Vtune Missing GPU Compute/Media Hotspot Analysis

[jwongx]

Hi,

I am using a i7-9700TE and want to profile Intel Graphics using Vtune. I selected the GPU Compute/Media Hotspot (preview) for the HOW section. 

However, I after the results finalized, there is no GPU Compute/Media Hotspot (preview) selection from the drop down list (attached image), thus there is no Memory Hierarchy Diagram.

I would like to know what am I missing here to enable the Memory Hierarchy Diagram. 

Thanks.

[JananiC_Intel]

Hi,

Thanks for posting in Intel forums.

We need to try from our end so could you share your workload and your vtune version?

Regards,

Janani Chandran

[jwongx]

Hi,

I am currently using Vtune 2021.1.1 (613804) FYI.

My workload is running in a docker environment so I was using Profile System target type. I also realized that the GPU programming API tracing is only limited to Launch process and Process Target type. 

In this case, is there any suggestion on how to profile a OpenCL and Media SDK program running in docker environment?

Regards,

J.W

[JananiC_Intel]

Hi,

We are forwarding the case to SME.

Regards,

Janani Chandran

[DMITRY_T_Intel]

Hi J.W,

I reproduced this issue and escalated it to development. I will let you know as soon as I get any information.

Regarding your question: "is there any suggestion on how to profile a OpenCL and Media SDK program running in docker environment?"

Perhaps you already seen these articles:

1. Profiling Docker* Containers https://software.intel.com/content/www/us/en/develop/documentation/vtune-cookbook/top/configuration-recipes/profiling-in-docker-container.html
2. GPU OpenCL™ Application Analysis https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/analyze-performance/code-profiling-scenarios/gpu-application-analysis/gpu-opencl-application-analysis-view.html

I realize that article 2 is not applicable since there is no GPU Compute/Media Hotspot (preview) selection in VTune. So lets wait reply from developers.

Thank you!

[DMITRY_T_Intel]

Hi J.W,

I got reply from VTune developers:

"GPU Memory Hierarchy Diagram is available only for results with GPU programming API traced, then for each compute-task VTune visualizes its performance metrics and memory traffic over the diagram. For system-wide analysis GPU RTs API tracing is not supported.

While we're interested to get better understanding of the customer use case and reasons behind the usage model, so far we did not do special enabling for GPU profiling for apps in a docker environment, so cannot yet comment on existing limitations and BKMs."

They advise try to use VTune inside container by following the steps in https://software.intel.com/content/www/us/en/develop/documentation/vtune-help/top/launch/containerization-support/run-from-container.html and see if GPU hotspots can be collected from inside container?

Could you please try and let me know results?

Thank you very much!

[jwongx]

Hi, Dmitry, 

Thanks for the insights. Well, I will let u know once I setup vtune in the docker and profile the application again. 

Regards, 

J.W

[DMITRY_T_Intel]

Hi J.W,

Please let us know results. Did you try to profile the application in Docker?

Thanks.

[DMITRY_T_Intel]

Hi J.W,

I think this issue is fixed or not relevant any more. If you still see this issue then just create a new case and we will investigate further.

Thank you!