topic Re: Re:Low EU active in dpc++ program calculated by vtune in Intel® oneAPI DPC++/C++ Compiler

Low EU active in dpc++ program calculated by vtune

Hannah_zhen — Fri, 01 Dec 2023 06:53:23 GMT

Hi community!
I was running a program under oneAPI base toolkit and using DPC++ for sycl parallization.

However, as the design of my program, the work group number can be from 480~60, while the work group size can be from 48~384.

I used vtune "vtune -collect gpu-offload" to analyse the EU ACTIVE RATE:

When the setting is work group number is 480 and size is 48, EU active rate in the kernel function is 7.2% .

When the setting is work group number is 60 and size is 384, EU active rate in the kernel function is 21.2%. 21.2% seems to be the highest in my program with parallel processing.

1-->What is the mapping rules from software to hardware WITH oneAPI?

2-->How can I improve the active rate as the SETTING
sycl::nd_range<3>(sycl::range<3>(1, 1, 60) * sycl::range<3>(1, 1, 384), sycl::range<3>(1, 1, localsize)) for work item and group is now the best performance but still with low active rate

Is there any solution to improve the active rate? Thanks in advance.

Below is my device information :

My environment is ubuntu22.04 with 6.2.0-36-generic and the processor is i5-1340P.

The oneAPI base toolkit is version 2023.1.0

Re:Low EU active in dpc++ program calculated by vtune

SeshaP_Intel — Wed, 06 Dec 2023 09:19:35 GMT

Hi,

Thank you for posting in Intel Communities.

Could you please try with the latest oneAPI 2024.0 and let us know the results?

Please provide us with the complete reproducer code and the steps you have followed to investigate the issue more from our end.

Thanks and Regards,

Pendyala Sesha Srinivas

Re:Low EU active in dpc++ program calculated by vtune

SeshaP_Intel — Fri, 15 Dec 2023 05:04:56 GMT

Hi,

We haven't heard back from you. Could you please provide an update on your issue?

Thanks and Regards,

Pendyala Sesha Srinivas

Re: Re:Low EU active in dpc++ program calculated by vtune

Hannah_zhen — Fri, 15 Dec 2023 05:41:37 GMT

Sorry for the late reply.

As the code is of security, I will post a new question next time the latest oneAPI 2024.0 is applied in our program.

Thanks.

Re: Low EU active in dpc++ program calculated by vtune

SeshaP_Intel — Mon, 18 Dec 2023 11:51:12 GMT

Hi,

For Community support, we would require you to submit a minimum reproduction sample code specific to your issue that provides us with the most relevant background information for triage.

In case you require privacy, and are unable to share the issue/sample with us publicly and If you are a licensed oneAPI product customer and/or member of Intel’s oneAPI Academic Program please submit a ticket for Priority support so that your application can be handled with the required data protection and privacy regulations.

>> As the code is of security, I will post a new question next time the latest oneAPI 2024.0 is applied in our program.

As per your reply, I am going ahead and closing this case. Please post a new question if you face any issues as this thread will no longer be monitored by Intel.

Thanks and Regards,

Pendyala Sesha Srinivas