Studio 2017, and the binary file. The library versions that the application was built against is also detailed in the repo:

https://github.com/mj8ac/trt-yolo-app_win64/blob/master/README.md

Any suggestions much appreciated.

Hi jmirza,

I can’t answer this, but all the codes and low layer libraries are all compiled under Linux base, maybe that’s why can’t work well on your Windows environment.

Hello,

Would it be worth trying to compile the application using the MinGW compiler? Is this the compiler that was used to compile the nvidia dll files?

I don’t understand what could be going wrong, the sample code was developed against the TRT5 (Linux) API, the TRT API is the same for Windows - is this correct?

Thanks,
Jonny

Hi jmirza,

We didn’t try that before, so no idea if could generate more issue or could work.