Problems on 4750U mobile APU

[simeonschaub]

As discussed on Slack a while back. I installed the packages rock-dkms-bin, hsa-runtime-git and rocm on Manjaro Linux and building AMDGPU.jl with JULIA_AMDGPU_DISABLE_ARTIFACTS=1 seems to work fine, but as soon as I try any operations on ROCArrays, I get the following error: https://gist.github.com/simeonschaub/d1bcee2600f69e08e82a229da490ffc6.

julia> versioninfo()
Julia Version 1.6.1
Commit 6aaedecc44 (2021-04-23 05:59 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: AMD Ryzen 7 PRO 4750U with Radeon Graphics
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-11.0.1 (ORCJIT, znver2)
Environment:
  JULIA_NUM_THREADS = 8

[vchuravy]

Can you try again, on current master?

[simeonschaub]

What's the current state regarding artifacts? Are they fully functional like the CUDA ones nowadays, or do I still need to install the hsa/rocm runtime through the system package manager?

[vchuravy]

On Julia 1.6 the artifacts "should" just work.

[simeonschaub]

Hmm, doing roc([1f0]) .+ roc([2f0]) does weird things to my screen but otherwise not a whole lot:

Is this art? 😅

[jpsamaroo]

@simeonschaub it might be worth updating your kernel to something recent and trying out #187 on Julia 1.7. That would get you a reasonably up-to-date ROCm, which probably has better APU support.

[simeonschaub]

I don't know what wizardry you have done here, but it actually seems to be mostly working with a 5.15 kernel! 🎉 Running the entire test suite, I believe there might still be a hang somewhere in the GPUArrays suite and if I try to cancel the process using Ctrl+c my laptop crashes. I am currently trying to narrow that down a bit more. I also haven't tested all the external libs like rocBLAS yet.

[simeonschaub]

I think I spoke too soon. The GPUArrays test suite actually passed except for the one requiring rocRAND. So I think I just underestimated the runtime of these tests and the crash might have just been a matter of AMDGPU not handling interrupts gracefully. I will consider this issue to be fixed now, I'll open a new one if I encounter a crash or any other issues again.