Random outputs when running on iGPU in asynchronous mode

[xxbird]

Hi all,

I was trying to run an inference on iGPU by the C++ async API.  When I compiled the model on CPU (core.compile_model(model, “CPU”)), it worked fine and the result was correct. However, when I changed the device to iGPU (core.compile_model(model, “GPU”)), the outputs became some random number.  

Does anyone occur the same issue? 

Thanks.

[Wan_Intel]

Hi Xxbird,

Thanks for reaching out to us.

I've validated Asynchronous Inference Request with Image Classification Async C++ Sample.

CPU:

GPU:

Are you using the latest version of the OpenVINO™ toolkit?

Regards,

Wan

[xxbird]

Hi Wan,

Thanks for your replying. 

I am using the 2022.3 LTS version. My device is 11th Gen Core i7-1195G7 and the system is Ubuntu 18.04.

I just tried the Classification Async C++ Sample  and the results were even more strange.

CPU:

 

GPU:

 

The results were totally different between CPU and GPU. When I tried to run on the same image on GPU multiple times, the results were changing everytime. Even sometimes the CPU had the same  problem ( not everytime).  

Best,

 

[Wan_Intel]

Hi Xxbird,

Thanks for your information.

Please share your model with us for further investigation.

Regards,

Wan

[Wan_Intel]

Hi Xxbird,

Please run the Image Classification Async C++ Sample using alexnet with the sample images in the attachment and see if the results of inferencing with CPU and GPU are similar.

Best,

Wan

[xxbird]

hi Wan,

I ran the sample using alexnet and the results were fine.  While I tried to use the mobilenet-v2 caffe version (https://docs.openvino.ai/latest/omz_models_model_mobilenet_v2.html), the results became weird on GPU (but it worked fine on CPU). 

Btw, I just used the "omz_converter" command to convert the model.

Best.

[xxbird]

I tested the models on my PC and the results were correct, so I am sure the model was converted in the right way.

[Wan_Intel]

Hi Xxbird,

Thanks for your information.

Glad to know the results were correct when the models were tested on your PC.

I've also validated the results of running Image Classification Async C++ Sample using mobilenet-v2 (FP16) with CPU and GPU.

CPU:

GPU:

Regards,

Wan

[xxbird]

Hi Wan,

I have found that it might be because the Opencl was not installed correctly.

Thanks for your help.

[bouachalazhar]

Hi, have you a link for install OpenCL on Windows 11 or Ubuntu 22.04 ?

[xxbird]

I just followed the official repo https://github.com/intel/compute-runtime. 

[Wan_Intel]

Hi Xxbird,

Thanks for providing your solution in the community.

We're glad to know that your issue has been resolved after installing OpenCL™ for GPU. Therefore, this thread will no longer be monitored since the issue has been resolved. If you need additional information from Intel, please submit a new question.

Regards,

Wan