Use oneAPI 2024.0 rather than 2023.2 to align with PyTorch.

[etiotto]

The current Triton compiler uses the oneAPI libraries. Currently it works when oneAPI is 2023.2 but fails with 2024.0. Investigate the reason(s) and address the problem(s).

[jinz2014]

ImportError: libmkl_sycl.so.3: cannot open shared object file: No such file or directory
ImportError: libsycl.so.6: cannot open shared object file: No such file or directory

File "/home/users//.local/lib/python3.10/site-packages/triton/third_party/xpu/init.py", line 400, in get_stream
return torch.xpu.current_stream(idx).sycl_queue
AttributeError: 'Stream' object has no attribute 'sycl_queue'

[tdeng5]

PyTorch team and Triton spirv path are using oneAPI 2024.0 for developing, we should align and use oneAPI 2024.0 for developing and testing.

[etiotto]

This is done now (PR #239)

[pbchekin]

Note that formally this issue is not complete, because we currently have to downgrade libigc1.

[quintinwang5]

@pbchekin Should we wait for a response from https://jira.devtools.intel.com/browse/IGC-8444? Or we can update genx branch ASAP? I think we need an assumption like: All public versions >= 1.0.15136.x use khronos intrinsic style. The problem is: Seems they use legacy style in the cmake file of their release branch, but they use khronos style in their released binary.

[pbchekin]

@pbchekin Should we wait for a response from https://jira.devtools.intel.com/browse/IGC-8444? Or we can update genx branch ASAP? I think we need an assumption like: All public versions >= 1.0.15136.x use khronos intrinsic style. The problem is: Seems they use legacy style in the cmake file of their release branch, but they use khronos style in their released binary.

Not sure if I am making a call here, but according to the commit they provided in the Jira issue they use Khronos translator by default on Linux (still disabled for Windows), so for me yes, if we want to use the latest release then we should use Khronos translator. @etiotto ?

[pbchekin]

Filed #358