A SYCL workload on Intel Arch A750

[JinyaoWang]

Hi All,

    I am trying to run my SYCL code on an Intel Arc A750 GPU. The code is based on a complex Monte Carlo method, so the workload for each kernel is large and it is difficult to split into multiple small kernels. Although NDRange is used to limit the size of the workgroup, the output screen will still be unresponsive when the code is running. What is the reason for this question? How to fix it?

    The drawings are the runtime output of the code. After printing the debug message "mcml ndrange global", the kernel starts running and the screen loses its response until the kernel runs end.

Thanks

-----------------------

[Andres_Intel]

Hello JinyaoWang,

Thank you for posting on the Intel communities. Since you are having issues running your SYCL code on your Intel® Arc™ A750 Graphics, the output screen will still be unresponsive when the code is running. Please answer the questions below to understand the issue in a better way:

• Is the graphics card an Intel Branded Card or is it from another brand?
• Does the issue happen with different graphics driver versions?
• Provide us with step-by-step instructions and the script file in case we need to replicate the issue.
• To know your system hardware and software information, run and attach the Intel® System Support Utility for Windows (SSU), you can find the steps on the link How to get the Intel® System Support Utility Logs on Windows. 

Regards,  

Andres P. 

Intel Customer Support Technician

[JinyaoWang]

The graphics card is from the Gunnir brand. 

This issue occured on both driver versions 31.0.101.4953 and 31.0.101.4887.

The device information output by SSU and my simple reproduction of the issue is attached. The program was compiled in debug mode using oneAPI Base ToolKit 2024.0 in vs2022. There are no specific values specified for the size of ndrange and the number of loops in the kernel. It only needs to be a large enough load on the kernel to cause the screen to become unresponsive.

[Jean_Intel]

Hello JinyaoWang,

We would like to let you know that there is a newer graphics driver version, so you should check if the issue remains with the latest fixes included in the driver 31.0.101.4972. However, before you install the new driver, you should run the Display Driver Uninstaller (DDU) to eliminate all previous driver files and leftovers. Sometimes corrupted files from previous versions can cause driver malfunctions.

After checking the system report, we noticed that your Operating System Build version is 22621; however, the latest build version is 22631. Based on this, we recommend you update your Operating System to the latest build.

Finally, we would like to thank you for the sample you have provided, but we want to ask you for more details about the options to run the software so we can replicate the issue:

• Is there a particular way to execute the code? 
• How do we increase the load on the kernel?
• How large does the load need to be?
• Is this a personal or work Project?

Best regards,

Jean O. 

Intel Customer Support Technician

[JinyaoWang]

I have uninstalled the original driver 31.0.101.4953 via DDU and installed driver 31.0.101.4972 and the problem persists. As for the OS version, I didn't receive an update push for build 22631.

• The code is currently only compiled with the debug mode of VS2022 and started with local Windows debugger. Attached is the Visual Studio solution and codes after streamlined that reproduces the issue.
• The variable "repeat_num" in the unew version (i.e. "renum" in the old version) is used to modify kernel load. 
• On my computer, the screen would be unresponsive during the code running when repeat_num >= 1000. The screen returns to normal at the end of the code run. The larger the repeat_num, i.e., the longer the kernel runs, the longer the screen will be unresponsive.
• This is a personal project.

[Andres_Intel]

Hello JinyaoWang,

Thank you for your response and clarification.

All the information provided has been really helpful. I will start with an investigation to provide you with the next steps as soon as possible.

Regards,  

Andres P. 

Intel Customer Support Technician

[Andres_Intel]

Hello JinyaoWang,

Thank you for your time.

I have been working on the investigation and in order to help you with your request, our recommendation is to open a new thread in the Developer Software Forums for further assistance.

For now, we will close this thread. If you need any additional information, please submit a new question as this thread will no longer be monitored.  

Best regards, 

Andres P. 

Intel Customer Support Technician