AMDGPU.jl icon indicating copy to clipboard operation
AMDGPU.jl copied to clipboard

AMDGPU.agents() doesn't see GPU

Open joelandman-jt opened this issue 3 years ago • 1 comments

Julia 1.7.1, latest AMDGPU.

2-element Vector{HSAAgent}:
 CPU-XX [Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz]
 CPU-XX [Intel(R) Xeon(R) Gold 6348 CPU @ 2.60GHz]

meanwhile rocminfo reports

Agent 3
*******
  Name:                    gfx90a
  Uuid:                    GPU-1e49a186eb42540e
  Marketing Name:
  Vendor Name:             AMD
  Feature:                 KERNEL_DISPATCH
  Profile:                 BASE_PROFILE```

This is a test MI210.  

New to AMDGPU, having used  CUDA in the past.

joelandman-jt avatar Jun 09 '22 21:06 joelandman-jt

Hi @joelandman-jt! This occurs because the version of the ROCm stack (specifically the ROCR library) that we ship with AMDGPU.jl is somewhat old and doesn't know about newer GPUs (his issue also occurs with the MI200). If you have a newer (>= 4.5.2) ROCm install on your system (or can get it installed via your distribution's package manager), then you can set the environment variable JULIA_AMDGPU_DISABLE_ARTIFACTS=1 and re-build AMDGPU.jl (using Pkg; Pkg.build("AMDGPU"), and then AMDGPU.jl will detect and use your system ROCm install.

I'll also be building JLLs for ROCR 4.5.2 soon, which should alleviate this issue.

jpsamaroo avatar Jun 13 '22 15:06 jpsamaroo

We now use much newer versions so this should be fine.

pxl-th avatar Sep 02 '23 20:09 pxl-th