Does oneAPI for GPU support SIMD likewise cm_sqrt function in the CM GPU development library?

[Yongfa]

Does oneAPI for GPU support SIMD likewise cm_sqrt function in the CM GPU development library?

[AbhishekD_Intel]

Hi,

Please give us more details regarding your use-case.

Regarding support of SIMD if the hardware supports then subgroup is the way to achieve SIMD, as subgroups map to SIMD hardware.

Warm Regards,

Abhishek

[Yongfa]

Thanks. I have found some documents to introduce the SIMD of oneAPI. But I still have some questions, could you please help to clarify, Thanks.

• Does oneAPI try to enable SIMD as possible since the SIMD width is 16 when I run the vtune to diagnose the GPU performance even I don't write any subgroup code?
• How to specify the SIMD width manually?
• What are the possible reasons to stop a higher SIMD width? the number of EU registers?

[AbhishekD_Intel]

Hi,

We are forwarding this issue to the Subject Matter Experts.

Thank you

[Sravani_K_Intel]

A new Explicit SIMD Programming Extension for DPC++ is coming up which helps with manually setting the SIMD width.