[Pkg-opencl-devel] Bug#769072: #769191, #770588: nvidia-opencl-icd breaking non-nvidia systems

[Rebecca N. Palmer]

Rebecca Palmerr wrote:
> The only other [than pyopencl]
> Depends or Recommends on opencl-icd in the current archive is bfgminer.
Sorry...only ones found by "path:debian/control opencl-icd" in 
sources.debian.net search (apt-cache rdepends doesn't work on virtual 
packages), which evidently doesn't search non-free as it missed that 
nvidia-libopencl1 Recommends: nvidia-opencl-icd.

Nathaniel Smith wrote:
> Looking through my apt history, it looks like the critical operation
> that gave me nvidia stuff was the installation of libboost[-all-dev] (!?):
libboost-all-dev Depends: libboost-mpi-dev Depends: libboost-mpi1.55-dev 
Depends: libboost-mpi1.55.0 Depends: libhwloc5 Recommends: 
libhwloc-plugins, which at the time had Depends: libopencl-1.1-1, a 
virtual package provided by (among other things) nvidia-libopencl1, 
which Recommends: nvidia-opencl-icd.

This has already been reported (#739409) and fixed: libhwloc-plugins now 
Depends: ocl-icd-libopencl1 | libopencl-1.1-1.  However, cutting the 
chain there doesn't remove nvidia-opencl-icd if already installed, hence 
this bug.

On 23/11/14 02:09, Andreas Beckmann wrote:
> I don't know how seriously the missing libcuda1 breaks
> nvidia-opencl-icd. I can see that this is being dlopen()ed, but at least
> clinfo still reports something about the GPU. I don't have a better
> testcase right now, suggestions welcome.
If you want a quick "does OpenCL work" test, try
python3 accuracy_speed_test.py
(from 
https://bugs.debian.org/cgi-bin/bugreport.cgi?msg=5;filename=accuracy_speed_test.py;att=1;bug=768090 
, Depends: python3-pyopencl, python3-scipy ).  (Note that some of those 
tests are expected to give high/NaN errors because not all the inputs 
used are valid for all the functions: for the present purpose we're 
mainly looking for crashes/exceptions.)

Given that we also don't want to break systems that are intentionally 
using nvidia-opencl-icd, a better fix might be for whatever sets nvidia 
as default graphics provider to only do so if the hardware is present, 
but I don't know whether that's practical.