Determine appropriate device architecture during compile stage

[alexbaden]

We have two new features which rely on the ocloc utility to query GPU architecture info (#1900) and compile native GPU code (#1792) during the compile stage. Ocloc needs a device parameter:

 -device <device_type>                     Target device.
                                            <device_type> can be: tgl, tgllp, rkl, adl-s, rpl-s, adl-p, rpl-p, adl-n, dg1, dg2-g10-a0, dg2-g10-a1, dg2-g10-b0, acm-g10, ats-m150, dg2-g10, dg2, dg2-g10-c0, dg2-g11-a0, dg2-g11-b0, acm-g11, ats-m75, dg2-g11, dg2-g11-b1, acm-g12, dg2-g12, dg2-g12-a0
, pvc-xl-a0, pvc-sdv, pvc-xl-a0p, pvc-xt-a0, pvc-xt-b0, pvc-xt-b1, pvc, pvc-xt-c0, pvc-vg, pvc-xt-c0-vg, mtl-u-a0, arl-s, arl-u, mtl-m, mtl-s, mtl-u, mtl, mtl-u-b0, mtl-h-a0, mtl-h, mtl-p, mtl-h-b0, arl-h-a0, arl-h, arl-h-b0, bmg-g21-a0, bmg-g21-a1, bmg-g21, bmg-g21-b0, lnl-a0, lnl-a1, lnl-m, l
nl-b0, xe, xe2, gen12lp, xe-hpc, xe-hpc-vg, xe-hpg, xe-lp, xe-lpg, xe-lpgplus, xe2-hpg, xe2-lpg, ip version  or hexadecimal value with 0x prefix

We currently get a name string from PyTorch:

'name': 'Intel(R) Data Center GPU Max 1100
'name': 'Intel(R) Graphics [0xe20c]',

we can try to use the name string, or we can ask PyTorch to implement the device architecture API from oneAPI - https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/ext/oneapi/experimental/device_architecture.hpp
However this API is changing between 2024 and 2025, so we will need to be careful to use the right enum in that case (particularly if we compile against one version but use a different version at runtime).

[whitneywhtsang]

@guangyey is going to check with the DPC++ team on whether there is a plan to move device architecture API out of the "experimental".

[EikanWang]

@guangyey , pls. keep the issue updated.

[guangyey]

It depends on 2025.0.

[EikanWang]

@alexbaden , @whitneywhtsang , we are upgrading PyTorch to fit 2025.0. We will submit the PR to add the feature as soon as 2025.0 in PyTorch is ready.

[guangyey]

PR is pytorch/pytorch#138186.
The compiler team replies that they have a plan to move architecture out of the experimental namespace, but they have no ETA.