Add rocm_smi_lib

[Krastanov]

rocm_smi_lib provides a library that reads virtual files
in the Linux file system in order to report GPU performance
parameters.

Both the library and the standard Python script that provides
a CLI for the library are included.

On a system with a functional AMD GPU and a Python 3
interpreter it can be tested with:

run(`python3 $(rocm_smi_lib_jll.rocm_smipy_path)`)

[Krastanov]

I felt the need to include the python files because:

• they are not python packages, just a single script that provides a CLI interface to the so file
• they have hardcoded paths in them which would require to patch them if we package them separately
• that way keeping the versions in sync is trivial
• they happen to not depend on anything else (only standard library)

[Krastanov]

And yes, long term we will probably just interface into the so library from Julia, without using the python script.

[vchuravy]

cc: @jpsamaroo

[giordano]

The file must be called build_tarballs.jl

[Krastanov]

That was embarrassing :D Should be fixed now.

[giordano]

Can build for more platforms? Or it's relevant only for x86-64 Linux?

[Krastanov]

AMD supports ROCm only on x86-64 Linux. It can probably built on any linux (it is a trivial C library that reads some virtual files), but those virtual files would never exist on another linux.

[giordano]

You should use the toolchain file https://docs.binarybuilder.org/stable/build_tips/#CMake-builds

[Krastanov]

oh, noted! I just copied this from rocminfo_jll. I will have to look into what these differences mean, as I do not have experience with cmake.

[jpsamaroo]

@Krastanov if you want to update this for ROCm 4.2.0, we can pull this JLL into JuliaGPU/AMDGPU.jl#187.

[Krastanov]

@jpsamaroo , I will do so by the end of the week.

[Krastanov]

Admittedly it is a single character diff, but I would like to check that it actually works on the hardware I have.

[Krastanov]

@jpsamaroo , here is the 4.2.0 #4350