Update profiling docs: rocprof v1, v3 and roctx

[Technici4n]

Hi, this PR is the result of trying to profile some Julia code on LUMI (default ROCm version 6.0.something, there is a build of 6.2.4 available, planned upgrade to 6.3 soon), and eventually succeeding after multiple days. Please let me know if any information is wrong and I will happily correct it. It would also be helpful if others with more experience could run the various commands on other versions of ROCm and see if the situation is different there.

Here is a test script that I used:

using AMDGPU

function rangePush(message)
    @ccall "libroctx64".roctxRangePushA(message::Ptr{Cchar})::Cint
end

function rangePop()
    @ccall "libroctx64".roctxRangePop()::Cint
end

N = 10000
mat = ROCArray(randn(N, N))
vec = ROCArray(randn(N))

tot = 0.0
for i in 1:10
    rangePush("Iteration $i")
    tot += sum(mat .* vec)
    rangePop()
end
println(tot)

PS: Given the existence of #801, I suppose that rocprofv3 is not expected to be working yet?

[luraess]

Thanks for the update! I will try it out on the CI machines and may add some infos about profiling with MPI as well.

WRT ROCTX, given that only these 2 functions seem to work for now, I wonder whether it would make sense to have them exposed by AMDGPU instead of relying on ROCTX.jl just for that?

[Technici4n]

given that only these 2 functions seem to work for now

Have you tried the others? I haven't myself as range push and pop were sufficient for my needs :)

whether it would make sense to have them exposed by AMDGPU

I don't know why NVTX is a separate package from CUDA, maybe the same reasoning applies here. It is possible that NVTX is a very light (and hard if the code is annotated) dependency whereas package authors prefer to leave CUDA support in an ext module?

[luraess]

Suggested change

	[rocprof](https://github.com/ROCm/rocm-systems/tree/develop/projects/rocprofiler)
	[rocprof](https://rocm.docs.amd.com/projects/rocprofiler/en/latest/index.html)

[luraess]

Links to the doc

[luraess]

One could add upfront here a comment stating that rocprofv3 is now to be used as others are now deprecated (unless one is on a system where it may not work).

[luraess]

Thanks for this addition!

[luraess]

I checked the various profiling tool version on the CI machine and with ROCm 6.4 all seem to work in reporting basic tracing info.

Maybe, before merging, one could simplify and streamline further the doc, focussing on v3 while still showing how to launch profiling with v2 and v1. Also, one could add to consistently for each version the -o prof or another output name. Some duplicate about link to Perfetto UI could also be removed.

After this, I would be happy to merge - thanks!