Does not find versioned libraries

[benz0li]

Discussion: JuliaGPU/CUDA.jl#2318

[benz0li]

Workarounds: Either

• install the -dev packages of the required libraries

or

• create unversioned symlinks for all required libraries

i.e.

libDir=/usr/local/cuda/lib64
for f in "$libDir/"*; do
  [ -e "$f" ] || [ -L "$f" ] || continue
  src="$(basename "$f" | grep "^lib.*.so.[[:digit:]]\+$")"
  if [ -s "$libDir/$src" ] && [ ! -e "$libDir/${src%.*}" ]; then
    ln -svT "$src" "$libDir/${src%.*}"
  fi
done

Ping @maleadt

[maleadt]

Should be fixed by e2c157b; can you try? I verified locally with a runtime container + cuda-cupti + cuda-nvcc + cuda-command-line-tools.

FYI, it should be possible to make ptxas or even nvdisasm optional and rely on the driver to assemble code, in case you would require a more lightweight container (seeing how cuda-nvcc pulls in an entire C++ development toolchain).

[benz0li]

Should be fixed by e2c157b; can you try? I verified locally with a runtime container + cuda-cupti + cuda-nvcc + cuda-command-line-tools.

Confirmed.

FYI, it should be possible to make ptxas or even nvdisasm optional and rely on the driver to assemble code

🆗

in case you would require a more lightweight container (seeing how cuda-nvcc pulls in an entire C++ development toolchain).

Currently not required, because the C++ development toolchain is installed in a later [image build] stage anyway.

[maleadt]

Thanks for confirming. I'll tag a release.