Is the current XMX optimization available for iGPU(ARL)?

[baihe-liu]

Describe the issue

Hi, I'm using ARL-H and I want to use XMX optimization. When I ran a inference I got UserWarning: Detect an unknown architecture, will treat it as no 2D Block feature support. (Triggered internally at /build/intel-pytorch-extension/csrc/gpu/runtime/Device.cpp:103.) and UserWarning: Detect an unknown architecture, will treat it as no XMX feature support. (Triggered internally at /build/intel-pytorch-extension/csrc/gpu/runtime/Device.cpp:58.)

OS: ubuntu 22.04 ipex version:2.3.110+xpu

Thank you!

[huiyan2021]

Hi @baihe-liu , please try ipex 2.5.10+xpu for ARL-H, you can see the device check here: https://github.com/intel/intel-extension-for-pytorch/blob/release/xpu/2.5.10/csrc/gpu/runtime/Device.cpp#L71

[baihe-liu]

@huiyan2021 Hi, Thank you very much! I'll try it.

[baihe-liu]

@huiyan2021 Hi, I've tried 2.5.10+xpu and found a worse performance than 2.3.110+xpu in my program. I got some trace by VTune, how can I check if it's effective? I only found a metric "XMX active".

[huiyan2021]

@baihe-liu , could you share the reproducer so that we can profile? thanks!