Bug#695883: nvidia improper OpenCL dependencies

[Andreas Beckmann]

Hi,

while implicitly recommending amd-opencl-icd is not such a bad idea (it
comes with a CPU implementation as well, so is usable for many people
even without proprietary graphics cards), perhaps it's better not to
Recommend a "random" icd by default and require explicit installation of
the required icd.

Package: ocl-icd-libopencl1
Suggests: opencl-icd
# require explicit installation of an icd

Package: vendor-libopencl1:
Recommends: vendor-opencl-icd, opencl-icd
# user probably wants exactly this if he installs vendor driver
# and switching loader vendors should not "orphan" foreign icds
  (yes, comma, not pipe)

Package: vendor-opencl-icd
Depends: ocl-icd-libopencl1 | vendor-libopencl1 | libopencl1
# yes, we want to get bug reports for ocl-icd loader working not
# properly with vendor icd implementations

or

Package: vendor-opencl-icd
Depends: vendor-libopencl1 | libopencl1
# use the well-tested driver from first vendor to be installed
# should avoid any possible interoperability problems
# probably more multiarch conflicts
  (install ocl-icd-libopencl1; install vendor-opencl-icd:foreign)

For clinfo:

Package: vendor-clinfo
Suggests: opencl-icd
# require explicit icd installation

or

Package: vendor-clinfo
Depends: ocl-icd-libopencl1 | libopencl1
Recommends: opencl-icd, vendor1-opencl-icd, vendor2-opencl-icd
# recommend all to get a better out-of-the-box result


Maybe we should provide some more virtual packages in the -icd packages:
opencl-icd-gpgpu // or is there a better term that covers
                 // Nvidia GTX xyz, Nvidia Tesla, Radeon HD wxyz
opencl-icd-gpgpu-$VENDOR
opencl-icd-cpu
opencl-icd-cpu-$CPUVENDOR


Would it be useful to make clinfo M-A: same co-installable? Rename the
binaries to clinfo-$ARCH or $ARCH-clinfo (fsvo ARCH, maybe better the
multiarch triplet) and a clinfo wrapper script that runs the one for the
primary arch (and thereafter for the secondary arches ?) Maybe amend the
output a bit if there are more than one binary to be run, but keep the
original output if there is only one.

Andreas